text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'The class of threshold functions is known to be characterizable by functional equations or, equivalently, by pairs of relations, which are called relational constraints. It was shown by Hellerstein that this class cannot be characterized by a finite number of such objects. In this paper, we investigate classes of threshold functions which arise as intersections of the class of all threshold functions with clones of Boolean functions, and provide a complete classification of such intersections in respect to whether they have finite characterizations. Moreover, we provide a characterizing set of relational constraints for each class of threshold functions arising in this way.'
address:
- |
LAMSADE – CNRS\
Université Paris-Dauphine\
Place du Maréchal de Lattre de Tassigny\
75775 Paris Cedex 16\
France
- |
University of Luxembourg\
Computer Science and Communications Research Unit\
6, rue Richard Coudenhove-Kalergi\
L–1359 Luxembourg\
Luxembourg
- |
University of Luxembourg\
Mathematics Research Unit\
6, rue Richard Coudenhove-Kalergi\
L–1359 Luxembourg\
Luxembourg
author:
- Miguel Couceiro
- Erkko Lehtonen
- Karsten Schölzel
title: A complete classification of equational classes of threshold functions included in clones
---
Introduction and preliminaries
==============================
Introduction
------------
Two approaches to characterize properties of Boolean functions have been considered recently: one in terms of functional equations [@EFHH], another in terms of relational constraints [@Pippenger]. As it turns out, these two approaches have the same expressive power in the sense that they characterize the same properties (classes) of Boolean functions, which can be described as initial segments of the so-called “minor” relation between functions: for two functions $f$ and $g$ of several variables, $f$ is said to be a minor of $g$ if $f$ can be obtained from $g$ by identifying variables, permuting variables, or adding inessential variables (see Subsection \[susec:MinorsConstraints\]). Furthermore, a class is characterizable by a finite number of functional equations if and only if it is characterizable by a finite number of relational constraints (see, e.g., [@CF; @Pippenger]). For the sake of simplifying the presentation of constructions and proofs, we will focus on the approach by relational constraints.
Several properties of functions can be charaterized by relational constraints (or, equivalently, by functional equations. In fact, uncountably many properties are expressible by such objects, even in the simplest interesting case of functions of several variables, i.e., the Boolean functions (see [@CP; @Pippenger]). Classical examples of such properties include idempotency, monotonicity and linearity. More contemporary examples include submodularity, supermodularity and the combination of the two, i.e., modularity (see, e.g., [@CouMar; @Lovasz; @Singer; @Topkis]).
Another noteworthy example is thresholdness that is the property of those Boolean functions whose true points can be separated from the false points by a hyperplane when considered as elements of the $n$-dimensional real space ${\mathbb{R}}^n$. Threshold functions have been widely studied in the literature on Boolean functions, switching theory, system reliability theory, game theory, etc.; for background see, e.g., [@Isbell; @Muroga; @Peleg1; @Peleg2; @Taylor; @Winder].
Despite being a property expressible by relational constraints, thresholdness cannot be captured by a finite set of relational constraints (see Hellerstein [@Hellerstein]). However, by imposing additional conditions such as linearity or preservation of componentwise conjunctions or disjunctions of tuples, the resulting classes of threshold functions may become characterizable by a finite number of relational constraints. In fact, these examples can be obtained from the class of threshold functions by intersecting it with certain clones, namely, those of linear functions, conjunctions and disjunctions, respectively. (Recall that a clone is a class of functions that contains all projections and is closed under functional composition.) Another noteworthy and well-known example of such an intersection is the class of “majority games”, which results as the intersection with the clone of self-dual monotone functions. The natural question is then: Is the class of majority games characterizable by a finite number of relational constraints?
In this paper we answer negatively to this question. In fact, we will determine, for each clone of Boolean functions, whether its intersection with the class of threshold functions is finitely characterizable by relational constraints. Moreover, we provide finite or infinite characterizing sets of relational constraints accordingly.
The paper is organized as follows. In the remainder of this section, we recall basic notions and results that will be needed throughout the paper. The main results are presented in Section \[sec:main\], in particular, the classification of all intersections $C \cap {T}$, where $C$ is a clone and ${T}$ is the class of all threshold functions, as well as the corresponding characterizing set of relational constraints. For the reader’s convenience, the constructions needed for the main results will be left for Section \[sec:constructions\]. One of the main tools in our proof is Taylor and Zwicker’s [@Taylor] theorem on the existence of a $k$-asummable function that is not $(k+1)$-asummable. In Section \[sec:magic\], we slightly refine Taylor and Zwicker’s result and show how the classes of functions characterizable by the relational constraints that arise in our current work are related to each other. Appendix \[App:Post\] provides a list of the clones of Boolean functions and relations characterizing them.
Boolean functions
-----------------
Throughout the paper, we denote the set $\{1, \dots, n\}$ by ${[{n}]}$ and the set $\{0, 1\}$ by ${\mathbb{B}}$. A *Boolean function* is a map $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ for some positive integer $n$ called the *arity* of $f$. Typical examples of Boolean functions include
- the $n$-ary $i$-th *projection* ($i \in {[{n}]}$) $e^{(n)}_i \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $(a_1, \dots, a_n) \mapsto a_i$;
- *negation* ${\overline{\cdot}} \colon {\mathbb{B}}\to {\mathbb{B}}$, ${\overline{0}} = 1$, ${\overline{1}} = 0$;
- *conjunction* $\wedge \colon {\mathbb{B}}^2 \to {\mathbb{B}}$, $x \wedge y = 1$ if and only if $x = y = 1$;
- *disjunction* $\vee \colon {\mathbb{B}}^2 \to {\mathbb{B}}$, $x \vee y = 0$ if and only if $x = y = 0$;
- *modulo-$2$ addition* $\oplus \colon {\mathbb{B}}^2 \to {\mathbb{B}}$, $x \oplus y = (x + y) \bmod 2$.
The set of all Boolean functions is denoted by $\Omega$ and the set of all projections is denoted by $I_c$.
The preimage ${{f}^{-1}(1)}$ of $1$ under $f$ is referred to as the set of *true points*, while ${{f}^{-1}(0)}$ is referred to as the set of *false points*.
The $i$-th variable of a Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ is said to be *essential* in $f$, or that $f$ depends on $x_i$, if there are $a_1, \ldots, a_{i-1},a_{i+1}, \ldots,a_n \in {\mathbb{B}}$ such that $$f(a_1, \ldots, a_{i-1},0,a_{i+1}, \ldots, a_n) \neq f(a_1, \ldots, a_{i-1}, 1, a_{i+1}, \ldots, a_n).$$
The *dual* of a Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ is the function ${{f}^{\mathrm{d}}} \colon {\mathbb{B}}^n \to {\mathbb{B}}$ given by $${{f}^{\mathrm{d}}}(x_1, \dots, x_n)
= {\overline{f({\overline{x}}_1, \dots, {\overline{x}}_n)}}.$$ A function $f$ is *self-dual* if $f = {{f}^{\mathrm{d}}}$.
If $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ and $g_1, \dots, g_n \colon {\mathbb{B}}^m \to {\mathbb{B}}$, then the *composition* of $f$ with $g_1, \dots, g_n$ is the function $f(g_1, \dots, g_n) \colon {\mathbb{B}}^m \to {\mathbb{B}}$ given by $$f(g_1, \dots, g_n)(\aa) = f(g_1(\aa), \dots, g_n(\aa))$$ for all $\aa \in {\mathbb{B}}^m$. A *clone* of Boolean functions is a subset $C$ of the set $\Omega$ of all Boolean functions that satisfies the following two conditions:
- $I_c \subseteq C$, i.e., $C$ contains all projections,
- if $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $g_1, \dots, g_n \colon {\mathbb{B}}^m \to {\mathbb{B}}$ and $f, g_1, \dots, g_n \in C$, then $f(g_1, \dots, g_n) \in C$, i.e., $C$ is closed under composition.
The clones of Boolean functions were completely described by Post [@Post], and they are often referred to as *Post’s classes.* We provide a list of all clones of Boolean functions in Appendix \[App:Post\].
Minors and relational constraints {#susec:MinorsConstraints}
---------------------------------
We will denote tuples in ${\mathbb{B}}^m$ by boldface letters and their entries with corresponding italic letters, e.g., $\mathbf{a} = (a_1, \dots, a_m)$. Tuples $\mathbf{a} \in {\mathbb{B}}^m$ may be viewed as mappings $\mathbf{a} \colon {[{m}]} \to {\mathbb{B}}$, $i \mapsto a_i$. With this convention, given a map $\sigma \colon {[{n}]} \to {[{m}]}$, we can write the tuple $(a_{\sigma(1)}, \dots, a_{\sigma(n)})$ as $\mathbf{a} \circ \sigma$, or simply $\mathbf{a} \sigma$.
A function $f \colon {\mathbb{B}}^m \to {\mathbb{B}}$ is a *minor* of another function $g \colon {\mathbb{B}}^n \to {\mathbb{B}}$ if there exists a map $\sigma \colon {[{n}]} \to {[{m}]}$ such that $f(\mathbf{a}) = g(\mathbf{a} \sigma)$ for all $\mathbf{a} \in {\mathbb{B}}^m$; in this case we write $f \leq g$. Functions $f$ and $g$ are *equivalent,* denoted $f \equiv g$, if $f \leq g$ and $g \leq f$. In other words, $f$ is a minor of $g$ if $f$ can be obtained from $g$ by permutation of arguments, addition and deletion of inessential arguments and identification of arguments. Functions $f$ and $g$ are equivalent if each one can be obtained from the other by permutation of arguments and addition and deletion of inessential arguments.
The minor relation $\leq$ is a quasi-order (i.e., a reflexive and transitive relation) on the set of all Boolean functions, and the relation $\equiv$ is indeed an equivalence relation. For further background see, e.g., [@CL1; @CL2; @CP; @EFHH; @Pippenger] In what follows, we shall consider minors of the following special form. Let $n \geq 2$, and let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. For any two-element subset $I$ of ${[{n}]}$, we define the function $f_I \colon {\mathbb{B}}^{n-1} \to {\mathbb{B}}$ by the rule $f_I(\aa) = f(\aa \delta_I)$ for all $\aa \in {\mathbb{B}}^{n-1}$, where $\delta_I \colon {[{n}]} \to {[{n-1}]}$ is given by the rule $$\label{eq:deltaI}
\delta_I(i) =
\begin{cases}
i, & \text{if $i < \max I$,} \\
\min I, & \text{if $i = \max I$,} \\
i - 1, & \text{if $i > \max I$.}
\end{cases}$$ In other words, if $I = \{i, j\}$ with $i < j$, then $$f_I(a_1, \dots, a_{n-1}) = f(a_1, \dots, a_{j-1}, a_i, a_j, \dots, a_{n-1}).$$ Note that $a_i$ occurs twice on the right side of the above equality: both at the $i$-th and at the $j$-th position. The function $f_I$ will be referred to as an *identification minor* of $f$.
It was shown by Pippenger [@Pippenger] that the classes of functions closed under taking minors are characterizable by so-called relational constraints. We will briefly survey some results which we will use hereinafter. An $m$-ary *relational constraint* is a couple $(R, S)$ of $m$-ary relations $R$ (the *antecedent*) and $S$ (the *consequent*) on ${\mathbb{B}}$ (i.e., $R, S \subseteq {\mathbb{B}}^m$). We denote the antecedent and the consequent of a relational constraint $Q$ by $R(Q)$ and $S(Q)$, respectively. If both $R(Q)$ and $S(Q)$ equal the binary equality relation, then $Q$ is called the binary *equality constraint*. Furthermore, we refer to constraints with empty antecedent and empty consequent as *empty constraints*, and to constraints where the antecedent and consequent are the full relation ${\mathbb{B}}^m$, for some $m\geq 1$, as *full constraints*. The set of all relational constraints is denoted by $\Theta$.
A function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ *preserves* an $m$-ary relational constraint $(R, S)$, denoted $f \triangleright (R, S)$, if for every $\mathbf{a}^{1}, \dots, \mathbf{a}^{n} \in R$, we have $f(\mathbf{a}^1, \dots, \mathbf{a}^n) \in S$. (Regarding tuples $\mathbf{a}^i$ as unary maps, $f(\mathbf{a}^1, \dots, \mathbf{a}^n)$ denotes the $m$-tuple whose $i$-th entry is $f(\mathbf{a}^1, \dots, \mathbf{a}^n)(i) = f(a^1_i, \dots, a^n_i)$.)
The preservation relation gives rise to a Galois connection between functions and relational constraints that we now briefly describe; for further background, see [@C; @CP; @Pippenger]. Define $\operatorname{cPol}\colon \mathcal{P}(\Theta) \to \mathcal{P}(\Omega)$, $\operatorname{cInv}\colon \mathcal{P}(\Omega) \to \mathcal{P}(\Theta)$ by $$\begin{aligned}
\operatorname{cPol}(\mathcal{Q}) &= \{f \in \Omega : \text{$f \triangleright Q$ for every $Q \in \mathcal{Q}$}\}, \\
\operatorname{cInv}(\mathcal{F}) &= \{Q \in \Theta : \text{$f \triangleright Q$ for every $f \in \mathcal{F}$}\}.\end{aligned}$$ We say that a set $\mathcal{F}$ of functions is *characterized* by a set $\mathcal{Q}$ of relational constraints if $\mathcal{F} = \operatorname{cPol}(\mathcal{Q})$. Dually, $\mathcal{Q}$ is *characterized* by $\mathcal{F}$ if $\mathcal{Q} = \operatorname{cInv}(\mathcal{F})$. In other words, sets of functions characterizable by relational constraints are exactly the fixed points of $\operatorname{cPol}\circ \operatorname{cInv}$, and, dually, sets of relational constraints characterizable by functions are exactly the fixed points of $\operatorname{cInv}\circ \operatorname{cPol}$.
Preservation of a relational constraint generalizes the notion of preservation of a relation, as in the classical $\mathrm{Pol}$–$\mathrm{Inv}$ theory of clones and relations, which establishes that the clones on finite sets are exactly the classes of functions that are characterized by relations (see [@BKKR; @Geiger]). In this framework, a function $f$ preserves a relation $R$ if and only if $f$ preserves the relational constraint $(R, R)$. Hence, clones are exactly the classes that are characterized by relational constraints of the form $(R, R)$ for some relation $R$.
The following result reassembles various descriptions of the Galois closed sets of functions, which can be found in [@CP; @EFHH; @Pippenger].
\[thm:characterizablefunctions\] Let $\mathcal{F}$ be a set of functions. The following are equivalent.
1. \[thm:characterizablefunctions:item1\] $\mathcal{F}$ is closed under taking minors.
2. \[thm:characterizablefunctions:item2\] $\mathcal{F}$ is characterizable by relational constraints.
3. \[thm:characterizablefunctions:item3\] $\mathcal{F}$ is of the form $$\operatorname{forbid}(A) := \{f \in \Omega : \text{$g \nleq f$ for all $g \in A$}\}$$ for some antichain $A$ with respect to the minor relation $\leq$.
It follows from the equivalence of \[thm:characterizablefunctions:item1\] and \[thm:characterizablefunctions:item2\] in Theorem \[thm:characterizablefunctions\] that the union and the intersection of classes that are characterizable by relational constraints are characterizable by relational constraints.
Note that the antichain $A$ in item \[thm:characterizablefunctions:item3\] of Theorem \[thm:characterizablefunctions\] is unique up to equivalence. In fact, $A$ can be chosen among the minimal elements of $\Omega \setminus \mathcal{F}$; the elements of $A$ are called *minimal forbidden minors for $\mathcal{F}$*.
As we will see, there are classes of functions that, even though characterizable by relational constraints, are not characterized by any finite set of relational constraints. A set of functions is *finitely characterizable* if it is characterized by a finite set of relational constraints.
The following theorem is a refinement of Theorem \[thm:characterizablefunctions\] and provides a description for finitely characterizable classes.
\[thm:finitelycharacterizablefunctions\] Let $\mathcal{F}$ be a set of functions. The following are equivalent.
1. $\mathcal{F}$ is finitely characterizable.
2. $\mathcal{F}$ is of the form $\operatorname{forbid}(A)$ for some finite antichain $A$ with respect to the minor relation $\leq$.
The Galois closed sets of relational constraints were likewise described by Pippenger [@Pippenger]; this description was extended to arbitrary, possibly infinite, underlying sets in [@CF]. We shall briefly survey Pippenger’s description of the Galois closed sets of constraints.
An $m$-ary relational constraint $(R,S)$ is a *simple minor* of an $(m+p)$-ary relational constraint $(R',S')$ if there is $h \colon \{1, \ldots, n\} \to \{1, \ldots, m+p\}$ such that $$\begin{aligned}
R
\left(\begin{array}{c}
x_1\\
\vdots \\
x_m\\
\end{array}\right)
&
\quad \iff \quad
\exists x_{m+1} \ldots \exists x_{m+p} \quad
R'
\left(\begin{array}{c}
x_{h(1)}\\
\vdots \\
x_{h(n)}\\
\end{array}\right)
\intertext{and}
S
\left(\begin{array}{c}
x_1\\
\vdots \\
x_m\\
\end{array}\right)
&
\quad \iff \quad
\exists x_{m+1} \ldots \exists x_{m+p} \quad
S'
\left(\begin{array}{c}
x_{h(1)}\\
\vdots \\
x_{h(n)}\\
\end{array}\right).
\end{aligned}$$
Note that simple minors subsume the notions of permutation, diagonalization and projection of arguments; for background see [@CF; @Pippenger].
A constraint $(R,S)$ is obtained from a constraint $(R',S)$ by *restricting the antecedent* if $R\subseteq R'$. Likewise, $(R,S)$ is obtained from a constraint $(R,S')$ by *extending the consequent* if $S\supseteq S'$. A constraint $(R,S\cap S')$ is said to be obtained from $(R,S)$ and $(R, S')$ by intersecting consequents.
A set $\mathcal{Q}$ of relational constraints is said to be *minor-closed* if it contains the binary equality constraint, the unary empty constraint, and it is closed under taking simple minors, restricting antecedents, and extending and intersecting consequents.
We can now state Pippenger’s [@Pippenger] description of the Galois closed sets of relational constraints.
\[thm:characterizableconstraints\] Let $\mathcal{Q}$ be a set of relational constraints. The following are equivalent.
1. \[thm:characterizableconstraints:item1\] $\mathcal{Q}$ is characterizable by some set of functions.
2. \[thm:characterizableconstraints:item2\] $\mathcal{Q}$ is minor-closed.
The following lemma provides a noteworthy tool for showing that certain classes of threshold functions are not finitely characterizable.
\[lemma:finiteDescendingChainIffFinitelyCharacterizable\] Let $C$ and $C_i$ for all $i \geq 1$ be classes of functions that are closed under taking minors, such that $C = \bigcap_{i \geq 1} C_i$, and $C_{i+1} \subseteq C_i$ for all $i \geq 1$. If $C$ is finitely characterizable by constraints, then there exists $\ell \in {\mathbb{N}}$ such that $C_j = C_\ell$ for all $j \geq \ell$.
By Theorem \[thm:characterizablefunctions\], each minor-closed class $C_i$ is characterized by some set $\mathcal{Q}_i$ of relational constraints, i.e., $C_i = \operatorname{cPol}\mathcal{Q}_i$ for all $i \geq 1$.
Assume that $C$ is finitely characterizable. Then there is some finite set $\mathcal{P}$ of constraints with $C = \operatorname{cPol}\mathcal{P}$, and thus $\operatorname{cInv}C = \operatorname{cInv}\operatorname{cPol}\mathcal{P}$. Since $$C = \bigcap_{i \geq 1} \operatorname{cPol}\mathcal{Q}_i = \operatorname{cPol}\bigcup_{i \geq 1} \mathcal{Q}_i$$ we can construct each $P \in \mathcal{P}$ from the constraints in $\bigcup_{i \geq 1} \mathcal{Q}_i$. Since the constraints are finite, all such constructions are finite. In particular, only a finite number of constraints from $\bigcup_{i \geq 1} \mathcal{Q}_i$ are used for each $P \in \mathcal{P}$. Since $\mathcal{P}$ is finite, this implies that only a finite number of constraints are needed to construct all $P \in \mathcal{P}$. Therefore $$C = \operatorname{cPol}\mathcal{P} \supseteq \operatorname{cPol}\bigcup_{i = 1}^l \mathcal{Q}_i
= \bigcap_{i=1}^l \operatorname{cPol}\mathcal{Q}_i
= \operatorname{cPol}\mathcal{Q}_l = C_l$$ holds for some $l \in {\mathbb{N}}$. Now this implies that for any $j \geq l$, $$C \subseteq C_j \subseteq C_l \subseteq C$$ and consequently $C = C_j = C_l$ for all $j \geq l$.
Main results: classification and characterizations of Galois closed sets of threshold functions {#sec:main}
===============================================================================================
Motivation
----------
A *threshold function* is a Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ such that there exist *weights* $w_1, \dots, w_n \in {\mathbb{R}}$ and a *threshold* $t \in {\mathbb{R}}$ fulfilling $$f(x_1, \dots, x_n) = 1 \iff \sum_{i=1}^n w_i x_i \geq t.$$ Another, equivalent, definition is the following. An $n$-ary Boolean function $f$ is called a *threshold function* if there is a hyperplane in ${\mathbb{R}}^n$ strictly separating the true points of $f$ from the false points of $f$, considered as elements of ${\mathbb{R}}^n$. The set of all threshold functions is denoted by ${T}$.
The class of threshold functions has remarkable invariance properties. For instance, it is closed under taking negations and duals (see Lemma \[lem:dualBl\]). Moreover, the class of threshold functions is also closed under taking minors of its members; hence it is characterizable by relational constraints by Theorem \[thm:characterizablefunctions\]. However, no finite set of relational constraints suffices.
\[thm:Hell\] The class of threshold functions is not finitely characterizable.
Imposing some additional conditions on threshold functions, we may obtain proper subclasses of ${T}$ that are finitely characterizable. Easy examples arise from the intersections of ${T}$ with the clones $L$, $\Lambda$, $V$ (see Appendix \[App:Post\]). However, as we have seen, other intersections $C \cap T$ may fail to be finitely characterizable, e.g., for $C = \Omega$.
This fact gives rise to the following problem.
Which clones $C$ of Boolean functions have the property that $C \cap {T}$ is finitely characterizable?
In the following subsection we present a solution to this problem.
Classification and characterizations of intersections of the class of threshold functions with clones
-----------------------------------------------------------------------------------------------------
We start by observing that $$L \cap {T}= \Omega(1), \qquad
\Lambda \subseteq {T}, \qquad
V \subseteq {T},$$ from which it follows that the intersection $C \cap {T}$ is a clone for any clone $C$ contained in one of $L$, $V$ and $\Lambda$. Hence, the characterization of $C \cap {T}$ for any such clone $C$ is given by the relational constraint $(R, R)$, where $R$ is the relation characterizing $C \cap {T}$ given in Appendix \[App:Post\].
We proceed to characterizing the intersections $C \cap {T}$ for the remaining clones $C$; as we will see, none of these is finitely characterizable. A characterization of the class ${T}$ of all threshold functions (i.e., for $C = \Omega$) is easily obtained with the help of the notion of asummability.
For $k \geq 2$, a Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ is *$k$-asummable* if for any $m \in \{2, \dots, k\}$ and for all $\aa_1, \dots, \aa_m \in f^{-1}(0)$ and $\bb_1, \dots, \bb_m \in f^{-1}(1)$, it holds that $$\aa_1 + \dots + \aa_m \neq \bb_1 + \dots + \bb_m.$$ (Addition here is standard vector addition in ${\mathbb{R}}^n$.) A function is *asummable* if it is $k$-asummable for all $k \geq 2$. It is well known that asummability characterizes threshold functions; see [@Chow; @Elgot; @Muroga].
\[thm:thresholdasummable\] A Boolean function is threshold if and only if it is asummable.
Define for $n \geq 1$, the $2n$-ary relational constraint $B_n$ as $$\begin{aligned}
R(B_n) & := \{ (x_1, \dots, x_{2n}) \in {\mathbb{B}}^{2n} : \sum_{i=1}^n x_i = \sum_{i=n+1}^{2n} x_i \} \\
S(B_n) & := {\mathbb{B}}^{2n} \setminus \{ (\underbrace{0,\dots,0}_{n},\underbrace{1,\dots,1}_n),
(\underbrace{1,\dots,1}_{n},\underbrace{0,\dots,0}_n) \}. \end{aligned}$$ Note that in the definition of $R(B_n)$ we employ the usual addition of real numbers. Denoting by $w(\aa)$ the *Hamming weight* of a tuple $\aa \in {\mathbb{B}}^n$ (i.e., the number of nonzero entries in $\aa$), we can equivalently define $R(B_n)$ as $\{(x_1, \dots, x_{2n}) \in {\mathbb{B}}^{2n} : w(x_1, \dots, x_n) = w(x_{n+1}, \dots, x_{2n})\}$.
\[lem:PolBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ and $\ell \geq 2$. Then $\aa_1 + \dots + \aa_\ell \neq \bb_1 + \dots + \bb_\ell$ for all $\aa_1, \dots, \aa_\ell \in f^{-1}(0)$ and $\bb_1, \dots, \bb_\ell \in f^{-1}(1)$ if and only if $f$ preserves $B_\ell$.
Assume first that $f$ does not preserve $B_\ell$. Then there exists a matrix $$M = \begin{pmatrix}
m_1^1 & m_2^1 & \dots & m_n^1 \\
m_1^2 & m_2^2 & \dots & m_n^2 \\
\vdots & \vdots & & \vdots \\
m_1^{2 \ell} & m_2^{2 \ell} & \dots & m_n^{2 \ell}
\end{pmatrix} = \begin{pmatrix}
M^1 \\
M^2 \\
\vdots \\
M^{2 \ell}
\end{pmatrix}
= (M_1, M_2, \dots, M_n),$$ i.e., $M^1, \dots, M^{2 \ell} \in {\mathbb{B}}^n$ are the rows of $M$, and $M_1, \dots, M_n \in {\mathbb{B}}^{2 \ell}$ are the columns of $M$, such that
- $M_1, \dots, M_n \in R(B_\ell)$, and
- $\zz := g(M_1, \dots, M_n) := \begin{pmatrix} g(M^1) \\ \vdots \\ g(M^{2 \ell}) \end{pmatrix} \notin S(B_\ell)$.
Thus $\zz \in \{ (\underbrace{0,\dots,0}_l,\underbrace{1,\dots,1}_l),
(\underbrace{1,\dots,1}_l,\underbrace{0,\dots,0}_l)\}$. As $B_\ell$ is invariant under swapping the first $\ell$ rows with the last $\ell$ rows, we can assume that $\zz = (\underbrace{0, \dots, 0}_\ell, \underbrace{1, \dots, 1}_\ell)$. Then $M^1, \dots, M^\ell \in f^{-1}(0)$ and $M^{\ell + 1}, \dots, M^{2 \ell} \in f^{-1}(1)$, and $M^1 + \dots + M^\ell = M^{\ell + 1} + \dots + M^{2 \ell}$ by the definition of $B_\ell$.
Assume then that there exist $\aa_1, \dots, \aa_\ell \in f^{-1}(0)$ and $\bb_1, \dots, \bb_\ell \in f^{-1}(1)$ such that $\aa_1 + \dots + \aa_\ell = \bb_1 + \dots + \bb_\ell$. Let $M$ be the $2 \ell \times n$ matrix whose rows are $\aa_1, \dots, \aa_\ell, \bb_1, \dots, \bb_\ell$. The columns of $M$ are tuples in $R(B_\ell)$, but $f(M) = (\underbrace{0, \dots, 0}_\ell, \underbrace{1, \dots, 1}_\ell) \notin S(B_\ell)$. We conclude that $f$ does not preserve $B_\ell$.
Now it is easy to define a set of relational constraints that characterizes $k$-asummable functions. For $k \geq 2$, let $\mathcal{A}_k := \{B_n : 2 \leq n \leq k\}$.
\[lem:k-asummable\] Let $k \geq 2$. A Boolean function $f$ is $k$-asummable if and only if $f \in \operatorname{cPol}(\mathcal{A}_k)$.
Follows immediately from the definition of $k$-asummability and Lemma \[lem:PolBl\].
Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. The following are equivalent.
1. \[cor:thr:item:thr\] $f$ is a threshold function.
2. \[cor:thr:item:Ak\] $f \in \bigcap_{k \geq 2} \operatorname{cPol}(\mathcal{A}_k)$.
3. \[cor:thr:item:Bn\] $f \in \operatorname{cPol}(\{B_n : n \geq 2\})$.
The equivalence of \[cor:thr:item:thr\] and \[cor:thr:item:Ak\] follows immediately from Theorem \[thm:thresholdasummable\] and Lemma \[lem:k-asummable\]. Conditions \[cor:thr:item:Ak\] and \[cor:thr:item:Bn\] are equivalent, because $$\bigcap_{k \geq 2} \operatorname{cPol}(\mathcal{A}_k)
= \operatorname{cPol}\Bigl( \bigcup_{k \geq 2} \mathcal{A}_k \Bigr)
= \operatorname{cPol}\Bigl( \bigcup_{k \geq 2} \{B_n : 2 \leq n \leq k\} \Bigr)
= \operatorname{cPol}(\{B_n : n \geq 2\}).
\qedhere$$
Since $\mathcal{A}_k \subseteq \mathcal{A}_k \cup \{B_{k+1}\} = \mathcal{A}_{k+1}$, it is clear that $\operatorname{cPol}(\mathcal{A}_{k+1}) \subseteq \operatorname{cPol}(\mathcal{A}_k)$ for all $k \geq 2$. Taylor and Zwicker have shown in [@Taylor] that for every $k \geq 2$, there exist $k$-asummable functions that are not $(k+1)$-asummable. Hence these inclusions are strict for every $k$.
\[theorem:strictAkInclusions\] For all $k \geq 2$, $\operatorname{cPol}(\mathcal{A}_{k+1}) \subset \operatorname{cPol}(\mathcal{A}_k)$.
\[thm:characterization\] The set $\operatorname{cPol}(\{ B_n : n \geq 2 \})$ is the class of all threshold functions. Moreover, for every clone $C$, the subclass $C \cap {T}$ of threshold functions is characterized by the set $\{B_n : n \geq 2\} \cup \mathcal{Q}_C$, where $\mathcal{Q}_C$ is the set of relational constraints characterizing the clone $C$, as given in Appendix \[App:Post\].
From Theorems \[theorem:strictAkInclusions\] and \[thm:characterization\] it follows that $${T}= \bigcap_{k \geq 2} \operatorname{cPol}(\mathcal{A}_k) \subset \dots \subset \operatorname{cPol}(\mathcal{A}_{\ell+1}) \subset \operatorname{cPol}(\mathcal{A}_\ell) \subset \dots \subset \operatorname{cPol}(\mathcal{A}_2)$$ holds for all $\ell \geq 3$, i.e., the sets $\operatorname{cPol}(\mathcal{A}_k)$ with $k \geq 2$ form an infinite descending chain, whose intersection is the set ${T}$ of all threshold functions.
Theorem \[thm:characterization\] provides an infinite set of relational constraints characterizing the set $C \cap {T}$ for each clone $C$. As Theorem \[thm:classification\] will reveal, the characterization provided is optimal for the clones not contained in $L$, $V$ or $\Lambda$ in the sense that for such clones $C$, the set $C \cap {T}$ is not finitely characterizable by relational constraints.
In order to proceed, we need the following lemma. Its proof is somewhat technical and is deferred to Section \[sec:constructions\].
\[lem:GCf\]
Let $f$ be a Boolean function, and let $C \in \{SM, M_cU_\infty, M_cW_\infty\}$. There exists a Boolean function $G_C(f)$ that satisfies the following conditions:
1. $G_C(f) \in C$,
2. for all $n \geq 2$, $f \in \operatorname{cPol}B_n$ if and only if $G_C(f) \in \operatorname{cPol}B_n$.
This brings together Corollaries \[cor:GSM\], \[cor:GMcUinfty\] and \[cor:GMcWinfty\], which will be proved in Section \[sec:constructions\].
\[rem:OldTheorem-theorem:strictCcapAkInclusions\] Lemma \[lem:GCf\] gives rise to a noteworthy refinement of Theorem \[theorem:strictAkInclusions\]. Indeed, by Theorem \[theorem:strictAkInclusions\], there is some $f \in \operatorname{cPol}(\mathcal{A}_k) \setminus \operatorname{cPol}(\mathcal{A}_{k+1})$ and, by Lemma \[lem:GCf\], there exists a function $G_E(f) \in E \subseteq C$ satisfying $G_{E}(f) \in \operatorname{cPol}(\mathcal{A}_k) \setminus \operatorname{cPol}(\mathcal{A}_{k+1})$. This implies that $G_{E}(f) \in (C \cap \operatorname{cPol}(\mathcal{A}_k)) \setminus (C \cap \operatorname{cPol}(\mathcal{A}_{k+1}))$ and thus $C \cap \operatorname{cPol}(\mathcal{A}_{k+1}) \subset C \cap \operatorname{cPol}(\mathcal{A}_k)$ for all $k \geq 2$.
This shows that if $C$ is a clone of Boolean functions satisfying $E \subseteq C$ for some $E \in \{ SM, M_cU_\infty, M_cW_\infty \}$, then $C \cap \operatorname{cPol}(\mathcal{A}_{k+1}) \subset C \cap \operatorname{cPol}(\mathcal{A}_k)$ for all $k \geq 2$.
= \[circle, fill=black,scale=0.5\] = \[scale=2,draw=none, fill=none, label distance=-4\] (P2) \[label=above:$\Omega$\] at (7.0 ,10.0) ; (T0) \[label=above left:$T_0$\] at (7.0-1.5,10.0-.5) ; (T1) \[label=above right:$T_1$\] at (7.0+1.0,10.0-.5) ; (T) at (7.0-.5,10.0-1.0) ; (M) \[label=above left:$M$\] at (6.76 ,9.38) ; (T0M) at (7.23-1.5, 9.51-.5) ; (T1M) at (6.76+1.0, 9.38-.5) ; (TM) at (7.23-.5, 9.51-1.0) ; (L) \[label=above right:$L$\] at (7.0 , 5.2) ; (T0L) at (7.0-1.5, 5.2-.5) ; (T1L) at (7.0+1.0, 5.2-.5) ; (TL) at (7.0-.5, 5.2-1.0) ; (AV) at (7.0 , 3.0) ; (T0AV) at (7.0-1.5, 3.0-.5) ; (T1AV) at (7.0+1.0, 3.0-.5) ; (TAV) at (7.0-.5, 3.0-1.0) ; (P21) \[label=right:[$\Omega(1)$]{}\] at (7.0,3.8) ; (S) \[label=right:$S$\] at (6.85,7.0) ; (SL) at (6.75,4.7) ; (SP21) at (6.65,2.5) ; (ST) at (6.5,6.5) ; (SM) \[label=below left:$SM$\] at (6.15,6.0) ;
/in [ P2/T0, P2/T1, T0/T, T1/T, M/T0M, M/T1M, T0M/TM, T1M/TM, P2/M, T0/T0M, T1/T1M, T/TM, L/T0L, L/T1L, T0L/TL, T1L/TL, P2/L, T0/T0L, T1/T1L, T/ST, ST/TL, AV/T0AV, AV/T1AV, T0AV/TAV, T1AV/TAV, L/P21, P21/AV, T0L/T0AV, T1L/T1AV, TL/TAV, P2/S, S/ST, S/SL, L/SL, SL/TL, SL/SP21, P21/SP21, SP21/TAV, ST/SM, SM/TAV]{} () – ();
(DiffT0xT) at ( 1.0,-0.7); (DiffT0xM) at ( 2.0,-0.3); (DiffT0xTM) at ( 3.0,-1.0); (T02) \[label=left:$U_2$\] at (-0.1, 8.5) ; (T02T) at ($ (T02) + (DiffT0xT) $) ; (T02M) at ($ (T02) + (DiffT0xM) $) ; (T02TM) at ($ (T02) + (DiffT0xTM) $) ; (T03) \[label=left:$U_3$\] at ($ (T02) - ( 0.0, 1.0) $) ; (T03T) at ($ (T03) + (DiffT0xT) $) ; (T03M) at ($ (T03) + (DiffT0xM) $) ; (T03TM) at ($ (T03) + (DiffT0xTM) $) ; (T0H) \[draw=none, fill=none, scale=0.1\] at ($ (T02) - ( 0.0, 1.5) $) ; (T0HT) \[draw=none, fill=none, scale=0.1\] at ($ (T0H) + (DiffT0xT) $) ; (T0HM) \[draw=none, fill=none, scale=0.1\] at ($ (T0H) + (DiffT0xM) $) ; (T0HTM) \[draw=none, fill=none, scale=0.1\] at ($ (T0H) + (DiffT0xTM) $) ; (T0e) \[label=left:$U_{\infty}$\] at ($ (T02) - ( 0.0, 2.3) $) ; (T0eT) at ($ (T0e) + (DiffT0xT) $) ; (T0eM) at ($ (T0e) + (DiffT0xM) $) ; (T0eTM) \[label=right:$M_cU_{\infty}$\] at ($ (T0e) + (DiffT0xTM) $) ;
\(A) \[label=below:$\Lambda$\] at (.5+3.5,3.7) ; (AT1) at ($ (A) + ( 1.0,-0.5) $) ; (AT0) at ($ (A) + (-1.5,-0.5) $) ; (AT) at ($ (A) + (-0.5,-1.0) $) ;
(DiffT1xT) at (-1.0,-0.7); (DiffT1xM) at (-2.0,-0.3); (DiffT1xTM) at (-3.0,-1.0); (T12) \[label=right:$W_2$\] at (13.6, 8.5) ; (T12T) at ($ (T12) + (DiffT1xT) $) ; (T12M) at ($ (T12) + (DiffT1xM) $) ; (T12TM) at ($ (T12) + (DiffT1xTM) $) ; (T13) \[label=right:$W_3$\] at ($ (T12) - ( 0.0, 1.0) $) ; (T13T) at ($ (T13) + (DiffT1xT) $) ; (T13M) at ($ (T13) + (DiffT1xM) $) ; (T13TM) at ($ (T13) + (DiffT1xTM) $) ; (T1H) \[draw=none, fill=none, scale=0.1\] at ($ (T12) - ( 0.0, 1.5) $) ; (T1HT) \[draw=none, fill=none, scale=0.1\] at ($ (T1H) + (DiffT1xT) $) ; (T1HM) \[draw=none, fill=none, scale=0.1\] at ($ (T1H) + (DiffT1xM) $) ; (T1HTM) \[draw=none, fill=none, scale=0.1\] at ($ (T1H) + (DiffT1xTM) $) ; (T1e) \[label=right:$W_{\infty}$\] at ($ (T12) - ( 0.0, 2.3) $) ; (T1eT) at ($ (T1e) + (DiffT1xT) $) ; (T1eM) at ($ (T1e) + (DiffT1xM) $) ; (T1eTM) \[label=left:$M_cW_{\infty}$\] at ($ (T1e) + (DiffT1xTM) $) ;
\(V) \[label=below:$V$\] at (9.5,3.7) ; (VT0) at ($ (V) + (-1.0,-0.5) $) ; (VT1) at ($ (V) + ( 1.5,-0.5) $) ; (VT) at ($ (V) + ( 0.5,-1.0) $) ;
/in [ T02/T02T, T02/T02M, T02T/T02TM, T02M/T02TM, T03/T03T, T03/T03M, T03T/T03TM, T03M/T03TM, T0e/T0eT, T0e/T0eM, T0eT/T0eTM, T0eM/T0eTM, T12/T12T, T12/T12M, T12T/T12TM, T12M/T12TM, T13/T13T, T13/T13M, T13T/T13TM, T13M/T13TM, T1e/T1eT, T1e/T1eM, T1eT/T1eTM, T1eM/T1eTM, T02/T03, T02T/T03T, T02M/T03M, T02TM/T03TM, T03/T0H, T03T/T0HT, T03M/T0HM, T03TM/T0HTM, T12/T13, T12T/T13T, T12M/T13M, T12TM/T13TM, T13/T1H, T13T/T1HT, T13M/T1HM, T13TM/T1HTM, T0eM/AT0, T0eTM/AT, T1eM/VT1, T1eTM/VT, A/AT0, A/AT1, AT0/AT, AT1/AT, V/VT0, V/VT1, VT0/VT, VT1/VT, A/AV, AT0/T0AV, AT1/T1AV, AT/TAV, V/AV, VT0/T0AV, VT1/T1AV, VT/TAV, T02TM/SM, T12TM/SM, T0/T02, T0M/T02M, T/T02T, TM/T02TM, T1/T12, T1M/T12M, T/T12T, TM/T12TM]{} () – (); (M) edge \[out=215, in=70\] (A); (T1M) edge \[out=215, in=90\] (AT1); (M) edge \[out=325, in=110\] (V); (T0M) edge \[out=325, in=90\] (VT0);
/in [ T0H/T0e, T0HT/T0eT, T0HM/T0eM, T0HTM/T0eTM, T1H/T1e, T1HT/T1eT, T1HM/T1eM, T1HTM/T1eTM]{} () – ();
plot coordinates[ ($(AT0) +(-1.5,-0.2)$) ($ (L) +( 0.1, 0.4)$) ($(VT1) +( 1.5,-0.2)$) ]{};
\[thm:classification\] Let $C$ be a clone of Boolean functions. The subclass $C \cap {T}$ of threshold functions is finitely characterizable if and only if $C$ is contained in one of the clones $L$, $V$, $\Lambda$.
This theorem is illustrated by Figure \[figure:PostsLattice\].
We have already observed that $C \cap {T}$ is finitely characterizable for every subclone $C$ of $L$, $V$ or $\Lambda$.
Now we consider all the other clones. Let $C$ be a clone such that $C \not\subseteq D$ for all $D \in \{L, V, \Lambda\}$. We can read off of Post’s lattice (see Figure \[figure:PostsLattice\]) that there is some $E \in \{SM, M_cU_\infty, M_cW_\infty\}$ such that $E \subseteq C$. It follows from Theorem \[theorem:strictAkInclusions\] and Lemma \[lem:GCf\] that for every $k \geq 2$, there exists a function $f_k \in E$ such that $f_k \in \operatorname{cPol}B_\ell$ whenever $2 \leq \ell \leq k$ and $f_k \notin \operatorname{cPol}B_{k+1}$. Note that $f_k \notin C \cap {T}$.
Suppose, on the contrary that $C \cap {T}$ is finitely characterizable. By Theorem \[thm:finitelycharacterizablefunctions\], $C \cap {T}$ is of the form $\operatorname{forbid}(A)$ for some finite antichain $A$ of minimal forbidden minors. Each one of the functions $f_k$ has a minor in $A$. Since $A$ is finite, there is an element $g \in A$ and an infinite set $S \subseteq {\mathbb{N}}$ such that $g \leq f_k$ for all $k \in S$. The function $g$ is not threshold, so there exists $p \in {\mathbb{N}}$ such that $p \geq 2$ and $g \notin \operatorname{cPol}B_p$. Being infinite, the set $S$ contains an element $q$ with $p \leq q$. Then we have $g \leq f_q$ and $f_q \in \operatorname{cPol}B_p$. We also have $g \in \operatorname{cPol}B_p$, because $\operatorname{cPol}B_p$ is closed under taking minors. This yields the desired contradiction.
Alternatively, Theorem \[thm:classification\] can be proved using Lemma \[lemma:finiteDescendingChainIffFinitelyCharacterizable\] and Remark \[rem:OldTheorem-theorem:strictCcapAkInclusions\].
As before if $C$ is a subclone of $L$, $V$ or $\Lambda$, then $C \cap {T}$ is finitely characterizable. As for any other clone $C$, we know (once again reading off of Post’s lattice) that there is some $E \in \{SM, M_cU_\infty, M_cW_\infty\}$ such that $E \subseteq C$. By Remark \[rem:OldTheorem-theorem:strictCcapAkInclusions\], we have $C \cap \operatorname{cPol}(\mathcal{A}_{k+1}) \subset C \cap \operatorname{cPol}(\mathcal{A}_k)$ for all $k \geq 2$. Furthermore, $$C \cap {T}= C \cap \bigcap_{n \geq 2} \operatorname{cPol}(\mathcal{A}_k)
= \bigcap_{n \geq 2} (C \cap \operatorname{cPol}(\mathcal{A}_k)),$$ i.e., we have an infinite descending chain the intersection of which equals $C \cap {T}$. By Lemma \[lemma:finiteDescendingChainIffFinitelyCharacterizable\] we thus conclude that $C \cap {T}$ is not finitely characterizable.
Constructions {#sec:constructions}
=============
In order to prove Lemma \[lem:GCf\], we will construct from a given Boolean function $f$, for each $C \in \{S, M_c, SM, U_\infty, M_cU_\infty, M_cW_\infty \}$, a Boolean function $G_C(f)$ that satisfies the following conditions:
1. $G_C(f) \in C$,
2. for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_C(f) \in \operatorname{cPol}B_\ell$.
We do this step by step. We first construct functions $G_S(f)$ and $G_{M_c}(f)$ with the desired properties. Using these two constructions as building blocks, we can construct $G_{SM}$ as $G_{M_c}(G_S(f))$. Then we construct $G_{U_\infty}(f)$, and, building upon this, we finally get $G_{M_cU_\infty}(f) := G_{U_\infty}(G_{M_c}(f))$ and $G_{M_cW_\infty}(f) := {{(G_{M_cU_\infty}(f))}^{\mathrm{d}}}$.
Construction of $G_{S}(f)$ {#subsec:ST}
--------------------------
Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. Then we define $G_{S}(f) \colon {\mathbb{B}}^{n+1} \to {\mathbb{B}}$ by $$G_{S}(f)(x_1, \dots, x_{n+1}) = (x_{n+1} \wedge f(x_1, \dots, x_n)) \vee ({\overline{x}}_{n+1} \wedge {{f}^{\mathrm{d}}}(x_1, \dots, x_n)).$$
\[lemma:anytoS\] For any $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, the function $G_{S}(f)$ is self-dual.
Let $g := G_{S}(f)$. Then $$\begin{aligned}
{{g}^{\mathrm{d}}}(\mathbf{x}, x_{n+1}) &=
{\overline{({\overline{x}}_{n+1} \wedge f({\overline{\mathbf{x}}})) \vee ({\overline{{\overline{x}}}}_{n+1} \wedge {{f}^{\mathrm{d}}}({\overline{\mathbf{x}}}))}} \\
&= (x_{n+1} \vee {{f}^{\mathrm{d}}}(\mathbf{x})) \wedge ({\overline{x}}_{n+1} \vee f(\mathbf{x})) \\
&= (x_{n+1} \wedge {\overline{x}}_{n+1}) \vee (x_{n+1} \wedge f(\mathbf{x})) \vee ({{f}^{\mathrm{d}}}(\mathbf{x}) \wedge {\overline{x}}_{n+1}) \vee ({{f}^{\mathrm{d}}}(\mathbf{x}) \wedge f(\mathbf{x})) \\
&= (x_{n+1} \wedge f(\mathbf{x})) \vee ({{f}^{\mathrm{d}}}(\mathbf{x}) \wedge {\overline{x}}_{n+1}) \\
&= g(\mathbf{x}, x_{n+1}),
\end{aligned}$$ where the second last equality holds since $${{f}^{\mathrm{d}}}(\mathbf{x}) \wedge f(\mathbf{x}) \leq (x_{n+1} \wedge f(\mathbf{x})) \vee ({{f}^{\mathrm{d}}}(\mathbf{x}) \wedge {\overline{x}}_{n+1})$$ for every $x_{n+1}$.
\[lemma:ST:fNotinBlimpliesgNotinBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. If $f \notin \operatorname{cPol}B_\ell$ for some $\ell \geq 2$, then $G_{S}(f) \notin \operatorname{cPol}B_\ell$.
Assume that $f \notin \operatorname{cPol}B_\ell$, and let $g := G_{S}(f)$. Then there are $\yy_1, \dots, \yy_n \in R(B_\ell)$ with $f(\yy_1, \dots, \yy_n) \notin S(B_\ell)$. Since $g(x_1, \dots, x_n, 1) = f(x_1, \dots, x_n)$, we have $$g(\yy_1,\dots,\yy_n,\11) = f(\yy_1,\dots,\yy_n) \notin S(B_\ell).$$ Since also $\11 \in R(B_\ell)$, we conclude that $g \notin \operatorname{cPol}B_\ell$.
\[lemma:ST:fInBlImpliesgInBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. If $f \in \operatorname{cPol}B_\ell$ for some $\ell \geq 2$, then $G_{S}(f) \in \operatorname{cPol}B_\ell$.
Let $g := G_{S}(f)$. Suppose, on the contrary, that $g \notin \operatorname{cPol}B_\ell$. Then there is some matrix $M$ given by $$M = \begin{pmatrix}
m_1^1 & m_2^1 & \dots & m_{n+1}^1 \\
m_1^2 & m_2^2 & \dots & m_{n+1}^2 \\
\vdots & \vdots & & \vdots \\
m_1^{2\ell} & m_2^{2\ell} & \dots & m_{n+1}^{2\ell}
\end{pmatrix} = \begin{pmatrix}
M^1 \\
M^2 \\
\vdots \\
M^{2\ell}
\end{pmatrix}
= (M_1,M_2,\dots,M_{n+1}),$$ i.e., $M^1,\dots,M^{2\ell} \in {\mathbb{B}}^{n+1}$ are the rows of $M$, and $M_1,\dots,M_{n+1} \in {\mathbb{B}}^{2\ell}$ are the columns of $M$, such that
- $M_1,\dots,M_{n+1} \in R(B_\ell)$, and
- $\zz := g(M_1,\dots,M_{n+1}) := \begin{pmatrix} g(M^1) \\ \vdots \\ g(M^{2\ell}) \end{pmatrix}
\notin S(B_\ell)$.
Thus $\zz \in \{ (\underbrace{0,\dots,0}_\ell,\underbrace{1,\dots,1}_\ell),
(\underbrace{1,\dots,1}_\ell, \underbrace{0,\dots,0}_\ell)\}$. As $B_\ell$ is invariant under swapping the first $\ell$ coordinates with the last $\ell$ coordinates, we can assume that $\zz = (\underbrace{0,\dots,0}_\ell, \underbrace{1,\dots,1}_\ell)$.
We now look at the last column $M_{n+1}$ of $M$. Since $\sum_{i=1}^\ell m_{n+1}^i = \sum_{i=\ell+1}^{2\ell} m_{n+1}^i$, and since $B_\ell$ is totally symmetric on the first $\ell$ rows and on the last $\ell$ rows, we can assume that $$M_{n+1} = (\underbrace{0,\dots,0}_\alpha, \underbrace{1,\dots,1}_\beta,
\underbrace{0,\dots,0}_\alpha, \underbrace{1,\dots,1}_\beta)$$ holds for some $\alpha,\beta \geq 0$ with $\alpha+\beta = \ell$.
We will now construct a matrix $K$ with $$K = \begin{pmatrix}
k_1^1 & k_2^1 & \dots & k_n^1 \\
k_1^2 & k_2^2 & \dots & k_n^2 \\
\vdots & \vdots & & \vdots \\
k_1^{2\ell} & k_2^{2\ell} & \dots & k_n^{2\ell}
\end{pmatrix} = \begin{pmatrix}
K^1 \\
K^2 \\
\vdots \\
K^{2\ell}
\end{pmatrix}
= (K_1,K_2,\dots,K_n),$$ that satisfies $K_1, \dots, K_n \in R(B_\ell)$ and $f(K_1, \dots, K_n) \notin S(B_\ell)$. This will yield the desired contradiction since we started with the assumption that $f \in \operatorname{cPol}B_\ell$.
We define $k_j^i$ for $1 \leq i \leq 2\ell$ and $1 \leq j \leq n$ by $$\def\arraystretch{1.2}
k_j^i = \left\{\begin{array}{llr@{{} \leq i \leq {}}l}
\overline{m}_j^{i+\ell} & \text{if } & 1 & \alpha \\
m_j^i & \text{if } & \alpha+1 & \ell \\
\overline{m}_j^{i-\ell} & \text{if } & \ell+1 & \ell+\alpha \\
m_j^i & \text{if } & \ell+\alpha+1 & 2\ell
\end{array}\right.$$ In other words, matrix $K$ is obtained from $M$ by omitting the last column, negating rows $1, \dots, \alpha$ and $\ell + 1, \dots, \ell + \alpha$, and then swapping rows $1, \dots, \alpha$ with rows $\ell + 1, \dots, \ell + \alpha$.
We need to show that $K_j \in R(B_\ell)$ for all $j \in [n]$. Let $j \in [n]$ be arbitrary, and let $$a := \sum_{i=1}^\alpha m_j^i, \qquad
b := \sum_{i=\alpha+1}^\ell m_j^i, \qquad
c := \sum_{i=\ell+1}^{\ell+\alpha} m_j^i, \qquad
d := \sum_{i=\ell+\alpha+1}^{2\ell} m_j^i.$$ Since $M_j \in R(B_\ell)$ we have $$\label{equation:ST:abEqualscd}
a+b = \sum_{i=1}^\ell m_j^i = \sum_{i=\ell+1}^{2\ell} m_j^i = c + d.$$ For $K_j$ we find the following: $$\begin{aligned}
\sum_{i=1}^\alpha k_j^i &
= \sum_{i=1}^\alpha \overline{m}_j^{i+\ell}
= \sum_{i=1}^\alpha (1-m_j^{i+\ell})
= \alpha - \sum_{i=\ell+1}^{\ell+\alpha} m_j^i
= \alpha - c, \displaybreak[0]\\
\sum_{i=\alpha+1}^\ell k_j^i & = \sum_{i=\alpha+1}^\ell m_j^i = b, \displaybreak[0]\\
\sum_{i=\ell+1}^{\ell+\alpha} k_j^i &
= \sum_{i=\ell+1}^{\ell+\alpha} \overline{m}_j^{i-\ell}
= \sum_{i=\ell+1}^{\ell+\alpha} (1-m_j^{i-\ell})
= \alpha - \sum_{i=1}^{\alpha} m_j^i
= \alpha - a, \displaybreak[0]\\
\sum_{i=\ell+\alpha+1}^{2\ell} k_j^i & = \sum_{i=\ell+\alpha+1}^{2\ell} m_j^i = d.
\end{aligned}$$ From this it follows that $$\sum_{i=1}^\ell k_j^i = \alpha - c + b \overset{\eqref{equation:ST:abEqualscd}}{=}
\alpha - a + d = \sum_{i=\ell+1}^{2\ell} k_j^i,$$ and thus $K_j \in R(B_\ell)$ for all $j \in [n]$.
We now show that $f(K^i) = 0$ if $1 \leq i \leq \ell$, and $f(K^i) = 1$ if $\ell+1 \leq i \leq 2\ell$. We need to consider four different cases for $i$:
- $1 \leq i \leq \alpha$. Then $(\overline{K}^i,0) = M^{i+\ell}$, and $$\overline{f(K^i)} = {{f}^{\mathrm{d}}}(\overline{K}^i) = g(\overline{K}^i, 0) = g(M^{i+\ell}) = 1.$$ Hence $f(K^i) = 0$.
- $\alpha+1 \leq i \leq \ell$. Then $(K^i,1) = M^i$, and $$f(K^i) = g(K^i,1) = g(M^i) = 0.$$
- $\ell+1 \leq i \leq \ell+\alpha$. Then $(\overline{K}^i,0) = M^{i-\ell}$, and $$\overline{f(K^i)} = {{f}^{\mathrm{d}}}(\overline{K}^i) = g(\overline{K}^i, 0) = g(M^{i-\ell}) = 0.$$ Hence $f(K^i) = 1$.
- $\ell+\alpha+1 \leq i \leq 2\ell$. Then $(K^i,1) = M^i$, and $$f(K^i) = g(K^i,1) = g(M^i) = 1.$$
Thus we have $$f(K_1,\dots,K_n) = \begin{pmatrix} f(K^1) \\ \vdots \\ f(K^\ell) \\ f(K^{\ell+1}) \\ \vdots \\ f(K^{2\ell}) \end{pmatrix}
= \begin{pmatrix} 0 \\ \vdots \\ 0 \\ 1 \\ \vdots \\1 \end{pmatrix} \notin S(B_\ell),$$ in contradiction to $f \in \operatorname{cPol}B_\ell$. We conclude that $g \in \operatorname{cPol}B_\ell$.
\[cor:GS\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_S(f) \in S$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_S(f) \in \operatorname{cPol}B_\ell$.
This brings together Lemmas \[lemma:anytoS\], \[lemma:ST:fNotinBlimpliesgNotinBl\] and \[lemma:ST:fInBlImpliesgInBl\].
Construction of $G_{M_c}(f)$ and $G_{SM}(f)$ {#subsec:SMT}
--------------------------------------------
Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. We define the Boolean function $G_{M_c}(f) \colon {\mathbb{B}}^{2n} \to {\mathbb{B}}$ by the following rules
- If $w(\xx) < n$, then $G_{M_c}(f)(\xx) := 0$.
- If $w(\xx) > n$, then $G_{M_c}(f)(\xx) := 1$.
- If $\xx = (\aa, \overline{\aa})$ for some $\aa \in {\mathbb{B}}^n$, then $G_{M_c}(f)(\xx) := f(\aa)$.
- If $w(\xx) = n$ and there exists $i \in [n]$ such that $x_i = x_{n+i}$ and $x_j \neq x_{n+j}$ for all $j < i$, then $G_{M_c}(f)(\xx) := x_i$.
It is easy to verify that the function $G_{M_c}(f)$ is defined on every tuple $\xx \in {\mathbb{B}}^{2n}$.
\[lemma:StoSM\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$.
1. \[lemma:StoSM:item1\] $G_{M_c}(f) \in M_c$, i.e., $G_{M_c}(f)$ is monotone and constant-preserving.
2. \[lemma:StoSM:item2\] If $f$ is self-dual, then $G_{M_c}(f)$ is self-dual.
Let $g := G_{M_c}(f)$.
Let $\xx,\yy \in {\mathbb{B}}^{2n}$ with $\xx < \yy$. Then $w(\xx) < w(\yy)$ and one of the following cases applies: $w(\xx) < n$ or $w(\yy) > n$. In the former case, we have $g(\xx) = 0 \leq g(\yy)$; in the latter case, we have $g(\xx) \leq 1 = g(\yy)$. We conclude that $g$ is monotone. Since $w(\00) = 0 < n$ and $w(\11) = 1 > n$, it holds that $f(\00) = 0$ and $f(\11) = 1$, i.e., $f$ preserves both constants.
Assume that $f$ is self-dual. Let $\xx \in {\mathbb{B}}^{2n}$.
If $w(\xx) > n$ then $w(\overline{\xx}) < n$, and thus $(g(\xx),g(\overline{\xx})) = (1,0)$. Similarly, if $w(\xx) < n$ then $w(\overline{\xx}) > n$, and thus $(g(\xx),g(\overline{\xx})) = (0,1)$.
If $\xx = (\aa, \overline{\aa})$ for some $\aa \in {\mathbb{B}}^n$, then $(g(\xx), g(\overline{\xx})) = (f(\aa), f(\overline{\aa})) \in \{ (0,1),(1,0) \}$ since $f$ is self-dual.
Otherwise, there is some $i \in [m]$ with $x_i = x_{m+i}$ and $x_j \neq x_{m+j}$ for all $j < i$. This holds also for the negation of $\xx$, and thus $(g(\xx), g(\overline{\xx})) \in \{ (0,1),(1,0) \}$.
We conclude that $g$ is self-dual.
\[lemma:SM:fnotinBlimpliesgnotinBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. If $f \notin \operatorname{cPol}B_\ell$ for some $\ell \geq 2$, then $G_{M_c}(f) \notin \operatorname{cPol}B_\ell$.
Let $f \notin \operatorname{cPol}B_\ell$ and $g := G_{M_c}(f)$. Then there are $\yy_1,\dots,\yy_n \in R(B_\ell)$ with $f(\yy_1,\dots,\yy_n) \notin S(B_\ell)$. Also $\overline{\yy}_1,\dots,\overline{\yy}_n \in R(B_\ell)$ and thus $$g(\yy_1,\dots,\yy_n,\overline{\yy}_1,\dots,\overline{\yy}_n) = f(\yy_1,\dots,\yy_n) \notin S(B_\ell).$$ Therefore $g \notin \operatorname{cPol}B_\ell$.
\[lemma:SM:finBlimpliesginBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$ with $f \in \operatorname{cPol}B_\ell$ for some $\ell \geq 2$. Then $G_{M_c}(f) \in \operatorname{cPol}B_\ell$.
Let $g := G_{M_c}(f)$. Suppose, on the contrary, that $g \notin \operatorname{cPol}B_\ell$. Then there is some matrix $M$ given by $$M = \begin{pmatrix}
m_1^1 & m_2^1 & \dots & m_{2n}^1 \\
m_1^2 & m_2^2 & \dots & m_{2n}^2 \\
\vdots & \vdots & & \vdots \\
m_1^{2\ell} & m_2^{2\ell} & \dots & m_{2n}^{2\ell}
\end{pmatrix} = \begin{pmatrix}
M^1 \\
M^2 \\
\vdots \\
M^{2\ell}
\end{pmatrix}
= (M_1,M_2,\dots,M_{2n}),$$ i.e., $M^1,\dots,M^{2\ell} \in {\mathbb{B}}^{2n}$ are the rows of $M$, and $M_1,\dots,M_{2n} \in {\mathbb{B}}^{2\ell}$ are the columns of $M$, such that
- $M_1,\dots,M_{2n} \in R(B_\ell)$, and
- $\zz := g(M_1,\dots,M_{2n}) := \begin{pmatrix} g(M^1) \\ \vdots \\ g(M^{2\ell}) \end{pmatrix}
\notin S(B_\ell)$.
Thus $\zz \in \{ (\underbrace{0,\dots,0}_\ell,\underbrace{1,\dots,1}_\ell),
(\underbrace{1,\dots,1}_\ell,\underbrace{0,\dots,0}_\ell)\}$. As $B_\ell$ is invariant under swapping the first $\ell$ coordinates with the last $\ell$ coordinates, we can assume that $\zz = (\underbrace{0,\dots,0}_\ell,\underbrace{1,\dots,1}_\ell)$.
We have the following possibilities for $M^i$ with $1 \leq i \leq 2\ell$:
1. \[enum:finQlimpliesginQl:smallweight\] $w(M^i) \neq n$;
2. \[enum:finQlimpliesginQl:equalpairs\] $w(M^i) = n$ and there is some $b \in [n]$ with $m^i_b = m^i_{n+b}$;
3. \[enum:finQlimpliesginQl:fromf\] $w(M^i) = n$ and $m^i_b \neq m^i_{n+b}$ for all $b \in [n]$, i.e., there is some $\aa_i \in {\mathbb{B}}^n$ with $M^i = (\aa_i,\overline{\aa_i})$.
We show that case \[enum:finQlimpliesginQl:smallweight\] cannot happen, since the weight of each row $M^i$ of $M$ is exactly $n$. Since $g(M^i) = 0$ for $1 \leq i \leq \ell$, we have $w(M^i) \leq n$ for $1 \leq i \leq \ell$. Similarly, we have $w(M^i) \geq n$ for $\ell+1 \leq i \leq 2\ell$. Thus $\sum_{i=1}^\ell w(M^i) \leq n\ell$ and $\sum_{i=\ell+1}^{2\ell} w(M^i) \geq n\ell$. Because $M_j \in R(B_\ell)$ for $1 \leq j \leq 2n$, we get $$\sum_{i=1}^\ell w(M^i) = \sum_{i=1}^\ell \sum_{j=1}^{2n} m_j^i
= \sum_{j=1}^{2n} \sum_{i=1}^\ell m_j^i
= \sum_{j=1}^{2n} \sum_{i=\ell+1}^{2\ell} m_j^i
= \sum_{i=\ell+1}^{2\ell} \sum_{j=1}^{2n} m_j^i
= \sum_{i=\ell+1}^{2\ell} w(M^i)$$ Therefore $\sum_{i=1}^\ell w(M^i) = \sum_{i=\ell+1}^{2\ell} w(M^i) = n\ell$, and $w(M^i) = n$ for $1 \leq i \leq 2\ell$. Thus the case \[enum:finQlimpliesginQl:smallweight\] cannot happen for $M^i$.
We will show that case \[enum:finQlimpliesginQl:equalpairs\] is also not possible. Suppose, on the contrary, that there is some $i \in [2\ell]$ and some $b \in [n]$ such that $m^i_b = m^i_{n+b}$, and $m^i_a \neq m^i_{n+a}$ for all $a < b$. We can assume that $b$ is the smallest number with this property.
Now we consider the weights of $M_b$ and $M_{n+b}$. Because $b$ is minimal, we have that $m^{i'}_a \neq m^{i'}_{n+a}$ for all $a < b$. Thus we have $(m^{i'}_b, m^{i'}_{n+b}) \in \{ (0,0),(0,1),(1,0) \}$ for $1 \leq i' \leq \ell$, and $(m^{i'}_b, m^{i'}_{n+b}) \in \{ (0,1),(1,0),(1,1) \}$ for $\ell+1 \leq i' \leq 2\ell$. Then $$\begin{aligned}
\sum_{i' = 1}^{\ell} (m^{i'}_b + m^{i'}_{n+b}) &
\leq n, \\
\sum_{i' = \ell+1}^{2\ell} (m^{i'}_b + m^{i'}_{n+b}) &
\geq n,
\end{aligned}$$ and at least one of these inequalities holds strictly. This implies that one of the following holds: $$\begin{aligned}
\sum_{i' = 1}^\ell m^{i'}_b & < \sum_{i' = \ell+1}^{2\ell} m^{i'}_b \quad \text{or} \\
\sum_{i' = 1}^\ell m^{i'}_{n+b} & < \sum_{i' = \ell+1}^{2\ell} m^{i'}_{n+b}.
\end{aligned}$$ This means that $M_b \notin R(B_\ell)$ or $M_{n+b} \notin R(B_\ell)$, in contradiction to the assumption. Thus no such $b$ exists, and case \[enum:finQlimpliesginQl:equalpairs\] cannot happen.
Thus case \[enum:finQlimpliesginQl:fromf\] applies for all $M^i$, i.e., $M^i = (\aa_i,\overline{\aa_i})$ for some $\aa_i \in {\mathbb{B}}^n$ holds for all $i \in [2\ell]$. By the definition of $g$ and since $f \in \operatorname{cPol}B_\ell$, we obtain $$\zz = g\begin{pmatrix} M^1 \\ \vdots \\ M^{2\ell} \end{pmatrix}
= f\begin{pmatrix} \aa_1 \\ \vdots \\ \aa_{2\ell} \end{pmatrix}
= f(M_1,\dots,M_n)
\in S(B_\ell).$$ But this is a contradiction to $\zz \notin S(B_\ell)$. Thus the matrix $M$ cannot exist, and we have $g \in \operatorname{cPol}B_\ell$.
\[cor:GMc\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_{M_c}(f) \in M_c$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_{M_c}(f) \in \operatorname{cPol}B_\ell$.
This brings together Lemmas \[lemma:StoSM\]\[lemma:StoSM:item1\], \[lemma:SM:fnotinBlimpliesgnotinBl\] and \[lemma:SM:finBlimpliesginBl\].
Let $G_{SM}(f) := G_{M_c}(G_{S}(f))$. Then we can conclude the following corollary from the preceding lemmas.
\[cor:GSM\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_{SM}(f) \in SM$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_{SM}(f) \in \operatorname{cPol}B_\ell$.
By Corollary \[cor:GS\], we have $G_{S}(f) \in S$, and by Lemma \[lemma:StoSM\], we get $G_{SM}(f) = G_{M_c}(G_S(f)) \in SM$.
By Corollary \[cor:GS\], the condition $f \in \operatorname{cPol}B_\ell$ is equivalent to $G_S(f) \in \operatorname{cPol}B_\ell$, which is in turn equivalent to $G_{SM}(f) = G_{M_c}(G_S(f)) \in \operatorname{cPol}B_\ell$ by Corollary \[cor:GMc\].
Construction of $G_{U_\infty}(f)$, $G_{M_cU_\infty}(f)$ and $G_{M_cW_\infty}(f)$ {#subsec:MCT}
--------------------------------------------------------------------------------
Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. Define $G_{U_\infty}(f) \colon {\mathbb{B}}^{n+1} \to {\mathbb{B}}$ by $$G_{U_\infty}(f)(x_1, \dots, x_{n+1}) = x_{n+1} \wedge f(x_1, \dots, x_n).$$
\[lemma:SMtoMcUinfty\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$.
1. \[lemma:SMtoMcUinfty:item1\] $G_{U_\infty}(f) \in U_\infty$.
2. \[lemma:SMtoMcUinfty:item2\] If $f$ is monotone, then $G_{U_\infty}(f)$ is monotone.
3. \[lemma:SMtoMcUinfty:item3\] If $f(\11) = 1$, then $G_{U_\infty}(f)$ preserves both constants.
4. \[lemma:SMtoMcUinfty:item4\] If $f \in M_c$, then $G_{U_\infty}(f) \in M_cU_\infty$.
Let $g := G_{U_\infty}(f)$.
\[lemma:SMtoMcUinfty:item1\] By the definition of $g$ we have that if $g(x_1,\dots,x_{n+1}) = 1$ then $x_{n+1} = 1$. Thus $g \in U_\infty$.
\[lemma:SMtoMcUinfty:item2\] Let $\xx, \yy \in {\mathbb{B}}^{n+1}$, and assume that $\xx < \yy$. If $x_{n+1} = 0$, then $g(\xx) = 0 \leq g(\yy)$. If $x_{n+1} = 1$, then also $y_{n+1} = 1$, and since $f$ is monotone, we have $$g(\xx) = f(x_1, \dots, x_n) \leq f(y_1, \dots, y_n) = g(\yy).$$ We conclude that $g$ is monotone.
\[lemma:SMtoMcUinfty:item3\] By the definition of $g$, we have $g(\00) = 0$. Furthermore, if $f(\11) = 1$, then we have $g(\11) = f(\11) = 1$.
\[lemma:SMtoMcUinfty:item4\] Follows immediately from the previous items.
\[lemma:McUinfty:fnotinBlimpliesgnotinBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. If $f \notin \operatorname{cPol}B_\ell$ for some $\ell \geq 2$, then $G_{U_\infty}(f) \notin \operatorname{cPol}B_\ell$.
The proof is exactly the same as the proof of Lemma \[lemma:ST:fNotinBlimpliesgNotinBl\].
\[lemma:McUinfty:finBlimpliesginBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. If $f \in \operatorname{cPol}B_\ell$ for some $\ell \geq 2$, then $G_{U_\infty}(f) \in \operatorname{cPol}B_\ell$.
Let $g := G_{U_\infty}(f)$.
Suppose, on the contrary, that $g \notin \operatorname{cPol}B_\ell$. Then there is some matrix $M$ given by $$M = \begin{pmatrix}
m_1^1 & m_2^1 & \dots & m_{n+1}^1 \\
m_1^2 & m_2^2 & \dots & m_{n+1}^2 \\
\vdots & \vdots & & \vdots \\
m_1^{2\ell} & m_2^{2\ell} & \dots & m_{n+1}^{2\ell}
\end{pmatrix} = \begin{pmatrix}
M^1 \\
M^2 \\
\vdots \\
M^{2\ell}
\end{pmatrix}
= (M_1,M_2,\dots,M_{n+1}),$$ i.e., $M^1,\dots,M^{2\ell} \in {\mathbb{B}}^{n+1}$ are the rows of $M$, and $M_1,\dots,M_{n+1} \in {\mathbb{B}}^{2\ell}$ are the columns of $M$, such that
- $M_1,\dots,M_{n+1} \in R(B_\ell)$, and
- $\zz := g(M_1,\dots,M_{n+1}) := \begin{pmatrix} g(M^1) \\ \vdots \\ g(M^{2\ell}) \end{pmatrix}
\notin S(B_\ell)$.
Thus $\zz \in \{ (\underbrace{0,\dots,0}_\ell,\underbrace{1,\dots,1}_\ell),
(\underbrace{1,\dots,1}_\ell,\underbrace{0,\dots,0}_\ell)\}$. As $B_\ell$ is invariant under swapping the first $\ell$ coordinates with the last $\ell$ coordinates, we can assume that $\zz = (\underbrace{0,\dots,0}_\ell,\underbrace{1,\dots,1}_\ell)$.
We now look at the last column $M_{n+1}$ of $M$. Since $\sum_{i=1}^\ell m_{n+1}^i = \sum_{i=\ell+1}^{2\ell} m_{n+1}^i$, and since $B_\ell$ is totally symmetric on the first $\ell$ rows and on the last $\ell$ rows, we can assume that $$M_{n+1} = (\underbrace{0,\dots,0}_\alpha, \underbrace{1,\dots,1}_\beta,
\underbrace{0,\dots,0}_\alpha, \underbrace{1,\dots,1}_\beta)$$ holds for some $\alpha,\beta \geq 0$ with $\alpha+\beta = \ell$.
If $\alpha > 0$ then $g(M^{\ell+1}) = g(m_1^{\ell+1},\dots,m_n^{\ell+1},0) = 0 \land f(m_1^{\ell+1},\dots,m_n^{\ell+1}) = 0$, in contradiction to $g(M^{\ell+1}) = 1$. Thus $\alpha = 0$, and $M_{n+1} = \11$. But then $f(M_1,\dots,M_n) = g(M_1,\dots,M_n,\11) = \zz \notin S(B_\ell)$, which implies that $f \notin \operatorname{cPol}B_\ell$. This contradicts the assumption $f \in \operatorname{cPol}B_\ell$, and we conclude that $g \in \operatorname{cPol}B_\ell$.
\[cor:GUinfty\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_{U_\infty}(f) \in U_\infty$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_{U_\infty}(f) \in \operatorname{cPol}B_\ell$.
This brings together Lemmas \[lemma:SMtoMcUinfty\]\[lemma:SMtoMcUinfty:item1\], \[lemma:McUinfty:fnotinBlimpliesgnotinBl\] and \[lemma:McUinfty:finBlimpliesginBl\].
Let $G_{M_cU_\infty}(f) := G_{U_\infty}(G_{M_c}(f))$ and $G_{M_cW_\infty}(f) := {{G_{M_cU_\infty}(f)}^{\mathrm{d}}}$.
\[cor:GMcUinfty\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_{M_cU_\infty}(f) \in M_cU_\infty$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_{M_cU_\infty}(f) \in \operatorname{cPol}B_\ell$.
By Corollary \[cor:GMc\], we have $G_{M_c}(f) \in M_c$, and by Lemma \[lemma:SMtoMcUinfty\], we get $G_{M_cU_\infty}(f) = G_{U_\infty}(G_{M_c}(f)) \in M_cU_\infty$.
By Corollary \[cor:GMc\], the condition $f \in \operatorname{cPol}B_\ell$ is equivalent to $G_{M_c}(f) \in \operatorname{cPol}B_\ell$, which in turn is equivalent to $G_{M_cU_\infty}(f) = G_{U_\infty}(G_{M_c}(f)) \in \operatorname{cPol}B_\ell$ by Corollary \[cor:GUinfty\].
Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$. We define the functions $\overline{f} \colon {\mathbb{B}}^n \to {\mathbb{B}}$ and $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, for $\uu \in {\mathbb{B}}^n$, as $$\begin{aligned}
\overline{f}(\aa) &= \overline{f(\aa)}, \\
f^\uu(\aa) &= f(\aa \oplus \uu).\end{aligned}$$ Note that ${{f}^{\mathrm{d}}} = \overline{f^\11}$, where $\11 := (1, \dots, 1) \in {\mathbb{B}}^n$.
\[lem:dualBl\] Let $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, and let $\ell \geq 2$. The following are equivalent:
1. $f \in \operatorname{cPol}B_\ell$,
2. $f^\uu \in \operatorname{cPol}B_\ell$ for any $\uu \in {\mathbb{B}}^n$,
3. $\overline{f} \in \operatorname{cPol}B_\ell$,
4. ${{f}^{\mathrm{d}}} \in \operatorname{cPol}B_\ell$.
$\text{(i)} \iff \text{(ii)}$: Let $\aa^1, \dots \aa^n \in R(B_\ell)$. Since $R(B_\ell)$ is invariant under taking negations of its members, we also have $\overline{\aa^1}, \dots \overline{\aa^n} \in R(B_\ell)$. Let $\uu \in {\mathbb{B}}^n$, and let $\bb^i := \aa^i$ if $u_i = 0$ and $\bb^i := \overline{\aa^i}$ if $u_i = 1$, for $i \in {[{n}]}$. If $f \in \operatorname{cPol}B_\ell$, then $$f^\uu(\aa^1, \dots, \aa^n) = f(\bb^1, \dots, \bb^n) \in S(B_\ell);$$ hence $f^\uu \in \operatorname{cPol}B_\ell$. The converse implication holds, since $(f^\uu)^\uu = f$.
$\text{(i)} \iff \text{(iii)}$: Assume that $f \in \operatorname{cPol}B_\ell$, and let $\aa^1, \dots \aa^n \in R(B_\ell)$. Then $f(\aa^1, \dots, \aa^n) \in S(B_\ell)$. Since $S(B_\ell)$ is invariant under taking negations of its members, we have $$\overline{f}(\aa^1, \dots, \aa^n)
= \overline{f(\aa^1, \dots, \aa^n)}
\in S(B_\ell);$$ hence $\overline{f} \in \operatorname{cPol}B_\ell$. The converse implication holds, since $\overline{\overline{f}} = f$.
$\text{(i)} \iff \text{(iv)}$: This follows immediately from the equivalence of (i), (ii) and (iii), because ${{f}^{\mathrm{d}}} = \overline{f^\11}$.
\[cor:GMcWinfty\] For any Boolean function $f \colon {\mathbb{B}}^n \to {\mathbb{B}}$, $G_{M_cW_\infty}(f) \in M_cW_\infty$ and for all $\ell \geq 2$, $f \in \operatorname{cPol}B_\ell$ if and only if $G_{M_cW_\infty}(f) \in \operatorname{cPol}B_\ell$.
Since $M_cW_\infty = \{{{f}^{\mathrm{d}}} : f \in M_cU_\infty\}$, the claim follows from Lemma \[lem:dualBl\] and Corollary \[cor:GMcUinfty\].
Simple games and magic squares revisited {#sec:magic}
========================================
In their proof of the existence of $k$-asummable functions that are not $(k+1)$-asummable (see Theorem \[theorem:strictAkInclusions\]), Taylor and Zwicker constructed a certain family of functions [@Taylor]. We recall their construction here, and then we will refine Theorem \[theorem:strictAkInclusions\] and determine how the sets $\operatorname{cPol}B_n$ are related to each other. We will also show that Taylor and Zwicker’s functions actually constitute an antichain of minimally non-threshold functions.
Fix an integer $k \geq 3$. For $p, q \in {[{k}]}$, define the $k \times k$ matrix $A^{p,q} = (a_{i,j})$ as follows: $$a_{i,j} =
\begin{cases}
k - 1, & \text{if $(i,j) = (p,q)$,} \\
1, & \text{if $i \neq p$ and $j \neq q$,} \\
0, & \text{otherwise.}
\end{cases}$$ For example, if $k = 4$, then $A^{2,3} = \begin{pmatrix} 1 & 1 & 0 & 1 \\ 0 & 0 & 3 & 0 \\ 1 & 1 & 0 & 1 \\ 1 & 1 & 0 & 1 \end{pmatrix}$. Let $B$ be the $k \times k$ matrix all of whose entries are equal to $k - 1$.
Let $S$ be a subset of ${[{k}]} \times {[{k}]}$. We refer to $S$ as the $i$-th *row* if $S = \{(i, j) : j \in {[{k}]}\}$, and we refer to $S$ as the $j$-th *column* if $S = \{(i, j) : i \in {[{k}]}\}$.
\[lem:roworcolumn\] Let $S \subseteq {[{k}]} \times {[{k}]}$. Then $\sum_{(p,q) \in S} A^{p,q} = B$ if and only if $S$ is a row or a column.
It is clear that if $S$ is a row or a column, then $\sum_{(p,q) \in S} A^{p,q} = B$.
Assume then that $\sum_{(p,q) \in S} A^{p,q} = B$. Clearly $S$ is nonempty, so choose an element $(p,q)$ of $S$; clearly $S$ contains another element $(p',q')$. If $p \neq p'$ and $q \neq q'$, then the entry on row $p$ column $q$ in the sum $\sum_{(p,q) \in S} A^{p,q}$ is at least $k$; hence the sum cannot be equal to $B$. Thus either $p = p'$ or $q = q'$. It is easy to see that in the former case, all remaining entries of $S$ must be on the $p$-th row, and all elements of the $p$-th row must be in $S$; in the latter case, all remaining entries of $S$ must be on the $q$-th column, and all elements of the $q$-th column must be in $S$. We conclude that $S$ is either a row or a column.
We define a function $\phi \colon {[{R}]}^{k \times k} \to {\mathbb{N}}$ that maps each $k \times k$ matrix with entries in ${[{R}]}$ to an integer, where $R$ is a sufficiently large integer that will be specified below. The function $\phi$ is defined as follows: for a matrix $M$, read the entries of $M$ from left to right and from top to bottom; the resulting string is the representation of $\phi(M)$ in base $R$. For $p, q \in {[{k}]}$, denote $w^{p,q} := \phi(A^{p,q})$ and $t := \phi(B)$. For example if $k = 4$, then $w^{2,3} = 1101003011011101_R$ and $t = 3333333333333333_R$. We must choose $R$ in such a way that when we add these numbers to form the sum $\sum_{(p,q) \in S} w^{p,q}$ for any $S \subseteq {[{k}]} \times {[{k}]}$, no carry will occur. Thus, the number $(k-1)^2 + (k-1) + 1 = k^2 - k + 1$, or anything larger, would be fine.
It is easy to see that the function $\phi$ has the following preservation property: for any $S \subseteq {[{k}]} \times {[{k}]}$, $\phi(\sum_{(p,q) \in S} A^{p,q}) = \sum_{(p,q) \in S} \phi(A^{p,q})$. It thus follows from Lemma \[lem:roworcolumn\] that for all $S \subseteq {[{k}]} \times {[{k}]}$, it holds that $\sum_{(p,q) \in S} w^{p,q} = t$ if and only if $S$ is a row or a column.
Fix a bijection $\beta \colon {[{k}]} \times {[{k}]} \to {[{k^2}]}$. The *characteristic tuple* of a subset $S$ of ${[{k}]} \times {[{k}]}$ is the tuple $\ee_S \in {\mathbb{B}}^{k^2}$, whose $i$-th entry is $1$ if $i = \beta(p,q)$ for some $(p,q) \in S$ and $0$ otherwise. With no risk of confusion, we will refer to the characteristic tuples of rows and columns also as *rows* and *columns,* respectively.
Let $\ww = (w^{\beta^{-1}(1)}, \dots, w^{\beta^{-1}(k^2)})$.
For any $n$-tuples $\aa, \bb \in {\mathbb{R}}^n$, the *dot product* is defined as $$\aa \cdot \bb = \sum_{i=1}^n a_i b_i.$$
Taylor and Zwicker’s function $f_k \colon {\mathbb{B}}^{k^2} \to {\mathbb{B}}$ is defined by the following rule: $f_k(\xx) = 1$ if and only if $\xx \cdot \ww > t$ or $\xx$ is a row. Note that for all $\xx \in {\mathbb{B}}^{k^2}$, $\xx \cdot \ww = t$ if and only if $\xx$ is a row or a column.
\[lem:fkBl\] Let $k \geq 3$ and $\ell \geq 2$. Then $f_k$ preserves $B_\ell$ if and only if $k$ is not a divisor of $\ell$.
If $\ell = mk$ for some integer $m$, then let $\aa^1, \dots, \aa^\ell$ comprise $m$ occurrences of each column, and let $\bb^1, \dots, \bb^\ell$ comprise $m$ occurrences of each row. Then, the $\aa^i$ are false points of $f_k$ and the $\bb^i$ are true points, and $\aa^1 + \dots + \aa^\ell = (m, \dots, m) = \bb^1 + \dots + \bb^\ell$. Thus $f_k$ is not $\ell$-asummable. Lemma \[lem:PolBl\] implies that $f_k$ does not preserve $B_\ell$.
Assume then that $k$ is not a divisor of $\ell$. Suppose, on the contrary, that $f_k$ does not preserve $B_\ell$. By Lemma \[lem:PolBl\], there exist $\aa^1, \dots, \aa^\ell \in f^{-1}(0)$ and $\bb^1, \dots, \bb^\ell \in f^{-1}(1)$ such that $\aa^1 + \dots + \aa^\ell = \bb^1 + \dots + \bb^\ell$. Since $\xx \cdot \ww \leq t$ for any false point $\xx$ of $f_k$, and $\xx \cdot \ww \geq t$ for any true point $\xx$, we have $$\sum_{i = 1}^\ell \aa^i \cdot \ww \leq \ell t
\qquad
\text{and}
\qquad
\sum_{i = 1}^\ell \bb^i \cdot \ww \geq \ell t.$$ On the other hand, since $\aa^1 + \dots + \aa^\ell = \bb^1 + \dots + \bb^\ell$, we have $$\sum_{i = 1}^\ell \aa^i \cdot \ww
= (\aa^1 + \dots + \aa^\ell) \cdot \ww
= (\bb^1 + \dots + \bb^\ell) \cdot \ww
= \sum_{i = 1}^\ell \bb^i \cdot \ww.$$ Consequently, $\aa^i \cdot \ww = t$ and $\bb^i \cdot \ww = t$ for all $i \in {[{\ell}]}$, and we conclude that each $\aa^i$ is a column and each $\bb^i$ is a row. Since $k$ is not a divisor of $\ell$, there necessarily exist two columns that have a different number of occurrences among $\aa^1, \dots, \aa^\ell$. Then $\phi^{-1}(\aa^1 + \dots + \aa^n)$ is a matrix that is constant along each column, but there are two columns with distinct values. This contradicts the fact that the matrix $\phi^{-1}(\bb^1 + \dots + \bb^n)$ is constant along each row. This completes the proof, and we conclude that $f_k$ preserves $B_\ell$.
\[lem:mod2sum\] The modulo-$2$ addition operation ${\oplus}$ preserves $B_\ell$ if and only if $\ell$ is odd.
The false points of $\oplus$ are $(0,0)$ and $(1,1)$, while the true points are $(0,1)$ and $(1,0)$. Hence the sum of any $\ell$ false points is of the form $(m,m)$ for some $m$ with $0 \leq m \leq \ell$. The sum of any $\ell$ true points is of the form $(m, \ell - m)$ for some $m$ with $0 \leq m \leq \ell$.
If $\ell$ is odd, then $m \neq \ell - m$ for any $m$. It follows that $\aa^1 + \dots + \aa^\ell \neq \bb^1 + \dots + \bb^\ell$ for any false points $\aa^1, \dots, \aa^\ell$ and any true points $\bb^1, \dots, \bb^\ell$. By Lemma \[lem:PolBl\], $\oplus$ preserves $B_\ell$.
If $\ell$ is even, say $\ell = 2k$, then $$\begin{gathered}
\underbrace{(0,0) + \dots + (0,0)}_k + \underbrace{(1,1) + \dots + (1,1)}_k = \\
\underbrace{(0,1) + \dots + (0,1)}_k + \underbrace{(1,0) + \dots + (1,0)}_k.\end{gathered}$$ By Lemma \[lem:PolBl\], $\oplus$ does not preserve $B_\ell$.
Let $\ell, m \geq 2$. Then $\operatorname{cPol}B_\ell \subseteq \operatorname{cPol}B_m$ if and only if $m$ divides $\ell$.
Assume first that $m$ does not divide $\ell$. If $m \neq 2$, then by Lemma \[lem:fkBl\], $f_m \in \operatorname{cPol}B_\ell$ but $f_m \notin \operatorname{cPol}B_m$. If $m = 2$, then by Lemma \[lem:mod2sum\], ${\oplus} \in \operatorname{cPol}B_\ell$ but ${\oplus} \notin \operatorname{cPol}B_m$. In either case, we conclude that $\operatorname{cPol}B_\ell \not\subseteq \operatorname{cPol}B_m$.
Assume then that $\ell = km$ for some integer $k$. Let $f \in \operatorname{cPol}B_\ell$. Let $\aa^1, \dots, \aa^n \in R(B_m)$. For each $i \in \{1, \dots, n\}$, define the tuple $\bb^i \in {\mathbb{B}}^\ell$ as $$\bb^i = (\underbrace{a^i_1, \dots, a^i_1}_k, \dots, \underbrace{a^i_m, \dots, a^i_m}_k, \underbrace{a^i_{m + 1}, \dots, a^i_{m + 1}}_k, \dots, \underbrace{a^i_{2m}, \dots, a^i_{2m}}_{k}).$$ It is clear that $\bb^i \in R(B_\ell)$. Let $\zz := f(\bb^1, \dots, \bb^n)$, that is, $$\zz = (\underbrace{f(a^1_1, \dots, a^n_1), \dots, f(a^1_1, \dots, a^n_1)}_k, \dots, \underbrace{f(a^1_{2m}, \dots, a^n_{2m}), \dots, f(a^1_{2m}, \dots, a^n_{2m})}_k).$$ Since $f \in \operatorname{cPol}B_\ell$, we have $\zz \in S(R_\ell)$. Then $$\zz \in {\mathbb{B}}^\ell \setminus \{(\underbrace{0, \dots, 0}_\ell, \underbrace{1, \dots, 1}_\ell), (\underbrace{1, \dots, 1}_\ell, \underbrace{0, \dots, 0}_\ell))\}.$$ This implies that $$f(\aa^1, \dots, \aa^n) \in {\mathbb{B}}^m \setminus \{(\underbrace{0, \dots, 0}_m, \underbrace{1, \dots, 1}_m), (\underbrace{1, \dots, 1}_m, \underbrace{0, \dots, 0}_m))\}.$$ Thus $f \in \operatorname{cPol}B_m$, and we conclude that $\operatorname{cPol}B_\ell \subseteq \operatorname{cPol}B_m$.
\[prop:tzantichain\] The functions $f_k$ ($k \geq 3$) are pairwise incomparable by the minor relation.
Let $m \neq n$, and consider the comparability of $f_m$ and $f_n$. Since all variables are essential in $f_m$ and in $f_n$, and the number of essential variables cannot increase when taking minors, we have that $f_m \not\leq f_n$ whenever $m > n$. If $m < n$, then $n$ is not a divisor of $m$ but $n$ is a divisor of itself. By Lemma \[lem:fkBl\], $f_n$ preserves $B_m$ and $f_m$ does not preserve $B_m$. Since every minor of $f_n$ preserves all relational constraints $f_n$ does, we must have that $f_m \not\leq f_n$ also in this case.
\[prop:tzmonotone\] For every $k \geq 3$, the function $f_k$ is monotone.
Let $\xx, \yy \in {\mathbb{B}}^{k^2}$. If $\xx < \yy$, then, since each $w^{p,q}$ is positive, $\xx \cdot \ww < \yy \cdot \ww$. Therefore one of the following conditions holds: $\xx \cdot \ww < t$ or $\yy \cdot \ww > t$. In the former case, $f(\xx) = 0 \leq f(\yy)$. In the latter case, $f(\xx) \leq 1 = f(\yy)$.
\[prop:tzminnonthr\] For every $k \geq 3$, the function $f_k$ is minimally non-threshold.
We need to show that every identification minor of $f_k$ is threshold. Let $(p,q)$ and $(p',q')$ be distinct elements of ${[{k}]} \times {[{k}]}$, let $I = \{\beta(p,q), \beta(p',q')\}$, and assume without loss of generality that $\beta(p,q) < \beta(p',q')$. We will show that $(f_k)_I$ is $\ell$-asummable for every $\ell \geq 2$ and hence threshold by Theorem \[thm:thresholdasummable\]. Let $\ell \geq 2$, and let $\aa^1, \dots, \aa^\ell \in ((f_k)_I)^{-1}(0)$, $\bb^1, \dots, \bb^\ell \in ((f_k)_I)^{-1}(1)$. Suppose, on the contrary, that $\aa^1 + \dots + \aa^\ell = \bb^1 + \dots + \bb^\ell$. Let $\vv \in {\mathbb{B}}^{k^2 - 1}$ be the tuple obtained from $\ww$ by replacing its $\beta(p,q)$-th entry by $w_{\beta(p,q)} + w_{\beta(p',q')}$ and deleting the $\beta(p',q')$-th entry. (Before proceeding, we ask the reader to recall the definition of $\delta_I$ from in Section \[susec:MinorsConstraints\].) It clearly holds that $\xx \cdot \vv = \xx \delta_I \cdot \ww$ for all $\xx \in {\mathbb{B}}^{k^2 - 1}$. Therefore $((f_k)_I)(\xx) = f_k(\xx \delta_I) = 1$ if and only if $\xx \cdot \vv = \xx \delta_I \cdot \ww > t$ or $\xx \delta_I$ is a row. Note that if $\xx \delta_I$ is a row or a column, then $\xx \cdot \vv = \xx \delta_I \cdot \ww = t$. In a similar way as we argued in the proof of Lemma \[lem:fkBl\], we have $$\ell t
\geq \sum_{i = 1}^\ell \aa^i \cdot \vv
= (\aa^1 + \dots + \aa^\ell) \cdot \vv
= (\bb^1 + \dots + \bb^\ell) \cdot \vv
= \sum_{i = 1}^\ell \bb^i \cdot \vv
\geq \ell t.$$ Hence $\aa^i \cdot \vv = t$ and $\bb^i \cdot \vv = t$ for all $i \in {[{\ell}]}$, that is, $\aa^i \delta_I$ is a column and $\bb^i \delta_I$ is a row for all $i \in {[{\ell}]}$.
Since $(p,q) \neq (p',q')$, we have $p \neq p'$ or $q \neq q'$. If $p \neq p'$, then none of the rows $\bb^i \delta_I$ is the $p$-th row; hence $\phi^{-1}(\bb^1 \delta_I + \dots + \bb^\ell \delta_I)$ is a matrix with a row full of $0$’s, while $\phi^{-1}(\aa^1 \delta_I + \dots + \aa^\ell \delta_I)$ has no row full of $0$’s. If $q \neq q'$, then none of the columns $\aa^i \delta_I$ is the $q$-th column; hence $\phi^{-1}(\aa^1 \delta_I + \dots + \aa^\ell \delta_I)$ is a matrix with a column full of $0$’s, while $\phi^{-1}(\bb^1 \delta_I + \dots + \bb^\ell \delta_I)$ has no column full of $0$’s. On the other hand, $$\aa^1 \delta_I + \dots + \aa^\ell \delta_I
= (\aa^1 + \dots + \aa^\ell) \delta_I
= (\bb^1 + \dots + \bb^\ell) \delta_I
= \bb^1 \delta_I + \dots + \bb^\ell \delta_I.$$ We have reached a contradiction.
We conclude that $(f_k)_I$ is $\ell$-asummable for every $\ell \geq 2$ and hence threshold.
Taylor and Zwicker’s functions $f_k$ constitute an infinite antichain of monotone, minimally non-threshold functions (Propositions \[prop:tzantichain\], \[prop:tzmonotone\], \[prop:tzminnonthr\]). It should be noted here that this antichain does not, however, characterize the set of monotone threshold functions in terms of forbidden minors, i.e., $M \cap {T}\neq \operatorname{forbid}(\{f_k : k \geq 3\})$. For example, there exist self-dual monotone non-threshold functions of arity $6$ (see, e.g., [@BioIba]), which clearly fail to have any of the $f_n$ as a minor.
Post classes {#App:Post}
============
We provide a concise description of all clones of Boolean functions as well as characterizing sets of relations $R$ – or, equivalently, relational constraints $(R,R)$ – for some clones; the characterizion of the remaining clones is easily derived by noting that if $C_1 = \operatorname{cPol}(\mathcal{Q}_1)$ and $C_2 = \operatorname{cPol}(\mathcal{Q}_2)$, then $C_1 \cap C_2 = \operatorname{cPol}(\mathcal{Q}_1 \cup \mathcal{Q}_2)$. We make use of notations and terminology appearing in [@FP] and [@JGK].
$\Omega$ denotes the clone of all Boolean functions. It is characterized by the empty relation.
$T_0$ and $T_1$ denote the clones of $0$- and $1$-preserving functions, respectively, i.e., $$T_0 = \{f \in \Omega : f(0, \dots, 0) = 0\}
\quad \text{and} \quad
T_1 = \{f \in \Omega : f(1, \dots, 1) = 1\}.$$ They are characterized by the unary relations $\{0\}$ and $\{1\}$, respectively.
$T_c$ denotes the clone of constant-preserving functions, i.e., $T_c = T_0 \cap T_1$.
$M$ denotes the clone of all monotone functions, i.e., $$M = \{f \in \Omega : \text{$f(\mathbf{a}) \leq f(\mathbf{b})$ whenever $\mathbf{a} \leq \mathbf{b}$}\}.$$ It is characterized by the binary relation ${\leq} := \{(0,0), (0,1), (1,1)\}$.
$M_0 = M \cap T_0$, $M_1 = M \cap T_1$, $M_c = M \cap T_c$.
$S$ denotes the clone of all self-dual functions, i.e., $$S = \{f \in \Omega : {{f}^{\mathrm{d}}} = f\}.$$ It is characterized by the binary relation $\{(0,1),(1,0)\}$.
$S_c = S \cap T_c$, $SM = S \cap M$.
$L$ denotes the clone of all linear functions, i.e., $$L = \{f \in \Omega : f = c_0 \oplus c_1 x_1 \oplus \dots \oplus c_n x_n\}.$$ It is characterized by the quaternary relation $\{(a,b,c,d) \in {\mathbb{B}}^4 : a \oplus b \oplus c = d\}$.
$L_0 = L \cap T_0$, $L_1 = L \cap T_1$, $LS = L \cap S$, $L_c = L \cap T_c$.
Let $a \in \{0,1\}$. A set $A \subseteq \{0,1\}^n$ is said to be *$a$-separating* if there is some $i\in [n]$ such that for every $(a_1, \dotsc, a_n) \in A$ we have $a_i = a$. A function $f$ is said to be *$a$-separating* if $f^{-1}(a)$ is $a$-separating. The function $f$ is said to be *$a$-separating of rank $k \geq 2$* if every subset $A \subseteq f^{-1}(a)$ of size at most $k$ is $a$-separating.
For $m \geq 2$, $U_m$ and $W_m$ denote the clones of all $1$- and $0$-separating functions of rank $m$, respectively. They are characterized by the $m$-ary relations ${\mathbb{B}}^m \setminus \{(0, \dots, 0)\}$ and ${\mathbb{B}}^m \setminus \{(1, \dots, 1)\}$, respectively.
$U_\infty$ and $W_\infty$ denote the clones of all $1$- and $0$-separating functions, respectively, i.e., $U_\infty = \bigcap_{k \geq 2} U_k$ and $W_\infty = \bigcap_{k \geq 2} W_k$.
$T_cU_m = T_c \cap U_m$ and $T_cW_m = T_c \cap W_m$, for $m = 2, \dotsc, \infty$.
$MU_m = M \cap U_m$ and $MW_m = M \cap W_m$, for $m = 2, \dotsc, \infty$.
$M_cU_m = M_c \cap U_m$ and $M_cW_m = M_c \cap W_m$, for $m = 2, \dotsc, \infty$.
$\Lambda $ denotes the clone of all conjunctions and constants, i.e., $$\begin{gathered}
\Lambda = \{f \in \Omega : f = x_{i_1} \wedge \dotsb \wedge x_{i_n}\} \cup
\{\mathbf{0}^{(n)}: n\geq 1\} \cup \{\mathbf{1}^{(n)}: n\geq 1\}.\end{gathered}$$ It is characterized by the ternary relation $\{(a,b,c) : a \wedge b = c\}$.
$\Lambda_0 = \Lambda \cap T_0$, $\Lambda_1 = \Lambda \cap T_1$, $\Lambda_c = \Lambda \cap T_c$.
$V$ denotes the clone of all disjunctions and constants, i.e., $$\begin{gathered}
V = \{f \in \Omega : f = x_{i_1} \vee \dotsb \vee x_{i_n}\} \cup
\{\mathbf{0}^{(n)}: n\geq 1\} \cup \{\mathbf{1}^{(n)}: n\geq 1\}.\end{gathered}$$ It is characterized by the ternary relation $\{(a,b,c) : a \vee b = c\}$.
$V_0 = V \cap T_0$, $V_1 = V \cap T_1$, $V_c = V \cap T_c$.
$\Omega (1)$ denotes the clone of all projections, negations, and constants. It is characterized by the ternary relation $\{(a,b,c) : \text{$a = b$ or $b = c$}\}$.
$I^* = \Omega(1) \cap S$, $I = \Omega(1) \cap M$.
$I_0 = I \cap T_0$, $I_1 = I \cap T_1$.
$I_c$ denotes the smallest clone containing only projections, i.e., $I_c = I \cap T_c$.
[99]{}
Bioch, J. C., Ibaraki, T.: Generating and approximating nondominated coteries, IEEE Trans. Parallel Distrib. Syst. **6**(9), 905–914 (1995)
Bodnarčuk, V. G., Kalužnin, L. A., Kotov, V. N., Romov, B. A.: Galois theory for Post algebras, I, II, Kibernetika **3**, 1–10, **5**, 1–9 (1969) (Russian). English translation: Cybernetics **5**, 243–252, 531–539 (1969)
Chow, C. K.: Boolean functions realizable with single threshold devices, Proc. Institute of Radio Engineers, Vol. 49, Jan. 1961, pp. 370–371
Couceiro, M.: On Galois connections between external functions and relational constraints: arity restrictions and operator decompositions, Acta Sci. Math. (Szeged) **72**, 15–35 (2006)
Couceiro, M., Foldes, S.: On closed sets of relational constraints and classes of functions closed under variable substitution, Algebra Universalis **54**, 149–165 (2005)
Couceiro, M., Lehtonen, E.: On the effect of variable identification on the essential arity of functions on finite sets, Internat. J. Found. Comput. Sci. **18**, 975–986 (2007)
Couceiro, M., Lehtonen, E.: Generalizations of Świerczkowski’s lemma and the arity gap of finite functions, Discrete Math. **309**, 5905–5912 (2009)
Couceiro, M., Marichal, J.-L.: Discrete integrals based on comonotonic modularity, Axioms **2**(3), 390–403 (2013)
Couceiro, M., Pouzet, M.: On a quasi-ordering on Boolean functions, Theoret. Comput. Sci. **396**, 71–87 (2008)
Denecke, K., Erné, M., Wismath, S. L. (eds.), Galois Connections and Applications, Kluwer Academic Publishers, Dordrecht (2004)
Ekin, O., Foldes, S., Hammer, P. L., Hellerstein, L.: Equational characterizations of Boolean function classes, Discrete Math. **211**, 27–51 (2000)
Elgot, C. C.: Truth functions realizable by single threshold organs, AIEE Conf. Paper 60-1311, Oct. 1960, revised Nov. 1960; also IEEE Symposium on Swithing Circuit Theory and Logical Design, Sept. 1961, pp. 225–245
Foldes, S., Pogosyan, G. R.: Post classes characterized by functional terms, Discrete Appl. Math. **142**, 35–51 (2004)
Geiger, D.: Closed systems of functions and predicates, Pacific J. Math. **27**, 95–100 (1968)
Hellerstein, L.: On generalized constraints and certificates, Discrete Math. **226**, 211–232 (2001)
Isbell, J. R.: A class of majority games, Q. J. Math. **7**(1), 183–187 (1956)
Jablonski, S. W., Gawrilow, G. P., Kudrjawzew, W. B.: Boolesche Funktionen und Postsche Klassen, Vieweg, Braunschweig, 1970.
Jeroslow, R. G.: On defining sets of vertices of the hypercube by linear inequalities, Discrete Math. **11**, 119–124 (1975)
Lau, D.: Function Algebras on Finite Sets, Springer-Verlag, Berlin, Heidelberg (2006)
Lay, S. R.: Convex sets and their applications, Dover (2007).
Lovász, L.: Submodular functions and convexity, Mathematical programming, 11th int. Symp., Bonn 1982, pp. 235–257 Muroga, S.: Threshold Logic and Its Applications, Wiley-Interscience, New York (1971)
Peleg, B.: A theory of coalition formation in committees, J. Math. Econom. **7**(2), 115–134 (1980)
Peleg, B.: Coalition formation in simple games with dominant players, Internat. J. Game Theory **10**, 11–33 (1981)
Pippenger, N.: Galois theory for minors of finite functions, Discrete Math. **254**, 405–419 (2002)
Pöschel, R.: Concrete representation of algebraic structures and a general Galois theory. In: Kautschitsch, H., Müller, W. B., Nöbauer, W. (eds.), Contributions to General Algebra (Proc. Klagenfurt Conf., 1978), pp. 249–272, Verlag Johannes Heyn, Klagenfurt (1979)
Post, E.: The Two-Valued Iterative Systems of Mathematical Logic, Annals of Mathematical Studies, vol. 5, Princeton University Press, Princeton, NJ (1941)
Singer, I.: Extensions of functions of 0-1 variables and applications to combinatorial optimization, Numer. Funct. Anal. Optim. **7**, 23–62 (1984)
Taylor, A., Zwicker, W.: Simple games and magic squares, J. Combin. Theory Ser. A **71** 67–88 (1995)
Topkis, D. M.: Minimizing a submodular function on a lattice, Oper. Res. **26**(2), 305–321 (1978)
Winder, R. O.: Threshold Logic, Ph.D. thesis, Mathematics Department, Princeton University (1962)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Latent block models are used for probabilistic biclustering, which is shown to be an effective method for analyzing various relational data sets. However, there has been no statistical test method for determining the row and column cluster numbers of latent block models. Recent studies have constructed statistical-test-based methods for stochastic block models, which assume that the observed matrix is a square symmetric matrix and that the cluster assignments are the same for rows and columns. In this study, we developed a new goodness-of-fit test for latent block models to test whether an observed data matrix fits a given set of row and column cluster numbers, or it consists of more clusters in at least one direction of the row and the column. To construct the test method, we used a result from the random matrix theory for a sample covariance matrix. We experimentally demonstrated the effectiveness of the proposed method by showing the asymptotic behavior of the test statistic and measuring the test accuracy.'
author:
- 'Chihiro Watanabe[^1]'
- 'Taiji Suzuki[^2]'
bibliography:
- 'paper.bib'
title: 'Goodness-of-fit Test for Latent Block Models'
---
Introduction
============
Block modeling [@Hartigan1972; @Arabie1978] is known to be effective in representing various relational data sets, such as the data sets of movie ratings [@Shan2008], customer-product transactions [@Shan2008], congressional voting [@Keribin2015], document-word relationships [@Dhillon2001], and gene expressions [@Pontes2015]. Latent block models or LBMs [@Govaert2003] are used for probabilistic biclustering of such relational data matrices, where rows and columns represent different objects. For instance, suppose that a matrix $A = (A_{ij})_{ij} \in \mathbb{R}^{n \times p}$ represents the relationship between users and movies, where entry $A_{ij}$ is the rating of the $j$-th movie by the $i$-th user. In LBMs, we assume a regular-grid block structure behind the observed matrix $A$; i.e., both rows (users) and columns (movies) of matrix $A$ are simultaneously decomposed into latent clusters. A block is defined as a combination of row and column clusters, and entries of the same block in matrix $A$ are supposed to be i.i.d. random variables.
An open problem in using LBMs is that there has been no statistical criterion for determining the numbers of row and column clusters. Recently, statistical-test-based approaches [@Bickel2016; @Lei2016; @Hu2017] have been proposed for estimating the cluster number of stochastic block models (SBMs) [@Holland1983]. SBMs are similar to LBMs in the sense that they assume a block structure behind an observed matrix; however, they are based on different assumptions from LBMs that an observed matrix is a square symmetric matrix and that the cluster assignments are the same for rows and columns [@Mariadassou2015]. In regard to the LBM setting, no statistical method has been constructed to determine row and column cluster numbers.
Aside from the test-based methods, several model selection approaches have been proposed based on cross-validation [@Chen2018] or an information criterion [@Keribin2012; @Peixoto2013; @Keribin2015]. However, these approaches have several limitations. (1) First, they cannot provide knowledge about the reliability of the result besides the finally estimated cluster numbers. Rather than minimizing the generalization error, in some cases, it is more appropriate to provide a probabilistic guarantee in reliability for the purpose of knowledge discovery. (2) Second, both the cross-validation-based and information-criterion-based methods depend on the clustering algorithm used. For instance, we can employ the Bayesian information criterion (BIC) for estimating the marginal likelihood only if the Fisher information matrix of the model is regular, which is not the case for block models. Constructing an information criterion that estimates the expectation of the generalization error for a wider class of models is generally difficult. (3) Finally, the above methods require relatively large computational complexity. Computation of an information criterion requires the process of approximating the posterior distribution by the Markov chain Monte Carlo (MCMC) method, and cross-validation requires the iterative calculation of the test error with different sets of partitions of the training and test data sets.
In this study, we proposed a new statistical test method for LBMs. To construct a hypothesis test with a theoretical guarantee, we used a result from random matrix theory. Recent studies on random matrix theory have revealed the asymptotic behavior of singular values of an $n \times p$ random matrix [@Geman1980; @Silverstein1985; @Yin1988; @Bai1993; @Johansson2000; @Johnstone2001; @Soshnikov2002; @Peche2009; @Pillai2014]. Here, we assume that each entry $Z_{ij}$ of matrix $Z$, which is given by $Z_{ij} = (A_{ij} - P_{ij})/\sigma_{ij}$ (which is computed by the original matrix $A$, its block-wise mean $P$ and standard deviation $\sigma$) follows a distribution with a sub-exponential decay. From the result in [@Pillai2014], the normalized maximum eigenvalue of $Z^{\top} Z$ converges in law to the Tracy-Widom distribution with index $1$, under the above sub-exponential condition. Based on this result, we constructed a goodness-of-fit test for a given set of row and column cluster numbers of an LBM, using the maximum singular value of matrix $\hat{Z}$, which is an estimator of the matrix $Z$. We proved that under the null hypothesis (i.e., observed matrix $A$ consists of a given set of row and column cluster numbers), the proposed test statistic $T$ converges in law to the Tracy-Widom distribution with index $1$ (Theorem \[th:realizable\]). We also showed that under the alternative hypothesis, test statistic $T$ increases in proportion to $m^{\frac{5}{3}}$ with a high probability, where $m$ is a number proportional to the matrix size (Theorems \[th:unrealizable\_lower\] and \[th:unrealizable\_upper\]).
The proposed method solves the limitations of other model selection approaches. (1) Our statistical test method enables us to obtain knowledge about the reliability of the test results. When testing a given set of row and column cluster numbers, we can explicitly set the probability of Type I error (or false positive) as a significance level $\alpha$. (2) Unlike the other model selection methods, the proposed method does not depend on the clustering algorithm as long as it satisfies the consistency condition (Section \[sec:method\]). It only uses the output of a clustering algorithm to test a given set of cluster numbers; there is no need to modify the test method according to the clustering algorithm. (3) The proposed test method requires relatively small computational complexity. It does not require the MCMC procedure or partitioning into the training and test data sets. For these reasons, the proposed test-based method can be widely used for the purpose of knowledge discovery.
The next sections consist of the detailed explanation of the proposed test method for LBMs. In Section \[sec:method\], we describe the proposed goodness-of-fit test and its theoretical guarantee with the assumptions required for the problem setting. Next, we briefly review the related works and their differences from the proposed method in Section \[sec:related\]. The main results are presented in Section \[sec:statistic\], where we prove the asymptotic properties of the proposed test statistic. In Section \[sec:experiments\], we experimentally demonstrate the effectiveness of the proposed test method by showing the asymptotic behavior of the test statistic and calculating the test accuracy. We discuss the results and limitations of the proposed method in Section \[sec:discussion\] and conclude the paper in Section \[sec:conclusion\].
Problem setting and statistical model for goodness-of-fit test for latent block models {#sec:method}
======================================================================================
Let $A \in \mathbb{R}^{n \times p}$ be an $n \times p$ observed matrix. We assume that each entry of matrix $A$ is independently generated, given its row and column clusters. Let $g^{(1)}_i$ be the cluster of the $i$-th row, and $g^{(2)}_j$ be the cluster of the $j$-th column of matrix $A$. To apply the result in [@Pillai2014], we assume that each entry of matrix $A$ is independently subject to a distribution with block-wise mean $P$ and block-wise standard deviation $\sigma$: $$\begin{aligned}
\label{eq:LBM}
&&P = (P_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
P_{ij} = B_{g^{(1)}_i g^{(2)}_j}. \nonumber \\
&&\sigma = (\sigma_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
\sigma_{ij} = S_{g^{(1)}_i g^{(2)}_j}. \nonumber \\
&&A = (A_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
\mathbb{E} [A_{ij}] = P_{ij}, \ \ \ \ \ \ \ \ \ \
\mathbb{E} [(A_{ij} - P_{ij})^2] = \sigma_{ij}^2, \end{aligned}$$ where $B_{k h}$ and $S_{k h} > 0$, respectively, are the mean and the positive standard deviation of entries in the $(k, h)$-th true block under the null hypothesis.
Let $(K, H)$ be the true set of cluster numbers for rows and columns of an observed matrix $A$, which is unknown in advance. In this paper, we propose a goodness-of-fit test for selecting these cluster numbers from observed matrix $A$. In such a test, we test whether $(K, H)$ is equal to a given set of cluster numbers $(K_0, H_0)$ or at least one of the given row and column cluster numbers $K_0$ or $H_0$ is smaller than the true cluster numbers $K$ or $H$. In other words, the null (N) and alternative (A) hypotheses are given by $$\begin{aligned}
\label{eq:hypothesis}
\mathrm{(N) :}\ (K, H) = (K_0, H_0), \ \ \ \ \ \ \ \ \ \
\mathrm{(A) :}\ K > K_0\ \mathrm{or}\ H > H_0. \end{aligned}$$ By sequentially testing the cluster numbers in the following order (Figure \[fig:order\_test\]), we can select the cluster numbers of a given observed matrix $A$.
1. Test $(K_0, H_0) = (1, 1)$.
2. Test $(K_0, H_0) = (1, 2), (2, 1)$.
3. Test $(K_0, H_0) = (1, 3), (2, 2), (3, 1)$.
4. $\cdots$
5. Test $(K_0, H_0) = (1, L), (2, L-1), \dots, (L, 1)$. Let $(\hat{K}, \hat{H})$ be the row and column cluster numbers where the null hypothesis is accepted and $\hat{K} + \hat{H} = L+1$ holds. The selected set of cluster numbers is $(\hat{K}, \hat{H})$.
Based on the above sequentially ordered test, selection of the cluster numbers requires $(K+H)(K+H-1)/2$ tests at most.
![The sequential order for testing row and column cluster numbers. For example, let the blue entry $(4, 3)$ be the true cluster numbers $(K, H)$. Based on this sequentially ordered test, the given cluster numbers $(K_0, H_0)$ are always unrealizable (that is, at least one of $K>K_0$ or $H>H_0$ holds), until it reaches to $(K, H)$. []{data-label="fig:order_test"}](order_test.pdf){width="50mm"}
#### Assumptions.
Throughout this paper, we make the following assumptions to derive the test statistics:
1. We assume that a distribution of $Z_{ij}$, which is given by $Z_{ij} = (A_{ij} - P_{ij})/\sigma_{ij}$ as in (\[eq:Z\_true\]) later, has a sub-exponential decay. That is, there exists some $\vartheta >0$ such that for $x>1$, $\mathrm{Pr} \left( \left| Z_{ij} \right| > x \right) \leq \vartheta^{-1}\exp (-x^{\vartheta})$. From this assumption, note that the fourth moment of a random variable $Z_{ij}$ is finite (i.e., $\mathbb{E} [Z_{ij}^4] < \infty$).
2. We denote the number of rows and columns of matrix $A$ as $n$ and $p$, respectively. We assume that both $n$ and $p$ increase in proportion to some sufficiently large number $m$ (i.e., $n, p \propto m$).
3. Let $K$ and $H$, respectively, be the minimum row and column cluster numbers to represent the block structure of observed matrix $A$ under the null hypothesis. We assume that both $K$ and $H$ are finite constants that do not increase with the matrix sizes $n$ and $p$. We also assume that the minimum row and column sizes of a block in the true block structure, which we denote as $n_{\mathrm{min}}$ and $p_{\mathrm{min}}$, respectively, satisfy $n_{\mathrm{min}} = \Omega_p (m)$ and $p_{\mathrm{min}} = \Omega_p (m)$, where we used the following notation: $$\begin{aligned}
X = \Omega_p \left( f(m) \right). \ \ \ \Leftrightarrow \ \ \
\forall \epsilon>0, \ \exists C>0, M>0, \ \forall m \geq M, \ \mathrm{Pr} \left(C f(m) \leq X \right) \geq 1-\epsilon. \end{aligned}$$ In other words, we assume that with high probability, there is no “too small” block in matrix $A$. \[asmp:block\_size\]
4. If the given set of cluster numbers $(K_0, H_0)$ is equal to the true cluster numbers $(K, H)$, then we call it a *realizable* case. Otherwise, we call it an *unrealizable* case ($K>K_0$ or $H>H_0$). Here, we only consider the cases where $K_0 \leq K$ and $H_0 \leq H$.
5. In the realizable case, we assume that a clustering algorithm is *consistent*, that is, the probability that it outputs the correct block structure converges to $1$, in the limit of $m \to \infty$. By using this assumption, the proposed method does not depend on a specific clustering algorithm. Several clustering algorithms including [@Flynn2012; @Ames2014; @Brault2016] have been proven to be consistent. \[asmp:consistency\]
Relation to existing works {#sec:related}
==========================
In this section, we briefly review the related works and explain the differences between them and the proposed method.
Model selection for block models
--------------------------------
#### Statistical-test-based methods (for SBM)
Recently, several methods have been proposed for testing the properties of a given observed matrix in relation to SBMs [@Bickel2016; @Lei2016; @Karwa2016; @Hu2017; @Yuan2018]. Particularly, the methods proposed in [@Bickel2016; @Lei2016; @Hu2017] have enabled us to estimate the number of blocks for SBMs. However, these methods differ from ours in the problem setting; they can be applied only to an SBM setting, where an observed matrix is a square symmetric matrix, and the cluster assignments are the same for rows and columns. There has been no method for estimating the block number for LBMs, where rows and columns (not necessarily square) of an observed matrix are simultaneously decomposed into clusters.
#### Cross-validation-based methods
Cross-validation is a widely used method for model selection, where a data set is first split into training and test data sets, and then the best model with the minimum test error is determined. Recently, cross-validation methods for matrix data have been proposed [@Dabbs2016; @Li2016; @Kawamoto2017; @Chen2018] to determine the number of clusters in network data. Although the purpose of these methods and our method is similar, these methods differ from ours in that their target is the network data, where the observed matrix is square and its rows and columns represent the same node sets. Thus, the block structure is symmetric regardless of whether the network itself is directed or undirected). Moreover, unlike a statistical test, these methods cannot provide quantitative knowledge about the reliability of the selected model. Furthermore, the computational cost of cross-validation is generally high because it requires the iterative calculation of the test error with different data set partitions.
#### Information-criterion-based methods
Another approach for determining the number of blocks in a matrix is to estimate the generalization error or marginal likelihood by some information criterion for given sets of block numbers. By using such information criteria, we can select a model in a statistically meaningful (non-heuristic) way. In regard to block models, many variants of BIC have been proposed [@Keribin2012; @Peixoto2013; @Keribin2015; @Hu2016; @Saldana2017]. Unlike our test-based method, which only requires a clustering algorithm to satisfy the consistency condition (Section \[sec:method\]), an information criterion for a theoretical guarantee should be carefully chosen according to the given clustering algorithm. For instance, BIC can be employed for estimating the marginal likelihood only if the Fisher information matrix of the model is regular, which is not the case for block models.
To solve this problem, as an alternative criterion to BIC, the integrated completed likelihood (ICL) criterion has been used in many studies for estimating the number of blocks in LBMs [@Lomet2012; @Wyse2014; @Corneli2015]. In ICL, we first derive a marginal likelihood for a given set of an observed matrix and block assignments and then substitute the set of estimated block assignments to approximate the marginal likelihood. However, since ICL is computed based on a single estimator of block assignments, there is no guarantee for the goodness of the approximation of marginal likelihood.
Similar to cross-validation-based methods, information-criterion-based methods cannot provide a probabilistic guarantee for the reliability of the selected model, which is a disadvantage for the purpose of knowledge discovery. The computational cost also becomes a problem because the computation of an information criterion requires the process of approximating the posterior distribution by MCMC.
#### Other model selection methods
Aside from the information criteria, several studies have proposed to determine the number of blocks in LBMs based on the co-clustering adjusted rand index [@Robert2017], the extended modularity for biclustering [@Labiod2011], or the expected posterior loss for a given loss function [@Rastelli2018]. Another approach is to define the posterior distribution not only on cluster assignments of rows and columns but also on row and column cluster numbers [@Wyse2012; @Passino2019]. Unlike the model selection approaches, such nonparametric Bayesian methods can estimate the distribution of the block numbers. The best-fitted number of the blocks can be determined based on the posterior distribution (e.g., we can choose a MAP estimator [@Passino2019]). However, in this case, the computational cost of MCMC is higher than that of the information-criterion-based methods because it requires a large number of iterations to approximate the posterior distribution both on the block assignments and the number of blocks.
Test statistic for determining the set of cluster numbers {#sec:statistic}
=========================================================
To derive the test statistic for the proposed goodness-of-fit test, we first normalize each entry $A_{ij}$ of an observed matrix $A$ by subtracting $P_{ij}$ and dividing it by $\sigma_{ij}$, where $P$ and $\sigma$, respectively, are the block-wise mean and standard deviation in (\[eq:LBM\]): $$\begin{aligned}
\label{eq:Z_true}
Z = (Z_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
Z_{ij} = \frac{A_{ij} - P_{ij}}{\sigma_{ij}}. \end{aligned}$$ By definition, each entry $Z_{ij}$ of matrix $Z$ in (\[eq:Z\_true\]) independently follows a distribution with zero mean and standard deviation of one. Therefore, according to the result in [@Pillai2014], if $n = n(p)$ and $n/p \to \gamma \neq 0, \infty$ in the limit of $p \to \infty$, the scaled maximum eigenvalue $T^*$ of matrix $Z^\top Z$ converges in law to the Tracy-Widom distribution with index $1$ ($TW_1$) in the limit of $p \to \infty$: $$\begin{aligned}
\label{eq:T_true}
T^* = \frac{\lambda_1 - a}{b}, \ \ \ \ \ \ \ \ \ \
T^* \rightsquigarrow TW_1 \ \mathrm{(Convergence\ in\ law)}, \end{aligned}$$ where $\lambda_1$ is the maximum eigenvalue of matrix $Z^\top Z$ and $$\begin{aligned}
\label{eq:ab}
a = (\sqrt{n} + \sqrt{p})^2, \ \ \ \ \ \ \ \ \ \
b = (\sqrt{n} + \sqrt{p}) \left( \frac{1}{\sqrt{n}} + \frac{1}{\sqrt{p}} \right)^{\frac{1}{3}}. \end{aligned}$$
In most cases, the true cluster numbers $(K, H)$ and the true cluster assignments $g^{(1)}$ and $g^{(2)}$ are unknown in advance. Therefore, we can only estimate the block structure based on the observed matrix $A$ and the given cluster numbers. Let $(K_0, H_0)$ be the given set of row and column cluster numbers, and $\hat{g}^{(1)}$ and $\hat{g}^{(2)}$, respectively, be the estimated cluster assignments for rows and columns. Based on such an estimated block structure $(\hat{g}^{(1)}, \hat{g}^{(2)})$, we estimate the block-wise mean and standard deviation by $$\begin{aligned}
\label{eq:BPS_hat}
&&\hat{B} = (\hat{B}_{k h})_{kh}, \ \ \ \ \ \ \ \ \ \
\hat{B}_{k h} = \frac{1}{|I_{k}| |J_{h}|} \sum_{i \in I_{k}, j \in J_{h}} A_{ij}, \nonumber \\
&&\hat{P} = (\hat{P}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
\hat{P}_{ij} = \hat{B}_{\hat{g}^{(1)}_i \hat{g}^{(2)}_j}, \nonumber \\
&&\hat{S} = (\hat{S}_{k h})_{kh}, \ \ \ \ \ \ \ \ \ \
\hat{S}_{k h} = \sqrt{ \frac{1}{|I_{k}| |J_{h}|} \sum_{i \in I_{k}, j \in J_{h}} \left( A_{ij} - \hat{P}_{ij} \right)^2}, \nonumber \\
&&\hat{\sigma} = (\hat{\sigma}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
\hat{\sigma}_{ij} = \hat{S}_{\hat{g}^{(1)}_i \hat{g}^{(2)}_j}, \end{aligned}$$ where $I_k$ is the set of row indices of matrix $A$ that are assigned to the $k$-th cluster, and $J_h$ is the set of column indices of matrix $A$ that are assigned to the $h$-th cluster: $$\begin{aligned}
\label{eq:I_k_J_h}
I_k = \left\{ i: \hat{g}^{(1)}_i = k \right\}, \ \ \ \ \ \ \ \ \ \
J_h = \left\{ j: \hat{g}^{(2)}_j = h \right\}. \end{aligned}$$ The consistency assumption \[asmp:consistency\] guarantees that if $(K_0, H_0) = (K, H)$, the probability that the cluster assignments $(I_k)_k$ and $(J_h)_h$ are correct converges to $1$ in the limit of $m \to \infty$.
We define an estimator of normalized matrix $Z$ in (\[eq:Z\_true\]) based on the estimated block-wise mean $\hat{P}$ and standard deviation $\hat{\sigma}$ in (\[eq:BPS\_hat\]): $$\begin{aligned}
\label{eq:Z_hat}
\hat{Z} = (\hat{Z}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \
\hat{Z}_{ij} = \frac{A_{ij} - \hat{P}_{ij}}{\hat{\sigma}_{ij}}. \end{aligned}$$
The test statistic $T$ for the proposed goodness-of-fit test is given by the scaled maximum eigenvalue of matrix $\hat{Z}^\top \hat{Z}$: $$\begin{aligned}
\label{eq:T_statistic}
T = \frac{\hat{\lambda}_1 - a}{b}, \end{aligned}$$ where $\hat{\lambda}_1$ is the maximum eigenvalue of matrix $\hat{Z}^\top \hat{Z}$, and $a$ and $b$ are given by (\[eq:ab\]).
We assume that the following condition holds: $n = n(p)$, $n/p \to \gamma \neq 0, \infty$ in the limit of $p \to \infty$. Under the consistency assumption \[asmp:consistency\] for the clustering algorithm, if $(K_0, H_0) = (K, H)$, $$\begin{aligned}
T \rightsquigarrow TW_1 \ \mathrm{(Convergence\ in\ law)},\end{aligned}$$ in the limit of $p \to \infty$, where $T$ is defined as in (\[eq:T\_statistic\]). \[th:realizable\]
We denote the operator norm by $\left\| \cdot \right\|_{\mathrm{op}}$, $$\begin{aligned}
\left\| A \right\|_{\mathrm{op}} = \sup_{u \in \mathbb{R}^p} \frac{\| Au \|}{\| u \|}. \end{aligned}$$
From now on, we use the following notation:
- $B_{kh}$ and $S_{kh}$, respectively, are the constant elements in the $(k, h)$-th **true** blocks of mean and standard deviation matrices $P$ and $\sigma$.
- $\tilde{B}_{kh}$ and $\tilde{S}_{kh}$, respectively, are the constant elements in the $(k, h)$-th **true** blocks of mean and standard deviation matrices $\tilde{P}$ and $\tilde{\sigma}$, which are calculated based on the observed matrix $A$ and the correct cluster assignments. That is, $\tilde{B}_{kh}$ and $\tilde{S}_{kh}$, respectively, are the sample mean and standard deviation of all the entries in the $(k, h)$-th **true** block in matrix $A$.
- $\hat{B}_{kh}$ and $\hat{S}_{kh}$, respectively, are the constant elements in the $(k, h)$-th **estimated** blocks of mean and standard deviation matrices $\hat{P}$ and $\hat{\sigma}$. Here, the estimated block structure is output by a clustering algorithm that satisfies the consistency assumption \[asmp:consistency\].
First of all, we derive the difference between $B_{kh}$ ($S_{kh}$) and $\tilde{B}_{kh}$ ($\tilde{S}_{kh}$). Since the number of entries in the block is proportional to $m^2$ by the assumption \[asmp:block\_size\], $\sqrt{m^2} \left( B_{kh} - \tilde{B}_{kh} \right)$ converges to $\mathcal{N} (0, S_{kh}^2)$ from the central limit theorem. Therefore, from Prokhorov’s theorem [@Vaart1998], we have $$\begin{aligned}
\label{eq:BBtildediff}
\tilde{B}_{kh} = B_{kh} + O_p \left( \frac{1}{m} \right). \end{aligned}$$ Also, the following equation holds (The proof is given in Appendix \[sec:ap\_sigma\_tilde\]): $$\begin{aligned}
\label{eq:SStildediff}
\tilde{S}_{kh} = S_{kh} + O_p \left( \frac{1}{m} \right). \end{aligned}$$
Let $\tilde{Z}$ be a matrix whose entries are calculated based on the observed matrix $A$, sample mean $\tilde{B}$ and standard deviation $\tilde{S}$ for the correct block structure. Formally, $$\begin{aligned}
\label{eq:tilde_Z}
&&\tilde{P} = (\tilde{P}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \ \tilde{P}_{ij} = \tilde{B}_{g^{(1)}_i g^{(2)}_j}, \nonumber \\
&&\tilde{\sigma} = (\tilde{\sigma}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \ \tilde{\sigma}_{ij} = \tilde{S}_{g^{(1)}_i g^{(2)}_j}, \nonumber \\
&&\tilde{Z} = (\tilde{Z}_{ij})_{ij}, \ \ \ \ \ \ \ \ \ \ \tilde{Z}_{ij} = \frac{A - \tilde{P}_{ij}}{\tilde{\sigma}_{ij}}. \end{aligned}$$ From here, we derive the difference between the maximum eigenvalue $\tilde{\lambda}_1$ of matrix $\tilde{Z}^{\top} \tilde{Z}$ and the maximum eigenvalue $\lambda_1$ of matrix $Z^{\top} Z$. From (\[eq:T\_true\]), we have $\lambda_1 = O_p (m)$. Therefore, the largest singular value of matrix $Z$, which is equal to $\left\| Z \right\|_{\mathrm{op}}$, is in the order of $O_p (\sqrt{m})$.
By the subadditivity of the operator norm, we have $$\begin{aligned}
\label{eq:ZZtildediff}
\left| \left\| Z \right\|_{\mathrm{op}} - \left\| \tilde{Z} \right\|_{\mathrm{op}} \right| \leq \left\| Z -\tilde{Z} \right\|_{\mathrm{op}}. \end{aligned}$$
Let $A^{(k, h)}$, $P^{(k, h)}$, $\tilde{P}^{(k, h)}$, $Z^{(k, h)}$ and $\tilde{Z}^{(k, h)}$, respectively, be the $(k, h)$-th **true** blocks of matrices $A$, $P$, $\tilde{P}$, $Z$ and $\tilde{Z}$. We also denote the row and column sizes of the $(k, h)$-th **true** block as $n^{(k, h)}$ and $p^{(k, h)}$, respectively. From the definitions in (\[eq:Z\_true\]) and (\[eq:tilde\_Z\]), we have $$\begin{aligned}
Z^{(k, h)} = \frac{A^{(k, h)} - P^{(k, h)}}{S_{kh}}, \ \ \ \ \ \ \ \ \ \
\tilde{Z}^{(k, h)} = \frac{A^{(k, h)} - \tilde{P}^{(k, h)}}{\tilde{S}_{kh}}. \end{aligned}$$ Combining this with (\[eq:BBtildediff\]), (\[eq:SStildediff\]) and the fact that the Frobenius norm upper bounds the operator norm, we have $$\begin{aligned}
\label{eq:ZZtilde_F}
&&\| Z^{(k, h)} - \tilde{Z}^{(k, h)} \|_{\mathrm{op}}
= \left\| \frac{A^{(k, h)} - P^{(k, h)}}{S_{kh}} - \frac{A^{(k, h)} - \tilde{P}^{(k, h)}}{\tilde{S}_{kh}} \right\|_{\mathrm{op}} \nonumber \\
&=& \left\| \frac{A^{(k, h)} - P^{(k, h)}}{S_{kh}} - \frac{A^{(k, h)} - P^{(k, h)}}{\tilde{S}_{kh}} + \frac{A^{(k, h)} - P^{(k, h)}}{\tilde{S}_{kh}} - \frac{A^{(k, h)} - \tilde{P}^{(k, h)}}{\tilde{S}_{kh}} \right\|_{\mathrm{op}} \nonumber \\
&\leq& \left\| \frac{A^{(k, h)} - P^{(k, h)}}{S_{kh}} - \frac{A^{(k, h)} - P^{(k, h)}}{\tilde{S}_{kh}} \right\|_{\mathrm{op}} + \left\| \frac{A^{(k, h)} - P^{(k, h)}}{\tilde{S}_{kh}} - \frac{A^{(k, h)} - \tilde{P}^{(k, h)}}{\tilde{S}_{kh}} \right\|_{\mathrm{op}} \nonumber \\
&=& \frac{\tilde{S}_{kh} - S_{kh}}{S_{kh} \tilde{S}_{kh}} \left\| A^{(k, h)} - P^{(k, h)} \right\|_{\mathrm{op}} + \frac{1}{\tilde{S}_{kh}} \left\| P^{(k, h)} - \tilde{P}^{(k, h)} \right\|_{\mathrm{op}} \nonumber \\
&=& \frac{\tilde{S}_{kh} - S_{kh}}{S_{kh} \tilde{S}_{kh}} \left\| A^{(k, h)} - P^{(k, h)} \right\|_{\mathrm{op}} + \frac{1}{\tilde{S}_{kh}} \left\| P^{(k, h)} - \tilde{P}^{(k, h)} \right\|_{\mathrm{F}} \nonumber \\
&=& \frac{\tilde{S}_{kh} - S_{kh}}{\tilde{S}_{kh}} \left\| Z^{(k, h)} \right\|_{\mathrm{op}} + \frac{1}{\tilde{S}_{kh}} \sqrt{n^{(k, h)} p^{(k, h)}} \left| B_{kh} - \tilde{B}_{kh} \right| \nonumber \\
&=& \frac{O_p (1/m)}{S_{kh} + O_p (1/m)} \left\| Z^{(k, h)} \right\|_{\mathrm{op}} + \frac{O_p (1/m)}{S_{kh} + O_p (1/m)} \sqrt{n^{(k, h)} p^{(k, h)}} \ \ \ \left(\because (\ref{eq:BBtildediff}), (\ref{eq:SStildediff}) \right) \nonumber \\
&=& \frac{O_p (1/m)}{S_{kh} + O_p (1/m)} O_p (\sqrt{m}) + \frac{O_p (1/m)}{S_{kh} + O_p (1/m)} \sqrt{n^{(k, h)} p^{(k, h)}} \ \ \ \left(\because (\ref{eq:T_true}) \right) \nonumber \\
&=& O_p \left( \frac{1}{\sqrt{m}} \right) + O_p (1) = O_p (1). \end{aligned}$$
Therefore, since the operator norm of a matrix is not larger than the sum of the operator norms of all of its blocks and the number of blocks are finite constants, we have $$\begin{aligned}
\left\| Z - \tilde{Z} \right\|_{\mathrm{op}} &\leq& \sum_{k, h} \left\| Z^{(k, h)} - \tilde{Z}^{(k, h)} \right\|_{\mathrm{op}} = O_p (1). \end{aligned}$$ By combining this with (\[eq:ZZtildediff\]), we obtain $$\begin{aligned}
\label{eq:Ztilde_op}
\left\| \tilde{Z} \right\|_{\mathrm{op}} = \left\| Z \right\|_{\mathrm{op}} + O_p (1). \end{aligned}$$
Next, we consider the joint probability of the event $F_m$ that the clustering algorithm outputs the correct block structure (i.e., $\tilde{Z} = \hat{Z}$) and the event $G_{m, C}$ that $\left| \left\| Z \right\|_{\mathrm{op}} - \left\| \tilde{Z} \right\|_{\mathrm{op}} \right| \leq C$ holds. Such a joint probability satisfies the following inequality: $$\begin{aligned}
\label{eq:jointp}
\mathrm{Pr} \left( F_m \cap G_{m, C} \right) \geq 1 - \mathrm{Pr} \left( F^{\mathrm{C}}_m \right) - \mathrm{Pr} \left( G^{\mathrm{C}}_{m, C} \right), \end{aligned}$$ where $A^{\mathrm{C}}$ is the complement of event $A$. The consistency assumption \[asmp:consistency\] guarantees that if $(K_0, H_0) = (K, H)$, $\mathrm{Pr} \left( F^{\mathrm{C}}_m \right)$ converges to $0$ in the limit of $m \to \infty$. By combining this fact with (\[eq:Ztilde\_op\]), we obtain $$\begin{aligned}
\label{eq:jointp2}
\forall \epsilon>0, \ \exists C>0, M>0, \ \forall m \geq M, \
\mathrm{Pr} \left( F_m \cap G_{m, C} \right) \geq 1 - \epsilon, \end{aligned}$$ which results in $$\begin{aligned}
\label{eq:Zhat_op}
\left\| \hat{Z} \right\|_{\mathrm{op}} = \left\| Z \right\|_{\mathrm{op}} + O_p (1). \end{aligned}$$
Here, note that $\left\| Z \right\|_{\mathrm{op}} = \sqrt{\lambda_1}$ and $\left\| \hat{Z} \right\|_{\mathrm{op}} = \sqrt{\hat{\lambda}_1}$. By combining these results and (\[eq:T\_true\]), from Slutsky’s theorem (note that both $a$ and $b$ are constants that only depend on the matrix size), $$\begin{aligned}
\frac{\hat{\lambda}_1 - a}{b} \rightsquigarrow TW_1 \ \mathrm{(Convergence\ in\ law)}. \end{aligned}$$ This is equivalent to the statement of Theorem \[th:realizable\].
Suppose $K_0 < K$ or $H_0 < H$. $$\begin{aligned}
T = \Omega_p \left( \frac{m^{\frac{5}{3}}}{K^2 H^2} \right), \end{aligned}$$ where $T$ is defined as in (\[eq:T\_statistic\]). \[th:unrealizable\_lower\]
Let $\bar{P}$ be a matrix that consists of the **estimated** block structure and whose entries are the population block-wise means, which can be calculated using $P$ (see also Figure \[fig:unrealizable\]).
![Difference between matrices $P$, $\bar{P}$, and $\hat{P}$ in an unrealizable case. []{data-label="fig:unrealizable"}](unrealizable.pdf){width="150mm"}
To derive the difference between matrices $P$ and $\hat{P}$, we first focus on the relationship between matrices $P$ and $\bar{P}$. In the unrealizable case (i.e., $K_0 < K$ or $H_0 < H$), we can assume $K_0 < K$ without loss of generality.
Let $n_k$ be the number of rows in the $k$-th **true** row cluster. For all the **true** row cluster indices $k \in \{1, \dots, K\}$, at least one **estimated** row cluster contains $n_k/K_0$ or more rows that are assigned to the $k$-th row cluster in the **true** block structure (otherwise, the total number of rows in the $k$-th **true** row cluster is smaller than $n_k$). Since $K_0 < K$, at least one estimated block contains two or more sets of rows whose **true** row clusters are mutually different, and both of which have the row sizes of at least $n_{\mathrm{min}}/K_0$, where $n_{\mathrm{min}}$ is the minimum row size of a block in the **true** block structure. On the other hand, for all the **true** column cluster indices $h \in \{1, \dots, H\}$, at least one **estimated** column cluster contains $p_h/H_0$ or more columns that are assigned to the $h$-th column cluster in the **true** block structure, where $p_h$ is the number of rows in the $h$-th **true** column cluster. By combining these facts, there exists at least one **estimated** block that contains two or more submatrices, both of which have the sizes of at least $(n_{\mathrm{min}}/K_0) \times (p_{\mathrm{min}}/H_0)$ and whose **true** blocks are mutually different.
Let $X_1$ and $X_2$ be such submatrices, whose **true** block-wise mean are $q_1$ and $q_2$, respectively. We can assume $q_1 > q_2$ without the loss of generality. In matrix $\bar{P}$, which has the **estimated** block structure, both of $X_1$ and $X_2$ have the same values $\bar{q}$. Here, if $\bar{q} \geq \frac{q_1 + q_2}{2}$ holds, $|q_2 - \bar{q}| \geq \frac{|q_1 - q_2|}{2}$, and $|q_1 - \bar{q}| \geq \frac{|q_1 - q_2|}{2}$ if otherwise. Therefore, for any $\bar{q}$, there exists at least one submatrix $\bar{X}$ (which is either $X_1$ or $X_2$) with a size of at least $(n_{\mathrm{min}}/K_0) \times (p_{\mathrm{min}}/H_0)$, where all the entries are $q$ (which is either $q_1$ or $q_2$) in matrix $P$ and $$\begin{aligned}
\label{eq:q_bar_q_bound}
|q - \bar{q}| \geq \frac{\min_{(k, h) \neq (k', h')} |B_{kh} - B_{k'h'}|}{2}. \end{aligned}$$
Let $(k_1, h_1)$ be the row and column cluster indices of the **estimated** block which contains the above submatrix $\bar{X}$. We denote the row and column sizes of the $(k_1, h_1)$-th **estimated** block as $\underline{n}_1$ and $\underline{p}_1$, respectively. Let $\underline{A}^{(k_1, h_1)}$, $\underline{P}^{(k_1, h_1)}$, $\underline{\bar{P}}^{(k_1, h_1)}$ and $\underline{\hat{P}}^{(k_1, h_1)}$, respectively, be the submatrices of $A$, $P$, $\bar{P}$ and $\hat{P}$ with the same row and column indices as the $(k_1, h_1)$-th **estimated** block. In regard to the difference between matrices $\bar{P}$ and $\hat{P}$ (both of which have the **estimated** block structure), we have $$\begin{aligned}
\label{eq:o_block_un}
&&|\hat{q} - \bar{q}| = \left| \frac{1}{\underline{n}_1 \underline{p}_1} \sum_{i, j} \left( \underline{\hat{P}}^{(k_1, h_1)}_{ij} - \underline{\bar{P}}^{(k_1, h_1)}_{ij} \right) \right| \nonumber \\
&=& \left| \frac{1}{\underline{n}_1 \underline{p}_1} \sum_{i, j} \left( \underline{A}^{(k_1, h_1)}_{ij} - \underline{P}^{(k_1, h_1)}_{ij} \right) \right| \ \ \ \left( \because \underline{\hat{P}}^{(k_1, h_1)}_{ij} = \frac{1}{\underline{n}_1 \underline{p}_1} \sum_{i, j} \underline{A}^{(k_1, h_1)}_{ij}, \underline{\bar{P}}^{(k_1, h_1)}_{ij} = \frac{1}{\underline{n}_1 \underline{p}_1} \sum_{i, j} \underline{P}^{(k_1, h_1)}_{ij}. \right) \nonumber \\
&=& \frac{1}{\underline{n}_1 \underline{p}_1} \left| \left\langle u_1, \left( \underline{A}^{(k_1, h_1)} - \underline{P}^{(k_1, h_1)} \right) u_2 \right\rangle \right|
\leq \frac{1}{\underline{n}_1 \underline{p}_1} \| u_1 \| \| u_2 \| \left\| \underline{A}^{(k_1, h_1)} - \underline{P}^{(k_1, h_1)} \right\|_{\mathrm{op}} \nonumber \\
&=& \frac{1}{\sqrt{\underline{n}_1 \underline{p}_1}} \left\| \underline{A}^{(k_1, h_1)} - \underline{P}^{(k_1, h_1)} \right\|_{\mathrm{op}}
\leq \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left\| \underline{A}^{(k_1, h_1)} - \underline{P}^{(k_1, h_1)} \right\|_{\mathrm{op}}
\leq \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left\| A - P \right\|_{\mathrm{op}} \nonumber \\ &\leq& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \sum_{k, h} \left\| A^{(k, h)} - P^{(k, h)} \right\|_{\mathrm{op}}
= \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \sum_{k, h} S_{kh} \left\| Z^{(k, h)} \right\|_{\mathrm{op}} \nonumber \\
&\leq& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} KH \max_{k, h} S_{kh} \left\| Z \right\|_{\mathrm{op}}
= O_p \left( \frac{KH}{\sqrt{m}} \right). \end{aligned}$$ where $A^{(k, h)}$, $P^{(k, h)}$ and $Z^{(k, h)}$, respectively, are the $(k, h)$-th **true** blocks of matrices $A$, $P$ and $Z$, and $u_1 = (1, 1, \dots, 1)^{\top} \in \mathbb{R}^{\underline{n}_1}$ and $u_2 = (1, 1, \dots, 1)^{\top} \in \mathbb{R}^{\underline{p}_1}$. To derive the final equation in (\[eq:o\_block\_un\]), we used the assumption that $n_{\mathrm{min}}, p_{\mathrm{min}} = \Omega_p (m)$ and the fact that $\left\| Z \right\|_{\mathrm{op}}$ is equal to the largest singular value of $Z$, which is $O_p (\sqrt{m})$ from (\[eq:T\_true\]).
Let $E_{m, C}$ be the event that $|q - \bar{q}| - CKH/\sqrt{m} \leq |q - \hat{q}|$ holds. For all $q$, $\bar{q}$, and $\hat{q}$, the following inequality holds: $$\begin{aligned}
\label{eq:q_qhat_qbar}
\Bigl| |q - \bar{q}| - |q - \hat{q}| \Bigr| \leq |\hat{q} - \bar{q}|. \end{aligned}$$ By combining (\[eq:o\_block\_un\]) and (\[eq:q\_qhat\_qbar\]), we obtain $$\begin{aligned}
\label{eq:q_qhat_qbar_prob}
\forall \epsilon>0, \ \exists C>0, M>0, \ \forall m \geq M, \ \mathrm{Pr} (E_{m, C}) \geq 1-\epsilon. \end{aligned}$$
From now on, we denote the row and column sizes of submatrix $\bar{X}$, respectively, by $n_1$ and $p_1$. Let $A^*$, $P^*$, $\bar{P}^*$, $\hat{P}^*$, $Z^*$ and $\hat{Z}^*$ respectively, be the submatrices of matrices $A$, $P$, $\bar{P}$, $\hat{P}$, $Z$ and $\hat{Z}$ with the same row and column indices as submatrix $\bar{X}$. We also denote the constant entries of the submatrices of $\sigma$ and $\hat{\sigma}$ with the same row and column indices as submatrix $\bar{X}$, respectively, as $\sigma^*$ and $\hat{\sigma}^*$. From the definition (\[eq:Z\_hat\]) and since the operator norm of a submatrix is not larger than that of the original matrix, we have $$\begin{aligned}
\label{eq:z_hat_op_partial}
\left\| \hat{Z} \right\|_{\mathrm{op}}
&\geq& \left\| \hat{Z}^* \right\|_{\mathrm{op}}
= \frac{1}{\hat{\sigma}^*} \left\| A^* - \hat{P}^* \right\|_{\mathrm{op}}
= \frac{1}{\hat{\sigma}^*} \left\| \left( A^* - P^* \right) + \left( P^* - \hat{P}^* \right) \right\|_{\mathrm{op}} \nonumber \\
&\geq& \frac{1}{\hat{\sigma}^*} \left| \left\| A^* - P^* \right\|_{\mathrm{op}} - \left\| P^* - \hat{P}^* \right\|_{\mathrm{op}} \right| \nonumber \\
&=& \frac{1}{\hat{\sigma}^*} \left| \sigma^* \left\| Z^* \right\|_{\mathrm{op}} - \left\| P^* - \hat{P}^* \right\|_{\mathrm{op}} \right|. \end{aligned}$$
First, the order of the estimated standard deviation $\hat{\sigma}^*$ is given by $$\begin{aligned}
\label{eq:hat_sigma_star_o}
\hat{\sigma}^* = O_p (KH). \end{aligned}$$ The proof of (\[eq:hat\_sigma\_star\_o\]) is in Appendix \[ap\_sigma\_star\].
The only non-zero (and thus, the largest) singular value of matrix $\left( P^* - \hat{P}^* \right)$ is $\sqrt{n_1 p_1} \left| q - \hat{q} \right|$. Since the largest singular value of a matrix is equal to its operator norm, we have $$\begin{aligned}
\left\| P^* - \hat{P}^* \right\|_{\mathrm{op}} = \sqrt{n_1 p_1} \left| q - \hat{q} \right|
\geq \sqrt{\frac{n_{\mathrm{min}}}{K_0} \frac{p_{\mathrm{min}}}{H_0}} \left| q - \hat{q} \right|. \end{aligned}$$ Therefore, by combining this fact with (\[eq:q\_bar\_q\_bound\]), if the statement of event $E_{m, C}$ holds, the following inequality also holds: $$\begin{aligned}
\label{eq:p_phat_op}
\sqrt{\frac{n_{\mathrm{min}}}{K_0} \frac{p_{\mathrm{min}}}{H_0}} \left( \frac{\min_{(k, h) \neq (k', h')} |B_{kh} - B_{k'h'}|}{2} - \frac{CKH}{\sqrt{m}} \right) \leq \left\| P^* - \hat{P}^* \right\|_{\mathrm{op}}, \end{aligned}$$ which results in that $\left\| P^* - \hat{P}^* \right\|_{\mathrm{op}} = \Omega_p \left( \delta m \right)$, where $\delta \equiv \min_{(k, h) \neq (k', h')} |B_{kh} - B_{k'h'}|$.
Also, from (\[eq:T\_true\]), we have $\left\| Z^* \right\|_{\mathrm{op}} \leq \left\| Z \right\|_{\mathrm{op}} =O_p (\sqrt{m})$. By substituting this fact, (\[eq:hat\_sigma\_star\_o\]) and (\[eq:p\_phat\_op\]) into (\[eq:z\_hat\_op\_partial\]), we finally obtain $$\begin{aligned}
\label{eq:apop_prob}
\left\| \hat{Z} \right\|_{\mathrm{op}}^2 = \Omega_p \left(\frac{\delta^2 m^2}{K^2 H^2} \right). \end{aligned}$$
Here, $\left\| \hat{Z} \right\|_{\mathrm{op}}^2$ is equal to the maximum eigenvalue $\hat{\lambda}_1$ of matrix $\hat{Z}^\top \hat{Z}$, and the test statistic is $T = \frac{\hat{\lambda}_1 - a}{b}$. Using the definition (\[eq:ab\]), we obtain $a = O_p (m)$ and $$\begin{aligned}
\label{eq:bm}
b &=& \left( \sqrt{n} + \sqrt{p} \right) \left( \frac{1}{\sqrt{n}} + \frac{1}{\sqrt{p}} \right)^{\frac{1}{3}}
= \left( \sqrt{\beta_1 m} + \sqrt{\beta_2 m} \right) \left( \frac{1}{\sqrt{\beta_1 m}} + \frac{1}{\sqrt{\beta_2 m}} \right)^{\frac{1}{3}} \nonumber \\
&=& m^{\frac{1}{3}} \left( \sqrt{\beta_1} + \sqrt{\beta_2} \right) \left( \frac{1}{\sqrt{\beta_1}} + \frac{1}{\sqrt{\beta_2}} \right)^{\frac{1}{3}}, \end{aligned}$$ where we used the definitions $\beta_1 \equiv n/m$ and $\beta_2 \equiv p/m$.
By combining these results and (\[eq:apop\_prob\]), we obtain $$\begin{aligned}
T m^{\frac{1}{3}} = \Omega_p \left(\frac{\delta^2 m^2}{K^2 H^2} \right)
\iff T = \Omega_p \left( \frac{\delta^2 m^{\frac{5}{3}}}{K^2 H^2} \right), \end{aligned}$$ which concludes the proof.
Suppose $K_0 < K$ or $H_0 < H$. Then, $$\begin{aligned}
T = O_p \left( m^{\frac{5}{3}} \right), \end{aligned}$$ where $T$ is defined as in (\[eq:T\_statistic\]). \[th:unrealizable\_upper\]
We define $P$, $\bar{P}$, and $\hat{P}$ as in Theorem \[th:unrealizable\_lower\]. Let $\underline{\hat{Z}}^{(k, h)}$, $\underline{A}^{(k, h)}$ and $\underline{\hat{P}}^{(k, h)}$, respectively, be the $(k, h)$-th **estimated** block of matrix $\hat{Z}$, $A$ and $\hat{P}$. We denote the row and column sizes of the $(k, h)$-th **estimated** block as $\underline{n}^{(k, h)}$ and $\underline{p}^{(k, h)}$, respectively. Since the operator norm of a matrix is not larger than the sum of the operator norms of all its blocks, we have $$\begin{aligned}
\label{eq:z_hat_op_upper}
\left\| \hat{Z} \right\|_{\mathrm{op}}
&\leq& \sum_{k, h} \left\| \underline{\hat{Z}}^{(k, h)} \right\|_{\mathrm{op}}
= \sum_{k, h} \frac{1}{\hat{S}_{kh}} \left\| \underline{A}^{(k, h)} - \underline{\hat{P}}^{(k, h)} \right\|_{\mathrm{op}} \nonumber \\
&=& \sum_{k, h} \frac{\sqrt{\underline{n}^{(k, h)} \underline{p}^{(k, h)}}}{\left\| \underline{A}^{(k, h)} - \underline{\hat{P}}^{(k, h)} \right\|_{\mathrm{F}}} \left\| \underline{A}^{(k, h)} - \underline{\hat{P}}^{(k, h)} \right\|_{\mathrm{op}}
\leq \sum_{k, h} \frac{\sqrt{\underline{n}^{(k, h)} \underline{p}^{(k, h)}}}{\left\| \underline{A}^{(k, h)} - \underline{\hat{P}}^{(k, h)} \right\|_{\mathrm{F}}} \left\| \underline{A}^{(k, h)} - \underline{\hat{P}}^{(k, h)} \right\|_{\mathrm{F}} \nonumber \\
&=& \sum_{k, h} \sqrt{\underline{n}^{(k, h)} \underline{p}^{(k, h)}}
\leq K_0 H_0 \sqrt{np} = O_p (m).\end{aligned}$$
The test statistic is $T = \frac{\hat{\lambda}_1 - a}{b}$, where $\hat{\lambda}_1 = \left\| \hat{Z} \right\|_{\mathrm{op}}^2 = O_p (m^2)$. Based on the same discussion as in Theorem \[th:unrealizable\_lower\], $a = O_p (m)$ and (\[eq:bm\]) hold. Consequently, we obtain $T = O_p (m^2 / m^{\frac{1}{3}}) = O_p (m^{\frac{5}{3}})$, which concludes the proof.
Based on the results in Theorems \[th:realizable\], \[th:unrealizable\_lower\] and \[th:unrealizable\_upper\], we propose a one-sided goodness-of-fit test for a given set of cluster numbers $(K_0, H_0)$ at the significance level of $\alpha$ by using the test statistic $T$: $$\begin{aligned}
\label{eq:rejection}
\mathrm{Reject\ null\ hypothesis}\ ((K, H) = (K_0, H_0)),\ \ \ \ \ \mathrm{if}\ T \geq t(\alpha), \end{aligned}$$ where $t(\alpha)$ is the $\alpha$ upper quantile of the Tracy-Widom distribution with index $1$. By applying the sequentially ordered test that we explained in Section \[sec:method\] based on the above rejection rule (\[eq:rejection\]), we can select a set of row and column cluster numbers $(\hat{K}, \hat{H})$ for a given observed matrix $A$.
Experiments {#sec:experiments}
===========
Realizable case: convergence of test statistic $T$ in law to Tracy-Widom distribution {#sec:exp_realizable}
-------------------------------------------------------------------------------------
First of all, we checked the convergence of the proposed test statistic $T$ in law to the Tracy-Widom distribution with index $1$, under the *realizable* setting, which has been stated in Theorem \[th:realizable\], by using a synthetic data that were generated based on three types of distributions:
- **Gaussian Latent Block Model**: The observed matrices were generated whose entries in the $(k, h)$-th block follows the normal distribution $\mathcal{N} (B_{kh}, S_{kh})$. In the Gaussian LBM setting, we used the following true model and parameters: $$\begin{aligned}
(K, H) = (4, 3), \ \ \ \ \
B =
\begin{pmatrix}
0.9 & 0.1 & 0.4 \\
0.2 & 0.7 & 0.3 \\
0.3 & 0.2 & 0.8 \\
0.6 & 0.9 & 0.1 \\
\end{pmatrix}, \ \ \ \ \
S =
\begin{pmatrix}
0.08 & 0.06 & 0.15 \\
0.14 & 0.12 & 0.07 \\
0.09 & 0.1 & 0.11 \\
0.16 & 0.13 & 0.05 \\
\end{pmatrix}. \end{aligned}$$
- **Bernoulli Latent Block Model** The observed matrices were generated whose entries in the $(k, h)$-th block follows the normal distribution $\mathrm{Bernoulli} (B_{kh})$. In the Bernoulli LBM setting, we used the following true model and parameters: $$\begin{aligned}
(K, H) = (4, 3), \ \ \ \ \
B =
\begin{pmatrix}
0.9 & 0.1 & 0.4 \\
0.2 & 0.7 & 0.3 \\
0.3 & 0.2 & 0.8 \\
0.6 & 0.9 & 0.1 \\
\end{pmatrix}. \end{aligned}$$
- **Poisson Latent Block Model** The observed matrices were generated whose entries in the $(k, h)$-th block follows the normal distribution $\mathrm{Pois} (B_{kh})$. In the Poisson LBM setting, we used the following true model and parameters: $$\begin{aligned}
(K, H) = (4, 3), \ \ \ \ \
B =
\begin{pmatrix}
9.0 & 1.0 & 4.0 \\
2.0 & 7.0 & 3.0 \\
3.0 & 2.0 & 8.0 \\
6.0 & 9.0 & 1.0 \\
\end{pmatrix}. \end{aligned}$$
Based on the above Latent Block Model, we randomly generated $1000$ observed matrices, estimated their block structures based on the Ward’s hierarchical clustering algorithm [@Ward1963], and computed the test statistic $T$. With respect to the matrix size, we tried the following $10$ settings: $(n, p) = (300 \times i, 225 \times i)$, $i = 1, \dots, 10$. When generating an observed matrix, the true cluster of each row was randomly chosen from the discrete uniform distribution on $\{1, 2, 3, 4\}$. Similarly, the true cluster of each column was randomly chosen from the discrete uniform distribution on $\{1, 2, 3\}$.
Figures \[fig:QQ\_normal\], \[fig:QQ\_bernoulli\], and \[fig:QQ\_poisson\], respectively, show the Q-Q plots of the test statistic $T$ and the $TW_1$ distribution in the settings of Gaussian, Bernoulli, and Poisson settings. Each plotted point corresponds to a sample of test statistic $T$, and the horizontal and vertical lines, respectively, show its theoretical and sample quantiles. These figures show that the test statistic converged in law to the $TW_1$ distribution.
Figure \[fig:preliminaryT\] shows the ratios of the trials where $T \geq t(0.01)$, $T \geq t(0.05)$, and $T \geq t(0.1)$ for the above three distributional settings, where $t(\alpha)$ is the $\alpha$ upper quantile of the $TW_1$ distribution. We used the approximated values $t(0.01) \approx 2.02345$, $t(0.05) \approx 0.97931$, and $t(0.1) \approx 0.45014$, according to Table $2$ in [@Tracy2009]. From Figure \[fig:preliminaryT\], we see that the tail probability of the test statistic $T$ also converged to those of the $TW_1$ distributions for all of the three distributional settings.
![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_300_225-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_600_450-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_900_675-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_1200_900-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_1500_1125-eps-converted-to.pdf "fig:"){width="0.192\hsize"}\
![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_1800_1350-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_2100_1575-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_2400_1800-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_2700_2025-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_normal_QQ_3000_2250-eps-converted-to.pdf "fig:"){width="0.192\hsize"}
\[fig:QQ\_normal\] ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_300_225-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_600_450-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_900_675-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_1200_900-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_1500_1125-eps-converted-to.pdf "fig:"){width="0.192\hsize"}\
![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_1800_1350-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_2100_1575-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_2400_1800-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_2700_2025-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_bernoulli_QQ_3000_2250-eps-converted-to.pdf "fig:"){width="0.192\hsize"}
\[fig:QQ\_bernoulli\] ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_300_225-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_600_450-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_900_675-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_1200_900-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_1500_1125-eps-converted-to.pdf "fig:"){width="0.192\hsize"}\
![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_1800_1350-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_2100_1575-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_2400_1800-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_2700_2025-eps-converted-to.pdf "fig:"){width="0.192\hsize"} ![Q-Q plot of test statistic $T$ against the $TW_1$ distribution in the setting of **Poisson case**. []{data-label="fig:QQ_poisson"}](preliminary_poisson_QQ_3000_2250-eps-converted-to.pdf "fig:"){width="0.192\hsize"}
![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_normal_T_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_bernoulli_T_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_poisson_T_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"}
\[fig:preliminaryT\] ![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_normal_Dr_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_bernoulli_Dr_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Test statistics $D\sqrt{r}$ of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. If the test statistic $D\sqrt{r}$ is larger than the significance level $\alpha$, then the null hypothesis that $T$ follows the $TW_1$ distribution is rejected, and if otherwise, the null hypothesis is accepted. []{data-label="fig:KStest"}](preliminary_poisson_Dr_all-eps-converted-to.pdf "fig:"){width="0.32\hsize"}
We also plotted the results of the Kolmogorov-Smirnov test [@Conover1999] for the test statistic $T$ in Figure \[fig:KStest\]. We tested whether the distribution of $T$ is the $TW_1$ distribution or not based on the test statistic $D\sqrt{r}$, where $D$ is the maximum absolute difference between the empirical distribution function of $T$ and the cumulative distribution function of the $TW_1$ distribution, and $r$ is the sample size, which is set at $1000$ in this experiment. Figure \[fig:KStest\] shows the convergence of the proposed test statistic $T$ in law to the $TW_1$ distribution under the realizable setting.
Unrealizable case: asymptotic behavior of test statistic $T$ {#sec:exp_unrealizable}
------------------------------------------------------------
Next, we checked the asymptotic behavior of the proposed test statistic $T$ under the *unrealizable* setting, which has been stated in Theorems \[th:unrealizable\_lower\] and \[th:unrealizable\_upper\], by using a synthetic data that were generated based on the same three types of distributions as in Section \[sec:exp\_realizable\]. By combining Theorems \[th:unrealizable\_lower\] and \[th:unrealizable\_upper\], we obtain the following theorem:
Suppose $K_0 < K$ or $H_0 < H$. Then, $$\begin{aligned}
\forall \epsilon>0, \ \exists C_1, C_2, M, \ \forall m \geq M, \ \mathrm{Pr} (C_1 m^{\frac{5}{3}} \leq T \leq C_2 m^{\frac{5}{3}}) \geq 1-\epsilon. \end{aligned}$$ \[th:unrealizable\_two\_sided\]
In other words, with high probability, the proposed test statistic $T$ increases in proportion to $p^{\frac{5}{3}}$ in the limit of $p \to \infty$, since we have assumed that $p \propto m$.
With respect to the true models and parameters, we used the same settings as in Section \[sec:exp\_realizable\] for all of the three distributional settings (i.e., Gaussian, Bernoulli, and Poisson LBMs). Based on such settings, we randomly generated $100$ observed matrices, estimated their block structures based on the Ward’s hierarchical clustering algorithm [@Ward1963], and computed the test statistic $T$. With respect to the matrix size, we tried the following $10$ settings: $(n, p) = (200 \times i, 150 \times i)$, $i = 1, \dots, 10$. When generating an observed matrix, the true cluster of each row was randomly chosen from the discrete uniform distribution on $\{1, 2, 3, 4\}$. Similarly, the true cluster of each column was randomly chosen from the discrete uniform distribution on $\{1, 2, 3\}$.
Figures \[fig:unrealizableT\] and \[fig:unrealizableT2\] show the asymptotic behavior of the proposed test statistic $T$ under the unrealizable setting. As shown in Theorem \[th:unrealizable\_two\_sided\], we see that $T$ increases in proportion to $m^{\frac{5}{3}}$, where $n, p \propto m$.
![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable_normal_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable_bernoulli_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable_poisson_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"}
\[fig:unrealizableT\] ![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable2_normal_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable2_bernoulli_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Mean test statistic $T$ divided by $n^{\frac{5}{3}}$ in the unrealizable case for $100$ trials. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:unrealizableT2"}](unrealizable2_poisson_T-eps-converted-to.pdf "fig:"){width="0.32\hsize"}
Accuracy of the proposed goodness-of-fit test
---------------------------------------------
Finally, we evaluated the proposed goodness-of-fit test in terms of its accuracy. By using a synthetic data that were generated based on the same three types of distributions as in Section \[sec:exp\_realizable\], we checked the ratio of trials where the selected set of cluster numbers $(K_0, H_0)$ is equal to the true one $(K, H)$. Here, we set the true set of cluster numbers at $(K, H) = (4, 3)$. For each distributional setting (i.e., Gaussian, Bernoulli, and Poisson LBMs), we tried $10$ settings with respect to the block-wise mean $B$. The concrete settings were as follows:
- **Gaussian Latent Block Model**: We used the following true parameters: $$\begin{aligned}
&&B' =
\begin{pmatrix}
0.9 & 0.1 & 0.4 \\
0.2 & 0.7 & 0.3 \\
0.3 & 0.2 & 0.8 \\
0.6 & 0.9 & 0.1 \\
\end{pmatrix}, \ \ \
\forall k, h,\ B_{kh} = \left( 1 - \frac{t}{10} \right) (B'_{kh} - 0.5) + 0.5, \ \ \ \mathrm{for}\ t = 0, \dots, 9, \nonumber \\
&&S =
\begin{pmatrix}
0.08 & 0.06 & 0.15 \\
0.14 & 0.12 & 0.07 \\
0.09 & 0.1 & 0.11 \\
0.16 & 0.13 & 0.05 \\
\end{pmatrix}. \end{aligned}$$
- **Bernoulli Latent Block Model** We used the following true parameters: $$\begin{aligned}
B' =
\begin{pmatrix}
0.9 & 0.1 & 0.4 \\
0.2 & 0.7 & 0.3 \\
0.3 & 0.2 & 0.8 \\
0.6 & 0.9 & 0.1 \\
\end{pmatrix}, \ \ \
\forall k, h,\ B_{kh} = \left( 1 - \frac{t}{10} \right) (B'_{kh} - 0.5) + 0.5, \ \ \ \mathrm{for}\ t = 0, \dots, 9. \end{aligned}$$
- **Poisson Latent Block Model** We used the following true parameters: $$\begin{aligned}
B' =
\begin{pmatrix}
9.0 & 1.0 & 4.0 \\
2.0 & 7.0 & 3.0 \\
3.0 & 2.0 & 8.0 \\
6.0 & 9.0 & 1.0 \\
\end{pmatrix}, \ \ \
\forall k, h,\ B_{kh} = \left( 1 - \frac{t}{10} \right) (B'_{kh} - 5) + 5, \ \ \ \mathrm{for}\ t = 0, \dots, 9. \end{aligned}$$
With respect to the matrix size, we tried the following $10$ settings for each distributional setting and for each setting of $B$: $(n, p) = (40 \times i, 30 \times i)$, $i = 1, \dots, 10$. When generating an observed matrix, the true cluster of each row was randomly chosen from the discrete uniform distribution on $\{1, 2, 3, 4\}$. Similarly, the true cluster of each column was randomly chosen from the discrete uniform distribution on $\{1, 2, 3\}$. In each of $3$ (Gaussian, Bernoulli, or Poisson LBM) $\times 10$ (for the setting of $B$) $\times 10$ (for the setting of matrix size) settings, we generated $1000$ observed matrices and applied the proposed sequential goodness-of-fit test, until the null hypothesis $(K, H) = (K_0, H_0)$ was accepted. For each observed matrix, we estimated their block structures based on the Ward’s hierarchical clustering algorithm [@Ward1963] under each setting of a hypothetical set of cluster numbers $(K_0, H_0)$, computed the test statistic $T$, and performed the proposed test for the given cluster numbers $(K_0, H_0)$ using a significance level of $\alpha = 0.01$. Figures \[fig:As\_normal\], \[fig:As\_ber\], \[fig:As\_pois\], respectively, show the examples of generated observed matrices of Gaussian, Bernoulli, and Poisson LBMs.
![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t1-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t2-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t3-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t4-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t5-eps-converted-to.pdf "fig:"){height="0.19\hsize"}\
![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t6-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t7-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t8-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t9-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_normal_t10-eps-converted-to.pdf "fig:"){height="0.19\hsize"}
\[fig:As\_normal\] ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t1-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t2-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t3-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t4-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t5-eps-converted-to.pdf "fig:"){height="0.19\hsize"}\
![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t6-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t7-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t8-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t9-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_bernoulli_t10-eps-converted-to.pdf "fig:"){height="0.19\hsize"}
\[fig:As\_ber\] ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t1-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t2-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t3-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t4-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t5-eps-converted-to.pdf "fig:"){height="0.19\hsize"}\
![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t6-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t7-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t8-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t9-eps-converted-to.pdf "fig:"){height="0.19\hsize"} ![Examples of true block structures of the **Poisson** LBM. []{data-label="fig:As_pois"}](A_poisson_t10-eps-converted-to.pdf "fig:"){height="0.19\hsize"}
Figure \[fig:accuracy\] shows the accuracy of the proposed test under $10$ different settings of block-wise mean $B$. From Figure \[fig:accuracy\], we see that the test accuracy increases with matrix size $n$ for a fixed block-wise mean $B$, and that it decreases with the smaller differences between the block-wise means for a fixed matrix size $n$.
![Accuracy of the proposed goodness-of-fit test under $10$ different settings of block-wise mean $B$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:accuracy"}](accuracy_normal-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Accuracy of the proposed goodness-of-fit test under $10$ different settings of block-wise mean $B$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:accuracy"}](accuracy_bernoulli-eps-converted-to.pdf "fig:"){width="0.32\hsize"} ![Accuracy of the proposed goodness-of-fit test under $10$ different settings of block-wise mean $B$. The left, center, and right figures, respectively, show the results for the settings of Gaussian, Bernoulli, and Poisson distributions. []{data-label="fig:accuracy"}](accuracy_poisson-eps-converted-to.pdf "fig:"){width="0.32\hsize"}
Discussion {#sec:discussion}
==========
In this section, we discuss the proposed test method in terms of the test statistic and the conditions for the generative model.
With respect to the asymptotic behavior, the proposed test has a favorable property in terms of the power. From Theorem \[th:unrealizable\_two\_sided\], under the alternative hypothesis, the test statistic $T$ increases in proportion to $m^{\frac{5}{3}}$ with high probability, where $n, p \propto m$. In other words, the probability that the test makes a type II error (i.e., $T < t(\alpha)$) converges to zero in the limit of $p \to \infty$. Based on this fact, in the asymptotic sense, we do not need to consider the correction for the multiple comparison when applying the proposed sequential testing. However, it has not been shown what occurs in the non-asymptotic setting. In general, practical data matrices have finite sizes, where there has been shown no theoretical guarantee like Theorems \[th:realizable\], \[th:unrealizable\_lower\], and \[th:unrealizable\_upper\]. On the other hand, for a Gaussian case (i.e., each entry of a matrix independently follows $\mathcal{N} (0, 1)$), the following statement holds [@Ma2012]: Suppose $n = n(p) > p$ and $n/p \to \gamma \in [1, \infty )$ in the limit of $p \to \infty$. Then, for any $s_0$, there exists $N_0 \in \mathbb{N}$ such that when $\max (n, p) \geq N_0$ and $\max (n, p)$ is even, for all $s \geq s_0$, $$\begin{aligned}
\label{eq:finite_sample_bound}
|\mathrm{Pr} (T^* \leq s) - F_1 (s)| \leq C(s_0) [\max (n, p)]^{-2/3} \exp \left( -\frac{s}{2} \right), \end{aligned}$$ where $T^*$ is defined as in (\[eq:T\_true\]) and $C(\cdot )$ is a continuous and non-increasing function. From the above inequality (\[eq:finite\_sample\_bound\]), *if the clustering algorithm outputs the correct block assignments,* the convergence rate of the normalized maximum eigenvalue $T^*$ of matrix $\tilde{Z}^\top \tilde{Z}$ (where $\tilde{Z}$ is defined as in (\[eq:tilde\_Z\])) to the Tracy-Widom distribution with index $1$ is $O(m^{-2/3})$. However, since the distribution of $T$ is unknown in the case where the correct block assignment is *not* obtained, the convergence rate of $T$ is also unknown. Deriving the convergence rate of $T$ by considering the above discussion is a future research topic.
In regard to the conditions for using the proposed test method, our proposed test is applicable to a wide range of practical distributional settings (e.g., Bernoulli distribution for binary data matrices and Poisson distribution for sparse ones). Nevertheless, it still requires some assumptions for the latent block structure of an observed matrix. For instance, the row and column cluster numbers $(K, H)$ should be constants that do not increase with the matrix sizes $n$ and $p$. Also, there should be no too small block (i.e., $n_{\mathrm{min}} = \Omega_p (m)$ and $p_{\mathrm{min}} = \Omega_p (m)$). In some practical cases, where it is more appropriate to assume that the number of blocks increases with the matrix size, it will be useful to construct a test which does not require the above conditions. Furthermore, there are proposed variants of latent block models with which we assume different block structures from a regular grid [@Roy2008; @Nakano2014]. To construct test methods for the above settings is an important topic for future research.
Conclusion {#sec:conclusion}
==========
Latent block models are effective tools for biclustering, where rows and columns of an observed matrix are simultaneously decomposed into clusters. Such a bicluster structure appears in various types of relational data, such as the customer-product transaction data or and the document-word relationship data. One open problem in using latent block models is that there has been no statistical test method for determining the number of blocks. In this study, we developed a goodness-of-fit test for latent block models based on a result from the random matrix theory. By defining the test statistic $T$ based on the estimators of the block-wise means and standard deviations, we have derived its asymptotic behavior in both realizable (i.e., $(K, H) = (K_0, H_0)$) and unrealizable (i.e., $K > K_0$ or $H > H_0$) cases. Particularly, it has been shown that the test statistic $T$ converges in law to Tracy-Widom distribution with index $1$ in the realizable case. Based on these results, it was made possible to test whether the given observed matrix had $K_0 \times H_0$ latent blocks or more ones. In the experiments, we showed the validity of the proposed test method in terms of both the asymptotic behavior of the test statistic and the test accuracy by using synthetic data matrices with ground truth block structures.
Acknowledgments {#acknowledgments .unnumbered}
===============
TS was partially supported by MEXT Kakenhi (26280009, 15H05707, 18K19793 and 18H03201), Japan Digital Design and JST-CREST.
Proof of $\tilde{S}_{kh} = S_{kh} + O_p \left( \frac{1}{m} \right)$. {#sec:ap_sigma_tilde}
=====================================================================
Let $n^{(k, h)}$ and $p^{(k, h)}$, respectively, be the row and column sizes of the $(k, h)$-th **true** block, and $A^{(k, h)}$, $P^{(k, h)}$ and $\tilde{P}^{(k, h)}$, respectively, be the $(k, h)$-th **true** blocks of matrices $A$, $P$ and $\tilde{P}$. Here, we prove the following lemma:
\[lm:sigma\_tilde\] Under the assumption that the fourth moment of the noise $Z_{ij}$ is bounded ($\mathbb{E} [Z_{ij}^4] < \infty$), $$\begin{aligned}
\label{lm:sigma_tilde_op}
\tilde{S}_{kh} = S_{kh} + O_p \left( \frac{1}{m} \right), \end{aligned}$$ where $\tilde{S}_{kh} = \sqrt{\frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - \tilde{B}_{kh} \right)^2}$.
From the above definition of $\tilde{S}_{kh}$, we have $$\begin{aligned}
\tilde{S}_{kh}^2
&=& \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - \tilde{B}_{kh} \right)^2 \nonumber \\
&=& \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - \left( B_{kh} - \tilde{B}_{kh} \right)^2. \end{aligned}$$ To derive the second equation, we used the fact that $\tilde{B}_{kh} = \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} A^{(k, h)}_{ij}$. Therefore, the following equation holds: $$\begin{aligned}
\label{eq:sigma_tilde_diff}
\tilde{S}_{kh}^2 - S_{kh}^2 = \left[ \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 \right] - \left( B_{kh} - \tilde{B}_{kh} \right)^2. \end{aligned}$$
The first term in (\[eq:sigma\_tilde\_diff\]) is given by $$\begin{aligned}
\label{eq:sigma_tilde_y}
\frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 &=& \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left[ \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 \right] \nonumber \\
&=& \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} Y^{(k, h)}_{ij}, \end{aligned}$$ where we defined that $Y^{(k, h)}_{ij} \equiv \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2$. Note that $(Y^{(k, h)}_{ij})_{ij}$ is independent. The expectation and the variance of $Y^{(k, h)}_{ij}$ are given by $$\begin{aligned}
\label{eq:sigma_tilde_ev}
\mathbb{E} \left[ Y^{(k, h)}_{ij} \right] &=& \mathbb{E} \left[ \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 \right] - S_{kh}^2 = 0, \nonumber \\
\mathbb{V} \left[ Y^{(k, h)}_{ij} \right] &=& \mathbb{E} \left[ \left( Y^{(k, h)}_{ij} \right)^2 \right]
= \mathbb{E} \left[ \left\{ \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 \right\}^2 \right] \nonumber \\
&=& S_{kh}^4 \left( \mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^4 \right] - 1 \right). \end{aligned}$$ From (\[eq:sigma\_tilde\_ev\]), we have $$\begin{aligned}
\label{eq:sigma_tilde_ev_mean}
\mathbb{E} \left[ \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} Y^{(k, h)}_{ij} \right] &=& 0, \nonumber \\
\mathbb{V} \left[ \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} Y^{(k, h)}_{ij} \right] &=& \frac{1}{n^{(k, h)} p^{(k, h)}} S_{kh}^4 \left( \mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^4 \right] - 1 \right). \end{aligned}$$ From (\[eq:sigma\_tilde\_y\]), (\[eq:sigma\_tilde\_ev\_mean\]) and Chebyshev’s inequality, for all $l>0$, $$\begin{aligned}
\mathrm{Pr} \left( \left| \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 \right| \geq l \frac{1}{\sqrt{n^{(k, h)} p^{(k, h)}}} \sqrt{ S_{kh}^4 \left( \mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^4 \right] - 1 \right) } \right) \leq \frac{1}{l^2}. \end{aligned}$$ Therefore, we have $$\begin{aligned}
\label{eq:sigma_tilde_term1}
\frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( A^{(k, h)}_{ij} - B_{kh} \right)^2 - S_{kh}^2 = O_p \left( \frac{1}{m} \right). \end{aligned}$$
On the other hand, the second term in (\[eq:sigma\_tilde\_diff\]) is given by $$\begin{aligned}
\label{eq:sigma_tilde_pp}
\left( B_{kh} - \tilde{B}_{kh} \right)^2
&=& \left( \mathbb{E} \left[ \tilde{B}_{kh} \right] - \tilde{B}_{kh} \right)^2 \nonumber \\
&=& \left[ \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} \left( P^{(k, h)}_{ij} - A^{(k, h)}_{ij} \right) \right]^2 \ \ \ \left(\because \tilde{B}_{kh} = \frac{1}{n^{(k, h)} p^{(k, h)}} \sum_{i, j} A^{(k, h)}_{ij}. \right) \nonumber \\
&=& \frac{S_{kh}^2}{\left( n^{(k, h)} \right)^2 \left( p^{(k, h)} \right)^2} \left( \sum_{i, j} Z^{(k, h)}_{ij} \right)^2. \end{aligned}$$
From (\[eq:sigma\_tilde\_pp\]), we have $$\begin{aligned}
\label{eq:sigma_tilde_e_pp}
\mathbb{E} \left[ \left( B_{kh} - \tilde{B}_{kh} \right)^2 \right]
= \frac{S_{kh}^2}{\left( n^{(k, h)} \right)^2 \left( p^{(k, h)} \right)^2} \mathbb{E} \left[ \left( \sum_{i, j} Z^{(k, h)}_{ij} \right)^2 \right]
= \frac{S_{kh}^2}{\left( n^{(k, h)} \right)^2 \left( p^{(k, h)} \right)^2} \mathbb{E} \left[ X_{kh}^2 \right], \end{aligned}$$ where $X_{kh} \equiv \sum_{i, j} Z^{(k, h)}_{ij}$. Here, $\mathbb{E} [X_{kh}] = 0$, and $\mathbb{E} \left[ X_{kh}^2 \right] = \mathbb{V} [X_{kh}] = n^{(k, h)} p^{(k, h)}$. By substituting this into (\[eq:sigma\_tilde\_e\_pp\]), we obtain $$\begin{aligned}
\label{eq:sigma_tilde_e_pp2}
\mathbb{E} \left[ \left( B_{kh} - \tilde{B}_{kh} \right)^2 \right] = \frac{S_{kh}^2}{n^{(k, h)} p^{(k, h)}}. \end{aligned}$$
From Markov’s inequality and (\[eq:sigma\_tilde\_e\_pp2\]), $$\begin{aligned}
\label{eq:sigma_tilde_markov}
&&\forall l>0, \mathrm{Pr} \left( \left( B_{kh} - \tilde{B}_{kh} \right)^2 \geq l \right) \leq \frac{S_{kh}^2}{n^{(k, h)} p^{(k, h)}} \frac{1}{l}. \nonumber \\
\iff &&\forall l'>0, \mathrm{Pr} \left( \left( B_{kh} - \tilde{B}_{kh} \right)^2 \geq \frac{S_{kh}^2}{n^{(k, h)} p^{(k, h)}} l' \right) \leq \frac{1}{l'}. \end{aligned}$$ Therefore, we have $$\begin{aligned}
\label{eq:sigma_tilde_term2}
\left( B_{kh} - \tilde{B}_{kh} \right)^2 = O_p \left( \frac{1}{m^2} \right). \end{aligned}$$
By combining (\[eq:sigma\_tilde\_diff\]), (\[eq:sigma\_tilde\_term1\]) and (\[eq:sigma\_tilde\_term2\]), $$\begin{aligned}
\label{eq:sigma_tilde_diff2}
\tilde{S}_{kh}^2 - S_{kh}^2 = O_p \left( \frac{1}{m} \right). \end{aligned}$$ The difference between $\tilde{S}_{kh}$ and $S_{kh}$ is given by $$\begin{aligned}
\label{eq:sigma_tilde_diff3}
\tilde{S}_{kh} - S_{kh} = \frac{\tilde{S}_{kh}^2 - S_{kh}^2}{\tilde{S}_{kh} + S_{kh}}. \end{aligned}$$ Here, from (\[eq:sigma\_tilde\_diff2\]), $m (\tilde{S}_{kh}^2 - S_{kh}^2)$ is bounded in probability. Therefore, $\tilde{S}_{kh}$ converges in probability to $S_{kh}$. By combining this fact with (\[eq:sigma\_tilde\_diff2\]) and (\[eq:sigma\_tilde\_diff3\]), we finally obtain $$\begin{aligned}
\tilde{S}_{kh} - S_{kh} = O_p \left( \frac{1}{m} \right), \end{aligned}$$ which concludes the proof.
Proof of $\hat{\sigma}^* = O_p (KH)$ in unrealizable case {#ap_sigma_star}
==========================================================
Throughout the proof, we use the following notations:
- $A^{(k, h)}$, $P^{(k, h)}$ and $Z^{(k, h)}$, respectively, are the $(k, h)$-th **true** blocks of matrices $A$, $P$ and $Z$
- $\underline{A}^{(k, h)}$, $\underline{P}^{(k, h)}$ and $\underline{\hat{P}}^{(k, h)}$, respectively, are the $(k, h)$-th **estimated** block of matrices $A$, $P$ and $\hat{P}$.
- We denote the row and column sizes of the $(k, h)$-th **estimated** block as $\underline{n}^{(k, h)}$ and $\underline{p}^{(k, h)}$, respectively.
- $(k_1, h_1)$ is the set of row and column cluster indices of submatrix $\bar{X}$ in the **estimated** block structure.
As for the order of the estimated standard deviation $\hat{\sigma}^*$, we have $\hat{\sigma}^* = \hat{S}_{k_1 h_1}$. Note that the block size $(n_1, p_1)$ of submatrix $\bar{X}$ is at least $(n_{\mathrm{min}}/K_0) \times (p_{\mathrm{min}}/H_0)$. Therefore, we have $$\begin{aligned}
\label{eq:sigma_hat}
\hat{\sigma}^* &=& \hat{S}_{k_1 h_1}
= \frac{1}{\sqrt{\underline{n}^{(k_1, h_1)} \underline{p}^{(k_1, h_1)}}} \left\| \underline{A}^{(k_1, h_1)} - \underline{\hat{P}}^{(k_1, h_1)} \right\|_{\mathrm{F}}
\leq \frac{1}{\sqrt{n_1 p_1}} \left\| \underline{A}^{(k_1, h_1)} - \underline{\hat{P}}^{(k_1, h_1)} \right\|_{\mathrm{F}} \nonumber \\
&\leq& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left\| \underline{A}^{(k_1, h_1)} - \underline{\hat{P}}^{(k_1, h_1)} \right\|_{\mathrm{F}}
\leq \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left\| A - \hat{P} \right\|_{\mathrm{F}}
= \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left\| A - P + P - \hat{P} \right\|_{\mathrm{F}} \nonumber \\
&\leq& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left( \left\| A - P \right\|_{\mathrm{F}} + \left\| P - \hat{P} \right\|_{\mathrm{F}} \right)
= \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left( \sqrt{\sum_{k, h} \left\| A^{(k, h)} - P^{(k, h)} \right\|_{\mathrm{F}}^2} + \left\| P - \hat{P} \right\|_{\mathrm{F}} \right) \nonumber \\
&=& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left( \sqrt{\sum_{k, h} S_{kh}^2 \left\| Z^{(k, h)} \right\|_{\mathrm{F}}^2} + \left\| P - \hat{P} \right\|_{\mathrm{F}} \right) \nonumber \\
&\leq& \sqrt{\frac{K_0 H_0}{n_{\mathrm{min}} p_{\mathrm{min}}}} \left[ \sqrt{KH} \left( \max_{k, h} S_{kh} \right) \left\| Z \right\|_{\mathrm{F}}+ \left\| P - \hat{P} \right\|_{\mathrm{F}} \right]. \end{aligned}$$
Here, for all $(i, j)$ in the $(k, h)$-th **true** block, $\left( Z^{(k, h)}_{ij} \right)^2$ independently follows the same distribution, and $\mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^2 \right] = 1$. We also have $\mathbb{V} \left[ \left( Z^{(k, h)}_{ij} \right)^2 \right] = \mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^4 \right] - 1 < \infty$, since we have assumed that $\mathbb{E} \left[ \left( Z^{(k, h)}_{ij} \right)^4 \right] < \infty$ from the sub-exponential assumption. Therefore, from the central limit theorem and Prokhorov’s theorem [@Vaart1998], we have $\frac{1}{\sqrt{n^{(k, h)} p^{(k, h)}}} \sum_{i \in I_k, j \in J_h} \left[ \left( Z^{(k, h)}_{ij} \right)^2 - 1 \right] = O_p (1)$. In other words, the following equation holds: $\sum_{i \in I_k, j \in J_h} \left( Z^{(k, h)}_{ij} \right)^2 = n^{(k, h)} p^{(k, h)} + O_p (m) = O_p (m^2)$. Based on this result, we obtain $$\begin{aligned}
\label{eq:Z_F}
\left\| Z \right\|_{\mathrm{F}}
&=& \sqrt{\sum_{k, h} \left\| Z^{(k, h)} \right\|_{\mathrm{F}}^2}
= \sqrt{\sum_{k, h} \sum_{i \in I_k, j \in J_h} \left( Z^{(k, h)}_{ij} \right)^2}
= \sqrt{\sum_{k, h} O_p (m^2)} = O_p (\sqrt{KH} m). \end{aligned}$$
Furthermore, we have $$\begin{aligned}
\label{eq:pp_F}
\left\| P - \hat{P} \right\|_{\mathrm{F}} &=& \sqrt{\sum_{i, j} \left(P_{ij} - \hat{P}_{ij} \right)^2}
= \sqrt{\sum_{i, j} \left(P_{ij} - \bar{P}_{ij} + \bar{P}_{ij} - \hat{P}_{ij} \right)^2} \nonumber \\
&\leq& \sqrt{\sum_{i, j} \left( \left|P_{ij} - \bar{P}_{ij} \right| + \left|\bar{P}_{ij} - \hat{P}_{ij} \right| \right)^2}
\leq \sqrt{\sum_{i, j} \left[ \left( \max_{i', j'} \left|P_{i' j'} - \bar{P}_{i' j'} \right| \right) + \left|\bar{P}_{ij} - \hat{P}_{ij} \right| \right]^2} \nonumber \\
&\leq& \sqrt{\sum_{i, j} \left[ \left(\max_{k, h, k', h'} \left| B_{kh} -B_{k' h'} \right| \right) + \left| \bar{P}_{ij} - \hat{P}_{ij} \right| \right]^2} \nonumber \\
&\leq& \sqrt{\sum_{i, j} \left[ \left(\max_{k, h, k', h'} \left| B_{kh} -B_{k' h'} \right| \right) + \left( \max_{i, j} \left| \bar{P}_{ij} - \hat{P}_{ij} \right| \right) \right]^2} \nonumber \\
&=& \sqrt{np} \left[ \left(\max_{k, h, k', h'} \left| B_{kh} -B_{k' h'} \right| \right) + \left( \max_{i, j} \left| \bar{P}_{ij} - \hat{P}_{ij} \right| \right) \right] \nonumber \\
&\leq& \sqrt{np} \left[ \left(\max_{k, h, k', h'} \left| B_{kh} -B_{k' h'} \right| \right) + O_p \left( \frac{KH}{\sqrt{m}} \right) \right]
= O_p (m +KH\sqrt{m}). \end{aligned}$$ Here, to derive the last inequality, we used the assumption that (\[eq:o\_block\_un\]) holds for the block with the maximum difference between $\bar{P}$ and $\hat{P}$. In the final equation, we used the fact that $\max_{k, h, k', h'} |B_{kh} -B_{k' h'}|$ is bounded by a finite constant.
By combining (\[eq:sigma\_hat\]), (\[eq:Z\_F\]) and (\[eq:pp\_F\]), we obtain $\hat{\sigma}^* = O_p (KH)$.
[^1]: chihiro\[email protected]
[^2]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present measurements of the X-ray observables of the intra-cluster medium (ICM), including luminosity $L_X$, ICM mass $M_{ICM}$, emission-weighted mean temperature $T_X$, and integrated pressure $Y_X$, that are derived from XMM-Newton X-ray observations of a Sunyaev-Zel’dovich Effect (SZE) selected sample of 59 galaxy clusters from the South Pole Telescope SPT-SZ survey that span the redshift range of $0.20 < z < 1.5$. We constrain the best-fit power law scaling relations between X-ray observables, redshift, and halo mass. The halo masses are estimated based on previously published SZE observable to mass scaling relations, calibrated using information that includes the halo mass function. Employing SZE-based masses in this sample enables us to constrain these scaling relations for massive galaxy clusters ($M_{500}\geq 3 \times10^{14}$ $M_\odot$) to the highest redshifts where these clusters exist without concern for X-ray selection biases. We find that the mass trends are steeper than self-similarity in all cases, and with $\geq 2.5{\sigma}$ significance in the case of $L_X$ and $M_{ICM}$. The redshift trends are consistent with the self-similar expectation, but the uncertainties remain large. Core-included scaling relations tend to have steeper mass trends for $L_X$. There is no convincing evidence for a redshift-dependent mass trend in any observable. The constraints on the amplitudes of the fitted scaling relations are currently limited by the systematic uncertainties on the SZE-based halo masses, however the redshift and mass trends are limited by the X-ray sample size and the measurement uncertainties of the X-ray observables.'
author:
- Esra Bulbul
- 'I-Non Chiu'
- 'Joseph J. Mohr'
- Michael McDonald
- Bradford Benson
- 'Mark W. Bautz'
- Matthew Bayliss
- Lindsey Bleem
- Mark Brodwin
- Sebastian Bocquet
- Raffaella Capasso
- 'Jörg P. Dietrich'
- Bill Forman
- 'Julie Hlavacek-Larrondo'
- 'W. L. Holzapfel'
- Gourav Khullar
- Matthias Klein
- Ralph Kraft
- 'Eric D. Miller'
- Christian Reichardt
- Alex Saro
- Keren Sharon
- Brian Stalder
- Tim Schrabback
- Adam Stanford
bibliography:
- 'literature.bib'
title: 'X-ray Properties of SPT Selected Galaxy Clusters at $0.2<\lowercase{z} <1.5$ Observed with [[*XMM-Newton*]{}]{}'
---
Introduction {#sec:intro}
============
The evolution of the mass function of clusters of galaxies is dependent on cosmology, making clusters unique probes of fundamental cosmological parameters— not only the normalization of the power spectrum $\sigma_{8}$ and mean matter density $\Omega_\mathrm{M}$, but also the equation of state parameter of the dark energy [@wang98; @haiman01]. The ability to select clusters out to high redshifts and to measure their masses is particularly important for constraints on the dark energy equation of state and the growth rate of cosmic structure.
The fully ionized intracluster medium (ICM) is heated to keV temperatures through gravitational acceleration and shocks as the cluster forms and grows. At these temperatures it emits X-rays through a combination of thermal bremsstrahlung and atomic line emission. Serendipitous X-ray surveys with XMM enabled the detection of $z>1$ galaxy clusters [@fassbender11], but the solid angle surveyed and the required optical and infrared imaging follow-up remain as challenges to this approach. The all sky X-ray survey with ROSAT [RASS; @voges99] has been used to define large samples of mostly low-redshift clusters [@boehringer04; @piffaretti11], and only now in combination with deep, large solid angle multi-wavelength optical surveys is beginning to deliver cluster samples extending to $z\approx1$ [@klein18].
The ICM also distorts the cosmic microwave background (CMB) through inverse Compton scattering, known as the Sunyaev-Zel’dovich Effect [SZE; @sunyaev72]. Large solid angle surveys employing the SZE have been carried out with the South Pole Telescope [SPT; @carlstrom11], [*Planck*]{} [@planck11-13], and the Atacama Cosmology Telescope [ACT; @fowler07]. The SZE-selected galaxy cluster sample from SPT is an approximately mass-selected sample (${\ensuremath{M_{500}}}\geq3\times10^{14}{\ensuremath{M_{\odot}}}$) of over 500 clusters that extends to the highest redshifts at which these clusters exist [@bleem15], and approximately 20 percent of the sample lies at $z>0.8$. To date, the highest redshift cluster identified in the 2500 deg$^2$ SPT-SZ survey has a redshift of $z =1.7\pm0.05$ (Strazzullo et al. in prep; Mantz et al. in prep).
X-ray observations of SZE-selected clusters provide low-scatter mass proxies which can be used to aid in the calibration of the SZE-based cluster masses and in the cosmological analysis of the SZE cluster samples. Pioneering observational studies have found low-scatter scaling relations that tie X-ray observables to cluster mass for X-ray selected low-redshift clusters [@mohr99; @finoguenov01; @reiprich02; @arnaud07; @pratt07; @vikhlinin09b; @mantz10b; @maughan12]. Through X-ray follow-up observations of these large samples of SZE-selected clusters, it has now become possible to extend these studies to high redshift. Moreover, by studying X-ray scaling relations in samples of SZE-selected clusters, it is possible to reduce the impact of selection-related biases that would have to be carefully corrected in studies of X-ray selected samples [@mantz10b].
In this work, we leverage the previous cosmological analyses of the SPT-SZ sample to characterize the X-ray observable–mass scaling relations by utilizing [[*XMM-Newton*]{}]{} follow-up observations of 59 SPT-selected clusters in the redshift range $0.20<z<1.5$. Here we focus on X-ray observables, which have direct implications for the structure evolution of the Universe. The halo masses we use in this analysis are derived from the observed SZE signal-to-noise ratio $\xi$ and redshift $z$ using the SZE mass–observable relation as calibrated within a self-consistent cosmological analysis that accounts for selection biases and systematic uncertainties on the masses [@bocquet15; @deHaan16]. Employing SZE masses allows us to extend studies of scaling relations to higher redshifts, enabling more robust studies of the redshift trends in these scaling relations. This cosmological analysis uses external mass information for a subset of clusters (i.e., weak lensing calibrated $Y_X$ measurements for 82 systems, as described in Section \[sec:M500\]), however inherently the cluster masses are based on the assumption that the cluster mass is well-described by the assumed functional form of the SZE-mass scaling relation and a general cluster mass function that can be well-fit to a $\Lambda$CDM cosmology. In this context, our results are comparable to other works that have performed similar analyses that jointly constrain observable–mass scaling relations in the context of a cosmological model (e.g., @mantz16), but in our our case using a different observable (i.e., SZE vs X-ray) and cluster sample (i.e., SPT-SZ vs RASS).
In this work, we are interested in comparing the measured X-ray observable–mass scaling relations to other results in the literature, including: the self-similar expectation, results based on direct-mass measurements, and results that include cosmological information. Agreement between results would indicate that cluster scaling relations are well-understood across a broad range of observables and assumptions, while differences could be indicative of tensions in the underlying assumptions or differences in the underlying cluster samples.
Robust observations of cluster scaling relations and their comparison to scaling relations from structure formation simulations then allow the baryonic physics and subgrid physics in the simulations to be tested and constrained. These constraints are crucial to accurately predicting the matter power spectrum [e.g., @springel18] and halo mass function [e.g., @bocquet16] needed to support forefront observational cosmological studies employing weak lensing, galaxy clustering and cluster counts.
The cluster sample and details of the [[*XMM-Newton*]{}]{} data reduction are given in Section \[sec:analysis\]. An explanation of the SZE-based halo masses and the measurements of the X-ray observables appears in Section \[sec:xrayobs\]. In Section \[sec:sclrelformfit\] we present our fitting procedure, and in Section \[sec:sclrel\] we present the X-ray scaling relations derived from this sample. Finally, we discuss our conclusions in Section \[sec:concl\].
All errors quoted throughout the paper correspond to $68\%$ (or $\Delta C$-stat=1) single-parameter confidence intervals unless otherwise stated. Throughout the paper, we adopt a standard, flat $\Lambda$CDM cosmology with the latest cosmological results from [@deHaan16]—$H_{0} =67.74$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{M}$= 0.304, and $\sigma_{8}=$0.82. In this work we refer to the cluster halo mass, [$M_{500}$]{}, as the total mass within a sphere of radius [$R_{500}$]{}. The overdensity radius [$R_{500}$]{} is defined as the radius within which the mean mass density of the cluster is 500 times the critical density of the Universe at that redshift.
Sample Selection and Data Reduction {#sec:analysis}
===================================
Sample Selection
----------------
SPT has detected 516 galaxy clusters via the SZE in the 2500 degree$^2$ SPT-SZ Survey at $0<z<1.8$ with masses ${\ensuremath{M_{500}}}\geq 3\times10^{14}{\ensuremath{M_{\odot}}}$ [@bleem15]. The redshifts of many of these clusters have also been reported in @ruel14 [@bayliss16]. [[*XMM-Newton*]{}]{} X-ray observations of 40 of these SZE-selected clusters have been performed through several programs (PIs: A. Andersson, B. Benson, J. Mohr, R. Suhada, E. Bulbul). An additional 33 clusters have been observed through various other non-SPT small programs. Five clusters have been excluded from this analysis, because one scattered below the detection threshold when better data were available (SPT-CL J2343$-$5521), and four observations are dominated by background flares (SPT-CL J0411$-$4819, SPT-CL J0013$-$4906, SPT-CL J0257$-$5732, SPT-CL J2136$-$6307).
![The distribution in SZE halo mass and redshift of the SPT-selected galaxy clusters observed with [[*XMM-Newton*]{}]{} is shown with each cluster appearing as a point with error bar. The inset shows the cluster redshift histogram.[]{data-label="fig:zhist"}](fig1){width="45.00000%"}
We exclude clusters at $z<0.20$ from the scaling relation analysis, because their SZE mass estimates obtained via the $\zeta$–[$M_{500}$]{} relation (Section \[sec:M500\]) are impacted by the filtering adopted to remove signal from the primary CMB [see e.g., @benson13]. From this sample of 68 clusters, 59 are at redshift z$>$0.2 and have a total of 1000 or more filtered source counts in MOS observations and are therefore included in our final sample. The details of the [[*XMM-Newton*]{}]{} observations of these clusters are given in Table \[table:obs\].
The final sample is shown in Figure \[fig:zhist\] in redshift-mass space with an inset redshift histogram. This cluster sample is not a complete SZE signal-to-noise selected cluster sample. It has a median mass and redshift of ${\ensuremath{M_{500}}}=4.77\times10^{14}{\ensuremath{M_{\odot}}}$ and $z_{\mathrm{med}}=0.45$, and five of the clusters lie at $z>1$. Nevertheless, the sample we study here has similar median mass, median redshift, and fraction of $z>1$ clusters as the SPT-SZ cosmology sample [@deHaan16] , which has a median mass ${\ensuremath{M_{500}}}=4.57\times10^{14}{\ensuremath{M_{\odot}}}$ (with roughly $6\%$ of the clusters at $z>1$), although it has a lower median redshift, 0.45 vs 0.55.
XMM-Newton Data Reduction {#sec:reduction}
-------------------------
Our [[*XMM-Newton*]{}]{} data reduction is described in detail in @Bulbul12a; here we summarize the main steps. [[*XMM-Newton*]{}]{} EPIC-MOS data analysis is carried out with Science Analysis System (SAS) version 16.0.0 and the latest available calibration files from Feb 2017. The Extended Source Analysis (ESAS) tools are used to reduce the data and extract the final data products [@snowden08]. The event files are filtered from the periods with elevated backgrounds through light-curve filtering. The good time interval files are produced and used to create cleaned event lists. The net exposure time after filtering the event lists for good time intervals is given in Table \[table:obs\]. There are three main detectors on board [[*XMM-Newton*]{}]{}: MOS1, MOS2, and PN. The back illuminated PN observations can be more sensitive to proton flares compared to MOS observations[^1]. As a result, the majority of the PN observation of some clusters in the sample is lost due to background filtering. Additionally, @schellenberger15 reports up to a $54\%$ bias in temperature measurements between [*Chandra*]{} and PN temperature measurements in the soft 0.7–2 keV band where the bulk of detected photon flux from the high redshift clusters appears. To avoid creating potential biases in the X-ray observables, we only use MOS observations in this analysis. We examine the individual chips which may be affected by an anomalous background level and exclude them from further analysis [@kuntz08].
The images are created in the 0.5–2 keV band from the filtered event files and used to detect point sources within the MOS field-of-view (FOV). The images are examined carefully for point sources missed by the CIAO algorithm [*wavdetect*]{}. An exposure map is created for each MOS detector and each pointing to account for chip gaps and mirror vignetting. The quiescent particle background (QBP) image is created from the filter-wheel closed data as described in @snowden08. The images and exposure maps of MOS1 and MOS2 detectors are combined prior to the background subtraction. The CIAO tool [*wavdetect*]{} convolved with the [[*XMM-Newton*]{}]{}’s point-spread-function (PSF) is used on the background-subtracted and exposure corrected images to detect point sources within the MOS FOV. All these point sources are excluded from further analysis.
We extract spectra using the ESAS tool [*mos-spectra*]{} within a radius of [$R_{500}$]{} for each cluster (see Section \[sec:xrayobs\] for the details of the [$R_{500}$]{} calculation). Redistribution matrix files (RMFs) and ancillary response files (ARFs) are created with [*rmfgen*]{} and [*arfgen*]{}, respectively. QPB is subtracted from the total spectra prior to the fitting. The spectral fitting of the source is done in the spectral fitting package XSPEC 12.9.0 [@arnaud96] with ATOMDB version 3.0.8 [@Smith2001; @Foster2012]. The adopted solar abundances are from @lodders09. The Galactic Column density is allowed to vary within $15\%$ of the measured @kalberla05 LAB value in our fits, following the approach described in @mcdonald16b. We use [*C-stat*]{} as a goodness of the fit estimator in XSPEC.
Spectra are extracted from two apertures of $<{\ensuremath{R_{500}}}$ and 0.15 [$R_{500}$]{}–[$R_{500}$]{} (again, see Section \[sec:xrayobs\] for discussion of [$R_{500}$]{}). The fits are performed in the 0.3–10 keV energy interval. The higher energy band 7–10 keV is used to constrain soft-proton contamination accurately. Soft-proton flares are largely removed by the light curve filtering. However, after the filtering some residuals may remain in the data. These are modeled by including an extra power-law model component to the total model and the MOS diagonal response matrices provided in the SAS distribution [@snowden08]. The cluster emission is fit with an absorbed single temperature [*apec*]{} model with free metallicity and temperature. Constraining metallicity is challenging for low-count observations of some of our high-z clusters. In these cases, we fixed the metallicity at 0.3$Z_{\odot}$, the typical value at both low and high redshifts [@tozzi03].
[lcccc]{}
We also consider the X-ray foreground emission, including Galactic halo, local hot bubble, cosmic X-ray background due to unresolved extragalactic sources, and solar wind charge exchange. The ROSAT All-Sky Survey background spectra[^2] extracted beyond $R_{\mathrm{vir}}$ (discussed in Section \[sec:xrayobs\]) are used to model the soft X-ray background as described in [@Bulbul12a]. The soft X-ray emission from the local hot bubble is modeled with a cool unabsorbed [*apec*]{} component with kT$\approx$0.1 keV, abundance of $Z_{\odot}$ at $z=0$, while the Galactic halo is modeled with a warmer absorbed thermal component kT$\approx$0.25 keV, abundance of $Z_{\odot}$ at $z=0$. The temperatures of the [*apec*]{} models are restricted, but the normalizations are allowed to vary in our fits. We model the cosmic X-ray background due to unresolved point sources using an absorbed power-law component with a spectral index of 1.4 [@Hickox2005] and normalization of $\approx$9$\times10^{-7}$ photons keV$^{-1}$ cm$^{-2}$ s$^{-1}$ at $\approx$1 keV [@kuntz08; @moretti2003]. The bright instrumental fluorescent lines Al–K (1.49 keV) and Si–K (1.74 keV) are not included in the MOS QBP files. Therefore, we model these instrumental lines by adding Gaussian models to our spectral fits to determine the best-fit energies, and normalizations.
Because of scattering in the [[*XMM-Newton*]{}]{} mirrors, some of the flux that originates from one area of the sky is detected in a different area of the detector. This is not a major concern if the gradient in plasma temperature from core to outskirts is smooth; however, it may be important for clusters with a strong cool core. Additionally, for high redshift clusters, 0.15 ${\ensuremath{R_{500}}}$ (discussed in Section \[sec:xrayobs\]) is comparable to the PSF for [[*XMM-Newton*]{}]{}, so this PSF effect is crucial and must be accounted for when making spectral fits. This radial cross-talk or contamination effect is treated as an additional model component in XSPEC. The cross-talk ARFs for the contribution of X-rays originating from a region on the sky to the another region on the detector are created using the SAS tool [*arfgen*]{} [@snowden08]. The cross-talk correction is applied to eliminate PSF effects for all clusters in our sample.
Cluster Masses and X-ray Observables {#sec:xrayobs}
====================================
The relationship between cluster X-ray observables (including emission-weighted mean temperature [${T_{\mathrm{X}}}$]{}, integrated pressure [$Y_{\mathrm{X}}$]{}, ICM mass [$M_{\mathrm{ICM}}$]{} and luminosity [$L_{\mathrm{X}}$]{}) and halo mass and cluster redshift exhibit a low scatter outside of the cluster center, where non-gravitational effects such as heating and cooling processes are less important [@fabian94c; @mohr97; @ohara06; @kravtsov06b; @nagai07]. We, therefore, measure all the X-ray observables both with and without the core region (except for the ICM mass [$M_{\mathrm{ICM}}$]{} where the core has no impact). Specifically, we extract observables within an aperture (0.15–1)[$R_{500}$]{} (core-excised marked as $cex$) and (0–1)[$R_{500}$]{} (core-included marked as $cin$). The cluster radius [$R_{500}$]{} is determined using the SZE-based halo mass [$M_{500}$]{} using $${\ensuremath{R_{500}}}= \left( \frac{3 {\ensuremath{M_{500}}}}{4\pi \times 500 \rho_{\mathrm{crit}}(z)}\right)^{1/3},$$ where the [$M_{500}$]{} masses are described in the next section, and $\rho_{\mathrm{crit}}(z)$ is the critical density of the Universe at the cluster redshift.
SZE-based mass $M_{500}$ {#sec:M500}
------------------------
We derive the cluster mass [$M_{500}$]{} based on the SZE signal-to-noise ratio $\xi$ and redshift $z$ as determined by SPT. The measured signal-to-noise $\xi$ is a biased observable subject to Gaussian noise that is extracted through a matched filter approach that employs a $\beta$ model with three degrees of freedom: sky location ($\alpha,\delta$) and core-radius $\theta_\mathrm{C}$. The mean value of the signal-to-noise $\left\langle \xi \right\rangle$ is related to the underlying unbiased signal-to-noise $\zeta$ as follows. $$\label{eq:sz_sr3}
\left\langle \xi \right\rangle = \sqrt{\zeta^{2} + 3} \, ,$$ for $\zeta>2$ [@deHaan16]. The $\zeta$–mass scaling relation is parametrized as follows: $$\label{eq:sz_sr1}
\zeta = A_{\mathrm{SZ}}\left( \frac{{\ensuremath{M_{500}}}}{4.3\times10^{14}{\ensuremath{M_{\odot}}}} \right)^{B_{\mathrm{SZ}}}
\left( \frac{E(z)}{E(z_{\mathrm{piv}})} \right)^{C_{\mathrm{SZ}}} \, ,$$ where the normalization is $A_{\mathrm{SZ}}$, the mass trend parameter is $B_{\mathrm{SZ}}$, the redshift trend parameter is $C_{\mathrm{SZ}}$, and there is log-normal intrinsic scatter in the observables at fixed mass of $\sigma_{\ln\zeta}$.
In this work we marginalize over the parameters of the $\zeta$–mass relation while fitting the parameters of the X-ray scaling relations that are investigated. This ensures that the final uncertainties in the X-ray observable–mass–redshift scaling relations include the systematic uncertainties associated with the imperfectly known SZE-based halo masses. In the interest of focusing on the X-ray scaling relations, we adopt priors on the parameters of the $\zeta$–mass relation that correspond to the fully marginalized posterior distributions reported in [@deHaan16] (Table \[tab:priors\]). This approach does not capture any covariances among the $\zeta$–mass scaling relation parameters, but these are indeed small [see @deHaan16 Figure 5]. The advantage is that the likelihood we must calculate in each iteration of the Markov chain involves our X-ray observables and the simple priors on the SZE $\zeta$–mass relation parameters.
[ll]{} Parameters & Priors\
\
$A_{\mathrm{SZ}}$ &$\mathcal{N}(4.842, 0.913^2)$\
$B_{\mathrm{SZ}}$ &$\mathcal{N}(1.668, 0.083^2)$\
$C_{\mathrm{SZ}}$ &$\mathcal{N}(0.550, 0.315^2)$\
$\sigma_{\ln\zeta}$ &$\mathcal{N}(0.199, 0.069^2)$\
\
$A_{{\ensuremath{{T_{\mathrm{X}}}}}}$ &$\mathcal{U}(0.1, 20)$ keV\
$A_{{\ensuremath{M_{\mathrm{ICM}}}}}$ & $\mathcal{U}(10^{12}, 2\times10^{14})$ [$M_{\odot}$]{}\
$A_{{\ensuremath{Y_{\mathrm{X}}}}}$ &$\mathcal{U}(5\times10^{12}, 2\times10^{15})$ keV[$M_{\odot}$]{}\
$A_{{\ensuremath{L_{\mathrm{X}}}}}$ &$\mathcal{U}(2\times10^{43}, 1.2\times10^{45})$ ergs s$^{-1}$\
$B_{{\ensuremath{\mathcal{X}}}}$ &$\mathcal{U}(0.1, 3.5)$\
$C_{{\ensuremath{\mathcal{X}}}}$ &$\mathcal{U}(-4, 4)$\
$\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ &$\mathcal{U}(0.005, 1.5)$\
$\gamma_{{\ensuremath{\mathcal{X}}}}$ &$\mathcal{U}(-4, 4)$\
$\delta_{{\ensuremath{\mathcal{X}}}}$ &$\mathcal{U}(-4,4)$\
The priors we adopt on the SZE observable mass relation are shown in Table \[tab:priors\], where $\mathcal{N}(\mu, \sigma^2)$ corresponds to a Gaussian with mean $\mu$ and dispersion $\sigma$. These SZE $\zeta$–mass parameter constraints emerge from a joint cosmology and mass calibration analysis that uses as input: (1) the SPT cluster distribution in $\xi$ and z (i.e. the number counts), (2) mass information from externally weak lensing calibrated [$Y_{\mathrm{X}}$]{} measurements for 82 systems, and (3) external cosmological parameter constraints [for more extensive discussion of SPT mass calibration see, e.g., @bocquet15; @chiu17]. For the baseline priors listed above, the external cosmological priors include a prior on the Hubble parameter [@riess11] and a prior on the baryon density parameter from Big Bang Nucleosynthesis [@cooke14].
Although the mass calibration presented in @deHaan16 includes information from Chandra X-ray observations of 82 clusters, we stress that the mass information is dominated by the cluster distribution in $\xi$ and redshift (i.e., the halo mass function information). That is, the $\zeta-$mass$-$redshift relation used to calculate SPT-SZ masses does not simply follow the employed ${\ensuremath{Y_{\mathrm{X}}}}-$mass$-$redshift relation, because it is a subdominant component of the mass information. Moreover, we adopt the resulting posteriors of the $\zeta-$mass relation as the priors in this work, effectively marginalizing over the systematic uncertainties of all ingredients used in calibrating the cluster mass. Modeling these priors as independent Gaussian distributions is appropriate, given the lack of strong covariances in the joint parameter constraints presented in @deHaan16 [see Figure 5]. It is important to note that the correlated intrinsic scatter between the mass proxies of SZE and X-ray does not impact the mass calibration with the current sample size [@deHaan16; @dietrich17], therefore, we can use the existing $\zeta-$mass$-$redshift relation with marginalized systematic uncertainties to investigate the X-ray observable-to-mass scaling relations.
To foreshadow an additional set of results that we present, we also adopt a separate set of priors derived from the second results column of Table 3 in @deHaan16, which include also an external cosmological prior coming from BAO distance measurements [@anderson14]. This set of results is consistent with the baseline results, but has smaller uncertainties (because the cosmological uncertainties typically dominate the posterior distributions of the SZE $\zeta$–mass parameters) and has a shift of $\Delta C_\mathrm{SZ}=+0.3$ that translates into a corresponding shift in the redshift trend parameters in the X-ray scaling relations.
Because we adopt similar four-parameter scaling relations for both the SZE and X-ray observables, we denote the targeted X-ray scaling relation (e.g., equation \[eq:FormA\]) as $r_{\ensuremath{\mathcal{X}}}= (A_{\ensuremath{\mathcal{X}}}, B_{\ensuremath{\mathcal{X}}}, C_{\ensuremath{\mathcal{X}}}, \sigma_{\ln {\ensuremath{\mathcal{X}}}})$ and the one used for estimating [$M_{500}$]{} as $r_{\zeta}=(A_{\mathrm{SZ}}, B_{\mathrm{SZ}}, C_{\mathrm{SZ}}, \sigma_{\ln\zeta})$. The notation $r_{\ensuremath{\mathcal{X}}}$ can be similarly extended to the five-parameter scaling relations for the X-ray observables, for which we define the functional forms in Section \[sec:form\].
We stress that the cluster masses in our work include corrections for selection biases (e.g. the Eddington bias, the Malmquist bias) and therefore they reflect the unbiased distribution of cluster mass [$M_{500}$]{} given the observable $\xi$ and redshift $z$ measured for each SZE-selected cluster.
[lrccccccccccc]{}
X-ray Observables {#sec:X-rayObservables}
-----------------
We measure the temperature, metallicity, and luminosity by fitting the spectra extracted in the apertures of the core-included region ($cin$, $r<{\ensuremath{R_{500}}}$) and core-excised region ($cex$, $0.15{\ensuremath{R_{500}}}<r<{\ensuremath{R_{500}}}$) with a single temperature thermal model. The best-fit core-included temperatures ([${T_{\mathrm{X,cin}}}$]{}), metallicity ([$Z_{\mathrm{X,cin}}$]{}), and luminosities ([$L_{\mathrm{X,cin}}$]{}) and core-excised temperatures ([${T_{\mathrm{X,cex}}}$]{}), metallicity ([$Z_{\mathrm{X,cex}}$]{}), and luminosities ([$L_{\mathrm{X,cex}}$]{}) are given in Table \[table:observables\]. In some clusters, the statistics of the observations are too poor to allow a determination of the global metallicity. In these cases, the metallicity is fixed to $0.3Z_{\odot}$ [@tozzi03; @mcdonald16b]. The metallicity constraints, and their evolution with redshift in this sample is extensively discussed in @mcdonald16b.
{width="49.00000%"} {width="49.00000%"}
The X-ray surface brightness is extracted from background-subtracted, exposure-corrected images within 1.5[$R_{500}$]{} in the fitting environment Sherpa in [*CIAO*]{} [@Freeman2001; @doe07]. We fit a 2-dimensional $\beta$ profile to determine the cluster centroids within software package [*Sherpa*]{}. This method also allows for precise measurements of X-ray centroids of the clusters in the sample. The X-ray surface brightness $S_{\mathrm{X}}$ (in units of erg s$^{-1}$ cm$^{-2}$ steradian$^{-1}$), produced by thermal Bremsstrahlung and line emission, is expressed as $$S_\mathrm{X} = \frac{1}{4\pi(1+z)^{4}}\int n_\mathrm{e} n_\mathrm{H} \Lambda_\mathrm{eH} (T_{\mathrm{X}}, Z)\ dl,$$
where $\Lambda_\mathrm{eH}(T_{\mathrm{X}}, Z)$ is the band averaged emissivity which is dependent on plasma temperature and metallicity, $dl$ is the integral along the line of sight, and $z$ is the cluster redshift. The electron and Hydrogen number densities ($n_\mathrm{e}$ and $n_\mathrm{H}$) have only weak dependence on plasma temperature and assumed abundance when derived from surface brightness in the 0.5–2 keV band [@mohr99].
We fit the surface brightness profiles using an analytic density model [@Bulbul2010 Bu10 hereafter]: $$n_{\mathrm{e}}(r) =n_{\mathrm{e}0}\left(\frac{1}{\left(\beta-2\right)} \frac{(1+r/r_{\mathrm{s}})^{\beta-2}-1}{r/r_{\mathrm{s}}(1+r/r_{\mathrm{s}})^{\beta-2}} \right)^{n}$$ where $n_{\mathrm{e}0}$ is the normalization of the electron density profile, $r_{\mathrm{s}}$ is the scale radius, $n$ is the slope of the density profile, and $\beta$ is the slope of the dark matter potential. We assume that the dark matter halos of the SPT selected sample follows the Navarro-Frenk-White (NFW) profile with a slope of $\beta=2$ [@navarro1997] and provides a good description of the electron density [Bu10; @Bonamente2012]. Application of the L’Hospital rule gives an electron density profile under the assumption of a NFW-like matter profile, $$n_{\mathrm{e}}(r) = \tau_\mathrm{cool}(r) \ n_{\mathrm{e}0}\left(\frac{\ln (1+r/r_{\mathrm{s}})}{r/r_{\mathrm{s}}} \right)^{n}.$$ The Bu10 density profile has been used for fitting both X-ray and SZE data [@Landry2013; @Romero2017]. The core taper function $\tau_\mathrm{cool}(r)$ is used to fit the surface brightness profiles of cool-core clusters [@vikhlinin06] $$\tau_\mathrm{cool}(r) = \frac{\alpha+(r/r_{\mathrm{cool}})^{\gamma}}{1+(r/r_{\mathrm{cool}})^{\gamma}}$$ For non-cool core clusters the parameter $\alpha$ is set to 1.
The Bu10 density model is projected along the line-of-sight and fit to the surface brightness profile obtained from background subtracted and exposure corrected X-ray images. The fitting is performed using a Markov chain Monte Carlo (MCMC) sampler within the [*emcee*]{} package in python [@foreman13]. The best-fit parameter values and their 1$\sigma$ uncertainties for non-cool core clusters (e.g., $n_{\mathrm{e}0}$, $n$, $r_{\mathrm{s}}$) and cool-core clusters (e.g., $n_{\mathrm{e}0}$, $n$, $\alpha$, $r_{\mathrm{s}}$, and $r_{cool}$) are determined using a maximum likelihood method. The surface brightness profile fit to the MOS observations of a non-cool core cluster SPT-CL J0304$-$4401 and a cool-core cluster SPT-CL J2217$-$6509 are shown in Figure \[fig:sbfit\].
To compute the ICM mass of a cluster within a given aperture of [$R_{500}$]{}, we use the enclosed ICM mass obtained by integrating the best-fit 3D ICM density profile, $$M_{\mathrm{ICM}} = 4\pi \mu_\mathrm{e} m_\mathrm{p}\int_{0}^{{\ensuremath{R_{500}}}} n_\mathrm{e}(r)\ r^{2}\ dr,$$ where $\mu_\mathrm{e}$ is the mean molecular weight of the electrons, and $m_\mathrm{p}$ is the proton mass. The ICM mass measurements within [$R_{500}$]{} for each cluster in the sample are given in Table \[table:observables\]. We use $\mu_\mathrm{e}= 1.17$ when determining the cluster ICM mass. The integrated Compton-$y$ parameter is the product of the ICM mass and temperature $${\ensuremath{Y_{\mathrm{X}}}}= {\ensuremath{M_{\mathrm{ICM}}}}\times {\ensuremath{{T_{\mathrm{X}}}}}\, ,$$ where [${T_{\mathrm{X}}}$]{} is the projected temperature measured within a 2D aperture either with or without the core and [$M_{\mathrm{ICM}}$]{} is integrated within a 3D sphere of radius [$R_{500}$]{}.
As described already in Section \[sec:M500\], there are remaining uncertainties in the SZE-based halo masses. This means that the extraction radius [$R_{500}$]{} used above is not a single value for each cluster, but a distribution of values. To include these uncertainties, we marginalize over them when studying the X-ray scaling relations. As described in Section \[sec:priors\], this means that we evaluate the X-ray observable at a range of radii [$R_{500}$]{} consistent with the SZE observable $\xi$ and redshift $z$. Specifically, we use the best-fit density profile to calculate the ICM mass in each fit iteration. For [$L_{\mathrm{X}}$]{} we extract the X-ray luminosity at a single radius—the baseline [$R_{500}$]{}—in this work, because we find the change in [$L_{\mathrm{X}}$]{} due to the radial range in the surface brightness fit is negligible. For [${T_{\mathrm{X}}}$]{} we have in general too few photons to make spectral fits beyond the baseline [$R_{500}$]{}, and so we adopt only a single radius for the temperature extraction. This means that for [$Y_{\mathrm{X}}$]{} we are properly including the variation of the [$M_{\mathrm{ICM}}$]{} component with [$R_{500}$]{} but not the [${T_{\mathrm{X}}}$]{} component.
Scaling Relation Form and Fitting {#sec:sclrelformfit}
=================================
Self-similar models, based on gravitational collapse in clusters, predict simple power-law relations between cluster properties [@kaiser86] that have been observed [using, e.g., ICM temperature, luminosity, ICM mass, X-ray isophotal size and total halo mass; @smith79; @mushotzky97; @mohr97; @mohr99; @arnaud99]. As previously noted, the observed scaling relations often depart from self-similiar behavior, and this has been interpreted as evidence of feedback into the ICM from star formation and AGN as well as radiative cooling in the cluster cores.
In this section, we describe how we determine the best-fit parameters of the X-ray observable–halo mass–redshift scaling relations for the sample of 59 SPT selected galaxy clusters observed with [[*XMM-Newton*]{}]{} at $0.2<z<1.5$.
Three Forms of each Scaling Relation {#sec:form}
------------------------------------
We use three functional forms to characterize the X-ray observable–mass–redshift scaling relations. In all cases, there are pivot masses and redshifts that should be chosen to be near the median values of the sample to reduce artificial covariances between the amplitude parameter and the mass and redshift trend parameters. For the X-ray observable [$\mathcal{X}$]{} to mass scaling relations, the pivot mass and pivot redshift are ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ and ${\ensuremath{z_{\mathrm{piv}}}}=0.45$, respectively.
The first form, similar to that used in [@vikhlinin09a] and many publications since, is defined as follows: $$\label{eq:FormA}
{\ensuremath{\mathcal{X}}}= A_{{\ensuremath{\mathcal{X}}}}\left( \frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}} \right)^{B_{{\ensuremath{\mathcal{X}}}}}
\left( \frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})} \right)^{C_{{\ensuremath{\mathcal{X}}}}},$$ where the normalization and trend parameters in mass and redshift are $A_{{\ensuremath{\mathcal{X}}}}$, $B_{{\ensuremath{\mathcal{X}}}}$ and $C_{{\ensuremath{\mathcal{X}}}}$, respectively, for the observable ${\ensuremath{\mathcal{X}}}$. Note that the redshift trend in this formulation is expressed as a function of the Hubble parameter $H(z)=H_0E(z)$, where $E^2(z)=\Omega_\mathrm{M}(1+z)^3+\Omega_\Lambda$ at late times in a flat $\Lambda$CDM Universe. That is, in this parametrization, the redshift evolution of the X-ray observable–mass relation is attributed an explicit cosmological dependence. In the case where the redshift evolution has a different cosmological dependence than adopted here (e.g., the evolution is non-self similar), then assuming this form will lead to biases in cosmological analyses. We refer to equation (\[eq:FormA\]) as Form I hereafter.
The second form includes the expected self-similar evolution of the observable with redshift, which depends on the cosmologically dependent evolution of the critical density, while modeling departures of the observable from self-similar evolution with a function $(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$. With this form we are adopting the view that the departures from self-similar evolution do not have a clearly understood cosmological dependence. Therefore, we model the departures with the cosmologically agnostic form $(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ that has been adopted in many previous works [e.g. @lin06]. This form is defined as follows: $$\begin{aligned}
\label{eq:FormB}
{\ensuremath{\mathcal{X}}}&= &A_{{\ensuremath{\mathcal{X}}}}\left( \frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}} \right)^{B_{{\ensuremath{\mathcal{X}}}}}
\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{C_{\mathrm{{\ensuremath{\mathcal{X}}},SS}}}
\left( \frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)^{\gamma_{{\ensuremath{\mathcal{X}}}}}\end{aligned}$$ where the normalization and mass trend are similarly characterized by the parameters $A_{{\ensuremath{\mathcal{X}}}}$ and $B_{{\ensuremath{\mathcal{X}}}}$, respectively. The redshift trend is modeled with [$C_\mathrm{{\ensuremath{\mathcal{X}}},SS}$]{} fixed to the self-similar expectation along with the factor $(1+z)^{\gamma_{{\ensuremath{\mathcal{X}}}}}$ to describe the departure of the redshift trend from the self-similar expectation. For instance, ${\ensuremath{C_\mathrm{{\ensuremath{\mathcal{X}}},SS}}}= \frac{2}{3}$ for the X-ray temperature–mass–redshift relation. In this way, the parameter $\gamma_{{\ensuremath{\mathcal{X}}}}$ directly quantifies the deviation from the self-similar redshift trend. This form of the scaling relation is easily distinguishable, because it has a parameter $\gamma_{\ensuremath{\mathcal{X}}}$ rather than $C_{\ensuremath{\mathcal{X}}}$. We refer to equation (\[eq:FormB\]) as Form II hereafter.
The third form we adopt is much like Form II above, but it includes a cross-term between cluster mass and redshift to characterize the possibility of having a redshift-dependent mass trend. Specifically, the third functional form is $$\begin{aligned}
\label{eq:FormC}
{\ensuremath{\mathcal{X}}}&= &A_{{\ensuremath{\mathcal{X}}}}\left( \frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}} \right)^{B'_{{\ensuremath{\mathcal{X}}}}}
\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{C_{\mathrm{{\ensuremath{\mathcal{X}}},SS}}}
\left( \frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)^{\gamma_{{\ensuremath{\mathcal{X}}}}}\end{aligned}$$ where the mass trend $B'_{{\ensuremath{\mathcal{X}}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln \left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ has a characteristic value of $B_{{\ensuremath{\mathcal{X}}}}$ at the pivot redshift and an additional rate of variation $\delta_{{\ensuremath{\mathcal{X}}}}$ with redshift. The normalization parameter $A_{{\ensuremath{\mathcal{X}}}}$ and the redshift trend $\gamma_{{\ensuremath{\mathcal{X}}}}$ are defined as in Form II. Specifically, the redshift trend is structured to capture the departures from the expected self-similar redshift evolution of the X-ray observable. It is worth mentioning that $\delta_{{\ensuremath{\mathcal{X}}}} = 0$ or statistically consistent with zero indicates there is no evidence for a redshift-dependent mass trend. We refer to equation (\[eq:FormC\]) as Form III hereafter.
For all three functional forms, we adopt log-normal intrinsic scatter in the observable at fixed mass, defined as $$\label{eq:intrinsic_scatter}
\sigma_{\ln{\ensuremath{\mathcal{X}}}} \equiv \sigma_{ \ln\left( {\ensuremath{\mathcal{X}}}| {\ensuremath{M_{500}}}\right) } .$$ In this way, each observable [$\mathcal{X}$]{} to mass scaling relation is parametrized by either $(A_{{\ensuremath{\mathcal{X}}}}, B_{{\ensuremath{\mathcal{X}}}}, C_{{\ensuremath{\mathcal{X}}}}, \sigma_{\ln{\ensuremath{\mathcal{X}}}})$, $(A_{{\ensuremath{\mathcal{X}}}}, B_{{\ensuremath{\mathcal{X}}}}, C_\mathrm{{\ensuremath{\mathcal{X}}},SS}, \gamma_{{\ensuremath{\mathcal{X}}}}, \sigma_{\ln{\ensuremath{\mathcal{X}}}})$ or $(A_{{\ensuremath{\mathcal{X}}}}, B_{{\ensuremath{\mathcal{X}}}}, C_\mathrm{{\ensuremath{\mathcal{X}}},SS}, \gamma_{{\ensuremath{\mathcal{X}}}}, \sigma_{\ln{\ensuremath{\mathcal{X}}}}, \\ \delta_{{\ensuremath{\mathcal{X}}}})$, and we denote these parameter sets by $r_{{\ensuremath{\mathcal{X}}}}$ hereafter for simplicity. Note that the expected self-similar redshift evolution parameter $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}$ is fixed, and so the first two parameterizations have four free parameters, and the last parametrization has five.
Fitting Procedure {#sec:fitting}
-----------------
We briefly introduce the likelihood and fitting framework below and refer the reader to previous publications for more details [@liu15a; @chiu16c]. This likelihood is designed to obtain the parameters of the targeted X-ray observable–mass–redshift scaling relation (e.g., $r_{{\ensuremath{\mathcal{X}}}}$) for a given sample that is selected using another observable (e.g., the SPT signal-to-noise $\xi$), for which the observable–mass–redshift relation is already known (e.g., equation (\[eq:sz\_sr1\]) used in this work). Specifically, the $i$-th term in the likelihood $\mathcal{L}_{i}$ contains the probability of obtaining the X-ray observable ${\ensuremath{\mathcal{X}}}_{i}$ for the $i$-th cluster at redshift $z_{i}$ with SZE signal-to-noise $\xi_{i}$, given the scaling relations $r_{\ensuremath{\mathcal{X}}}$ and $r_\mathcal{\xi}$. $$\label{eq:l15_likelihood1}
\begin{split}
\mathcal{L}_{i}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) &=
P({\ensuremath{\mathcal{X}}}_{i} | \xi_{i}, z_{i}, r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta} ) \\
&=
\frac{
\int\mathrm{d}{\ensuremath{M_{500}}}\ P({\ensuremath{\mathcal{X}}}_{i}, \xi_{i}| z_{i}, r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta} )\ n({\ensuremath{M_{500}}}, z_{i})
}{
\int\mathrm{d}{\ensuremath{M_{500}}}\ P( \xi_{i}| z_{i}, r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta} ) \ n({\ensuremath{M_{500}}}, z_{i})} \, ,
\end{split}$$ where $n({\ensuremath{M_{500}}}, z_{i})$ is the mass function whose inclusion allows the Eddington bias correction to be included when determining the mass corresponding to the SZE observable $\xi$ at redshift $z$. The integrals are over the relevant range of the mass $M_{500}$ used in the mass function. The @tinker08 mass function is used with fixed cosmological parameters in calculations of $n({\ensuremath{M_{500}}}, z_{i})$, although given the mass range of the SPT sample the use of a mass function determined from hydrodynamical simulations would make no difference [@bocquet16].
We ignore correlated scatter between the X-ray observable ${\ensuremath{\mathcal{X}}}$ and SZE observable $\xi$ in our analysis. This will not impact our results, because in previous studies of the SPT-SZ sample no evidence of correlated scatter between the X-ray [$Y_{\mathrm{X}}$]{}, X-ray based [$M_{\mathrm{ICM}}$]{} and the SZE signal-to-noise has emerged [@deHaan16; @dietrich17]. In future analyses with much larger X-ray samples, we plan to explore again the evidence for correlated scatter in the X-ray and SZE properties of the clusters. As discussed in @liu15a, in this limit of no correlated X-ray and SZE scatter, there are no selection effects to be accounted for in the X-ray scaling relation.
Based on Bayes’ Theorem, the best-fit scaling relation parameters $r_{\mathcal{X}}$ and $r_{\mathcal{X}}$ are obtained by maximizing the probability, $$\label{eq:prob_baye}
P(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) \propto \mathcal{L}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) \mathcal{P}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) \, ,$$ where $\mathcal{P}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta})$ is the prior on $r_{{\ensuremath{\mathcal{X}}}}$ and $r_{\zeta}$ (see Table \[tab:priors\]), and the likelihood $\mathcal{L}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta})$ is evaluated using equation (\[eq:l15\_likelihood1\]) as follows.
$$\label{eq:full_like}
\mathcal{L}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) = \prod_{i=1}^{N_{\mathrm{cl}}}~\mathcal{L}_{i}(r_{{\ensuremath{\mathcal{X}}}}, r_{\zeta}) \, ,$$
where $i$ runs over the $N_{\mathrm{cl}}$ clusters. We use the python package `emcee` to explore the parameter space. The intrinsic scatter and measurement uncertainties of $\xi_{i}$ for each cluster are taken into account while evaluating equation (\[eq:full\_like\]). We have verified that this likelihood recovers unbiased scaling relation parameters by testing it against large mocks ($>1300$ clusters). Moreover, it has been further optimized in the goal of obtaining the parameters of scaling relations in a high dimensional space [@chiu16c].
We note that in each iteration of the chain we use the current value of [$R_{500}$]{} for each cluster to recalculate the [$M_{\mathrm{ICM}}$]{} (see Section \[sec:X-rayObservables\]). For the temperature [${T_{\mathrm{X}}}$]{} and the luminosity [$L_{\mathrm{X}}$]{} we extract only once at the [$R_{500}$]{} appropriate for the model [$\zeta-M_{500}-z$]{} parameter values in our priors (see Table \[tab:priors\]), because the impact of adjusting [$R_{500}$]{} at each iteration is small.
Priors adopted during fitting {#sec:priors}
-----------------------------
As discussed in Section \[sec:M500\], we marginalize over the parameters of the $\zeta$–[$M_{500}$]{}-$z$ while fitting the X-ray observable [$\mathcal{X}$]{}–[$M_{500}$]{}-$z$ relations (i.e., $r_{\zeta}$ and $r_{\mathcal{X}}$, respectively). Specifically, we adopt informative priors on $r_{\zeta}$, which have been obtained in a joint cosmology and mass calibration analysis described in [@deHaan16 see Table 3]. Our baseline priors on $r_{\zeta}$ are listed in Table \[tab:priors\] and correspond to the posterior distributions for each parameter reported in the first results column of @deHaan16 [Table 3 ]. We explore a second set of priors on $r_{\zeta}$ corresponding to the posterior parameter distributions from the second results column of @deHaan16 [Table 3], and we report those results in Table \[tab:sclrel\_col2\].
Our approach allows us to effectively marginalize over the remaining uncertainties in our [$M_{500}$]{} estimates. In each iteration of the chain, each cluster has a different halo mass [$M_{500}$]{} and associated radius [$R_{500}$]{}. The X-ray observables [$M_{\mathrm{ICM}}$]{} and [$Y_{\mathrm{X}}$]{} defined in Section \[sec:X-rayObservables\] are then extracted at this radius [$R_{500}$]{} and used to determine the likelihood for this iteration. Final uncertainties on the X-ray observable scaling relation parameters $r_{\mathcal{X}}$ therefore include not only those due to measurement uncertainties but also due to the (largely systematic) uncertainties in the underlying halo masses.
In the fitting, we apply the uniform priors listed in Table \[tab:priors\] on $r_{{\ensuremath{\mathcal{X}}}}$ during the likelihood maximization. With this approach we evaluate the scaling relations Form I, II, and III. In Table \[tab:priors\] we present the parameter in the first column and the form of the prior in column two. In this table $\mathcal{N}$ denotes a normal or Gaussian distribution, and $\mathcal{U}$ represents a uniform or flat distribution between the two values presented.
In a final step, we report the parameters of the Form III relation also while fixing the parameters of $r_{\zeta}$ to the best-fit values derived in @deHaan16 (i.e., the central values listed in Table \[tab:priors\]). Through the comparison to the results when marginalizing over the posterior distributions with those when fixing the $r_{\zeta}$ parameters we can gauge the impact of the remaining systematic uncertainties on the SZE-based halo masses.
We note that the [@deHaan16] priors we adopt when estimating cluster halo masses are derived using the cluster mass function information (distribution of clusters in signal-to-noise $\xi$ and $z$) together with a sample of 82 [$Y_{\mathrm{X}}$]{} measurements that have been calibrated to mass first through hydrostatic masses [@vikhlinin09a] and later through weak lensing [@hoekstra15]. We note that the mass information from the [$Y_{\mathrm{X}}$]{} measurements is subdominant in comparison to that from the mass function information [see prior and posterior distributions on $r_\mathcal{\xi}$ parameters in Figure 5 of @deHaan16].
Moreover, the follow-up studies using weak lensing masses of 32 SPT-SZ clusters [@dietrich17] and using dynamical masses from 110 SPT-SZ clusters [@capasso17] have provided independent mass calibration of the [$\zeta-M_{500}-z$]{} relation and cross-checks of cluster masses, and they are all in excellent agreement with the cluster masses in [@deHaan16] as we adopt for our study. Ongoing work with DES weak lensing will further improve our knowledge of the [$\zeta-M_{500}-z$]{} relation, allowing even more accurate cluster halo mass estimates in the future [e.g., @stern18].
{width="95.00000%"} -0.2in
{width="95.00000%"} -0.2in
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------- ------------------------------------------------------ ------------------------------------------------------ ------------------------------------------ --------------------------------------- ---------------------------------------
Scaling Relation $A_{{\ensuremath{\mathcal{X}}}}$ $B_{{\ensuremath{\mathcal{X}}}}$ $C_{{\ensuremath{\mathcal{X}}}}$ $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ $\gamma_{{\ensuremath{\mathcal{X}}}}$ $\delta_{{\ensuremath{\mathcal{X}}}}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.36^{+0.70}_{-0.64}$ $0.80\pm 0.10$ $0.33\pm 0.27$ $0.18\pm 0.04$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $6.41^{+0.64}_{-0.66}$ $0.79^{+0.08}_{-0.12}$ $2\over3$ $0.18^{+0.05}_{-0.04}$ $-0.36^{+0.27}_{-0.26}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $6.48^{+0.58}_{-0.69}$ $0.79^{+0.09}_{-0.10}$ $2\over3$ $0.18^{+0.04}_{-0.04}$ $-0.22^{+0.29}_{-0.35}$ $0.81^{+0.56}_{-0.46}$
\[3pt\] III with fixed SZE params $6.41\pm 0.22$ $0.78^{+0.08}_{-0.09}$ $2\over3$ $0.16^{+0.04}_{-0.03}$ $-0.20^{+0.23}_{-0.25}$ $0.77^{+0.57}_{-0.47}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.17^{+0.71}_{-0.63}$ $0.83^{+0.09}_{-0.10}$ $0.28^{+0.28}_{-0.23}$ $0.13^{+0.05}_{-0.05}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $6.09^{+0.76}_{-0.51}$ $0.80^{+0.11}_{-0.08}$ $2\over3$ $0.13^{+0.04}_{-0.05}$ $-0.33^{+0.23}_{-0.28}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $6.31^{+0.57}_{-0.69}$ $0.81^{+0.09}_{-0.08}$ $2\over3$ $0.13^{+0.05}_{-0.04}$ $-0.30^{+0.27}_{-0.28}$ $0.35^{+0.53}_{-0.41}$
\[3pt\] III with fixed SZE params $6.17^{+0.20}_{-0.17}$ $0.79^{+0.10}_{-0.06}$ $2\over3$ $0.12^{+0.04}_{-0.03}$ $-0.29^{+0.19}_{-0.25}$ $0.38^{+0.41}_{-0.43}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}=1$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}=0$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.80^{+1.09}_{-0.87}$ $1.260^{+0.10}_{-0.11}$ $0.17^{+0.28}_{-0.29}$ $0.12^{+0.04}_{-0.08}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{0}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $7.37^{+0.76}_{-1.35}$ $1.26^{+0.12}_{-0.09}$ $0$ $0.10^{+0.04}_{-0.07}$ $0.18^{+0.30}_{-0.31}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $7.09^{+0.91}_{-1.11}$ $1.26^{+0.11}_{-0.09}$ $0$ $0.10^{+0.05}_{-0.07}$ $0.16^{+0.33}_{-0.31}$ $0.16^{+0.47}_{-0.44}$
\[3pt\] III with fixed SZE params $7.02^{+0.21}_{-0.27}$ $1.26^{+0.09}_{-0.07}$ $0$ $0.07\pm 0.05$ $0.20^{+0.20}_{-0.22}$ $0.26^{+0.42}_{-0.51}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={5\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.70\pm 1.1$ $2.00^{+0.19}_{-0.14}$ $0.44^{+0.46}_{-0.54}$ $0.15^{+0.05}_{-0.12}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.60\pm 1.1$ $1.99^{+0.17}_{-0.15}$ $2\over3$ $0.16^{+0.05}_{-0.12}$ $-0.21^{+0.50}_{-0.45}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.52^{+1.23}_{-0.91}$ $2.00^{+0.16}_{-0.17}$ $2\over3$ $0.16^{+0.07}_{-0.10}$ $-0.28^{+0.56}_{-0.40}$ $0.77^{+0.74}_{-0.53}$
\[3pt\] III with fixed SZE params $4.57^{+0.25}_{-0.21}$ $1.98^{+0.16}_{-0.10}$ $2\over3$ $0.07^{+0.09}_{-0.05}$ $-0.09^{+0.34}_{-0.32}$ $1.01^{+0.61}_{-0.71}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={5\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.31\pm 0.96$ $2.01^{+0.20}_{-0.13}$ $0.44\pm 0.49$ $0.16^{+0.04}_{-0.11}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.50^{+1.0}_{-1.1}$ $2.02^{+0.16}_{-0.17}$ $2\over3$ $0.11^{+0.07}_{-0.08}$ $-0.17^{+0.47}_{-0.50}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.54^{+1.09}_{-0.98}$ $2.01^{+0.18}_{-0.14}$ $2\over3$ $0.13^{+0.07}_{-0.08}$ $-0.20^{+0.52}_{-0.47}$ $0.55^{+0.78}_{-0.56}$
\[3pt\] III with fixed SZE params $4.40^{+0.23}_{-0.22}$ $2.04^{+0.10}_{-0.15}$ $2\over3$ $0.04^{+0.08}_{-0.03}$ $-0.14^{+0.33}_{-0.32}$ $0.67^{+0.66}_{-0.74}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={1}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2}$
I: ${\ensuremath{\mathcal{X}}}(z)\propto E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.20^{+0.91}_{-0.92}$ $1.93^{+0.16}_{-0.20}$ $1.72^{+0.53}_{-0.46}$ $0.25^{+0.10}_{-0.10}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(z)\propto E(z)^{2}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.12^{+0.91}_{-0.94}$ $1.89^{+0.23}_{-0.13}$ $2$ $0.27^{+0.08}_{-0.12}$ $-0.20^{+0.51}_{-0.49}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.39^{+0.82}_{-0.99}$ $1.93^{+0.19}_{-0.18}$ $2$ $0.28^{+0.07}_{-0.11}$ $-0.13^{+0.63}_{-0.46}$ $0.71^{+0.89}_{-0.72}$
\[3pt\] III with fixed SZE params $3.96^{+0.22}_{-0.24}$ $1.95^{+0.14}_{-0.18}$ $2$ $0.24^{+0.08}_{-0.06}$ $-0.02^{+0.32}_{-0.48}$ $0.84^{+0.81}_{-0.80}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={1}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $2.84^{+0.60}_{-0.53}$ $1.60^{+0.17}_{-0.13}$ $1.86^{+0.47}_{-0.43}$ $0.27^{+0.07}_{-0.10}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $2.84^{+0.53}_{-0.50}$ $1.60^{+0.16}_{-0.15}$ $2$ $0.27^{+0.07}_{-0.11}$ $-0.10^{+0.47}_{-0.42}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $2.89^{+0.55}_{-0.51}$ $1.56^{+0.18}_{-0.16}$ $2$ $0.28^{+0.07}_{-0.08}$ $0.10^{+0.35}_{-0.60}$ $0.30^{+0.86}_{-0.62}$
\[3pt\] III with fixed SZE params $2.66^{+0.17}_{-0.11}$ $1.60^{+0.14}_{-0.16}$ $2$ $0.26^{+0.05}_{-0.05}$ $-0.01^{+0.33}_{-0.42}$ $0.60^{+0.79}_{-0.75}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={4\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={7\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $15.4^{+2.8}_{-3.3}$ $2.15^{+0.24}_{-0.19}$ $1.90^{+0.55}_{-0.53}$ $0.29^{+0.09}_{-0.13}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{7\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $14.8^{+3.5}_{-2.7}$ $2.19^{+0.21}_{-0.17}$ $7\over3$ $0.29^{+0.08}_{-0.13}$ $-0.14^{+0.62}_{-0.57}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $13.8^{+3.2}_{-3.9}$ $2.12^{+0.23}_{-0.18}$ $7\over3$ $0.31^{+0.08}_{-0.12}$ $-0.26^{+0.58}_{-0.60}$ $1.53^{+0.31}_{-1.11}$
\[3pt\] III with fixed SZE params $14.94^{+0.65}_{-1.01}$ $2.24^{+0.13}_{-0.15}$ $7\over3$ $0.22^{+0.08}_{-0.10}$ $-0.17^{+0.43}_{-0.33}$ $1.67^{+0.28}_{-0.97}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={4\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={7\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $10.2^{+2.6}_{-2.1}$ $1.89^{+0.17}_{-0.18}$ $2.01^{+0.53}_{-0.44}$ $0.29^{+0.07}_{-0.12}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{7\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $10.7\pm 2.3$ $1.88^{+0.19}_{-0.17}$ $7\over3$ $0.27^{+0.07}_{-0.13}$ $-0.26^{+0.53}_{-0.43}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $10.4^{+2.4}_{-2.2}$ $1.86^{+0.21}_{-0.16}$ $7\over3$ $0.28^{+0.07}_{-0.09}$ $0.02^{+0.48}_{-0.58}$ $0.76^{+0.76}_{-0.71}$
\[3pt\] III with fixed SZE params $9.93^{+0.58}_{-0.49}$ $1.90^{+0.13}_{-0.18}$ $7\over3$ $0.25^{+0.07}_{-0.06}$ $-0.18^{+0.48}_{-0.32}$ $0.80^{+0.93}_{-0.57}$
\[3pt\]
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------- ------------------------------------------------------ ------------------------------------------------------ ------------------------------------------ --------------------------------------- ---------------------------------------
Scaling Relation Constraints {#sec:sclrel}
============================
In this section, we describe the results of the fits and compare them to the self-similar expectation and to previous results in the literature. We present the scaling relations involving [${T_{\mathrm{X}}}$]{}, then followed by [$M_{\mathrm{ICM}}$]{}, [$Y_{\mathrm{X}}$]{} and [$L_{\mathrm{X}}$]{}. For all X-ray observables aside from [$M_{\mathrm{ICM}}$]{} we present both core-included and -excised scaling relations.
Best-fit parameters and uncertainties are presented in Table \[tab:sclrel\], where the parameter constraints for each specific X-ray observable are presented in separate, delineated vertical subsections of the table. Within each table subsection the first line identifies the scaling relation and presents the self-similar expectation for the mass and redshift trends. Thereafter, the best-fit parameters are presented for the scaling relation Forms I, II, III and then III with fixed SZE scaling relation parameters. From left to right in the table we present the scaling relation and then the parameters for the normalization $A_{\ensuremath{\mathcal{X}}}$, mass trend $B_{\ensuremath{\mathcal{X}}}$, redshift trend $C_{\ensuremath{\mathcal{X}}}$ parametrized using $E(z)$, log-normal intrinsic scatter $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ of the observable at fixed [$M_{500}$]{}, departure from self-similar redshift scaling $\gamma_{\ensuremath{\mathcal{X}}}$, and redshift evolution of the mass trend $\delta_{\ensuremath{\mathcal{X}}}$.
[$T_{\mathrm{X}}-M_{500}-z$]{} Relation {#sec:tm}
---------------------------------------
Before cluster mass measurements were available, the emission-weighted ICM temperature was viewed as the most robust mass proxy available and was therefore employed in early studies of cluster scaling relations [@smith79; @mushotzky97; @mohr97; @mohr99]. Early attempts to constrain the [${T_{\mathrm{X}}}$]{}–mass relation using hydrostatic masses were carried out first for low temperature clusters using spatially resolved spectroscopy from *ROSAT* [@david93] and then later for clusters with a broad range of temperature using the *ASCA* observatory [@finoguenov01].
By combining the virial condition ($GM/R\sim T$) and the definition of the virial radius (${\ensuremath{R_{500}}}\sim [{\ensuremath{M_{500}}}/\rho_{\mathrm{crit}}]^{1/3}$) one can show that the self similar expectation for the [$T_{\mathrm{X}}-M_{500}-z$]{} relation is $${\ensuremath{{T_{\mathrm{X}}}}}\propto {{\ensuremath{M_{500}}}}^{2/3}\ E(z)^{2/3} \, .$$ As noted in Section \[sec:fitting\], we examine the scaling relations with and without the core region, and for three different scaling relation functional forms.
{width="95.00000%"}
![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the case of the core-included [$T_{\mathrm{X,cin}}-M_{500}-z$]{} relation. []{data-label="fig:triangle_txi"}](fig6.pdf){width="45.00000%"}
### Parameter constraints
We present the parameters associated with the [$T_{\mathrm{X,cin}}-M_{500}-z$]{} and [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relations relations in Table \[tab:sclrel\]. The marginalized posteriors of the parameters and joint parameter confidence regions using the core-excised and -included observables are contained in Figures \[fig:triangle\_txe\] and \[fig:triangle\_txi\]. Here we provide the best-fit [$T_{\mathrm{X,cin}}-M_{500}-z$]{} and [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relations for the Form II scaling relation. For the core-included X-ray emission-weighted mean temperature, $$\begin{aligned}
\label{eq:txi}
{\ensuremath{{T_{\mathrm{X,cin}}}}}&=
&6.41^{+0.64}_{-0.66}~\mathrm{keV}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{0.80^{+0.09}_{-0.12}} \nonumber \\
& & \left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{2}{3}} \left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.36^{+0.27}_{-0.26}},\end{aligned}$$ with the intrinsic scatter of $0.18^{+0.05}_{-0.04}$. For the core-excised X-ray temperature [${T_{\mathrm{X,cex}}}$]{}, the best-fit relation is $$\begin{aligned}
\label{eq:txe}
{\ensuremath{{T_{\mathrm{X,cex}}}}}&=
&6.09^{+0.76}_{-0.51}~\mathrm{keV}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{0.80^{+0.11}_{-0.08}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{2}{3}}\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.33^{+0.23}_{-0.28}},\end{aligned}$$ with intrinsic scatter of $0.13^{+0.05}_{-0.04}$. As for all relations, the mass and redshift pivots are ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ and ${\ensuremath{z_{\mathrm{piv}}}}=0.45$.
The mass trend parameters of the [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relations using forms I, II, and III (see equations \[eq:FormA\], \[eq:FormB\] and \[eq:FormC\]) are $0.83^{+0.09}_{-0.10}$, $0.80^{+0.11}_{-0.08}$ and $0.81^{+0.09}_{-0.08}$, respectively, showing consistency at better than the $1\sigma$ confidence level. All derived mass trends are consistent with the self-similar expectation at $\approx1.6\sigma$ level. Note that there is neither a significant redshift-dependence in the mass trend ($\delta_{{\ensuremath{{T_{\mathrm{X,cex}}}}}}=0.35^{+0.53}_{-0.41}$) nor a strong deviation ($\gamma_{{\ensuremath{{T_{\mathrm{X,cex}}}}}}=-0.33^{+0.23}_{-0.28}$) from a self-similar redshift trend. In addition, as can be seen in Figure \[fig:triangle\_txe\], there is a mild degeneracy between the slope and redshift evolution, such that the mass trend can be pushed back closer to the self-similar expectation with a stronger deviation of the redshift evolution from self-similarity. Fixing the SZE $\zeta$–mass relation does not change the best-fit parameters but reduces the uncertainty on the normalizations $A_{{\ensuremath{{T_{\mathrm{X}}}}}}$ by a factor of between two and three (see Table \[tab:sclrel\]). This indicates that only the normalizations of the [$T_{\mathrm{X}}-M_{500}-z$]{} relations are dominated by the systematic uncertainties in the SZE-based halo masses.
This behavior is clearly visible in Figure \[fig:triangle\_txe\], where we show the joint confidence contours and fully marginalized posterior distributions for each of the [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relation variables for all three forms of the relation. In the figure, the self-similar parameter expectations are marked with dashed red lines. The preference for the mass trend to deviate by $\approx2\sigma$ and the redshift trend to deviate by $\approx1\sigma$ can be seen both in the joint parameter constraint panels and the fully marginalized single parameter distributions. Evidence for parameter covariance is clearest in the parameter pair $C_{\ensuremath{{T_{\mathrm{X,cex}}}}}$ and $\delta_{\ensuremath{{T_{\mathrm{X,cex}}}}}$. We include the four [$\zeta-M_{500}-z$]{} relation parameters to show the strong positive correlations among the corresponding parameters $A_\mathrm{SZ}$ and $A_{\ensuremath{{T_{\mathrm{X,cex}}}}}$, $B_\mathrm{SZ}$ and $B_{\ensuremath{{T_{\mathrm{X,cex}}}}}$, $C_\mathrm{SZ}$ and $C_{\ensuremath{{T_{\mathrm{X,cex}}}}}$. Also a strong negative correlation among the scatter parameters $\sigma_{\ln\zeta}$ and $\sigma_{\ln{\ensuremath{{T_{\mathrm{X,cex}}}}}}$ are indicated by the same parameters. This is as expected and follows from the importance of the SZE-based masses in the [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relation and the fact that the quadrature sum of the scatter in [$\zeta-M_{500}-z$]{} and [$T_{\mathrm{X,cex}}-M_{500}-z$]{} is constrained by the measurement-error corrected scatter in the data about the best-fit [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relation. In all other cases that follow, we exclude the [$\zeta-M_{500}-z$]{} parameters from the plots to conserve space, but the correlations persist, as expected. Finally, this plot makes clear (black lines) that the improvement from fixing the [$\zeta-M_{500}-z$]{} scaling relation parameters at their best-fit values is a dramatic decrease in the uncertainties of $A_{\ensuremath{{T_{\mathrm{X,cex}}}}}$ but only a modest impact on the other parameters.
For the core-included [$T_{\mathrm{X,cin}}-M_{500}-z$]{} relation, the mass and redshift trends as well as the normalization are consistent with those for the core-exclude case within the quoted $1\sigma$ uncertainties. The discernible difference in the two relations comes from intrinsic scatter. The log-normal intrinsic scatter for the core-excised observable at fixed mass $\sigma_{\ln{\ensuremath{{T_{\mathrm{X,cex}}}}}}$ is $\approx0.12$, approximately a factor of 1.5 smaller than in the case of the core-included observable. Figure \[fig:triangle\_txi\] contains the joint and single parameter constraints for this relation.
{width="46.00000%"} {width="46.00000%"} {width="46.00000%"} {width="46.00000%"}
### Comparison to previous results
We show the mass and redshift trends in the [${T_{\mathrm{X}}}$]{} scaling relations in the top panels of Figs. \[fig:mtrends\] and \[fig:ztrends\], respectively. In both figures, the core-excised measurements are on the left and core-included on the right. In the case of the mass trends, all X-ray observable measurements are corrected to the pivot redshift ${\ensuremath{z_{\mathrm{piv}}}}=0.45$ using the best-fit redshift trend from the Form II scaling relation, and in the case of the redshift trends, all are corrected to the pivot mass ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ using the best-fit mass trend from Form II. Also shown are the self-similar expectations (red dashed line) and the gray region corresponds to the one sigma allowed region for the relation. One of the outlier clusters in the sample is SPT-CLJ0217-5245, whose temperature is high compared to the expected temperature from the luminosity scaling relations (see Section \[sec:lm\]). We note that the noise dominated spectrum of this cluster makes it challenging to determine its temperature from the shape of the continuum Bremsstrahlung emission.
Our analysis shows steeper mass trends than those measured before. @vikhlinin09a found a self-similar slope ($B_{T_x} = 0.65\pm0.03$) in the [$T_{\mathrm{X,cex}}-M_{500}-z$]{} relation for X-ray selected clusters observed with [[*Chandra*]{}]{} in the redshift range of $0.02<z<0.9$ and with hydrostatic mass measurements in the range $10^{14}{\ensuremath{M_{\odot}}}\lesssim{\ensuremath{M_{500}}}\lesssim10^{15}{\ensuremath{M_{\odot}}}$, while a similar mass slope of $B_{T_X} = 0.67\pm0.07$ was reported in @arnaud05 covering the [[*XMM-Newton*]{}]{} observations of low redshifts clusters at $z\lesssim0.15$ with a hydrostatic mass range of $9\times10^{13}{\ensuremath{M_{\odot}}}\lesssim{\ensuremath{M_{500}}}\lesssim8.4\times10^{14}{\ensuremath{M_{\odot}}}$. Our result for the mass slope is $1.6\sigma$ and $1.1\sigma$ away, respectively, from these results. In [@mahdavi13], the mass slopes of $0.51^{+0.42}_{-0.16}$ and $0.70^{+0.11}_{-0.08}$ were derived using weak lensing and hydrostatic masses, respectively, for a sample of 50 galaxy clusters at $z\lesssim0.5$ with a similar mass range to those we study here; no tension with our results is seen. A recent result based on the 100 brightest clusters selected in the XXL survey [@lieu16] gave slopes of $\approx0.56^{+0.12}_{-0.10}$, and this slope became $\approx0.60\pm0.05$ if combined with the lower mass groups. Given their preference for a shallower than self-similar relation, these results are in tension at 1.5 and 2.0$\sigma$ with ours. In [@mantz16], a mass trend of $\approx0.66\pm0.05$ was reported for X-ray selected clusters with redshift range of $0.07<z<1.06$ and mass range of $3\times10^{14}{\ensuremath{M_{\odot}}}\leq{\ensuremath{M_{500}}}\lesssim2\times10^{15}{\ensuremath{M_{\odot}}}$. This result is in $1.4\sigma$ tension with ours.
In the upper-left panel of Figure \[fig:comps\], we further compare our results of ${\ensuremath{{T_{\mathrm{X,cex}}}}}$ to the simulated clusters at $z=0.1$ from the C-Eagle cosmological hydrodynamical simulations [@Barnes2017], together with the nearby clusters from [@pratt09] and the clusters at $z\lesssim0.5$ from [@vikhlinin09a]. In addition, we over-plot the best-fit relation from [@mantz16] (in blue dashed line) and the self-similar prediction with the normalization anchored to our best-fit value (in grey dashed line).
It is important to note that there exist non-negligible systematic differences among these studies, especially in the estimation of cluster masses. In [@pratt09], the cluster mass is estimated using the ${\ensuremath{Y_{\mathrm{X,cex}}}}$-[$M_{500}$]{} relation derived from hydrostatic mass estimates in nearby, relaxed clusters. For the sake of consistency, we take the ${\ensuremath{Y_{\mathrm{X}}}}$-inferred masses from [@vikhlinin09a]. We scale up the cluster masses from [@pratt09] and [@vikhlinin09a] by a factor of $1.12$ to account for the offsets between hydrostatic masses and our masses [@bocquet15]. For [@mantz16], the cluster masses are calibrated using weak lensing, for which we do not expect significant systematic offsets with our results [@deHaan16; @dietrich17; @schrabback18]. For the simulated clusters in [@Barnes2017], we directly use the true halo masses.
To make the figure, we re-scale each reported [${T_{\mathrm{X,cex}}}$]{} from the literature studies to the pivotal redshift [$z_{\mathrm{piv}}$]{} by multiplying by $\left(E({\ensuremath{z_{\mathrm{piv}}}})/E(z)\right)^{\frac{2}{3}}$, because we observe that the core-excised temperature is evolving as predicted by the self-similar evolution. As seen in Figure \[fig:comps\], our results are consistent in terms of normalization and mass trends with the simulations [@Barnes2017] and other observed clusters over the common mass range, except that we observe a shift in normalization in comparison with [@mantz16].
In summary, the previous results show mass trends that are in agreement with the self-similar prediction (i.e., the value of $2/3$), while the fully marginalized posterior of our mass trend parameter is steeper than self-similar at $\approx1.6\sigma$ and in tension with these previous results at a similar or lower level. One difference between our work and these others is that we simultaneously fit the mass and redshift trends of the scaling relation, exploiting the fact that our SZE-selected sample is approximately mass selected out to redshift $\approx1.4$. Most of these previous analyses have assumed self-similar redshift evolution, because their datasets tend to cover very different mass ranges at low and high redshift, introducing strong degeneracies in the mass and redshift trend parameters. Thus, our sample provides the first direct constraint on the deviation from self-similarity for massive clusters out to $z\approx1.4$ that accounts for both mass and redshift trends. Only the analysis of larger samples with improved halo mass estimates will allow us to definitively determine departures from self-similarity in the mass trends of the [$T_{\mathrm{X}}-M_{500}-z$]{} scaling relations.
[$M_\mathrm{ICM}-M_{500}-z$]{} Relation {#sec:mm}
---------------------------------------
The [$M_\mathrm{ICM}-M_{500}-z$]{} scaling relation and its redshift trends has important implications for ICM mass fractions and baryon fraction within clusters, because a majority of the baryons reside within the ICM [@lin03b; @chiu16a]. The expression for the self-similar scaling of [$M_\mathrm{ICM}-M_{500}-z$]{} relation is: $${\ensuremath{M_{\mathrm{ICM}}}}\propto {\ensuremath{M_{500}}}.$$ That is, in the simplest Universe with no feedback or radiative processes the ICM mass fraction would be expected to be identical in halos of all mass and at all redshifts.
![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the case of [$M_\mathrm{ICM}-M_{500}-z$]{} scaling relations.[]{data-label="fig:triangle_micm"}](fig8.pdf){width="45.00000%"}
### Parameter constraints
We present the best-fit parameters of the [$M_\mathrm{ICM}-M_{500}-z$]{} relations using the scaling relation forms I, II, and III (equations \[eq:FormA\], \[eq:FormB\] and \[eq:FormC\]) in Table \[tab:sclrel\], and the marginalized posteriors of the single and joint parameter constraints are presented in Figure \[fig:triangle\_micm\]. We do not present core-excised values for the ICM mass, because the central core region contains only a negligible portion of the ICM. The best-fit [$M_\mathrm{ICM}-M_{500}-z$]{} scaling relation using Form II is $$\begin{aligned}
\label{eq:micm}
{\ensuremath{M_{\mathrm{ICM}}}}&=
&7.37^{+0.76}_{-1.35}\times10^{13}{\ensuremath{M_{\odot}}}\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{1.26^{+0.12}_{-0.09}} \times \nonumber \\
&&
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{0.18^{+0.30}_{-0.31}} \, ,\end{aligned}$$ with intrinsic scatter of $0.10^{+0.04}_{-0.07}$. As before, the mass and redshift pivots are ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ and ${\ensuremath{z_{\mathrm{piv}}}}=0.45$.
We find that the mass trend parameter is $B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.26^{+0.12}_{-0.09}$, which is steeper than the self-similar scaling at the $\approx2.9\sigma$ level. Although the uncertainties are large, there are no significant redshift trends observed. The data provide no evidence for a redshift-dependent mass slope, given that $\delta_{{\ensuremath{M_{\mathrm{ICM}}}}}$ of $0.16^{+0.47}_{-0.44}$. The normalization $A_{{\ensuremath{M_{\mathrm{ICM}}}}}$ of $7.37^{+0.76}_{-1.35}\times10^{13}{\ensuremath{M_{\odot}}}$ suggests an ICM mass fraction of $\approx(16.0\pm2)\%$ at the pivot mass and redshift. A consistent picture is suggested by all three functional forms. Furthermore, fixing the SZE parameters $r_{\zeta}$ does not shift the best-fit parameters, but reduces the uncertainty of the normalization by a factor of four and the uncertainties on the mass and redshift trends by a factor of two.
### Comparison to previous results
We show the redshift and mass trends of [$M_{\mathrm{ICM}}$]{} in the second row from the top of Figs. \[fig:mtrends\] and \[fig:ztrends\], respectively. As for the case of the other X-ray observables shown in this plot, we scale the measurements to the pivot redshift ${\ensuremath{z_{\mathrm{piv}}}}=0.45$ or pivot mass ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ using the best-fit redshift and mass trends from the Form II relation (see Table \[tab:sclrel\]).
Our measured mass trends are consistent with that found by @zhang12 [$B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.38\pm0.36$], where a sample of 19 clusters ($z<0.1$ and $2\times 10^{14}{\ensuremath{M_{\odot}}}\lesssim{\ensuremath{M_{500}}}\lesssim 2\times10^{15}{\ensuremath{M_{\odot}}}$) selected by their X-ray fluxes was studied, and also in the study of the 100 brightest galaxy clusters and groups at redshift range of 0.05–1.1 and and mass range of $10^{13}-10^{15}{\ensuremath{M_{\odot}}}$) selected from the XXL survey [$B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.21^{+0.11}_{-0.10}$; @eckert16]. The mass trends derived from low-redshift clusters [@arnaud07; @pratt09 $1.24\pm 0.06$ and $1.21\pm 0.03$, respectively] using hydrostatic masses are also consistent with our measurement. Our derived mass trends are in good agreement with those derived based on the SPT clusters observed with [[*Chandra*]{}]{} [@chiu16b; @chiu17 $B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.33\pm0.07$]. The agreement between our results and previous [[*Chandra*]{}]{}-based works of SPT-selected clusters indicates that [$M_\mathrm{ICM}-M_{500}-z$]{} relation is relatively insensitive to the instrumental systematics.
In two other works, mass trends more consistent with self-similar behavior have been found. Our results are in tension with the @mahdavi13 [$B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.04\pm0.10$] weak lensing analysis at $\approx1.6\sigma$ and with @mantz16 [$B_{{\ensuremath{M_{\mathrm{ICM}}}}}=1.004\pm0.015$] analysis of massive, RASS selected clusters at $\approx2.9\sigma$. The tension with the @mantz16 results is particularly strong, but in general we find that our results are in excellent agreement with those from past studies carried out either with weak lensing or hydrostatic masses. Most previous studies have been carried out over a narrower range of lower redshifts. And indeed, as already noted, our data provide no evidence for a redshift dependent mass slope.
Similarly, In Figure \[fig:comps\] we also compare our results with simulated clusters [@Barnes2017], together with [@pratt09], [@vikhlinin09a] and [@mantz16]. Note that in addition to shifting the hydrostatic mass based halo masses as described in the previous section, here we also scale up the [$M_{\mathrm{ICM}}$]{} measurements in the literature by $3.8\%$ because [$M_{\mathrm{ICM}}$]{} is increasing linearly with cluster radius (i.e., ${\ensuremath{M_{\mathrm{ICM}}}}\propto{\ensuremath{R_{500}}}\propto{{\ensuremath{M_{500}}}}^{\frac{1}{3}}$). In the case of the ICM mass, our results show good agreement with the simulations and with the previous results, although with a steeper mass trend in comparison to [@mantz16] (see the discussion above).
It is worth noting that the intrinsic scatter $\sigma_{\ln{\ensuremath{M_{\mathrm{ICM}}}}}$ in [$M_{\mathrm{ICM}}$]{} at fixed halo mass is at the $\approx10\%$ level, indicating that the ICM mass is among the highest quality cluster mass proxies available.
[$Y_{\mathrm{X}}-M_{500}-z$]{} Relation {#sec:ym}
---------------------------------------
The X-ray estimated integrated pressure $Y_{\mathrm{X}}$ is of interest because of relatively low intrinsic scatter, its connection to the SZE observable and its relative insensitivity to the influence of feedback from AGN and star formation [@kravtsov06a; @nagai07; @bonamente08; @vikhlinin09a; @andersson11; @benson13]. The self-similar expectation of the $Y_{\mathrm{X}}$ to mass scaling relations is; $$Y_{\mathrm{X}} \propto {\ensuremath{M_{500}}}^{5/3}\ E(z)^{2/3},$$ which results from [$Y_{\mathrm{X}}$]{} being the product of [$M_{\mathrm{ICM}}$]{} and [${T_{\mathrm{X}}}$]{} together with the dependence of the [$T_{\mathrm{X}}-M_{500}-z$]{} relation on the evolution of the critical density.
![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the core-excised [$Y_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-included [$Y_{\mathrm{X,cex}}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_yx"}](fig9a.pdf "fig:"){height="0.37\textheight"} ![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the core-excised [$Y_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-included [$Y_{\mathrm{X,cex}}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_yx"}](fig9b.pdf "fig:"){height="0.37\textheight"}
### Parameter constraints
Similar to previous sections, we present the [$Y_{\mathrm{X}}-M_{500}-z$]{} relation derived using both the core-included [$Y_{\mathrm{X,cin}}$]{} and -excised [$Y_{\mathrm{X,cex}}$]{} observables for the scaling relations forms I, II, and III (equations \[eq:FormA\], \[eq:FormB\] and \[eq:FormC\], respectively). The best-fit parameters and uncertainties of the scaling relations are listed in Table \[tab:sclrel\] for both [$Y_{\mathrm{X}}$]{} observables, and the marginalized posteriors of the single and joint parameters are presented in Figure \[fig:triangle\_yx\].
The best-fit [$Y_{\mathrm{X}}-M_{500}-z$]{} scaling relation using functional form II in the core-included case is $$\begin{aligned}
\label{eq:yxi}
{\ensuremath{Y_{\mathrm{X,cin}}}}&=
&4.6^{+1.1}_{-1.1}\times10^{14}~{\ensuremath{M_{\odot}}}\mathrm{keV}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{1.99^{+0.17}_{-0.15}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{2}{3}}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.21^{+0.50}_{-0.45}},\end{aligned}$$ with intrinsic scatter of $0.16^{+0.05}_{-0.12}$. For the core-excised observable [$Y_{\mathrm{X,cex}}$]{} the best-fit relation is $$\begin{aligned}
\label{eq:yxe}
{\ensuremath{Y_{\mathrm{X,cex}}}}&=
&4.50^{+1.00}_{-1.10}\times10^{14}~{\ensuremath{M_{\odot}}}\mathrm{keV}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{2.02^{+0.16}_{-0.17}} \times \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{2}{3}}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.17^{+0.47}_{-0.50}} \, ,\end{aligned}$$ with intrinsic scatter $0.11^{+0.07}_{-0.08}$. As for all other cases, the mass and redshift pivots are ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ and ${\ensuremath{z_{\mathrm{piv}}}}=0.45$.
For [$Y_{\mathrm{X}}-M_{500}-z$]{} relations, we observe the mass trend $B_{{\ensuremath{Y_{\mathrm{X}}}}}$ that is in tension with the self-similar prediction at the $\approx2\sigma$ level for the core-included and -excised X-ray observables. On the other hand, the redshift trends for all three functional forms are consistent with self-similarity within the quoted $1\sigma$ uncertainty. There is no evidence for a redshift-dependent mass trend. Fixing the SZE parameters $r_{\zeta}$ leads to no major parameter shifts, but does reduce the parameter uncertainties on the normalization by a factor of about four and, interestingly, leads to a reduction in the estimate of the intrinsic scatter. With intrinsic scatter at the $\approx10\%$ level as with the [$M_{\mathrm{ICM}}$]{}, the [$Y_{\mathrm{X}}$]{} observable with or without core excision offers an outstanding single cluster mass proxy.
### Comparison to previous results
We show the redshift and mass trends of [$Y_{\mathrm{X}}$]{} in the third row from the top of Figs. \[fig:mtrends\] and \[fig:ztrends\], respectively. As for the case of the other X-ray observables shown in this plot, we scale the measurements to the pivot redshift ${\ensuremath{z_{\mathrm{piv}}}}=0.45$ or pivot mass ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ using the best-fit redshift and mass trends from the Form II relation (see Table \[tab:sclrel\]).
Given that we are adopting halo masses from the [$\zeta-M_{500}-z$]{} scaling relation calibrated in the analysis of @deHaan16, we note that the slope of the [$Y_{\mathrm{X}}-M_{500}-z$]{} relation was found in that work to favor a scaling steeper than its self similar predicted value (i.e., $\approx$2 vs 1.67). In this work, we measure X-ray observables ([${T_{\mathrm{X}}}$]{}, [$M_{\mathrm{ICM}}$]{}, [$Y_{\mathrm{X}}$]{}, and [$L_{\mathrm{X}}$]{}) for the SPT-SZ cluster sample using a different set of observations from the XMM-Newton satellite. While these data are independent of the data used in @deHaan16, one would nevertheless expect that given the results of that earlier analysis using [[*Chandra*]{}]{} data, that we should see a [$Y_{\mathrm{X}}-M_{500}-z$]{} relation that is steeper than self-similar, as indeed we do.
In comparison to other previously published results, the constraints on the mass trend of the full sample is steeper than the reported value in @arnaud07 [$B_{{\ensuremath{Y_{\mathrm{X}}}}}=1.83\pm0.09$], a difference of $\approx1\sigma$. Other studies employing X-ray hydrostatic masses also resulted in shallower slopes [@vikhlinin09a; @Lovisari2015 $B_{{\ensuremath{Y_{\mathrm{X}}}}}=1.75\pm0.09$ and $B_{{\ensuremath{Y_{\mathrm{X}}}}}=1.67\pm0.08$, respectively], which also show weak tension with our results at 1.4 and 1.8$\sigma$ significance, respectively. The weak lensing based study of @mahdavi13 also found a weaker mass trend of $B_{{\ensuremath{Y_{\mathrm{X}}}}}=1.79\pm0.22$ that is nonetheless statistically consistent with our results. The tension between our result and the @mantz16 analysis ($B_{{\ensuremath{Y_{\mathrm{X}}}}}=1.61\pm0.04$) is at the 2.3$\sigma$ level.
In Figure \[fig:comps\] we also compare our core-excised [$Y_{\mathrm{X,cex}}$]{} with simulated clusters [@Barnes2017] and the observations from [@pratt09] and [@vikhlinin09a]. Similar to the case of ICM mass, we also scale up the [$Y_{\mathrm{X,cex}}$]{} by $3.8\%$ because of ${\ensuremath{Y_{\mathrm{X,cex}}}}\equiv{\ensuremath{{T_{\mathrm{X,cex}}}}}{\ensuremath{M_{\mathrm{ICM}}}}$. Our results are broadly consistent with both the simulated and observed clusters but with a preference for a slope that is steeper than the self-similar prediction.
As with the [$T_{\mathrm{X}}-M_{500}-z$]{} relations presented previously, our measured [$Y_{\mathrm{X}}$]{} mass trends are steeper and exhibit greater tension with self-similar behavior than previous works. This can be understood as the combination of the [${T_{\mathrm{X}}}$]{} and [$M_{\mathrm{ICM}}$]{} mass trends—each steeper than self-similar—presented in the last two sections. However, while the [$M_\mathrm{ICM}-M_{500}-z$]{} relation mass trend we measure is in good agreement with previous analyses, it is our [$T_{\mathrm{X}}-M_{500}-z$]{} relation that appears steeper. Whether this is due to our unique SZE-selection, leading to an approximately mass-limited sample over a very large redshift range, or due to systematic differences in our mass estimates that include Eddington and Malmquist bias corrections that are typically not considered in earlier works, this must be clarified with a larger sample of clusters and with the ongoing improvements in mass calibration of our own sample.
[$L_\mathrm{X}-M_{500}-z$]{} Relation {#sec:lm}
-------------------------------------
We extract the X-ray luminosity obtained from the core-included aperture of $<{\ensuremath{R_{500}}}$ in the 0.5–2 keV (i.e., the soft-band luminosity $L_{\mathrm{X}}$) and the 0.01:100 keV band (i.e., the bolometric luminosity $L_{\mathrm{X,bol}}$) to study the [$L_\mathrm{X}-M_{500}-z$]{} scaling relations. In previous studies, the [$L_\mathrm{X}-M_{500}-z$]{} scaling relations have tended to exhibit larger scatter if cluster cores are included in the analysis [@pratt09] due to the complex cool-core phenomenon that impacts the central regions of clusters. Indeed, it was argued long ago that the primary driver of the ${\ensuremath{L_{\mathrm{X}}}}-{\ensuremath{{T_{\mathrm{X}}}}}$ relation scatter was this cool core phenomenon [@fabian94c], and with the availability of cluster samples extending to high redshift it was shown to be true out to $z\approx0.8$ [@ohara06]. Therefore, we also additionally extract the X-ray luminosities obtained from the core-excised aperture of $(0.15-1){\ensuremath{R_{500}}}$ in both soft and bolometric bands. As a result, we derive four [$L_\mathrm{X}-M_{500}-z$]{} scaling relations—(1) core-included and soft-band luminosity to mass [$L_{\mathrm{X,cin}}$]{}–[$M_{500}$]{}, (2) core-included and bolometric luminosity to mass [$L_{\mathrm{X,cin,bol}}$]{}–[$M_{500}$]{}, (3) core-excised and soft-band luminosity to mass [$L_{\mathrm{X,cex}}$]{}–[$M_{500}$]{}, and (4) core-excised and bolometric luminosity to mass [$L_{\mathrm{X,cex,bol}}$]{}–[$M_{500}$]{} scaling relations. The self-similar expectation of the [$L_\mathrm{X}-M_{500}-z$]{} scaling relation is $$\begin{aligned}
L_{\mathrm{X}} &\propto &{\ensuremath{M_{500}}}\ E(z)^{2} \, . \nonumber \\
L_{\mathrm{X,bol}} &\propto &{\ensuremath{M_{500}}}^{4/3} \ E(z)^{7/3} \, . \nonumber \\\end{aligned}$$ for the the soft-band and bolometric luminosities, respectively, where for the soft-band we have assumed that the emissivity is temperature independent [see discussion in @mohr99].
![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the 0.5:2.0 keV core-included luminosity [$L_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-excised luminosity [$L_\mathrm{X,cex}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_soft_lx"}](fig10a.pdf "fig:"){height="0.37\textheight"} ![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the 0.5:2.0 keV core-included luminosity [$L_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-excised luminosity [$L_\mathrm{X,cex}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_soft_lx"}](fig10b.pdf "fig:"){height="0.37\textheight"}
![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the core-included bolometric luminosity [$L_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-excised bolometric luminosity [$L_\mathrm{X,cex}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_bol_lx"}](fig11a.pdf "fig:"){height="0.37\textheight"} ![Similar to Figure \[fig:triangle\_txe\] but containing constraints for the cases of the core-included bolometric luminosity [$L_{\mathrm{X,cin}}-M_{500}-z$]{} (the upper panel) and core-excised bolometric luminosity [$L_\mathrm{X,cex}-M_{500}-z$]{} (the lower panel) scaling relations.[]{data-label="fig:triangle_bol_lx"}](fig11b.pdf "fig:"){height="0.37\textheight"}
### Parameter constraints
The resulting best-fit scaling relation parameters and uncertainties are listed in Table \[tab:sclrel\], and the marginalized posteriors of the single and joint parameters constraints for the core-included and -excised observables appear in Figure \[fig:triangle\_soft\_lx\] (0.5–2.0 keV) and Figure \[fig:triangle\_bol\_lx\] (bolometric).
For the core-included, soft-band 0.5–2.0 keV X-ray luminosity [$L_{\mathrm{X,cin}}$]{}, the best-fit relation is $$\begin{aligned}
\label{eq:lxi-soft}
{\ensuremath{L_{\mathrm{X,cin}}}}&=
&4.12^{+0.91}_{-0.94}\times10^{44}\mathrm{erg/s}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{1.89^{+0.23}_{-0.13}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{2}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.20^{+0.51}_{-0.49}} \, ,\end{aligned}$$ with intrinsic scatter of $0.27^{+0.08}_{-0.12}$. For the [$L_\mathrm{X,cex}-M_{500}-z$]{} relation, the best-fit is $$\begin{aligned}
\label{eq:lxe-soft}
{\ensuremath{L_{\mathrm{X,cex}}}}&=
&2.84^{+0.53}_{-0.50}\times10^{44}\mathrm{erg/s}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{1.60^{+0.16}_{-0.15}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{2}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.10^{+0.47}_{-0.42}},\end{aligned}$$ with intrinsic scatter of $0.27^{+0.07}_{-0.11}$. As before, the mass and redshift pivots are ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ and ${\ensuremath{z_{\mathrm{piv}}}}=0.45$.
The soft-band, core-excised [$L_\mathrm{X}-M_{500}-z$]{} relation shows a mass trend that is $\approx4\sigma$ higher than the self-similar trend ($B_{{\ensuremath{L_{\mathrm{X}}}}} = 1$), while the core-included relation is steeper and exhibits a tension of $\approx6.8\sigma$ with the self-similar behavior. The redshift trends for both core-included and -excised luminosities are consistent with the self-similar trend of $C_{{\ensuremath{L_{\mathrm{X}}}}} = 2$. There is no evidence for a redshift-dependent mass slope in either soft-band [$L_{\mathrm{X}}$]{} measurement. Fixing the SZE parameters $r_{\zeta}$ (the black curves in Figure \[fig:triangle\_soft\_lx\]) does not change the overall picture except that the uncertainties of the normalization are reduced by about a factor of four.
The characteristic luminosities at the pivot mass and redshift for core-included clusters are a factor of $\approx45\%$ higher than the core-excised luminosities, a difference of $\approx 1\sigma$. Interestingly, the scatter of the two relations is similar at $\approx27\%$.
Similarly, for the bolometric luminosities the best-fit [$L_{\mathrm{X,cin,bol}}-M_{500}-z$]{} and [$L_{\mathrm{X,cex,bol}}-M_{500}-z$]{} relations are $$\begin{aligned}
\label{eq:lxi-bol}
{\ensuremath{L_{\mathrm{X,cin}}}}&=
&14.8^{+3.5}_{-2.7}\times10^{44}\mathrm{erg/s}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{2.19^{+0.21}_{-0.17}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{7}{3}}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.14^{+0.62}_{-0.57}} \, ,\end{aligned}$$ and $$\begin{aligned}
\label{eq:lxe-bol}
{\ensuremath{L_{\mathrm{X,cex}}}}&=
&10.7^{+2.3}_{-2.3}\times10^{44}\mathrm{erg/s}
\left(\frac{{\ensuremath{M_{500}}}}{{\ensuremath{M_{\mathrm{piv}}}}}\right)^{1.88^{+0.19}_{-0.17}} \nonumber \\
&&\left(\frac{E(z)}{E({\ensuremath{z_{\mathrm{piv}}}})}\right)^{\frac{7}{3}}
\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}}\right)^{-0.26^{+0.53}_{-0.43}},\end{aligned}$$ with intrinsic scatter of $0.29^{+0.08}_{-0.13}$ and $0.27^{+0.07}_{-0.13}$, respectively. The same pivot mass and redshift as before are used.
As expected, the bolometric luminosity relations have steeper mass trends than those of the soft band luminosities. Similar to the soft band, the bolometric luminosity to mass scaling relations have mass trends that are steeper than self-similar ($B_{{\ensuremath{L_{\mathrm{X}}}}}={4\over3}$) with a significance of $\approx3.2\sigma$ and $\approx5.1\sigma$ for the core-excised and core-included luminosities, respectively. The redshift trends of the scaling relations are all consistent with the self-similar trend $C_{{\ensuremath{L_{\mathrm{X}}}}}={7\over3}$, and there is a preference for a redshift dependent mass trend in the core-included luminosity scaling relation. Fixing the SZE parameters $r_{\zeta}$ does not result in significant differences except by decreasing the uncertainties of the normalization by a factor of three to five, and it also slightly reduces the scatter in the core-included relation. Both relations exhibit intrinsic scatter at around the $27\%$ level, which is comparable to that in the soft band.
### Comparison to previous results
We show the redshift and mass trends of [$L_{\mathrm{X}}$]{} in the two bottom rows in Figures \[fig:mtrends\] and \[fig:ztrends\], respectively. As for the case of the other X-ray observables shown in this plot, we scale the measurements to the pivot redshift ${\ensuremath{z_{\mathrm{piv}}}}=0.45$ or pivot mass ${\ensuremath{M_{\mathrm{piv}}}}=6.35\times10^{14}{\ensuremath{M_{\odot}}}$ using the best-fit redshift and mass trends from the Form II relation (see Table \[tab:sclrel\]).
Our core-excised bolometric luminosities are consistent with the bolometric luminosities reported from [[*XMM-Newton*]{}]{} observations of the low-[*z*]{} REXCESS clusters [@pratt09 with a slope of $1.77\pm0.05$]. Additionally, our core-excised soft-band luminosities from [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} observations of the 15 SPT selected clusters [@andersson11 with a slope of $1.45\pm0.29$], [[*Chandra*]{}]{} observations of massive clusters @vikhlinin09b [with a slope of $1.61\pm 0.14$], [[*Chandra*]{}]{} observations of 115 clusters [@maughan07 with a slope of $1.63\pm0.08$], and the [[*XMM-Newton*]{}]{} observations of HIFLUGCS sample [@Lovisari2015 with a slope of $1.61\pm0.19$] at $z<0.05$. We note that all these results in the literature depart from the self-similar expectation. However, the slope of the mass trend of the core-excised soft-band luminosity ($B_{{\ensuremath{L_{\mathrm{X}}}}}\approx1.60\pm0.15$) is steeper than the value reported in @mantz16 [$B_{{\ensuremath{L_{\mathrm{X}}}}}=1.02 \pm 0.09$] at the $3.4\sigma$ level. Our slope is consistent with the low-redshift ($z<0.2$) HIFLUGCS Cosmology (HICOSMO) sample [$B_{{\ensuremath{L_{\mathrm{X}}}}}=1.35\pm0.07$; @schellenberger15] at $\approx1.5\sigma$ level. Overall, in terms of mass trends our study demonstrates a much steeper than self-similar mass trend in agreement with most previously published analyses.
Our constraints on the redshift trend of the core-excised, soft-band [$L_\mathrm{X}-M_{500}-z$]{} is $L_{{\ensuremath{L_{\mathrm{X}}}}}\propto E(z)^{1.72^{+0.53}_{-0.46}}$, which is in good agreement with that found by @mantz16 [$C_{{\ensuremath{L_{\mathrm{X}}}}}=1.82\pm0.35$]. @vikhlinin09b reports a redshift trend of $C_{{\ensuremath{L_{\mathrm{X}}}}}=1.85\pm0.40$, which is also consistent with our results. In addition, our soft band measurements follow a similar trend with that seen in the C-Eagle cosmological hydrodynamical simulations of clusters [@Barnes2017].
In Figure \[fig:comps\], we over-plot our results of core-excised soft-band luminosity [$L_{\mathrm{X,cex}}$]{} with the ones from simulated clusters [@Barnes2017] and other observational studies [@pratt09; @vikhlinin09a; @mantz16]. Although our SPT clusters are sampling the relatively high-mass end, our results show no significant tension in the mass trend with previous work extending to the low mass regime. With the exception of the analysis of [@mantz16], the [$L_{\mathrm{X,cex}}$]{} from both simulated and observed clusters all show steeper mass trends with respect to the self-similar prediction (the grey dashed line). We note also that the scatter in the simulated C-Eagle clusters is 0.30, which is larger than, but statistically consistent with, our measurement of $\sigma_{\ln{\ensuremath{L_{\mathrm{X}}}}}=0.27^{+0.07}_{-0.11}$. This is also true for the values of 0.17 and 0.25 found in the REXCESS [@pratt09] and HIFLUGCS samples [@Lovisari2015]. An interesting element of our result is that the scatter is similar in both the core-included and -excised luminosity measurements.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------- ------------------------------------------------------ ------------------------------------------------------ ------------------------------------------ --------------------------------------- --------------------------------------- -- -- -- --
Scaling Relation $A_{{\ensuremath{\mathcal{X}}}}$ $B_{{\ensuremath{\mathcal{X}}}}$ $C_{{\ensuremath{\mathcal{X}}}}$ $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ $\gamma_{{\ensuremath{\mathcal{X}}}}$ $\delta_{{\ensuremath{\mathcal{X}}}}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.50\pm 0.66$ $0.78^{+0.10}_{-0.09}$ $0.44^{+0.27}_{-0.26}$ $0.18^{+0.04}_{-0.05}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $6.51^{+0.58}_{-0.70}$ $0.81^{+0.09}_{-0.10}$ $2\over3$ $0.18^{+0.04}_{-0.04}$ $-0.21^{+0.24}_{-0.22}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $6.42^{+0.67}_{-0.65}$ $0.77^{+0.10}_{-0.09}$ $2\over3$ $0.17^{+0.05}_{-0.03}$ $-0.14^{+0.33}_{-0.26}$ $0.60^{+0.57}_{-0.47}$
\[3pt\] III with fixed SZE params $6.44^{+0.24}_{-0.23}$ $0.79^{+0.08}_{-0.09}$ $2\over3$ $0.17^{+0.04}_{-0.03}$ $-0.05^{+0.25}_{-0.27}$ $0.81^{+0.49}_{-0.53}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.26^{+0.57}_{-0.68}$ $0.80^{+0.10}_{-0.08}$ $0.43^{+0.23}_{-0.22}$ $0.13^{+0.04}_{-0.06}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $6.25^{+0.49}_{-0.75}$ $0.81^{+0.08}_{-0.09}$ $2\over3$ $0.14^{+0.04}_{-0.06}$ $-0.20^{+0.21}_{-0.22}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $6.18^{+0.63}_{-0.62}$ $0.81^{+0.10}_{-0.10}$ $2\over3$ $0.14^{+0.04}_{-0.05}$ $-0.17^{+0.28}_{-0.23}$ $0.35^{+0.41}_{-0.45}$
\[3pt\] III with fixed SZE params $6.14^{+0.22}_{-0.17}$ $0.81^{+0.08}_{-0.08}$ $2\over3$ $0.13^{+0.04}_{-0.04}$ $-0.13\pm 0.23$ $0.45^{+0.41}_{-0.48}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}=1$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}=0$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $6.60^{+1.1}_{-1.0}$ $1.27^{+0.08}_{-0.11}$ $0.43^{+0.24}_{-0.23}$ $0.11^{+0.04}_{-0.08}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{0}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $7.14^{+1.00}_{-1.01}$ $1.26^{+0.09}_{-0.10}$ $0$ $0.11^{+0.04}_{-0.08}$ $0.39^{+0.25}_{-0.24}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $6.89^{+1.15}_{-0.91}$ $1.27^{+0.10}_{-0.10}$ $0$ $0.11^{+0.03}_{-0.08}$ $0.40^{+0.23}_{-0.24}$ $0.19^{+0.40}_{-0.54}$
\[3pt\] III with fixed SZE params $6.93^{+0.28}_{-0.19}$ $1.27^{+0.08}_{-0.08}$ $0$ $0.08^{+0.04}_{-0.06}$ $0.41^{+0.21}_{-0.22}$ $0.11^{+0.44}_{-0.47}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={5\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.5^{+1.0}_{-1.1}$ $2.00^{+0.16}_{-0.14}$ $0.80^{+0.42}_{-0.35}$ $0.15^{+0.08}_{-0.09}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.7^{+1.1}_{-1.0}$ $2.01^{+0.16}_{-0.14}$ $2\over3$ $0.15^{+0.07}_{-0.11}$ $0.15^{+0.39}_{-0.34}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.9\pm 1.1$ $2.03^{+0.13}_{-0.15}$ $2\over3$ $0.17^{+0.06}_{-0.10}$ $0.32^{+0.39}_{-0.41}$ $1.16^{+0.44}_{-0.77}$
\[3pt\] III with fixed SZE params $4.59^{+0.22}_{-0.27}$ $1.97\pm 0.11$ $2\over3$ $0.07^{+0.07}_{-0.06}$ $0.23^{+0.35}_{-0.33}$ $0.73^{+0.77}_{-0.55}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={5\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.25^{+0.96}_{-1.02}$ $1.99^{+0.14}_{-0.16}$ $0.77^{+0.37}_{-0.42}$ $0.13^{+0.05}_{-0.10}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.35^{+0.90}_{-1.10}$ $2.00^{+0.16}_{-0.15}$ $2\over3$ $0.05^{+0.12}_{-0.03}$ $0.09^{+0.44}_{-0.29}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.70^{+1.1}_{-1.2}$ $1.99^{+0.18}_{-0.14}$ $2\over3$ $0.13^{+0.05}_{-0.10}$ $0.35^{+0.38}_{-0.41}$ $0.47^{+0.63}_{-0.65}$
\[3pt\] III with fixed SZE params $4.31^{+0.30}_{-0.18}$ $1.98\pm 0.12$ $2\over3$ $0.04^{+0.09}_{-0.03}$ $0.27^{+0.33}_{-0.35}$ $0.72^{+0.57}_{-0.69}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={1}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $4.13^{+0.87}_{-0.98}$ $1.93^{+0.15}_{-0.18}$ $2.01^{+0.44}_{-0.37}$ $0.28^{+0.07}_{-0.12}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $4.15^{+1.10}_{-0.81}$ $1.91^{+0.18}_{-0.15}$ $2$ $0.25^{+0.08}_{-0.13}$ $0.20^{+0.41}_{-0.43}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $4.33^{+1.11}_{-0.89}$ $1.85^{+0.21}_{-0.16}$ $2$ $0.28^{+0.08}_{-0.11}$ $0.21^{+0.53}_{-0.44}$ $0.82^{+0.61}_{-0.94}$
\[3pt\] III with fixed SZE params $3.86^{+0.22}_{-0.20}$ $1.94\pm 0.15$ $2$ $0.23^{+0.08}_{-0.07}$ $0.33^{+0.42}_{-0.37}$ $0.79^{+0.83}_{-0.65}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={1}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={2}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $2.83^{+0.53}_{-0.52}$ $1.57^{+0.17}_{-0.14}$ $2.17^{+0.34}_{-0.43}$ $0.26^{+0.07}_{-0.09}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{2}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $2.82^{+0.61}_{-0.49}$ $1.63^{+0.13}_{-0.16}$ $2$ $0.26^{+0.08}_{-0.09}$ $0.20^{+0.41}_{-0.34}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $2.77^{+0.55}_{-0.51}$ $1.57^{+0.18}_{-0.16}$ $2$ $0.28^{+0.07}_{-0.07}$ $0.24^{+0.51}_{-0.32}$ $0.47^{+0.93}_{-0.57}$
\[3pt\] III with fixed SZE params $2.67^{+0.16}_{-0.15}$ $1.58\pm 0.15$ $2$ $0.26^{+0.06}_{-0.05}$ $0.25\pm 0.38$ $0.75^{+0.76}_{-0.74}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={4\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={7\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $13.8^{+3.6}_{-3.2}$ $2.21^{+0.15}_{-0.20}$ $2.28^{+0.46}_{-0.41}$ $0.29^{+0.09}_{-0.12}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{7\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $14.3\pm 3.2$ $2.18\pm 0.19$ $7\over3$ $0.26^{+0.09}_{-0.13}$ $0.03^{+0.54}_{-0.37}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $14.8^{+3.3}_{-2.9}$ $2.16^{+0.19}_{-0.16}$ $7\over3$ $0.28^{+0.09}_{-0.11}$ $0.19^{+0.44}_{-0.50}$ $1.14^{+0.69}_{-0.76}$
\[3pt\] III with fixed SZE params $14.55^{+0.98}_{-0.70}$ $2.23\pm 0.15$ $7\over3$ $0.19^{+0.08}_{-0.10}$ $0.35^{+0.39}_{-0.42}$ $1.34^{+0.49}_{-0.74}$
\[3pt\] $B_\mathrm{{\ensuremath{\mathcal{X}}},SS}={4\over3}$ $C_\mathrm{{\ensuremath{\mathcal{X}}},SS}={7\over3}$
I: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{C_{\ensuremath{\mathcal{X}}}}$ $10.6^{+2.5}_{-2.2}$ $1.87^{+0.19}_{-0.16}$ $2.31^{+0.45}_{-0.35}$ $0.26^{+0.08}_{-0.13}$ – –
\[3pt\] II: ${\ensuremath{\mathcal{X}}}(M,z)\propto {\ensuremath{M_{500}}}^{B_{{\ensuremath{\mathcal{X}}}}}E(z)^{7\over3}(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$ $10.5^{+2.1}_{-2.5}$ $1.87^{+0.20}_{-0.14}$ $7\over3$ $0.27^{+0.07}_{-0.14}$ $0.10^{+0.47}_{-0.35}$ –
\[3pt\] III: as II with $B'_{\ensuremath{\mathcal{X}}}=B_{{\ensuremath{\mathcal{X}}}} + \delta_{{\ensuremath{\mathcal{X}}}}\ln\left(\frac{1+z}{1+{\ensuremath{z_{\mathrm{piv}}}}} \right)$ $9.9^{+2.2}_{-2.1}$ $1.80^{+0.17}_{-0.22}$ $7\over3$ $0.30^{+0.07}_{-0.09}$ $0.23^{+0.45}_{-0.49}$ $0.73^{+0.78}_{-0.70}$
\[3pt\] III with fixed SZE params $10.06^{+0.54}_{-0.61}$ $1.86^{+0.15}_{-0.16}$ $7\over3$ $0.25^{+0.05}_{-0.07}$ $0.36^{+0.32}_{-0.49}$ $0.88^{+0.76}_{-0.57}$
\[3pt\]
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------- ------------------------------------------------------ ------------------------------------------------------ ------------------------------------------ --------------------------------------- --------------------------------------- -- -- -- --
SZE-based halo masses with external cosmological priors from BAO {#sec:column2}
----------------------------------------------------------------
Currently, the redshift trend parameter on the SZE [$\zeta-M_{500}-z$]{} relation is the least well constrained, and this leads to additional uncertainty in understanding the X-ray observable mass relations. In @deHaan16 [the second column of Table 3], an analysis within the context of a flat $\Lambda$CDM model was undertaken where additional external cosmological priors from BAO were added. This helped reduce the cosmological parameter space consistent with the SPT cluster sample distribution in $\xi$ and redshift, tightening up [$\zeta-M_{500}-z$]{} parameter uncertainties. In addition, the redshift evolution parameter was shifted upward from $C_\zeta=0.55\pm0.3$ to $C_\zeta=0.80\pm0.15$. The combination of the shift and reduction in uncertainties have motivated us to present the scaling relations derived using the X-ray observables together with these SZE-based halo masses. Table \[tab:sclrel\_col2\] contains the results of these relations. We recommend that those particularly interested in obtaining precise redshift trends in the scaling relations should use these results.
Conclusions {#sec:concl}
===========
We present here measurements of the X-ray observables in a sample of 59 SZE selected galaxy clusters with redshifts $0.20<z<1.5$ that have been observed with [[*XMM-Newton*]{}]{}. We use these measurements together with SZE-based halo masses [$M_{500}$]{} to study the scaling relations between X-ray observables, halo mass and redshift. A strength of our work is the ability to directly constrain the redshift and mass trends based on an SZE-selected cluster sample spanning a wide range of redshift. This selection is approximately equivalent to a mass selection, and this sample spans a mass range of $3\times10^{14}{\ensuremath{M_{\odot}}}\leq{\ensuremath{M_{500}}}\leq1.8\times10^{15}{\ensuremath{M_{\odot}}}$. The biasing effects in X-ray selected samples due to the X-ray cool core phenomenon are significantly reduced and perhaps even completely removed. This simplifies the interpretation of the results from our analysis.
We use the [[*XMM-Newton*]{}]{} observations to derive X-ray observables [${T_{\mathrm{X}}}$]{}, [$M_{\mathrm{ICM}}$]{}, [$Y_{\mathrm{X}}$]{}, and rest frame 0.5–2.0 keV and bolometric [$L_{\mathrm{X}}$]{}. For all these observables—save for the [$M_{\mathrm{ICM}}$]{}—we extract both core-included and core-excised quantities, where we define the core to be the region within 0.15[$R_{500}$]{}. The cluster halo masses are derived from the SPT [$\zeta-M_{500}-z$]{} scaling relation and are corrected for selection effects, such as Eddington and Malmquist biases as described in detail in other publications [@bocquet15]. As discussed in detail in Section \[sec:priors\], we adopt priors on the [$\zeta-M_{500}-z$]{} scaling relation from the @deHaan16 joint cosmology and mass calibration analysis, which have since been validated using weak lensing masses of 32 SPT-SZ clusters [@dietrich17] and dynamical masses of 110 SPT-SZ clusters [@capasso17]. These SZE-based halo masses are characteristically uncertain at the $\approx25\%$ level (statistical and systematic uncertainties combined in quadrature).
We fit our data to three different power-law models (see equations \[eq:FormA\], \[eq:FormB\] and \[eq:FormC\]) and derive the best-fit normalization $A_{\ensuremath{\mathcal{X}}}$, mass trend $B_{\ensuremath{\mathcal{X}}}$, $E(z)$ redshift trend $C_{\ensuremath{\mathcal{X}}}$, departure from self-similar redshift trend $\gamma_{\ensuremath{\mathcal{X}}}$, log-normal intrinsic scatter in the X-ray observable at fixed halo mass $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$, and also a redshift dependence to the mass trend $\delta_{\ensuremath{\mathcal{X}}}$. While all three scaling relation forms are adequate to fit the data, we recommend that those interested in cosmological studies adopt Form II, because it models the departure from self-similar evolution with redshift using a cosmologically agnostic form $(1+z)^{\gamma_{\ensuremath{\mathcal{X}}}}$. We marginalize over the uncertainties in the SZE-based halo masses, adjusting the radius [$R_{500}$]{} as appropriate in each iteration in the chain and re-extracting the X-ray observables in a self-consistent manner. Thus, the final parameter uncertainties of the X-ray observable–mass scaling relations include both measurement and systematic halo mass uncertainties (see Table \[tab:sclrel\]).
The halo mass scaling relations for [${T_{\mathrm{X}}}$]{}, [$M_{\mathrm{ICM}}$]{}, [$Y_{\mathrm{X}}$]{} and [$L_{\mathrm{X}}$]{} are steeper, but statistically consistent (within $2\sigma$ confidence) with the results from the literature. However, we observe significant departures from the @mantz16 soft band core-excised luminosity at $3.4\sigma$ level, ICM mass at $2.9\sigma$ level, and [$Y_{\mathrm{X}}$]{} at $2.3\sigma$ level. The mass trends we find in all our scaling relations are steeper than the self-similar behavior at $\gtrsim1.6\sigma$ confidence. In the case of [$M_{\mathrm{ICM}}$]{} and [$L_{\mathrm{X}}$]{} the mass trends we measure (${\ensuremath{M_{\mathrm{ICM}}}}\propto{\ensuremath{M_{500}}}^{1.26\pm0.10}$ and ${\ensuremath{L_{\mathrm{X,cex}}}}\propto{\ensuremath{M_{500}}}^{1.60\pm0.15}$ in soft band) are consistent with most previously published results that employ X-ray selected samples and a mix of weak lensing and hydrostatic masses. However, for [${T_{\mathrm{X}}}$]{} and [$Y_{\mathrm{X}}$]{} our mass trends (${\ensuremath{{T_{\mathrm{X,cex}}}}}\propto{\ensuremath{M_{500}}}^{0.80\pm0.09}$ and ${\ensuremath{Y_{\mathrm{X,cex}}}}\propto{\ensuremath{M_{500}}}^{2.0\pm0.16}$) are steeper than most previous work at $\approx1.6\sigma$ (see parameter $B_{\ensuremath{\mathcal{X}}}$ in Tables \[tab:sclrel\] and \[tab:sclrel\_col2\]).
In addition, we probe for a redshift-dependent mass trend (Form III, equation \[eq:FormC\]) and find that the data currently provide no evidence for such a trend, with the highest significance departure from no evolution being in the core-included [$Y_{\mathrm{X}}$]{} and [$L_{\mathrm{X}}$]{} (see parameter $\delta_{{\ensuremath{\mathcal{X}}}}$ in Tables \[tab:sclrel\] and \[tab:sclrel\_col2\]).
We examine the redshift trends in all scaling relations, finding no significant departures from the self-similar behavior that arises simply due to the evolution of the critical density with redshift. There is no tension between our results and those from previous studies, although many previous studies were not in a position to examine redshift trends, given the limitations of their samples and the availability of halo mass measurements (see parameter $\gamma_{\ensuremath{\mathcal{X}}}$ in Tables \[tab:sclrel\] and \[tab:sclrel\_col2\]).
We report the intrinsic scatter in X-ray observable at fixed halo mass $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ for all scaling relations. These indicate exquisite scatter at the $\approx10\%$ level for [$M_{\mathrm{ICM}}$]{} and core-excised integrated pressure [$Y_{\mathrm{X,cex}}$]{}, somewhat higher scatter of $\approx13\%$ for core-excised temperature [${T_{\mathrm{X,cex}}}$]{}, and scatter of $\approx27$ percent for X-ray luminosities [$L_{\mathrm{X}}$]{} (see parameter $\sigma_{\ln{\ensuremath{\mathcal{X}}}}$ in Tables \[tab:sclrel\] and \[tab:sclrel\_col2\]). We do not account for correlated scatter among the SZE and X-ray observables, because previous analyses of larger SPT-SZ selected samples have failed to detect these effects [@deHaan16; @dietrich17], and therefore they are too small to have an impact on our results.
In all cases, our baseline results are presented in Table \[tab:sclrel\], and the mass and redshift trends for each observable are highlighted in Figs. \[fig:mtrends\] and \[fig:ztrends\]. In addition, we present an alternative set of results in Table \[tab:sclrel\_col2\] that have somewhat better defined redshift trends that come from adopting a calibration of the SZE [$\zeta-M_{500}-z$]{} relation that includes external cosmological priors from BAO (see discussion in Section \[sec:column2\]).
One of the reasons for the steeper mass trends in [${T_{\mathrm{X}}}$]{} and [$Y_{\mathrm{X}}$]{} found in this work could be due to calibration differences affecting the temperatures differently in [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{}. In previous studies of low redshift, high flux clusters, it has been shown that [[*XMM-Newton*]{}]{} temperature estimates lie below [[*Chandra*]{}]{} temperatures in a manner that increases as a function of cluster temperature [@schellenberger15]. However, our sample contains also high redshift systems where the known calibration differences would have less of an impact. Moreover, the [[*XMM-Newton*]{}]{} observations at higher redshift in our sample tend to be lower signal to noise, and in the limit of low signal to noise the background subtraction systematics will tend to be more important than the effective area systematics. Thus, overall we do not expect that the effective area systematics at high energies between [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} are playing an important role in the mass trends of the [${T_{\mathrm{X}}}$]{} and [$Y_{\mathrm{X}}$]{} observables.
Our results are broadly consistent with recent numerical simulations @Barnes2017 at the 1-2$\sigma$ level. A departure from self-similarity in a scaling relation could well indicate that non-gravitational effects in the galaxy clusters are important, and disagreement between simulated and observed scaling relations provides a direct test of the accuracy of the subgrid physics adopted in the simulations. However, one must always be cautious about halo mass systematics as well.
Another concern is a bias in the calibration of the SZE [$\zeta-M_{500}-z$]{} relation, because a bias in the mass trend of the SZE mass–observable relation would indeed be reflected in biased trends in the X-ray observable–mass relations. Here we note only that this SPT calibrated [$\zeta-M_{500}-z$]{} relation offers a unique capability of delivering $\approx25\%$ accurate single cluster masses that have been self-consistently calibrated within a cosmological context that uses the SPT cluster distribution in signal-to-noise and redshift in combination with external mass information. Cross-checks of SZE-based masses with weak lensing [@dietrich17] and dynamical [@capasso17] masses have so far provided no evidence for biases in our masses. Work continues to improve this calibration using weak lensing information from the Dark Energy Survey [e.g., @stern18]. We remind the reader that this work is among the first to extend scaling relation studies to redshifts, which have not yet been covered by the previous X-ray studies. Using SZE-selected clusters and SZE-based halo masses in scaling relations allows us to explore the evolution of massive structures out to higher redshifts.
This work shows the potential of [[*XMM-Newton*]{}]{} observations in deriving X-ray observables of massive, SZE-selected clusters extending to redshifts $z>1$. With the deployment of the next generation SZE experiments [e.g., SPT-3G, CMB-S4, Advanced ACTPOL; @benson14; @CMB-S4-17; @thornton16] and X-ray surveys with eRosita [@merloni12], a large number of new high-redshift clusters will be discovered. Moreover, with deep, multi-wavelength optical surveys like DES, it is already possible to use even the shallower RASS survey to probe the $z\approx1$ Universe [@klein18]. X-ray follow-up observations with [[*XMM-Newton*]{}]{} of these new clusters will provide high quality X-ray spectroscopy for a mass-complete sample at $z>1$ and would enable significant improvements in our understanding of the formation and evolution of the most massive collapsed structures in the Universe.
Acknowledgements {#acknowledgements .unnumbered}
================
Authors thank the anonymous referee and David Rapetti for helpful comments on the draft. We acknowledge the support by the DFG Cluster of Excellence “Origin and Structure of the Universe”, the DLR award 50 OR 1205 that supported I. Chiu during his PhD project, and the Transregio program TR33 “The Dark Universe”. The South Pole Telescope is supported by the National Science Foundation through grant PLR-1248097. Partial support is also provided by the NSF Physics Frontier Center grant PHY-1125897 to the Kavli Institute of Cosmological Physics at the University of Chicago, the Kavli Foundation and the Gordon and Betty Moore Foundation grant GBMF 947.
This paper made use of the package `ChainConsumer` [@hinton16]. This work made use of the IPython package [@PER-GRA:2007], SciPy [@jones_scipy_2001], TOPCAT, an interactive graphical viewer and editor for tabular data [@2005ASPC..347...29T], matplotlib, a Python library for publication quality graphics [@Hunter:2007], Astropy, a community-developed core Python package for Astronomy , NumPy [@van2011numpy].
[^1]: https://heasarc.gsfc.nasa.gov/docs/xmm/uhb/epicextbkgd.html
[^2]: https://heasarc.gsfc.nasa.gov/cgibin/Tools/xraybg/xraybg.pl
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Physics, Syracuse University, Syracuse, NY 13244\
E-mail: [email protected]
author:
- 'Steven R. Blusk'
title: 'Cabibbo-Allowed and Doubly Cabibbo Suppressed $D\to K\pi$ Decays'
---
Introduction
============
For over two decades, the $D^0\to K^-\pi^+$ has served as a workhorse in charm and beauty physics. However, there is general interest in measuring all the $D\to K\pi$ branching fractions. In particular, while it is often assumed that ${\cal{B}}(D\to K^0_S\pi)={\cal{B}}(D\to K^0_L\pi)$, interference between $D\to K^0\pi$ and $D\to \overline{K^0}\pi$ can break this equality[@bigi]. Although this asymmetry is expected, measuring it has alluded experiments because of the challenge of reconstructing the $K^0_L$. Another $K\pi$ mode which has alluded experiments is the doubly Cabibbo-suppressed (DCS) $D^+\to K^+\pi^0$. The difficult stems primarily from the low rate, but also for hadron machines, the lack of a detectable displaced vertex and the large $\pi^0$ combinatorial background make this mode extremely difficult to detect. These difficult $D\to K\pi$ modes are accessible at CLEO-c due to the low-multiplicity environment and threshold production of $D\bar{D}$.
The analyses presented are based on a 281 $\ipb$ sample of data collected at the peak of the $\psi(3770)$ ($\sqrt{s}=3774$ MeV). The resonance is just above threshold for production of $D\bar{D}$, and therefore the final state is in a coherent C=-1 state. For $D^0\bar{D^0}$, these quantum correlations produce deviations in measured branching fractions[@asner_sun], which are maximal when CP eigenstates, $S_{\rm CP}$, are involved. For example, the rate for ($D^0\to S_{\pm},\bar{D}^0\to S_{\pm}$) is zero, and ($D^0\to S_{\pm},\bar{D^0}\to S_{\mp}$) is twice as large with respect to the values obtained when quantum correlations are absent. Four cases of interest that enter into the analyses presented here are: ($D^0\to S_{CP\pm},\bar{D^0}\to X$) and ($D^0\to S_{\pm},\bar{D^0}\to f$), where $f$ represents a flavored final state and $X$ is an unspecified final state. Because of the quantum correlations, the branching fractions are modified as shown in Table \[tab:cpcorr\][@asner_sun], where $x$ and $y$ are the mixing parameters, $r_f e^{-i{\delta_f}}\equiv<f|\bar{D^0}>/<f|D^0>$, and $z_f\equiv\cos\delta_f$. In untagged analyses we can easily correct branching fractions using the word-average $y=0.008\pm0.005$[@pdg04]. We also note that yields in these and other combinations of final states can be used to measure the $D^0\bar{D}^0$ mixing parameters and the strong phase $\delta_{K\pi}$[@asner_sun].
In reconstructing $D$ mesons, we use two kinematic variables: $\de\equiv E_{D}-\ebeam$ and $\mbc\equiv\sqrt{\ebeam^2-p_D^2}$, where $E_D$ is the energy of the $D$ candidate and $p_D$ its momentum. [*Untagged*]{} analyses reconstruct $D$ mesons in exclusive final states using all charged particles and showers in the event. [*Tagged*]{} analyses start with events that already have a $D$ candidate [*ie., a tag*]{}, and seek to reconstruct the second $D$ meson (referred to as the [*signal*]{}). Because of the highly constrained kinematics, the signal $D$ may contain undetected particles, such as a $K^0_L$ (or a $\nu$), which are inferred by energy/momentum conservation. In particular, for the decay $D\to K^0_L\pi$, the signal is a peak in the missing-mass squared, defined using the measured four-momenta as: $\mmiss^2 = (p_{\rm event}-p_{tag}-p_{\pi})^2$.
$\bf(D^0\to K^0_{S,L}\pi^0)$
============================
We first measure $\bf(D^0\to K^0_S\pi^0)$ using an untagged analysis. Candidates are formed by combining $K^0_S\to\pi^+\pi^-$ and $\pi^0$ candidates and requiring $\de$ and $\mbc$ to be within 3 standard deviations of 0 and $M_{D^0}$, respectively. Combinatorial background and cross-feed from $D^0\to\pi^+\pi^-\pi^0$ are estimated using $\de$ and $K^0_S\to\pi^+\pi^-$ mass sideband regions, respectively. Combining the signal yield of $7487\pm99$ events with the efficiency of 29.0% and $N_{D^0\bar{D^0}}=1.015\times10^6$, we find: $\bf(D^0\to K^0_S\pi^0)=(1.260\pm0.02\pm0.054)\%$. Of the 4.2% systematic uncertainty, 3.8% is from the $\pi^0$ detection efficiency, which cancels when comparing $K_S^0\pi^0$ and $K_L^0\pi^0$.
Measurement of $\bf(D^0\to K^0_L\pi^0)$ requires a tagged analysis, and since $K^0_L\pi^0$ is a CP+ eigenstate, it requires that we determine the factor $1+2r_fz_f+r_f^2$ (which is unknown, since $\delta_f$ is unknown). However, by measuring $\bf(D^0\to K^0_S\pi^0)$ in tagged events, and comparing to the value in untagged events, we can determine $(1-2r_fz_f+r_f^2)$. Along with the measured values of $r_f$, this enables us to compute the factor we want, $(1+2r_fz_f+r_f^2)$. We therefore need $\bf(D^0\to K^0_S\pi^0)$ in flavor-tagged events.
The tagged $D^0\to K^0_S\pi^0$ tagged analysis starts with events containing a reconstructed $D$-tag in $\bar{D^0}\to K^+\pi^-$, $\bar{D^0}\to K^+\pi^-\pi^0$ or $\bar{D^0}\to K^+\pi^-\pi^+\pi^-$, and then seeks to reconstruct $D^0\to K^0_S\pi^0$ candidates as described in the untagged analysis. The yields, efficiencies and corresponding products $\bf(D^0\to K_S^0\pi^0)(1-2r_fz_f+r_f^2)$ are shown in Table \[tab:kspi0\_sum\]. Using the measured value of $\bf(D^0\to K^0_S\pi^0)$ from the untagged analysis, we also compute $(1-2r_fz_f+r_f^2)$ and subsequently $(1+2r_fz_f+r_f^2)$ using the most recent $r_f$ values[@pdg04; @zhang; @tian]. That these factors are not unity is a direct consequence of the quantum coherence of the final state.
The measurement of $K^0_L\pi^0$ is slightly more complicated. It starts with the same sample of D-tag’s as in the $K^0_S\pi^0$ tagged analysis, and, for each candidate, we require the presence of one and only one additional $\pi^0$ candidate, and no extra tracks or $\eta\to\gamma\gamma$ candidates. In these events, we form $\mmiss^2$, which for $K^0_L\pi^0$ events peaks at $M_{K_L^0}^2$. Backgrounds such as $K_S^0\pi^0$ and $\eta\pi^0$ are highly suppressed by the selection requirements, but do peak under the signal. These backgrounds are estimated using simulation. Other backgrounds are estimated using $\mmiss^2$ sidebands in data in combination with shapes from simulation. The distribution of $\mmiss^2$ is shown in Fig. \[fig:klpi0\_mmiss\] for all tag modes combined; the data are the points with error bars, the solid line is the simulation, and the dashed lines show various background contributions. The data are peaked toward slightly lower missing-mass than simulation. This effect is traced to a 0.5% difference in the energy scale of $\pi^0$’s, which has only a minor effect in this analysis. Yields, efficiencies and the branching fractions, $\bf(D^0\to K_L^0\pi^0)$, are shown in Table \[tab:klpi0\_sum\], where the branching fractions have been corrected by the factor, $(1+2r_fz_f+r_f^2)$. After averaging the three tag modes, we obtain $\bf(D^0\to K_L^0\pi^0)=(0.986\pm0.049\pm0.047)\%$, where the last uncertainty is systematic and dominated by the $\pi^0$ efficiency (3.8%).
Defining an asymmetry: $$R(D) = \frac{\bf(D\to K_S^0\pi)-\bf(D\to K_L^0\pi)}{\bf(D\to K_S^0\pi)+\bf(D\to K_L^0\pi)},$$
we find that $R(D^0)=0.122\pm0.025\pm0.019$, establishing the inequality of these branching fractions. Using general arguments involving the contributing Feynman diagrams, one would expect this asymmetry to be $R(D^0)=2\tan^2\theta_C=0.109\pm0.001$, where $\theta_C$ is the Cabibbo angle. This expectation is in good agreement with our measurement.
![Distribution of $\mmiss^2$ for $D^0\to K^0_L\pi^0$ candidates in tagged events. The points with error bars are data, the solid line is the total simulation, and the dashed lines are various backgrounds. \[fig:klpi0\_mmiss\]](klpi0_mmiss.eps){height=".23\textheight"}
$\bf(D^+\to K^0_{S,L}\pi^+)$
============================
We look to measure the same asymmetry in charged $D$ decays. The branching fraction, $\bf(D^+\to K^0_S\pi^+)$ has been measured in a separate analysis[@dhad]. The measurement of $\bf(D^+\to K^0_L\pi^+)$ requires a tagged analysis, and is strategically similar to the $\bf(D^0\to K^0_L\pi^0)$ measurement. We reconstruct a $D^-$ tag in 6 tag modes: $D^-\to K^+\pi^-\pi^-$, $K^+\pi^-\pi^-\pi^0$, $K^0_S\pi^-$, $K^0_S\pi^-\pi^0$, $K^0_S\pi^-\pi^+\pi^-$, and $K^+K^-\pi^-$, by requiring $\de$ consistent with zero. Selecting events within $\sim$3$\sigma$ of $M_{D^-}$, we obtain 165,00 $D^-$ tags. For each tag, we query the remainder of the event and require exactly 1 extra charged track, consistent with a pion hypothesis, and no extra $\pi^0$’s. Using the $D^-$ tag and the pion, we compute $\mmiss^2$, which is shown in Fig. \[fig:klpi\_mmiss\] for all tag modes combined. The points with error bars show the data, and the colored line passing through the points shows the fit. The prominent $K^0_L$ peak is evident as well as a high-side shoulder from $D^+\to\eta\pi^+$ (this analysis does not veto $\eta\to\gamma\gamma$). The other lines show the individual contributions from $K^0_L\pi^+$ (signal), and various backgrounds, such as $K^0_S\pi^+$, which peaks under $K^0_L\pi^+$; $D^+\to\pi^+\pi^0$ and $D^+\to\mu^+\nu_{\mu}$, which peak near zero; and other non-peaking backgrounds. A total of 2023$\pm$54 $D^+\to K^0_L\pi^+$ signal events are observed from an initial tagged sample of 165,000 $D^-$ tags.
The branching fraction is computed for each tag mode and then the results are combined using a weighted average. The efficiency varies slightly depending on the tag mode, but is typically about 82%. The average branching fraction is found to be: $\bf(D^+\to K^0_L\pi^+)=1.46\pm0.040\pm0.035\pm0.004)$, where the last systematic is due the uncertainty in $\bf(D^+\to K^0_S\pi^+)$.
![Distribution of $\mmiss^2$ for $D^+\to K^0_L\pi^+$ candidates in tagged events. The points with error bars are data, the solid line are signal and background contributions as described in the text. \[fig:klpi\_mmiss\]](klpi_mmiss.eps){height=".23\textheight"}
Using $\bf(D^+\to K^0_S\pi^+)=(1.552\pm0.022\pm0.029)\%$[@dhad], we measure an asymmetry, $R(D^+)=0.031\pm0.016\pm0.016$. This asymmetry is consistent with zero. Because of the larger number of additional Feynman diagrams which contribute to this decay, no simple prediction of this asymmetry can be made. Both this analysis and the $D^0\to K^0_{S,L}\pi^0$ will be submitted for publication soon.
$\bf(D^+\to K^+\pi^0)$
======================
Until recently, the DCS $D^+$ decays were limited to modes with only charged particles due to the low rate and large combinatorial background associated with $\pi^0$ reconstruction. The threshold production of $D\bar{D}$ events in CLEO-c make this measurement accessible[@cleo_dcs]. CLEO searches for this decay using an untagged analysis by combining $K^+$ and $\pi^0$ candidates and requiring $-40<\de<35$ MeV. We find a yield of $148\pm23$ events. We use $D^+\to K^-\pi^+\pi^+$ as a normalizing mode, for which there are 79612 decays. The efficiencies of the DCS and normalizing mode are 42.30% and 52.16%, respectively, yielding a branching fraction, $\bf(D^+\to K^+\pi^0)=(2.28\pm0.36\pm0.15\pm0.08)\times10^{-4}$. This measurement is of substantially better precision than a recent measurement by BaBar[@babar], which used data collected near the $\Upsilon$(4S) with $\sim$1000 times larger integrated luminosity than CLEO-c.
We gratefully acknowledge the effort of the CESR staff in providing us with excellent luminosity and running conditions, and the National Science Foundation for support of this work.
[9]{}
I. I. Bigi and H. Yamamoto, Phys. Lett. B [**349**]{}, 363 (1995).
D. M. Asner and W. M. Sun, Phys. Rev. D [**73**]{}, 034024 (2006).
S. Eidelman, [*et. al.*]{}, Phys. Lett. B [**592**]{}, 1 (2004)
L. M. Zhang [*et. al.*]{} (Belle Collaboration), Phys. Rev. Lett. [**96**]{}, 151801 (2006).
X. C. Tian [*et. al.*]{} (Belle Collaboration), Phys. Rev. Lett. [**95**]{}, 231801 (2005).
Q. He [*et al.*]{} (CLEO Collaboration), Phys. Rev. Lett. [**95**]{}, 121801 (2005). Update based on 281 $\ipb$ to be submitted to Phys. Rev. D.
S. Dytman [*et. al.*]{} (CLEO Collaboration), Phys. Rev. D [**74**]{}, 071102 (2006).
B. Aubert [*et. al.*]{} (BaBar Collaboration), Phys. Rev. D [**74**]{}, 0111107(R) (2006).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Bayesian On-line Changepoint Detection is extended to on-line model selection and non-stationary spatio-temporal processes. We propose spatially structured Vector Autoregressions ([[[<span style="font-variant:small-caps;">VARs</span>]{}]{}]{}) for modelling the process between changepoints ([[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}) and give an upper bound on the approximation error of such models. The resulting algorithm performs prediction, model selection and [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}detection on-line. Its time complexity is linear and its space complexity constant, and thus it is two orders of magnitudes faster than its closest competitor. In addition, it outperforms the state of the art for multivariate data.'
bibliography:
- 'library.bib'
---
Introduction {#Section_introduction}
============
Real-world spatio-temporal processes are often poorly modelled by standard inference methods that assume stationarity in time and space. A variety of techniques have been developed for modelling non-stationarity in time via changepoints ([[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}), ranging from methods for Gaussian Processes ([[[<span style="font-variant:small-caps;">GPs</span>]{}]{}]{}) [@Osborne], the Lasso [@lin2017sharp] or the Ising model [@fazayeli2016generalized] over approaches using density ratio estimation [@liu2013change] and kernel-based methods exploiting M-statistics [@li2015m] to framing [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}detection as time series clustering [@khaleghi2014asymptotically]. In contrast, [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}inference allowing for non-stationarity in space [@Flaxman] has received comparatively little attention.
0.1in
-0.2in
We offer the first on-line solution to this problem by modeling non-stationarity in both space and time. [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}are used to model non-stationarity in time, and the use of spatially structured Bayesian Vector Autoregressions ([[[<span style="font-variant:small-caps;">SSBVAR</span>]{}]{}]{}) circumvents the assumption of stationarity in space. We unify @BOCD and @FearnheadOnlineBCD into an inference procedure for **on-line** **prediction**, **model selection** and **[[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}detection**, see Fig. \[picture\_demo\]. [ Our construction exploits that both algorithms use Product Partition Models [@PPM], which assume independence of parameters conditional on the [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}and independence of observations conditional on these parameters. ]{}
[ Our method can be seen as modified on-line version of @MurphyMVTSBCP. In their method, inference is off-line, the model universe $\mathcal{M}$ is built during execution and multivariate dependencies are restricted to decomposable graph. In contrast, our procedure specifies $\mathcal{M}$ before execution, but runs on-line and does not restrict dependencies. The closest competing on-line procedure in the literature thus far is the work of @GPBOCD, which develops Gaussian Process ([[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}) [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}models for Bayesian On-line Changepoint Detection ([[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}). Though our results suggest that parametric models may be preferable to [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}models, the latter can still be integrated into our method as elements of the model universe $\mathcal{M}$ without any further modifications. ]{}
[ In summary, we make three contributions: Firstly, we substantially augment the existing work on [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}by allowing for model uncertainty. Unlike previous extensions of the algorithm [e.g. @BOCD; @GPBOCD], this avoids having to guess a single best model a priori. Secondly, we introduce [[[<span style="font-variant:small-caps;">SSBVARs</span>]{}]{}]{}as the first class of models for multivariate inference within [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}. Thirdly, we demonstrate that using a collection of parametric models can outperform nonparametric [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}models in terms of prediction, [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}detection and computational efficiency. ]{}
The structure of this paper is as follows: Section \[Algorithm\] generalizes the [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}algorithm of @BOCD, henceforth [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}, by integrating it with the approach of @FearnheadOnlineBCD, henceforth [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}. In so doing, we arrive at [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}with Model Selection, henceforth [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}. Section \[VAR\] proposes [[[<span style="font-variant:small-caps;">VAR</span>]{}]{}]{}models for non-stationary processes within the [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}framework. This motivates populating the model universe $\mathcal{M}$ with spatially structured [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}(SSBVAR) models. Sections \[hyperpar\_opt\]–\[computational\_complexity\] address computational aspects. Section \[results\] demonstrates the algorithm’s advantages on real world data.
BOCPDMS {#Algorithm}
=======
Let $\{\*Y_t\}_{t=1}^{\infty}$ be a data stream with an unknown number of CPs. Focusing on univariate data, [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}and [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}tackled inference by tracking the posterior distribution for the most recent [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}. While [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}allow the data to be described by different models between [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}, [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}only allow for a single model. However, [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}perform one-step-ahead predictions, whereas [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}do not. Instead, they propose a Maximum A Posteriori ([[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}) segmentation for [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}and models. In the remainder of this section, we unify both inference approaches. We call the resulting algorithm [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}with model selection ([[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}), as it performs prediction, [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}segmentation and model selection on-line.
Run-length & model universe {#run-length-model-univ-subsec}
---------------------------
The *run-length* $r_t$ at time $t$ is defined as the time since the most recent [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}at time $t$, so $r_t = 0$ corresponds to a [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}at time $t$. Suppose that data between successive [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}can be described by Bayesian models collected in the *model universe* $\mathcal{M}$. For the process $\{\*Y_t\}$ on $\mathbb{R}^S$, a model $m \in \mathcal{M}$ with finite memory of length $L\in\mathbb{N}_0$ consists of an observation density $f_m(\*Y_t=\*y_t|\*\theta_m, \*y_{(t-L):(t-1)})$ on $\mathbb{R}^S$ and a parameter prior $\pi_m(\*\theta_m)$ on $\*\Theta_m$ depending on hyperparameters $\*\nu_m$. The notion of $\mathcal{M}$ is due to [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}and allows for model uncertainty amongst models developed for [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}. For instance, $m \in \mathcal{M}$ could be a [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}[@GPBOCD], a time-deterministic regression [@FearnheadSignalProcessing] or a mixture distribution [@CaronDoucet].
model universe $\mathcal{M}$; hazard $H$; prior $q$ next observation $\*y_t$ $t$: $\widehat{\*y}_{(t+1):(t+h_{\max})}$, $S_t$, $p(m_t|\*y_{1:t})$\
Initialize $p(\*y_{1:t}, r_t=0, m_t=m)$ with prior Update $p(\*y_{1:t}, r_t, m_t=m)$ via , Prune model-specific run-length distribution Perform hyperparameter inference via Obtain joint distribution over $\mathcal{M}$ via – Compute – **Output:** $\widehat{\*y}_{(t+1):(t+h_{\max})}, S_t, p(m_t|\*y_{1:t})$
Probabilistic formulation & recursions
--------------------------------------
Denote by $m_t$ the model describing $\*y_{(t-r_t):t}$, i.e. the data since the last [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}. Given hazard function $H:\mathbb{N} \to [0,1]$, and model prior $q:\mathcal{M} \to [0,1]$, the prior beliefs are
[l]{} p(r\_t|r\_[t-1]{}) =
1 - H(r\_[t-1]{} + 1) & r\_t=r\_[t-1]{}+1\
H(r\_[t-1]{} + 1) & r\_t=0\
0 &
\[run\_length\_prior\]\
q(m\_t|m\_[t-1]{},r\_t) =
\_[m\_[t-1]{}]{}(m\_t) & r\_t=r\_[t-1]{}+1\
q(m\_t) & r\_t=0.\
\[model\_prior\]
Eq. implies that the model at time $t$ will be equal to the model at time $t-1$ unless a [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}occured at $t$, in which case the next model $m_t$ will be a random draw from $q$. At time $t$, the algorithm requires for all possible models $m$ and run-lengths $r_t$ the computation of the posterior predictives
[lll]{} & & f\_m(y\_[t]{}| y\_[1:(t-1)]{}, r\_t)\
& = & \_[\_[m]{}]{} f\_m(y\_[t]{}|\_[m]{})\_m(\_[m]{}|y\_[(t-L-r\_t):(t-1)]{}) d\_[m]{}.
To make the evaluation of this integral efficient, one can use conjugate models [@MurphyMVTSBCP] or approximations [@TurnerVB; @CHAMP], which make the following recursion efficient, too:
[rCl]{} & & p(y\_[1:t]{}, r\_[t]{}, m\_[t]{}) =\
& & \_[m\_[t-1]{}]{}\_[r\_[t-1]{}]{}{ f\_[m\_t]{}(y\_t|y\_[1:(t-1)]{},r\_[t]{}) q(m\_t|y\_[1:(t-1)]{}, r\_t,m\_[t-1]{})\
&& p(r\_t|r\_[t-1]{}) p(y\_[1:(t-1)]{}, r\_[t-1]{},m\_[t-1]{})}. \[recursion\]
The recursion in [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}is the special case for $|\mathcal{M}| = 1$. For $|\mathcal{M}| > 1$, $q(m_t|m_{t-1}, r_t, \*y_{1:(t-1)})$ arises as a new term, which for $\*{1}_{a}$ as the indicator function of $a$ is given by
[C]{}
\_[m\_[t-1]{}]{}(m\_t)q(m\_[t-1]{}|y\_[1:(t-1)]{}, r\_[t-1]{}) & r\_t=r\_[t-1]{}+1\
q(m\_t) & r\_t=0.
\[model\_posterior\_probability\]
Next, define the *growth-* and *changepoint probabilities* as
[lll]{} && p(y\_[1:t]{}, r\_[t]{}=r\_[t-1]{}+1, m\_[t]{}) =\
& & f\_[m\_t]{}(y\_t| y\_[1:(t-1)]{},r\_t) p(y\_[1:(t-1)]{}, r\_[t-1]{}, m\_[t-1]{}) \[growth\_probability\]\
&& (1-H(r\_t)) q(m\_[t-1]{}|y\_[1:(t-1)]{}, r\_t),\
&& p(y\_[1:t]{}, r\_[t]{}=0, m\_[t]{}) =\
& & f\_[m\_t]{}(y\_t| y\_[1:(t-1)]{}, r\_[t]{})q(m\_t) \[cp\_probability\]\
& & \_[m\_[t-1]{}]{}\_[r\_[t-1]{}]{}{ H(r\_[t-1]{} + 1) p(y\_[1:(t-1)]{}, r\_[t-1]{}, m\_[t-1]{}) }.
The evidence can then be calculated via Eq. , which in turn allows calculating the joint model-and-run-length distribution , the model posterior , as well as the model-specific and global run-length distributions:
[rCl]{} p(y\_[1:t]{}) & = & \_[m\_t]{}\_[r\_t]{} p(y\_[1:t]{}, m\_t, r\_t) \[evidence\]\
p(r\_t,m\_t| y\_[1:t]{}) & = & p(y\_[1:t]{}, r\_t, m\_t)/p(y\_[1:t]{}) \[model\_and\_run\_length\]\
p(m\_t|y\_[1:t]{}) & = & \_[r\_t]{}p(r\_t,m\_t| y\_[1:t]{}) \[model\_posterior\]\
p(r\_t|m\_t, y\_[1:t]{}) & = & p(r\_t, m\_t|y\_[1:t]{})/p(m\_t|y\_[1:t]{}) \[model\_specific\_run\_length\]\
p(r\_t|y\_[1:t]{}) & = &\_[m\_t]{}p(r\_t,m\_t| y\_[1:t]{}) \[global\_run\_length\]\
q(m\_[t-1]{}|y\_[1:(t-1)]{}, r\_[t-1]{}) & = &. \[conditional\_model\_posterior\]
Eq. is the conditional model posterior from Eq. . Eq. is arrived at directly in [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}and used for on-line [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}segmentation. By framing our derivations in the run-length framework of [[[<span style="font-variant:small-caps;">AM</span>]{}]{}]{}, we additionally obtain –, thus enabling on-line prediction and model selection at the same computational cost.
On-line algorithm outputs
-------------------------
**Prediction:** Recursive $h$-step-ahead forecasting uses :
[lll]{} &&p(Y\_[t+h]{}|y\_[1:t]{})\
& = & \_[r\_t,m\_t]{}{ p(Y\_[t+h]{}|y\_[1:t]{},\_t\^h, r\_t, m\_t) p(r\_t,m\_t| y\_[1:t]{}) }, \[posterior\_predictive\]
where $\widehat{\*y}_t^h = \emptyset$ if $h=1$ and $\widehat{\*y}_t^h = \widehat{\*y}_{(t+1):(t+h-1)}$ otherwise, with $\widehat{\*y}_{t+h} = \mathbb{E}(\*Y_{t+h}|\*y_{1:t},\widehat{\*y}_t^h)$ the recursive forecast.
**Tracking the model posterior/Bayes Factors:** One of the novel capabilites of the algorithm is on-line monitoring of the model posterior via Eq. . This is attractive when structural changes in the data happen slowly and are not captured well by [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}. In this case, $\mathbb{P}(m_t|\*y_{1:t})$ can be used to identify periods of change, see Fig. \[Temperatures\]. For pairwise comparisons, Bayes Factors can be monitored, too:
[rCl]{} \_t & = & .\[BayesFactors\]
**Maximum A Posteriori ([[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}) segmentation:** For $\text{MAP}_t$ the density of the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}-estimate of models and [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}before $t$ and $\text{MAP}_0 = 1$, [[[<span style="font-variant:small-caps;">FL</span>]{}]{}]{}’s recursive estimator is given by
[rCl]{} \_t = \_[r,m]{}{ p(y\_[1:t]{}, r\_t=r, m\_t=m) \_[t-r-1]{} }. \[MAP\_estimator\]
For $r^{\ast}_t, m^{\ast}_t$ maximizers for time $t$, the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}segmentation is $S_t = S_{t-r^{\ast}_t-1} \cup \{(t-r^{\ast}_t, m^{\ast}_t) \}$, $S_0 = \emptyset$, where $(t',m_{t'}) \in S_t$ means a [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}at $t'\leq t$, with $m_{t'} \in \mathcal{M}$ the model for $\*y_{t':t}$.
Building a spatio-temporal model universe {#VAR}
=========================================
The last section derived [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}for arbitrary data streams $\{\*Y_t\}$. Next, we propose models for $\mathcal{M}$ if $\{\*Y_t\}$ can be mapped into a space $\mathbb{S}$. Let $\mathcal{S}$ with $|\mathcal{S}| = S$ be a set of spatial locations in $\mathbb{S}$ with measurements $\*Y_t = (Y_{t,1}, Y_{t,2}, \dots, Y_{t, S})^T$ recorded at times $t=1,2,\dots$
Bayesian [[[<span style="font-variant:small-caps;">VAR</span>]{}]{}]{}([[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{})
--------------------------------------------------------------------------------------------------------------------------------------
Inference on $\{\*Y_t\}$ can be drawn using conjugate Bayesian Vector Autoregressions ([[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}) with lag length $L$ and $E$ additional variables $\*Z_t$ as elements of model universe $\mathcal{M}$:
[rCl]{} \^2 & \~& (a,b) \[BVAR\_eq\_1\]\
\_t|\^2 & [\~]{} & (0, \^2 ) \[BVAR\_eq\_2\]\
c|\^2 & \~& (0, \^2 V\_c)\[BVAR\_eq\_3\]\
Y\_t & = & + B Z\_t + \_[l = 1]{}\^L A\_l Y\_[t-l]{} + \_t. \[BVAR\_eq\_4\]
Here, $\*A_l, \*B$ are $S\times S$, $S \times E$ matrices, $\*c = (\*\alpha, \text{vec}(\*B), \text{vec}(\*A_1), \text{vec}(\*A_2), \dots \text{vec}(\*A_L))^T$ is a vector of $S \cdot (LS + 1 + E)$ model parameters. Scalars $a,b>0$, matrix $\*V_c$, and diagonal matrix $\*\Omega$ are hyperparameters.
Approximating processes using [[[<span style="font-variant:small-caps;">VARs</span>]{}]{}]{} {#section_approx_var}
--------------------------------------------------------------------------------------------
Modelling $\{\*Y_t\}$ as [[[<span style="font-variant:small-caps;">VAR</span>]{}]{}]{}is attractive, as many complex non-linear processes have [[[<span style="font-variant:small-caps;">VAR</span>]{}]{}]{}representations, including [[[<span style="font-variant:small-caps;">HMMs</span>]{}]{}]{}, time-stationary [[[<span style="font-variant:small-caps;">GPs</span>]{}]{}]{}as well as multivariate [[[<span style="font-variant:small-caps;">GARCH</span>]{}]{}]{}and fractionally integrated [[[<span style="font-variant:small-caps;">VARMA</span>]{}]{}]{}processes [@BaxterIE2; @VARFARIMA]. Performance guarantees for [[[<span style="font-variant:small-caps;">VAR</span>]{}]{}]{}approximations to such processes are derived using Baxter’s Inequalitiy with multivariate versions of results in @HannanKavalieris.
\[VAR\_Thm\] Let $\{\*Y_t\}$ be a time-stationary spatio-temporal process with spectral density satisfying regularity condition **A** in the Appendix, $||\cdot ||$ a matrix norm, $\mathbb{E}(\*Y_t) = 0$, $\mathbb{E}(\*Y_t\*Y_t^T)<\infty$, $\sum_{h=-\infty}^{\infty}(1 + |h|)^3 ||\mathbb{E}[\*Y_{t}\*Y_{t+h}']|| < \infty$. Then (1)–(3) hold.\
- $\*Y_t = \sum_{i=1}^{\infty}\*A_i\*Y_{t-i} + \*\varepsilon_t$ for matrices $\{\*A_l\}_{l\in\mathbb{N}}$ and $\mathbb{E}(\varepsilon_t)= 0$, $\mathbb{E}(\*\varepsilon_t\*\varepsilon_t')= \*D$, $\*D$ diagonal.\
- For $\*Y_t = \sum_{l=1}^{L}{\*A}^L_l\*Y_{t-l} + e_t$ with $\{{\*A}^L_l\}_{l=1}^L$ the best linear projection coefficients, $\exists L_0:\forall L>L_0$, $\sum_{l=1}^L(1+|l|)^3||\*A^L_l - \*A_l|| \leq C \cdot \sum_{l=L+1}^{\infty}(1+|l|)^3||\*A_l||$ with $C$ constant.\
- Using $T$ observations with $L = \mathcal{O}([T/\ln(T)]^{1/6})$ to estimate ${\*A}^L_l$ as [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}$\widehat{\*A}^L_l$ of –, it holds that $L(T)^2\sum_{l=1}^{L(T)}||\widehat{\*A}^{L(T)}_l - {\*A}^{L(T)}_l|| \overset{P}{\to} 0$ as $T \to \infty$.\
Part (1) is shown in @VARFARIMA, part (2) in Lemma 3.1 of @VARSieve. Part (3) follows by their [Remark 3.3]{} if we can prove that the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}estimator $\hat{\*c}(L(T))$ of $\*c$ equals its Yule-Walker estimator ([[[<span style="font-variant:small-caps;">YWE</span>]{}]{}]{}) as $T \to \infty$. Let $\*B=\*0$, $\*\alpha =\*0$ and note that [[[<span style="font-variant:small-caps;">YWE</span>]{}]{}]{}equals [[[<span style="font-variant:small-caps;">OLS</span>]{}]{}]{}as $T \to \infty$. With $\*X_{1:T}$ the regressor matrix of $\*Y_{t-L(T):t}$, $\hat{\*c}(L(T)) = (\*X_{1:T}'\*X_{1:T} + \*V_c^{-1})^{-1}(\*X_{1:T}'\*Y_{1:T})$. Then, part (3) holds as [[[<span style="font-variant:small-caps;">OLS</span>]{}]{}]{}$\overset{P}{\to}\mathbb{E}(\*X_{1:T}'\*X_{1:T})^{-1}\mathbb{E}(\*X_{1:T}'\*Y_{1:T})$ and
[rCl]{} (L(T)) & = & (X\_[1:T]{}’X\_[1:T]{} + V\_c\^[-1]{})\^[-1]{}(X\_[1:T]{}’Y\_[1:T]{})\
& = & ( X\_[1:T]{}’X\_[1:T]{} + V\_c\^[-1]{})\^[-1]{} (X\_[1:T]{}’Y\_[1:T]{})\
& & (X\_[1:T]{}’X\_[1:T]{})\^[-1]{}(X\_[1:T]{}’Y\_[1:T]{}).\
In Thm. \[VAR\_Thm\], assuming $\mathbb{E}(\*Y_t) = \*0$ is without loss of generality: If $\mathbb{E}(\*Y_t) = \*\alpha + \*B\*Z_t$, define $\*Y_t^{\ast} = \*Y_t - (\*\alpha + \*B\*Z_t)$ and apply the theorem to $\{\*Y_t^{\ast}\}$. Moreover, the results do *not* require stationarity in space. Lastly, part [(3)]{} suggests a principled way of picking lag lengths $\mathcal{L}$ for [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}models based on functions $L(T) = C\cdot(T/\ln(T))^{1/6}$, with $C$ a constant: If between $T_1$ and $T_2$ observations are expected between [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}, $\mathcal{L} = \{L \in \mathbb{N}: L(T_1) \leq L\leq L(T_2) \}$. In our experiments, we employ this strategy using $T_1 = 1, T_2 = T$.
Modeling spatial dependence {#nbh_sequences_section}
---------------------------
While Thm. \[VAR\_Thm\] motivates approximating spatio-temporal processes between [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}with –, the matrices $\{{\*A}^L_l\}_{l=1}^L$ have $S (LS + 1 + E)$ parameters. This increases model complexity and ignores spatial information. We remedy both issues through neighbourhood systems on $\mathcal{S}$.
For a set of locations $\mathcal{S}$ with the sets $N_i(s) \subseteq \mathcal{S}$ as the [$i$-th neighbourhoods of $s$]{} for $0 \leq i \leq n$ and all $s \in \mathcal{S}$, let $N_i(s) \cap N_j(s) = \emptyset$, $s' \in N_i(s) \Longleftrightarrow s \in N_i(s')$ and $N_0(s) = \{ s \}$. Then, the corresponding [neighbourhood system]{} is ${N}(\mathcal{S}) = \left\{\{N_i(s)\}_{i=1}^{n}: s \in \mathcal{S}, 0 \leq i \leq n\right\}$.
In the remainder of the paper, smaller indices $i$ imply that the neighbourhoods $N_i(s)$ are closer to $s$. For a [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}model of lag length $L$, the decay of spatial dependence is encapsulated through $\Pi:\{1,\dots,L\} \to \{0,\dots, n\}$. In particular, only $s' \in N_i(s)$ with $i\leq \Pi(l)$ are modeled as affecting $s$ after $l$ time periods.
= \[circle,draw,fill=gray!40,minimum size=10\] =\[circle,draw,fill=red!40,minimum size=10\] =\[circle,draw,fill=blue!40,minimum size=10\] = \[circle, draw, fill=orange!40, minimum size = 10\] = \[circle, draw, fill=white!40, minimum size = 10\] (tmtwo) at (0.85, 2.5) [$t-2$]{}; (1) at (0,1.5) [1\[1\]]{}; (2) at (0.75,1.5) [2\[2\]]{}; (3) at (1.5,1.5) [3\[3\]]{}; (4) at (0,0.75) [4\[4\]]{}; (5) at (0.75,0.75) [5\[5\]]{}; (6) at (1.5,0.75) [6\[6\]]{}; (7) at (0,0) [7\[7\]]{}; (8) at (0.75,0) [8\[8\]]{}; (9) at (1.5,0) [9\[9\]]{};
(tmone) at (3.85, 2.5) [$t-1$]{}; (1\_) at (3,1.5) [1\[1\_\]]{}; (2\_) at (3.75,1.5) [2\[2\_\]]{}; (3\_) at (4.5,1.5) [3\[3\_\]]{}; (4\_) at (3,0.75) [4\[4\_\]]{}; (5\_) at (3.75,0.75) [5\[5\_\]]{}; (6\_) at (4.5,0.75) [6\[6\_\]]{}; (7\_) at (3,0) [7\[7\_\]]{}; (8\_) at (3.75,0) [8\[8\_\]]{}; (9\_) at (4.5,0) [9\[9\_\]]{};
\(t) at (7.2, 2.5) [$t$]{}; (1\_\_) at (6,1.5) [1\[1\_\_\]]{}; (2\_\_) at (6.75,1.5) [2\[2\_\_\]]{}; (3\_\_) at (7.5,1.5) [3\[3\_\_\]]{}; (4\_\_) at (6,0.75) [4\[4\_\_\]]{}; (5\_\_) at (6.75,0.75) [5\[5\_\_\]]{}; (6\_\_) at (7.5,0.75) [6\[6\_\_\]]{}; (7\_\_) at (6,0) [7\[7\_\_\]]{}; (8\_\_) at (6.75,0) [8\[8\_\_\]]{}; (9\_\_) at (7.5,0) [9\[9\_\_\]]{};
\(2) edge \[red!40, line width = 1.0\] (4); (2) edge \[red!40, line width = 1.0\] (6); (4) edge \[red!40, line width = 1.0\] (8); (6) edge \[red!40, line width = 1.0\] (8); (2\_) edge \[red!40, line width = 1.0\] (4\_); (2\_) edge \[red!40, line width = 1.0\] (6\_); (4\_) edge \[red!40, line width = 1.0\] (8\_); (6\_) edge \[red!40, line width = 1.0\] (8\_); (1\_) edge \[blue!40, out = 50, in = 130,line width = 1.0\] (3\_); (1\_) edge \[blue!40, out = 225, in = 140, line width = 1.0\] (7\_); (3\_) edge \[blue!40,in = 40, out = -40 , line width = 1.0\] (9\_); (7\_) edge \[blue!40, out = -50, in = -130, line width = 1.0\] (9\_);
\(2) edge \[out = 10, in =125, red!40, line width = 2.0\] (5\_\_); (2\_) edge \[out = 20, in =125, red!40, line width = 2.0\] (5\_\_); (5) edge \[out = -25, in =225, orange!40, line width = 2.0\] (5\_\_); (5\_) edge \[out = -50, in =225, orange!40, line width = 2.0\] (5\_\_); (3\_) edge \[out = -40, in =180, blue!40, line width = 2.0\] (5\_\_);
-0.2in
Spatializing [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}
---------------------------------------------------------------------------
In principle, given $N(\mathcal{S})$, sparsification of the [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}model – is possible in two ways: As restriction on the *contemporaneous* dependence via the covariance matrix of the error term $\*\varepsilon_t$, or as restriction on the *conditional* dependence via the coefficient matrices $\{\*A_l\}_{l=1}^L$. We choose the latter for three reasons: Firstly, linear effects have more interesting interpretations than error covariances. Secondly, using $\{\*A_l\}_{l=1}^L$ to encode spatial dependency allows us to work with arbitrary neighbourhoods. In contrast, modelling dependent errors under conjugacy limits dependencies to decomposable graphs [@MurphyMVTSBCP]. Since not even a regular grid is decomposable, this is problematic for spatial data. Thirdly, modelling errors as contemporaneous is attractive for low-frequency data where the resolution of temporal effects is coarse, but the situation reverses for high-frequency data. Since the algorithm runs on-line, we expect $\{\*Y_t\}$ to be observed with high frequency.
For process $\{\*Y_t\}$ on $\mathcal{S}$ and $(L, N(\mathcal{S}), \Pi(\cdot))$, define the matrices $\{\widetilde{\*A}_l\}_{l=1}^L$ by imposing that $[\widetilde{\*A}_l]_{(s, s')} = 0 \Longleftrightarrow s' \notin N_i(s)$ for any $i \leq \Pi(l)$. Let $\widetilde{\*A}_l^{\neq 0}$ be the vector of non-zero entries in $\widetilde{\*A}_l$ and $\widetilde{\*c} = (\*\alpha, \text{vec}(\*B),\widetilde{\*A}_1^{\neq 0}, \widetilde{\*A}_2^{\neq 0}, \dots \widetilde{\*A}_L^{\neq 0})^T$. The [[[<span style="font-variant:small-caps;">SSBVAR</span>]{}]{}]{}model on $\{\*Y_t\}$ induced by $(L, N(\mathcal{S}), \Pi(\cdot))$ is obtained by combining – with
[rCl]{} |\^2 & \~& (0, \^2 V\_)\[BVAR\_eq\_3\_prime\]\
Y\_t & = & + B Z\_t + \_[l = 1]{}\^L \_l Y\_[t-l]{} + \_t. \[BVAR\_eq\_4\_prime\]\
Fig. \[SSBVAR\_graph\] illustrates this idea. Further sparsification is possible by modelling neighbourhoods jointly, i.e. $[\widetilde{\*A}_l]_{(s, s')} = a_i(s), \forall s' \in N_i(s)$, reducing the number of parameters to $S \cdot \sum_{l=1}^L\Pi(l)$. If one imposes $a_i(s) = a_i(s') = \dots = a_i$, this number drops to $\sum_{l=1}^L\Pi(l)$.
Building [[[<span style="font-variant:small-caps;">SSBVARs</span>]{}]{}]{}: choosing $L, N(\mathcal{S}), \Pi(\cdot)$
--------------------------------------------------------------------------------------------------------------------
For the choice of lag lengths $L$, part *(3)* of Thm. \[VAR\_Thm\] suggests $L \in \{L' \in \mathbb{N}: L(T_1) \leq L'\leq L(T_2) \}$ if one expects $T_1$ to $T_2$ observations between [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}. For any data stream $\{\*Y_t\}$ on a space $\mathbb{S}$, there are different ways of constructing neighbourhood structures $N(\mathcal{S})$. For example, when analysing pollutants in London’s air in section \[results\], $N(\mathcal{S})$ could be constructed from Euclidean or Road distances. By filling $\mathcal{M}$ with [[[<span style="font-variant:small-caps;">SSBVARs</span>]{}]{}]{}constructed using competing versions of $N(\mathcal{S})$, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}provides a way of dealing with such uncertainty about spatial relations. In fact, it can dynamically discern changing spatial relationships on $\mathbb{S}$. Lastly, $\Pi(\cdot)$ should usually be decreasing to reflect that measurements affect each other less when further apart.
Hyperparameter optimization {#hyperpar_opt}
===========================
Hyperparameter inference on $\*\nu_m$ can be addressed either by introducing an additional hierarchical layer [@HazardLearningBOCD] or using type-II [[[<span style="font-variant:small-caps;">ML</span>]{}]{}]{}. The latter is obtained by maximizing the model-specific evidence
[rCl]{}\[hyperpar\_opt\_Turner\] p(y\_[1:T]{}|\_m) & = & \_[t=1]{}\^T p(y\_t|\_m, y\_[1:(t-1)]{}).
Computation of the righthand side requires evaluating the gradients $\nabla_{\*\nu_m} p(\*y_{1:t}, r_t| \*\nu_m)$, which are obtained efficiently and recursively [@ABOCD]. @GPBOCD use $\*y_{1:T'}$ as a test set, and run [[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}$K$ times to find $\widehat{\*\nu}_m = \arg\max_{\*\nu_m}\left\{ p(\*y_{1:T'}|\*\nu_m) \right\}$. [ Most other on-line [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}approaches also require substantial recomputations for hyperparameter learning [e.g., @OnlineGP2]. ]{} In contrast, @CaronDoucet propose on-line gradient descent updates via
[rCl]{} \_[m,t+1]{} & = & \_[m,t]{} + \_t \_[\_[m,t]{}]{}p(y\_[t+1]{}| y\_[1:t]{}, \_[m\_[1:t]{}]{}). \[gradient\_descent\_caron\]
The latter is preferable for two reasons: Firstly, inference and type-II [[[<span style="font-variant:small-caps;">ML</span>]{}]{}]{}are executed simultaneously (rather than sequentially) and thus enable cold-starts of [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}. Secondly, neither the on-line nature nor the computational complexity of [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}is affected.
Computation & Complexity {#computational_complexity}
========================
While tracking $|\mathcal{M}|$ models, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}has linear time complexity. Step 1 in the pseudocode is the bottleneck, but looping over $\mathcal{M}$ can be parallelized: With $N$ threads, it executes in $\mathcal{O}\left(\lceil |\mathcal{M}|/N \rceil \cdot \max_{M \in \mathcal{M}} \text{CmpTime}(M)\right)$. Step 2 takes $\mathcal{O}(|R(t)||\mathcal{M}|)$, for $R(t)$ all run-lengths at time $t$.
Pruning the run-length distribution {#RLD}
-----------------------------------
In a naive implementation, all run-lengths are retained and $R(t) = \{1,2,\dots,t\}$. This implies execution time of order $\mathcal{O}(t)$ for processing $\*y_t$, but can be made time-constant by pruning: [ If one discards run-lengths whose posterior probability is $\leq 1/R_{\max}$ or only keeps the $R_{\max}$ most probable ones, $|R(t)| \leq R_{\max}$ [@BOCD]. A third way is [Stratified Rejection Control ([[[<span style="font-variant:small-caps;">SRC</span>]{}]{}]{})]{} [@FearnheadOnlineBCD], which @CaronDoucet and the current paper found to perform as well as the other approaches. In our experiments, we prune by keeping the $R_{\max}$ most probable model-specific run-lengths $p(r_t|m_t,\*y_{1:t})$ for each model. ]{}
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}updates
---------------------------------------------------------------------
The bottleneck when updating a [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}model in $\mathcal{M}$ is step I.B.1 in the pseudocode of [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}, when updating the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}estimate $\*c(r,t) = \*F(r,t)\*W(r,t)$ of the coefficient vector , where $\*F(r,t) = (\*X_{(t-r):t}'\*X_{(t-r):t} + \*V_{\widetilde{c}})^{-1}$ and $\*W(r,t) = \*X_{(t-r):t}'\*Y_{(t-r):t}$ for all $r \in R(t)$. Since $\*W(r,t) = \*W(r-1,t-1) + \*X_t'\*Y_t$, updates are $\mathcal{O}(kS)$. $\*F(r-1,t-1)$ can be updated to $\*F(r,t)$ using rank-$k$ updates to its [[[<span style="font-variant:small-caps;">QR</span>]{}]{}]{}-decomposition in $\mathcal{O}( k^3)$ or using Woodbury’s formula, in $\mathcal{O}(S^3)$, implying an overall complexity of $\mathcal{O}(|R(t)| \min\{k^3, S^3\})$ at time $t$.
Comparison with [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based approaches
---------------------------------------------------------------------------------------------
Define $k_{\max}$ as the largest number of regressors of any [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}model in $\mathcal{M}$. From the previous paragraphs, it follows that if all models in $\mathcal{M}$ are [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}, the overhead $C =\lceil N/|\mathcal{M}| \rceil \cdot \min\{k_{\max}^3, S^3\}$ is time-constant. Thus, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}runs in $\mathcal{O}(T R_{\max})$ on $T$ observations. In contrast, the models of @GPBOCD run in $\mathcal{O}(T R_{\max}^3)$. [ The experiments in section \[results\] confirm this: Using the software of @TurnerThesis on the Nile data, fitting one [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}model takes $42$ seconds compared to $12$ seconds when fitting three models in [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}, so a [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}model is $>10\times$ faster to process. Per inferred parameter, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}is $>60\times$ faster than [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}; and this factor is much larger for multivariate data (e.g., $>270$ for the $30$ Portfolio data). More details on the run-times can be found in the Appendix. ]{}
-0.2in
-0.2in
Experimental results {#results}
====================
We evaluate the performance of the algorithm in two parts. First, we compare it to benchmark performances of [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based models on real world data reported by @GPBOCD. This shows that as implied by Thm. \[VAR\_Thm\], [[[<span style="font-variant:small-caps;">VARs</span>]{}]{}]{}are excellent approximations for a large variety of data streams. Next, we showcase [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}’ novelty in the multivariate setting. All computations can be reproduced with code available on the first author’s website. We use uniform model priors $q$, a constant Hazard functions $H$ and gradient descent for hyperparameter optimization as in Section \[hyperpar\_opt\]. The lag lengths of models in $\mathcal{M}$ are chosen based on Thm. \[VAR\_Thm\] (3) and the rates of @HannanKavalieris for [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}and Bayesian Autoregressions ([[[<span style="font-variant:small-caps;">BARs</span>]{}]{}]{}), respectively.
Comparison with [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based approaches
---------------------------------------------------------------------------------------------
As in @GPBOCD, [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}will refer to the non-linear [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based [[[<span style="font-variant:small-caps;">AR</span>]{}]{}]{}model, [[[<span style="font-variant:small-caps;">GPTSCP</span>]{}]{}]{}to the time-deterministic model, and [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}to the non-stationary [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}allowing hyper-parameters to change at every [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}. @GPBOCD compute the mean squared error ([[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{}) as well as the negative log predictive likelihood ([[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{}) of the one-step-ahead predictions for three data sets: The water height of the Nile between $622-1284$ AD, the snowfall in Whistler (Canada) over a 37 year period and the $3$-dimensional time series ($x$-, $y$-coordinate and headangle) of a honey bee during a waggle dance sequence. In @TurnerThesis, all of the models except [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}were also compared on daily returns for $30$ industry portfolios from $1975-2008$. In Table \[benchmark\_table\], [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}is compared to these benchmarks for $\mathcal{M}$ consisting of [[[<span style="font-variant:small-caps;">BAR</span>]{}]{}]{}and [[[<span style="font-variant:small-caps;">SSBVAR</span>]{}]{}]{}models.
0.1in
---------------------------------------------------------------------------- --------------------------------------------------------------- --------------------------------------------------------------- --------------------------------------------------------------- ---------------------------------------------------------------
Method [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{}
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $0.553 $ $1.15$ $0.750$ $-0.604$
$(0.0962) $ $(0.0555) $ $(0.0315)$ $(0.0385)$
[[[<span style="font-variant:small-caps;">GPTSCP</span>]{}]{}]{} $0.583$ $1.19 $ $0.689$ $1.17 $
$(0.0989)$ $(0.0548) $ $(0.0294)$ $(0.0183) $
[[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{} $0.585$ $1.15$ $\mathbf{0.618}$ $\mathbf{-1.98}$
$(0.0988)$ $(0.0655)$ $(0.0242)$ $(0.0561)$
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} $\mathbf{0.550}$ $\mathbf{1.13}$ $0.681$ $0.923$
$(0.0948)$ $(0.0684)$ $(0.0245)$ $(0.0231)$
Method [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{} [[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{}
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $2.62$ $4.07$ $29.95$ $39.55$
$(0.195)$ $(0.150)$ $(0.50)$ $(0.22)$
[[[<span style="font-variant:small-caps;">GPTSCP</span>]{}]{}]{} $3.13$ $4.54$ $30.17$ $\mathbf{39.44}$
$(0.241)$ $(0.188)$ $(0.51)$ $(0.22)$
[[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{} $3.17$ $4.19$ – –
$(0.230)$ $(0.212)$ – –
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} $\mathbf{1.74}$ $\mathbf{3.57}$ $\mathbf{25.93}$ $48.32$
$(0.222)$ $(0.166) $ $(0.906)$ $(0.964)$
---------------------------------------------------------------------------- --------------------------------------------------------------- --------------------------------------------------------------- --------------------------------------------------------------- ---------------------------------------------------------------
: One-step-ahead predictive [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{}and [[[<span style="font-variant:small-caps;">NLL</span>]{}]{}]{}of [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}compared to [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based techniques, with $95\%$ error bars. All [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}results are taken from @GPBOCD and @TurnerThesis.[]{data-label="benchmark_table"}
### Designing $\mathcal{M}$
[ Both the Nile and the snowfall data are univariate, so $\mathcal{M}$ consists of [[[<span style="font-variant:small-caps;">BARs</span>]{}]{}]{}with varying lag lengths. For the $3$-dimensional bee data, $\mathcal{M}$ additionally contains unrestricted [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}. Lastly, [[[<span style="font-variant:small-caps;">SSBVARs</span>]{}]{}]{}are used on the $30$ Portfolio data. Two neighbourhood systems are constructed from distances in the spaces of pairwise contemporaneous correlations and autocorrelations [prior]{} to $1975$, a third using the *Standard Industrial Classification ([[[<span style="font-variant:small-caps;">SIC</span>]{}]{}]{})*, with $\Pi(\cdot)$ decreasing linearly. ]{}
### Findings
**Predictive performance and fit:** In terms of [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{}, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}clearly outperforms all [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models on multivariate data. Even on univariate data, the only exception to this is the snowfall data, where [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}does better. However, [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}requires grid search or Hamiltonian Monte Carlo sampling for hyperparameter optimization at each observation [@GPBOCD]. Overall, there are three main reasons why [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}performs better: Firstly, being able to change lag lengths between [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}seems more important to predictive performance than being able to model non-linear dynamics. Secondly, unlike the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models, we allow the time series to communicate via $\{\*A_l^L\}$. Thirdly, the hyperparameters of the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}have a strong influence on inference. In particular, the noise variance $\sigma$ is treated as a hyperparameter and optimized via type-II [[[<span style="font-variant:small-caps;">ML</span>]{}]{}]{}. Except for the [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}, this is only done during a training period. Thus, the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models cannot adapt to the observations after training, leading to overconfident predictive distributions that are too narrow [see @TurnerThesis p. 172]. This in turn leads them to be more sensitive to outliers, and to mislabel them as [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}. In contrast, – models $\sigma$ as part of the inferential Bayesian hierarchy, and hyperparameter optimization is instead applied at one level higher. Consequently, our predictive distributions are wider, and the algorithm is less confident about the next observations, making it more robust to outliers. This is also responsible for the overall smaller standard errors of the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models in Table \[benchmark\_table\], since the [[[<span style="font-variant:small-caps;">GPs</span>]{}]{}]{}interpret outliers as [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}and immediately adapt to short-term highs or lows. **[[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}Detection:** A good demonstration of this finding is the Nile data set, where the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}segmentation finds a single [[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}, corresponding to the installation of the nilometer around $715$ CE, see Fig \[Nile\]. In contrast, @GPBOCD report $18$ additional [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}corresponding to outliers. The same phenomenon is also reflected in the run-length distribution ([[[<span style="font-variant:small-caps;">RLD</span>]{}]{}]{}): While the probability mass in Figs. \[30PF\_comparison\], \[30PF\_FinCrisis\] and \[Nile\] are spread across the retained run-lengths, the [[[<span style="font-variant:small-caps;">RLD</span>]{}]{}]{}reported in @GPBOCD is more concentrated and even degenerate for the $30$ Portfolio data set. On the other hand, such enhanced sensitivity to change can be advantageous. For instance, in the bee waggle dance, the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-based techniques are better at identifying the true [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}. The reason is twofold: Firstly, the variance for the bee waggle data is homogeneous across time, so treating it as fixed helps inference. Secondly, the [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}in this data set are subtle, so having narrower predictive distributions is of great help in detecting them. However, it adversely affects performance when changes in the error variance are essential, as for financial data: In particular, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}finds the ground truths labelled in @GPBOCD, and discovers even more, see Fig. \[30PF\_comparison\]. This is especially apparent in times of market turmoil where changes in the variance of returns are significant. We show this using the example of the subprime mortgage financial crisis: While the [[[<span style="font-variant:small-caps;">RLD</span>]{}]{}]{}of @GPBOCD identified only $2$ [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}with ground truth and a third unlabelled one during the height of the crisis, [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}detects a large number of [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}corresponding to ground truths, see Fig. \[30PF\_FinCrisis\].
Lastly, we note that segmentations obtained off-line for both the bee waggle dance and the $30$ Portfolios are reported in @MurphyMVTSBCP. Compared to the on-line segmentations produced by [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}, these are closer to the truth for the bee waggle data, but not for the $30$ Portfolio data set.
-0.2in
**Model selection:** In most of the experiments where abrupt changes model the non-stationarity well, the model posterior is fairly concentrated and periods of model uncertainty are short. This is different when changes are slower, see Fig. \[Temperatures\]. The implicit model complexity penalization Bayesian model selection performs provides [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}with an Occam’s Razor mechanism: Simple models are typically favoured until evidence for more complex dynamics accumulates. For the bee waggle and the $30$ Portfolio data set, [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}are preferred to [[[<span style="font-variant:small-caps;">BARs</span>]{}]{}]{}. For the $30$ Portfolio data, the [[[<span style="font-variant:small-caps;">MAP</span>]{}]{}]{}segmentation only selects [[[<span style="font-variant:small-caps;">SSBVARs</span>]{}]{}]{}with neighbourhoods constructed from contemporaneous correlation and autocorrelations. Neighbourhoods using [[[<span style="font-variant:small-caps;">SIC</span>]{}]{}]{}codes are not selected, reflecting that this classification from $1937$ is out of date.
0.05in
-0.2in
Performance on spatio-temporal data
-----------------------------------
**European Temperature:** Monthly temperature averages $01/01/1880-01/01/2010$ for the $21$ longest-running stations across Europe are taken from [http://www.ecad.eu/]{}. We adjust for seasonality by subtracting monthly averages for each station. Station longitudes and latitudes are available, so $N(\mathcal{S})$ is based on concentric rings around the stations using Euclidean distances. Two different decay functions $\Pi(\cdot), \Pi^{+}(\cdot)$ are used, with $\Pi^{+}(\cdot)$ using larger neighbourhoods and slower decaying. Temperature changes are poorly modeled by [[[<span style="font-variant:small-caps;">CPs</span>]{}]{}]{}and more likely to undergo slow transitions. Fig. \[Temperatures\] shows the way in which the model posterior captures such longer periods of change in dynamics. The values on the bottom panel are calculated by considering $\widehat{m}_t = \arg\max_{m_t \in \mathcal{M}}p(m_t|\*y_{1:t})$ as $|\mathcal{M}|$-dimensional multinomial random variable. Its Standardized Generalized Variance ([[[<span style="font-variant:small-caps;">SGV</span>]{}]{}]{}) [@GV; @SGV] is calculated as $|\mathcal{M}|$-th root of the covariance matrix determinant. We plot the log of the [[[<span style="font-variant:small-caps;">SGV</span>]{}]{}]{}computed using the model posteriors for the last $8$ years. This provides an informative summary of the model posterior dispersion. **Air Pollution:** Finally, we analyze Nitrogen Oxide ([[[<span style="font-variant:small-caps;">NOX</span>]{}]{}]{}) observed at $29$ locations across London $17/08/2002 - 17/08/2003$. The quarterhourly measurements are averaged over $24$ hours. Weekly seasonality is accounted for by subtracting week-day averages for each station. $\mathcal{M}$ is populated with [[[<span style="font-variant:small-caps;">SSBVAR</span>]{}]{}]{}models whose neighbourhoods are constructed from both road- and Euclidean distances. As $17/02/2003$ marks the introduction of London’s first ever congestion charge, we find structural changes in the dynamics around that date. A model with shorter lag length but identical neighbourhood structure is preferred after the congestion charge. In Fig. \[AirPollution\], Bayes Factors ([[[<span style="font-variant:small-caps;">BFs</span>]{}]{}]{}) capture the shift: @BF classify logs of [[[<span style="font-variant:small-caps;">BFs</span>]{}]{}]{}as very strong evidence if their absolute value exceeds $5$.
Conclusion {#Discussion}
==========
We have extended Bayesian On-line Changepoint Detection ([[[<span style="font-variant:small-caps;">BOCPD</span>]{}]{}]{}) to multiple models by generalizing @FearnheadOnlineBCD and @BOCD, arriving at [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}. For inference in multivariate data streams, we propose [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}with closed form distributions that have strong theoretical guarantees summarized in Thm. \[VAR\_Thm\]. We sparsify [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}based on neighbourhood systems, thus making [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}especially amenable to spatio-temporal inference. To demonstrate the power of the resulting framework, we apply it to multivariate real world data, outperforming the state of the art. In future work, we would like to add and remove models from $\mathcal{M}$ on-line. This could lower the computational cost for the case where $|\mathcal{M}|$ is significantly larger than the number of threads.
0.1in
-0.2in
Acknowledgements {#acknowledgements .unnumbered}
================
We want to thank N. Karampatziakis for his help with making the method computationally more efficient. JK is funded by [[[<span style="font-variant:small-caps;">EPSRC</span>]{}]{}]{}grant EP/L016710/1. Further, this work was supported by The Alan Turing Institute for Data Science and AI under [[[<span style="font-variant:small-caps;">EPSRC</span>]{}]{}]{}grant EP/N510129/1 and the Lloyds Register Foundation programme on Data Centric Engineering.
Condition of Theorem 1
======================
*Denoting the spectrum of a matrix $\*B$ (i.e., the set of its eigenvalues) by $\sigma(\*B)$, the following condition is a restatement of the relevant part in condition **A** of Meyer & Kreiss (2015)*:
Let $\*W$ be the spectral density matrix of the purely non-deterministic stochastic process $\{\*Y_t\}_{t=1}^{\infty}$ satisfying the conditions of Theorem 1. We assume that the spectral density matrix is bounded, i.e. there is a constant $c>0$ so that
[rCl]{} ( (W()) ) & & c
for all frequencies $\lambda \in (-\pi, \pi]$, i.e. the eigenvalues of the spectral density matrix are uniformly bounded away from zero.
Empirical evaluation of computation time
========================================
For this comparison, we use the original code of Turner (2012) for the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models. As the [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{}is smallest for [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}for all data sets except for the snowfall data, we compare [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}against the arguably best [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}[[[<span style="font-variant:small-caps;">CP</span>]{}]{}]{}model. We note that while [[[<span style="font-variant:small-caps;">NSGP</span>]{}]{}]{}performs better on the snowfall data than [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}, its requirement to do Hamiltonian Monte Carlo sampling will make it significantly slower. We also note that [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}models inside [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}outperformed the [[[<span style="font-variant:small-caps;">MSE</span>]{}]{}]{}of the [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}model for all data sets considered. All computations were performed on a 3.1 [[[<span style="font-variant:small-caps;">GHz</span>]{}]{}]{}Intel [[i7]{}]{}[ ]{}with 16[[[<span style="font-variant:small-caps;">GB</span>]{}]{}]{}[[[<span style="font-variant:small-caps;">RAM</span>]{}]{}]{}.
Table \[benchmark\_table\] summarizes the results. It is clear that [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}outperforms [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}computationally: e.g., the computation time per parameter is between $60$ (Nile data) and $585$ (Bee data) times faster for [[[<span style="font-variant:small-caps;">BOCPDMS</span>]{}]{}]{}with [[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{}models. Computation times are faster per model, too. The only exception to this is the $30$ Portfolio data set, where the deployed [[[<span style="font-variant:small-caps;">SSBVAR</span>]{}]{}]{}models are orders of magnitude more parameter-rich than the [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{}-model. Related to this, we also note that comparing the computation time per parameter makes sense for two reasons: Firstly, [[[<span style="font-variant:small-caps;">BVARs</span>]{}]{}]{}model the $d$ time series jointly, thus requiring $d^2$ parameters in the posterior covariance matrix of $\*y_t$. In contrast, the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models ignore any dependence between the series, resulting in $d$ parameters of the (diagonal) posterior covariance matrix for $\*y$. Secondly, the parameters of the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}’s kernel arguably making its parameter space $\Theta$ infinite-dimensional are not actually learnt on-line at all. Instead, they are optimized for a training period of $T'$ observations and then fixed, see section 4 in the main paper. Hence, the parameter space the [[[<span style="font-variant:small-caps;">GP</span>]{}]{}]{}-models can learn in is finite-dimensional.
0.1in
---------------------------------------------------------------------------- ---------------------- -----------------
Time/$|\mathcal{M}|$ Time/$|\Theta|$
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $42.2$ $21.0$
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} $\mathbf{4.03}$ $\mathbf{0.35}$
Time/$|\mathcal{M}|$ Time/$|\Theta|$
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $284$ $142$
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} $\mathbf{157}$ $\mathbf{4.25}$
Time/$|\mathcal{M}|$ Time/$|\Theta|$
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $164$ $23.4$
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} $\mathbf{97.3}$ $\mathbf{0.04}$
Time/$|\mathcal{M}|$ Time/$|\Theta|$
\[1.5pt\] [[[<span style="font-variant:small-caps;">ARGPCP</span>]{}]{}]{} $\mathbf{12077}$ $403$
[[[<span style="font-variant:small-caps;">BVAR</span>]{}]{}]{} ${34183}$ $\mathbf{1.48}$
---------------------------------------------------------------------------- ---------------------- -----------------
: Computation time in seconds per model and per parameter in the space $\Theta = \cup_{m \in \mathcal{M}}\Theta_m$[]{data-label="benchmark_table"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we relate the deformation theory of Ginzburg Calabi-Yau algebras to negative cyclic homology. We do this by exhibiting a DG-Lie algebra that controls this deformation theory and whose homology is negative cyclic homology. We show that the bracket induced on negative cyclic homology coincides with Menichi’s string topology bracket. We show in addition that the obstructions against deforming Calabi-Yau algebras are annihilated by the map to periodic cyclic homology.
In the commutative we show that our DG-Lie algebra is homotopy equivalent to $(T^{{\operatorname{poly}}}[[u]],-u{\operatorname{div}})$.
address: 'Departement WNI, Universiteit Hasselt, 3590 Diepenbeek, Belgium.'
author:
- Louis de Thanhoffer de Völcsey
- Michel Van den Bergh
title: 'Calabi-Yau Deformations and Negative Cyclic Homology'
---
Introduction
============
Throughout $k$ is a field of characteristic zero. In this paper we discuss the deformation theory of Calabi-Yau $k$-algebras in the sense of Ginzburg [@Gi]. Recall that a $k$-algebra $A$ is $d$-Calabi-Yau if it is perfect as $A^e$-module and there is an isomorphism[^1] in $D(A^e)$ $$\label{ref-1.1-0}
\eta:{\operatorname{RHom}}_{A^e}(A,A^e){\longrightarrow}\Sigma^{-d}A$$ In the rest of this introduction we fix a $d$-CY algebra $A$. Here and throughout the paper we take the point of view that $\eta$ is part of the structure of $A$.
The definition of a $d$-Calabi-Yau algebra can be “relativized” without any difficulty. Hence there is an associated deformation theory. Our first result in this paper is the construction of a DG-Lie algebra which controls this deformation theory.
To be more precise: let ${\operatorname{Nilp}}$ be the category of commutative, finite dimensional, local $k$-algebras $(R,m)$ such that $R/m=k$. For $R\in
{\operatorname{Nilp}}$ let ${\operatorname{Def}}_{A,\eta}(R)$ be the category of $R$-algebras $B$ which are $d$-Calabi-Yau (with respect to $R$) and which are in addition equipped with an isomorphism $B\otimes_R k\cong A$ respecting $\eta$. We view ${\operatorname{Def}}_{A,\eta}$ as a pseudo-functor from ${\operatorname{Nilp}}$ to the category of groupoids ${\operatorname{Gd}}$.
For a nilpotent DG-Lie algebra $\mathfrak{h}$ let ${\operatorname{MC}}(\frak{h})$ be the groupoid of solutions to the Maurer-Cartan equation in $\mathfrak{h}$ (see §\[ref-7-32\]). For an arbitrary DG-Lie algebra $\mathfrak{g}$ we have an associated “deformation functor” $${\operatorname{\mathcal{MC}}}(\mathfrak{g}):{\operatorname{Nilp}}{\longrightarrow}{\operatorname{Gd}}:(R,m)\mapsto {\operatorname{MC}}(\mathfrak{g}\otimes_k m)$$ In this paper we introduce a DG-Lie algebra $\mathfrak{D}^\bullet(A,\eta)$ (see §\[ref-8-36\]) which controls the deformation theory of $(A,\eta).$[^2]
(a combination of Prop. \[ref-6.2-30\] and Thm \[ref-8.1-37\] below) There is a morphism of pseudo-functors $\pi:{\operatorname{\mathcal{MC}}}(\mathfrak{D}^\bullet(A,\eta)) {\longrightarrow}{\operatorname{Def}}_{A,\eta}$ which when evaluated on an arbitrary $R\in {\operatorname{Nilp}}$ is essentially surjective on objects and surjective on morphisms.
We obtain in particular for $R\in {\operatorname{Nilp}}$ a bijection between ${\operatorname{MC}}(\mathfrak{D}^\bullet(A)\otimes_k m)/{\cong}$ and ${\operatorname{Def}}_{A,\eta}(R)/{\cong}$. In this sense the deformation theory of $A$ is controlled by the DG-Lie algebra $\mathfrak{D}^\bullet(A,\eta)$.
$\mathfrak{D}^\bullet(A,\eta)$ is constructed as a twisted semi-direct product of the Hochschild cochain complex with the negative cyclic chain complex of $A$. So the construction is similar in spirit to [@TerTrad] which treats finite dimensional $A_\infty$-algebras with a non-degenerate inner product. However our algebras are not finite dimensional and they do not carry an inner product.
The construction of $\mathfrak{D}^\bullet(A,\eta)$ yields a morphism $$\phi:\mathfrak{D}^\bullet(A,\eta) {\longrightarrow}\bar{\mathfrak{C}}^\bullet(A)$$ where $\bar{\mathfrak{C}}^\bullet(A)$ is the (shifted) Hochschild cochain complex of $A$. As is well-known, $\bar{\mathfrak{C}}^\bullet(A)$ controls the deformation theory of $A$ as algebra. The morphism $\phi$ corresponds to “forgetting $\eta$” as is explained in §\[ref-9-47\].
The next result is the construction of an explicit quasi-isomorphism of complexes $$\label{ref-1.2-1}
\mathfrak{D}^\bullet(A,\eta)\xrightarrow{\cong} \Sigma^{-d+1}{\operatorname{CC}}^-_\bullet(A)$$ between $\mathfrak{D}^\bullet(A,\eta)$ and the shifted negative cyclic complex ${\operatorname{CC}}^-_\bullet(A)$. As a result we obtain the following information about the deformation theory of Calabi-Yau algebras.
1. The tangent space to the deformation space of a $d$-Calabi-Yau algebra is ${\operatorname{HC}}^-_{d-2}(A)$.
2. The obstructions against deforming a $d$-Calabi-Yau algebra are in $\ker({\operatorname{HC}}^-_{d-3}(A)\rightarrow {\operatorname{HC}}^{\text{per}}_{d-3}(A))$.
The first statement is a formal consequence of . The second statement is Theorem \[ref-12.1-70\] below. It follows in particular that if ${\operatorname{HC}}_{d-3}^-(A)\rightarrow
{\operatorname{HC}}^{\text{per}}_{d-3}(A)$ is injective then the deformation theory of $A$ as Calabi-Yau algebra is unobstructed. This happens for example if $d\le 3$ (see Corollary \[ref-12.5-74\] and Lemma \[ref-12.6-75\] below).
Our next result is the description of the Lie bracket on ${\operatorname{HC}}_\bullet(A)$ induced by :
(Theorem \[ref-10.2-51\] below) The Lie bracket on negative cyclic homology induced by is the “string topology” Lie bracket introduced in [@Menichi] by Menichi.
Let us now specialize to the case where $A$ is commutative. Let $T^{{\operatorname{poly}},\bullet}(A)$ be the Lie algebra of poly-vector fields on $A$. Then $\eta$ in may be interpreted as a volume form (see §\[ref-11-59\] below). Let ${\operatorname{div}}$ be the divergence operator on $T^{{\operatorname{poly}},\bullet}(A)$ associated to $\eta$. Using Willwacher’s “formality for cyclic chains” [@Willwacher1] (see also [@Dolgushev; @Shoikhet2; @Tsygan]) we show that there is an isomorphism $$(T^{{\operatorname{poly}},\bullet}[[u]],-u{\operatorname{div}})\xrightarrow{\cong} \mathfrak{D}^\bullet(A,\eta)
\qquad \text{($|u|=2$)}$$ in the homotopy category of DG-Lie algebras which fits in a commutative diagram $$\label{ref-1.3-2}
\xymatrix{
(T^{{\operatorname{poly}},\bullet}(A)[[u]],-u{\operatorname{div}})\ar[r]^-\cong\ar[d]_{u\mapsto 0}& \mathfrak{D}^\bullet(A)\ar[d]^\phi\\
T^{{\operatorname{poly}},\bullet}(A)\ar[r]_\cong&\bar{\mathfrak{C}}^\bullet(A)
}$$ The lower arrow is a globalized version of Kontsevich’s formality quasi-isomorphism [@Ko3]. This diagram gives a conceptual explanation of Dolgushev’s result [@Dolgushev2] that the Kontsevich $\ast$-product associated to a divergence free Poisson bracket is Calabi-Yau.[^3]
Acknowledgement
===============
We like to thank Damien Calaque and Boris Tsygan for help with references.
Notation and conventions
========================
All rings and ring homomorphisms are unital. We mix homological and cohomological indices, using the convention $X_i=X^{-i}$.
Preliminaries on the Hochschild and cyclic complexes {#ref-4-3}
====================================================
In this section remind the reader about the basic operations on the Hochschild and cyclic complexes. The reason for putting this section first is that it also allows us to introduce some notation. Readers vaguely familiar with the material may safely skip to the next section.
Notation
--------
Let $R$ be a commutative ring and assume that $B$ is an $R$-algebra. Let ${\operatorname{C}}_\bullet(B)$ and ${\operatorname{C}}^\bullet(B)$ denote the usual relative Hochschild chain and cochain complexes of $B/R$. Thus $$\begin{aligned}
{\operatorname{C}}^\bullet(B)&= \bigoplus_{n}{\operatorname{Hom}}_R(\Sigma B^{\otimes n},B)\\
{\operatorname{C}}_\bullet(B)&= \bigoplus_n B\otimes(\Sigma B)^{\otimes n}\end{aligned}$$ where here and below, *all unadorned tensor products are over $R$*. We also use $$\begin{aligned}
\mathfrak{C}^\bullet(B)&= \Sigma{\operatorname{C}}^\bullet(B)\label{cohom}\\
&=\bigoplus_{n}{\operatorname{Hom}}_R(\Sigma B^{\otimes n},\Sigma B)\end{aligned}$$ We will also consider the *normalized versions* of these objects. $$\begin{aligned}
\bar{{\operatorname{C}}}^\bullet(B)&= \bigoplus_{n}{\operatorname{Hom}}_R(\Sigma (B/R)^{\otimes n},B)\\
\bar{{\operatorname{C}}}_\bullet(B)&= \bigoplus_n B\otimes \Sigma (B/R)^{\otimes n}\end{aligned}$$ and a similar definition for $\bar{\mathfrak{C}}^\bullet(B)$. It is well-known that the obvious maps $\bar{{\operatorname{C}}}^\bullet(B){\longrightarrow}{\operatorname{C}}^\bullet(B)$ and ${\operatorname{C}}_\bullet(B){\longrightarrow}\bar{{\operatorname{C}}}_\bullet(B)$ are quasi-isomorphisms [@weibel Thm 8.3.8, Lemma 8.3.7].
If $x\in \mathfrak{C}^n(B)$ then we write $|x|=n-1$. Thus $|x|$ refers to the cohomological degree of $x$.
The Hochschild cochain complex {#ref-4.2-4}
------------------------------
The standard algebraic structures on the Hochschild cochain complex can all be deduced from its structure as a *brace algebra* [@VG]. Recall that the braces are maps $${\ensuremath{\mathfrak{C}}}^{\bullet}(B) {\otimes}\ldots {\otimes}{\ensuremath{\mathfrak{C}}}^{\bullet}(B) {\longrightarrow}{\ensuremath{\mathfrak{C}}}^{\bullet}(B):x\otimes x_1\otimes\cdots\otimes
x_m\mapsto x\{x_1,\ldots,x_m\}$$ defined by $$\begin{aligned}
&x\{x_1, \ldots, x_m\}(b_1, \ldots b_n)=\\
& \sum_{0 \le i_1 \ldots \le i_m \le n} (-1)^\epsilon x(b_1, \ldots b_{i_1},x_1(b_{i_1+1}, \ldots b_{i_1+\vert x_1 \vert+1}), \ldots, b_{i_m}, x_m(b_{i_m+1}, \ldots ,b_{i_m+\vert x_m \vert+1}), \ldots b_n)\end{aligned}$$ where $\epsilon= \sum_1^m\vert x_k\vert i_k$ The corresponding *Gerstenhaber Lie bracket* on $\mathfrak{C}^\bullet(B)$ is $$[x,y]= x\{y\} -(-1)^{\vert x \vert\vert y \vert}y \{x\}$$ Let $\mu\in \mathfrak{C}^1(B)={\operatorname{Hom}}(\Sigma B\otimes \Sigma B,\Sigma B)$ denote the *“inverse” multiplication* $\mu(b_1,b_2)=-b_1b_2$. Then $[\mu,\mu]=0$ and hence $$\label{ref-4.1-5}
dx=[\mu,x]$$ defines a differential of degree one on $\mathfrak{C}^\bullet(B)$.
The *cupproduct* on $\mathfrak{C}^\bullet(B)$ is defined by $$x\cup y=(-1)^{|x|}\mu\{x,y\}$$ This is an associative product of degree one on $\mathfrak{C}^\bullet(B)$, or equivalently an associative product of degree zero on ${\operatorname{C}}^\bullet(B)$. One has [@VG]
1. $({\ensuremath{\mathfrak{C}}}^{\bullet}(B),d, [\ ,\ ])$ is a DG-Lie algebra
2. $({\operatorname{C}}^{\bullet}(B), d,\cup)$ is a DG-algebra, commutative up to homotopy
3. More generally: (${\operatorname{C}}^{\bullet}(B),d,[\ ,\ ],\cup$) is a so-called DG-“Gerstenhaber algebra” up to homotopy
The Hochschild chain complex {#ref-4.3-6}
----------------------------
When combining the Hochschild cochain complex with the Hochschild chain complex things becomes much more intricate [@CR; @TT]. We will content ourselves by giving formulas for the basic operations and stating some relations between them. We refer to [@TT] for more details.
The first basic operation is the *contraction*. $$i_x(b_0{\otimes}\ldots {\otimes}b_n):=b_0x(b_1, \ldots b_d) {\otimes}b_{d+1}{\otimes}\ldots {\otimes}b_n$$ for $x\in \mathfrak{C}^\bullet(B)$ and $b_0{\otimes}\ldots {\otimes}b_n\in
{\operatorname{C}}_\bullet(B)$. This is an action of $\mathfrak{C}^\bullet(B)$ on ${\operatorname{C}}_\bullet(B)$ satisfying $|i_x|=|x|+1$ and $$\label{ref-4.2-7}
i_xi_y=(-1)^{(|x|+1)(|y|+1)}i_{y\cup x}$$ The contraction is often written as a *capproduct*: $i_x(-)=x\cap -$.
The second basic operation is the *Lie derivative* $$\begin{aligned}
L_x(b_0{\otimes}\ldots {\otimes}b_n):&=\sum_{i=0}^{n-\vert x\vert-1} (-1)^{\vert x\vert i}b_0{\otimes}\ldots {\otimes}b_i {\otimes}x(b_{i+1}, \ldots , b_{i+\vert x\vert+1 }){\otimes}\ldots {\otimes}b_n\\
&+\sum_{i=n-\vert x\vert}^n (-1)^{n(i+1)+|x|}x(b_{i+1}, \ldots b_n, b_0, \ldots , b_{\vert x\vert -n+i}){\otimes}\ldots {\otimes}b_i\end{aligned}$$ The Lie derivative defines a graded Lie action of $\mathfrak{C}^\bullet(B)$ on ${\operatorname{C}}_\bullet(B)$. Explicitly: $|L_x|=|x|$ and $$\label{ref-4.3-8}
[L_x,L_y]=L_{[x,y]}$$ The *Hochschild differential* on $\mathfrak{C}_\bullet(B)$ is defined as $\mathsf{b}=L_\mu$. From and one obtains $$\label{ref-4.4-9}
[\mathsf{b},L_x]=L_{dx}$$ Hence $({\operatorname{C}}_\bullet(B),\mathsf{b})$ is a DG-Lie module over $\mathfrak{C}^\bullet(B)$. One also has compatibility with the contraction: $$\label{ref-4.5-10}
[\mathsf{b},i_x]+i_{dx}=0$$
The last basic operation we need is the *Connes differential*. $$\mathsf{B}:{\operatorname{C}}_\bullet(B){\longrightarrow}{\operatorname{C}}_\bullet(B)$$ with formula $$\begin{aligned}
\mathsf{B}(b_0{\otimes}\ldots {\otimes}b_n)&= \sum_{i=0}^n (-1)^{ni} 1{\otimes}b_i {\otimes}\ldots {\otimes}b_n {\otimes}b_0 {\otimes}\ldots {\otimes}b_{i-1}\\
\\&\quad +
\sum_0^n (-1)^{n(i+1)} b_{i-1}{\otimes}1{\otimes}b_i {\otimes}\ldots {\otimes}b_n {\otimes}b_0 {\otimes}\ldots {\otimes}b_{i-2}\end{aligned}$$ It is well-known that $|\mathsf{B}|=-1$, $\mathsf{b}\mathsf{B}+\mathsf{B}\mathsf{b}=0$, $\mathsf{B}^2=0$.
Some of the following identies hold only for normalized chains/cochains. Note that if $x\in \bar{\mathfrak{C}}^\bullet(B)$ then $i_x,L_x$ are well-defined operations on $\bar{{\operatorname{C}}}_\bullet(B)$.
Assume $x\in \bar{\mathfrak{C}}^\bullet(B)$. Then on ${\operatorname{C}}_\bullet(B)$ we have $$\label{ref-4.6-11}
[\mathsf{B},L_x]=0$$
The formula does not hold for unnormalized cochains.
The negative cyclic complex
---------------------------
Let $u$ be a variable of degree two and put $$\overline{{\operatorname{CC}}}^{\,-}_\bullet(B)=\bar{{\operatorname{C}}}_\bullet(B)[[u]]$$ Equipped with the *cyclic differential* $\mathsf{b}+u\mathsf{B}$, this is the *normalized negative cyclic complex*. In the sequel operations on $\bar{{{\operatorname{C}}}}_\bullet(B)$ will be (tacitly) extended to $\overline{{\operatorname{CC}}}^{\,-}_\bullet(B)$ by making them $u$-linear. This applies in particular to $i_x$ and $L_x$. Combining we obtain $$\label{ref-4.7-12}
[\mathsf{b}+u\mathsf{B},L_x]=L_{dx}$$ The compatibility of $i_x$ with the cyclic differential is more subtle. In [@TT] (see also [@Getzler1]) Tamarkin and Tsygan define for $x\in \bar{\mathfrak{C}}^\bullet(B)$ a graded endomorphism $S_x$ of $\bar{{\operatorname{C}}}_\bullet(B)$ (depending linearly on $x$) such that $|S_x|=|x|-1$ and such that the following identity holds $$\label{ref-4.8-13}
[\mathsf{b}+u\mathsf{B}, i_x+uS_x]+i_{dx}+uS_{dx}=uL_{x}$$ on $\overline{{\operatorname{CC}}}^{\,-}_\bullet(B)$. This identity will be important for us in the sequel. Note that it implies .
The following is a special case of [@TT Prop. 3.3.4].
\[ref-4.2-14\] Let $x,y\in\bar{\mathfrak{C}}^\bullet(B)$ be such that $dx=dy=0$. Then $[L_x,i_y+uS_y]$ is homotopic to $(-1)^{|x|}(i_{[x,y]}+uS_{[x,y]})$.
A comment on base change {#ref-4.5-15}
------------------------
If $A$ is a $k$-algebra and $R$ is a commutative $k$-algebra then for $B=A\otimes_k R$ it is clear that ${\operatorname{C}}_{R,\bullet}(B)\cong{\operatorname{C}}_\bullet(A)\otimes_k R$ (where contrary to our usual conventions we have now made the base ring explicit in the notation). Since the negative cyclic complex involves a product this is not true for ${\operatorname{CC}}^-_{R,\bullet}(B)$. However it is true if $R$ is finite dimensional. Similarly in that case we have ${\operatorname{C}}^\bullet_R(B)\cong {\operatorname{C}}^\bullet(A)\otimes_k R$. In the sequel we will not mention these base change isomorphisms explicitly.
Some comments on signs {#ref-4.6-16}
----------------------
In the previous sections the operations $i_x,L_x,S_x,\mathsf{b},\mathsf{B}$ of degree $|x|+1$, $|x|$, $|x|-1$, $1$, $-1$ were defined as acting on $\bar{{\operatorname{C}}}_\bullet(B)$. We define corresponding operations on shifts $\Sigma^r
\bar{{\operatorname{C}}}_\bullet(B)$ in the usual way. $$\begin{aligned}
i_x(s^r b)&=(-1)^{r(|x|+1)}s^r i_x(b)\\
L_x(s^r b)&=(-1)^{r|x|}s^r L_x(b)\\
S_x(s^r b)&=(-1)^{r(|x|-1)}s^r S_x(b)\\
\mathsf{b}(s^rb)&=(-1)^r s^r \mathsf{b}(b)\\
\mathsf{B}(s^rb)&=(-1)^r s^r \mathsf{b}(B)\end{aligned}$$ where $s$ is the *degree change operator* $|sb|=|b|-1$.
The relations between $i_x,L_x,S_x,\mathsf{b},\mathsf{B}$ stated in §\[ref-4.3-6\],§\[ref-4.3-6\] carry over to all shifts $\Sigma^r \bar{{\operatorname{C}}}_\bullet(B)$ without any sign changes, since all terms in the identities (necessarily) have the same degree.
Preliminaries on Calabi-Yau algebras
====================================
In this section we extend Ginzburg’s definition of Calabi-Yau algebras to the relative case.
Let $R$ be a commutative ring. For an $R$-algebra $B$ we put $B^e=B\otimes_R B^\circ$. We use without further comment the standard equivalences between the categories of left $B^e$-modules, right $B^e$-modules and $B$-bimodules which are $R$-central.
A $B^e$-module is called *perfect* if it has a finite resolution by finitely generated projective $B^e$-modules. If $B$ is $R$-flat and $B$ is a perfect $B^e$-module then we say that $B$ is *homologically smooth* over $R$. The implicit flatness hypothesis ensures that $B^e=B\otimes_R B^\circ$ is the correct definition from a derived point of view. We could have avoided this hypothesis by first replacing $B$ by an $R$-flat DG-resolution but for simplicity we have chosen not to do this.
(Ginzburg [@Gi]) \[ref-5.1-17\] An $R$-Calabi-Yau algebra of dimension $d$ is a pair $(B,\eta)$ where
1. $B$ is an $R$-algebra $B$ which is homologically smooth over $R$;
2. $\eta$ is an isomorphism ${\operatorname{RHom}}_{B^e}(B,B^e) {\longrightarrow}\Sigma^{-d} B$ in $D(B^e)$.
Note that the amount of freedom for $\eta$ is quite limited. If $(B,\eta)$, $(B,\eta')$ are Calabi-Yau then there exists $z\in Z(B)$ such that $\eta'=z\eta$ (see [@davison]).
Recall that if $M$ is a complex of $B^e$-module then its *Hochschild homology and cohomology* are respectively defined as $$\begin{aligned}
{\operatorname{HH}}_i(B,M)&=H^{-i}(M\otimes^L B)\\
{\operatorname{HH}}^i(B,M)&=H^i(\operatorname{RHom}_{B^e}(B,M))\end{aligned}$$ As usual we write ${\operatorname{HH}}_i(B)={\operatorname{HH}}_i(B,B)$ and similarly ${\operatorname{HH}}^i(B)={\operatorname{HH}}^i(B,B)$. One has $${\operatorname{HH}}_i(B)=H^{-i}({\operatorname{C}}_{\bullet}(B))$$ and if $B$ is a projective $R$-module then $${\operatorname{HH}}^i(B)=H^{i}({\operatorname{C}}^{\bullet}(B))$$ The operations $[\ ,\ ],\cup,\cap,L,\mathsf{B}$ introduced in §\[ref-4-3\] descend to homology and make the pair $({\operatorname{HH}}^\bullet(B),{\operatorname{HH}}_\bullet(B))$ into a so-called *calculus* structure [@TT]. Up to suitable, and for us irrelevant, signs $\cup$ is the Yoneda products on ${\operatorname{HH}}^\bullet(B)={\operatorname{Ext}}^\bullet_{B^e}(B,B)$ and $\cap$ is the action of ${\operatorname{HH}}^\bullet(B)$ on ${\operatorname{HH}}_\bullet(B)=H^{-\bullet}(B\otimes^L_{B^e} B)$ through its action on the second factor (see e.g. [@CRVdB Prop 11.1, 12.1]).
\[ref-5.3-18\] Let $B$ be a homologically smooth algebra. Then for $M$ a perfect $B^e$-module there is a canonical isomorphism $$\label{ref-5.1-19}
{\operatorname{HH}}_i(B,M) \cong {\operatorname{Hom}}_{B^e}(\Sigma^i {\operatorname{RHom}}_{B^e}(M,B^e),B)$$ in $D(R)$.
Since $M$ is perfect we may replace it with a complex of finitely generated projective $B^e$-modules. In this way we reduce to $M=B^e$ which is an easy verification.
Let $B$ be a homologically smooth algebra $R$ and $\eta \in {\operatorname{HH}}_d(B)$. We say that $\eta$ is *nondegenerate* if its image under is an isomorphism.
This allows us to redefine a $d$-Calabi-Yau algebra over $R$ as a couple $(B,\eta)$ where $B$ is a homologically smooth $R$-algebra and $\eta$ is a non-degenerate element of ${\operatorname{HH}}_d(B)$. Below we will massage this new definition further. Recall the following
(“Poincare duality”) \[ref-5.5-20\] Assume that $(B,\eta)$ is a $d$-Calabi-Yau $R$-algebra. Then for each $i$, the map $$\label{ref-5.2-21}
{\operatorname{HH}}^i(B) {\longrightarrow}{\operatorname{HH}}_{d-i}(B): \mu {\mapsto}\mu\cap \eta$$ is an isomorphism
The existence of the isomorphism was first stated in [@VdB29] without the explicit formula . The formula is folklore. For completeness we include a possible proof.
The proof of Lemma \[ref-5.3-18\] shows that there is a canonical isomorphism in $D(R)$ $$\label{ref-5.3-22}
{\operatorname{RHom}}_{B^e}({\operatorname{RHom}}_{B^e}(B,B^e),\overset{\downarrow}{B})\cong B\otimes^L_{B^e}\overset{\downarrow}{B}$$ which is compatible with the ${\operatorname{RHom}}_{B^e}(B,B)$-actions on the marked copies of $B$.
By definition $\eta\in H^{-d}(B\otimes^L_{B^e}B)$ corresponds under to an isomorphism $\eta^+:{\operatorname{RHom}}_{B^e}(B,B^e)\rightarrow
\Sigma^{-d}B$. This yields an isomorphism $$\label{ref-5.4-23}
{\operatorname{RHom}}_{B^e}(\Sigma^{-d}B,\overset{\downarrow}{B}){\longrightarrow}{\operatorname{RHom}}_{B^e}({\operatorname{RHom}}_{B^e}(B,B^e),\overset{\downarrow}{B})
:\theta\mapsto \theta\circ \eta^+$$ also compatible with the marked ${\operatorname{RHom}}_{B^e}(B,B)$-actions. Composing gives us an isomorphism $$\xi:{\operatorname{RHom}}_{B^e}(\Sigma^{-d}B,\overset{\downarrow}{B}){\longrightarrow}B\otimes^L_{B^e}\overset{\downarrow}{B}$$ which sends $\operatorname{Id}_B$ to $\eta$.
According to the discussion preceding Lemma \[ref-5.3-18\], the compatibility with the ${\operatorname{RHom}}_{B^e}(B,B)$-actions implies that $\xi$ transforms $\cup$ into $\cap$ on the level of cohomology. More precisely $$\xi(\mu\cup \sigma)=\pm \mu\cap \xi(\sigma)$$ The lemma now follows by taking $\sigma=\operatorname{Id}_B$.
\[ref-5.6-24\] Assume that $(B,\eta)$ is a $d$-Calabi-Yau $R$-algebra. Then $$\begin{aligned}
{\operatorname{HH}}^i(B)=0&\qquad\text{for $i\not \in [0,d]$}\\
{\operatorname{HH}}_i(B)=0&\qquad\text{for $i\not \in [0,d]$}\qed\end{aligned}$$
As before let ${\operatorname{CC}}^-_\bullet(B)=({\operatorname{C}}_\bullet(B)[[u]],\mathsf{b}+u\mathsf{B})$ be the negative cyclic complex and denote its corresponding homology by ${\operatorname{HC}}_\bullet^-(B)$.
\[ref-5.7-25\] Let $(B,\eta)$ be a $d$-Calabi-Yau $R$-algebra. Then ${\operatorname{HC}}^-_i(B)=0$ for $i>d$ and furthermore the map $$\pi:{\operatorname{CC}}^-_{\bullet}(B)
{\longrightarrow}{\operatorname{C}}_{\bullet}(B): \sum b_iu^i {\mapsto}b_0$$ induces an isomorphism $
{\operatorname{HC}}^-_d(B) \cong {\operatorname{HH}}_d(B)
.$
We use a spectral sequence argument. We view ${\operatorname{CC}}^-_\bullet(B)$ as a double complex with $\mathsf{b}$ pointing vertically upwards and $u\mathsf{B}$ pointing horizontally to the right. By Corollary \[ref-5.6-24\] we have ${\operatorname{HH}}_i(B)=0$ for $i>d$. Hence if we filter ${\operatorname{CC}}^-_\bullet(B)$ by column degree then the $E^1$ term of the resulting spectral sequence looks like $$\xymatrix{
0& {\operatorname{HH}}_{d-2}(B)\ar[r]^-{u\mathsf{B}} & u{\operatorname{HH}}_{d-1}(B)\ar[r]^-{u\mathsf{B}} &u^2{\operatorname{HH}}_d(B) \\
0& {\operatorname{HH}}_{d-1}(B)\ar[r]^-{u\mathsf{B}} & u{\operatorname{HH}}_d(B) & 0\\
0&{\operatorname{HH}}_d(B)& 0 & 0\\
0& 0 & 0 & 0
}$$ From this the result follows.
Let $B$ be a homologically smooth $R$-algebra. We say that an element $\eta \in {\operatorname{HC}}^-_d(B)$ is non-degenerate if $\pi(\eta)$ is non-degenerate.
The leads to the following redefinition of a Ginzburg $d$-Calabi-Yau $R$-algebra which we use below.
(Restatement of Definition \[ref-5.1-17\].) \[ref-5.9-26\] A *Calabi-Yau algebra of dimension $d$ over $R$* is a couple $(B,\eta)$ where $B$ is a homologically smooth $R$-algebra and $\eta$ is a non-degenerate element of ${\operatorname{HC}}_d^-(B)$.
We have shown that this definition is equivalent to Ginzburg’s original definition. In the more general setting of DG-algebras this is no longer the case. It is generally believed that Definition \[ref-5.9-26\] is the “correct” definition for a $d$-Calabi-Yau algebra in the DG-case. This is the point of view of Kontsevich-Soibelman in [@KS2] and also of Keller [@Kl1].
Deformations of Calabi-Yau algebras {#ref-6-27}
===================================
In this section we fix a $d$-Calabi-Yau $k$-algebra $(A,\eta_0)$ as in Definition \[ref-5.9-26\]. We will study the deformations of $A$ as a Calabi-Yau algebra.
Let ${\operatorname{Nilp}}$ be the category of commutative, finite dimensional, local $k$-algebras $(R,m)$ such that $R/m=k$. For $(R,m)\in
{\operatorname{Nilp}}$ we define a groupoid ${\operatorname{Def}}_{A,\eta_0}(R)$ as follows: the objects in ${\operatorname{Def}}_{A,\eta_0}(R)$ are triples $(B,s,\eta)$ such that $B$ is an $R$-flat $R$-algebra, $s:B\rightarrow A$ is an $R$-algebra map inducing an isomorphism $B\otimes_R k\rightarrow A$ and $\eta$ is an element of ${\operatorname{HC}}^-_d(B)$ such that $s(\eta)=\eta_0$.
A morphism $(B_1,s_1,\eta_1)\rightarrow (B_2,s_2,\eta_2)$ is a commutative diagram $$\xymatrix{
B_1\ar[rr]^\phi\ar[dr]_{s_1} & &B_2\ar[dl]^{s_2}\\
& A
}$$ such $\eta_2=\phi(\eta_1)$. One sees that ${\operatorname{Def}}_{A,\eta_0}$ becomes a pseudo-functor ${\operatorname{Nilp}}\rightarrow {\operatorname{Gd}}$ in the obvious way.
To be able to rightfully claim that ${\operatorname{Def}}_{A,\eta_0}$ describes the Calabi-Yau deformations of $(A,\eta_0)$ we need the following elementary lemma.
Assume that $(B,s,\eta)\in {\operatorname{Def}}_{A,\eta_0}(R)$. Then $(B,\eta)$ is $d$-Calabi-Yau.
We have to show that $B$ is a perfect $B^e$-module and that $\eta$ induces an isomorphism $\eta^+:{\operatorname{RHom}}_{B^e}(B,B^e)\rightarrow \Sigma^{-d} B$.
Since $R$ is finite dimensional every flat $R$-module is $R$-projective. This applies in particular to $B$ and $B^e$. Let $$0\rightarrow P_u\rightarrow\cdots\rightarrow P_0\rightarrow A\rightarrow 0$$ be a finite resolution of $A$ by finitely generated projective $A^e$-modules. It is easy to see that this resolution can be lifted step by step to a resolution $$0\rightarrow Q_u\rightarrow\cdots\rightarrow Q_0\rightarrow B\rightarrow 0$$ where the $Q_i$ are finitely generated projective $B^e$-modules satisfying $Q_i\otimes_R k\cong P_i$. In particular $B$ is perfect.
It also follows that $H=\operatorname{cone}\eta^+$ is perfect. It it easy to that $\eta^+\otimes^L k\cong \eta_0^+$ and hence $(\operatorname{cone}\eta^+)\otimes^L k\cong \operatorname{cone}(\eta^+\otimes^L k)\cong
\operatorname{cone}\eta_0^+=0$. If now suffices to note that if $H$ is perfect and $H\otimes^L k=0$ then $H=0$.
We now introduce a variant of the groupoid ${\operatorname{Def}}_{A,\eta_0}(R)$ which is easier to describe cohomologically. We remind the reader of the base change convention exhibited in §\[ref-4.5-15\] which we will use throughout. As in §\[ref-4.2-4\] let $-\mu_0\in \mathfrak{C}^1(A)$ be the multiplication map on $A$ and let $\hat{\eta}_0$ be a lift of $\eta_0$ to $\overline{{\operatorname{CC}}}^{\,-}_d(A)$. We define an associated groupoid ${\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R)$ as follows. The objects are couples $(\mu,\eta)$ where
1. $\mu\in \mathfrak{C}^1(A)\otimes_k R$ is such that $-\mu$ defines a unital associative multiplication on $A\otimes_k R$;
2. $\mu\mod m=\mu_0$;
3. $\eta\in \overline{{\operatorname{CC}}}^{\,-}_d(A)\otimes_k R$;
4. $(L_\mu+u\mathsf{B})(\eta)=0$;
5. $\eta\mod m=\hat{\eta}_0$.
For (4) recall that $L_{\mu}+u\mathsf{B}$ is the cyclic differential for the algebra $(A\otimes_k R, \mu)$. A morphism $(\mu_1,\eta_1)\rightarrow (\mu_2,\eta_2)$ in ${\operatorname{Def}}^\flat_{A,\mu_0,\hat{\eta}_0}(R)$ is a couple $(\phi,\xi)$ where
1. $\phi$ is a unital map of $R$-algebras $\phi: (A\otimes_k R, -\mu_1) {\longrightarrow}(A\otimes_k R, -\mu_2)$;
2. $\phi$ is the identity modulo $m$;
3. $\xi$ is an element of $\overline{{\operatorname{CC}}}^{\,-}_{d+1}(A)\otimes_k m$;
4. \[ref-4-28\] $
(L_{\mu_2}+u\mathsf{B})(\xi)=\phi(\eta_1)-\eta_2
$.
The composition of morphisms $$(\mu_1,\eta_1)\xrightarrow{(\phi',\xi')} (\mu_2,\eta_2)
\xrightarrow{(\phi,\xi)}(\mu_2,\eta_2)$$ is defined by $$\label{ref-6.1-29}
(\phi,\xi) \circ
(\phi',\xi')= (\phi \circ \phi',\phi(\xi')+\xi)$$ Below we will often use the notation $\bar{\eta}$ for the cohomology class of a cocycle $\eta$.
\[ref-6.2-30\] The morphism of groupoids $$\begin{aligned}
\operatorname{Ob}({\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R)){\longrightarrow}\operatorname{Ob}({\operatorname{Def}}_{A,\eta_0}(R))&:(\mu,\eta)\mapsto ((A\otimes_k R,-\mu),``\operatorname{mod} m",\bar{\eta})\\
\operatorname{Mor}({\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R)){\longrightarrow}\operatorname{Mor}({\operatorname{Def}}_{A,\eta_0}(R))&:(\phi,\xi)\mapsto \phi\end{aligned}$$ is essentially surjective on objects and surjective on morphisms.
We first prove essential surjectivity. Let $(B,s,\psi)
\in {\operatorname{Def}}_{A,\eta_0}(R)$. Then since $R$ is finite dimensional local and $B$ is $R$-flat we have an isomorphism $B\cong A\otimes_k R$ as $R$-modules and it is easy to see that this isomorphism may be chosen to make the following diagram commutative $$\xymatrix{
B\ar[rr]^\phi\ar[dr]_s && A\otimes_k R \ar[dl]^{\mod m}\\
& A
}$$ We now transfer the multiplication on $B$ to $A\otimes_k R$ where it becomes an element of $-\mu\in\mathfrak{C}^1(A)\otimes_k R$ which modulo $m$ is equal to $-\mu_0$. We do the same with $\psi\in {\operatorname{HC}}^-_d(B)$ and we choose an element $\eta\in \overline{{\operatorname{CC}}}^{\,-}_d(A)\otimes_k
R$ such that $(L_\mu+u\mathsf{B})(\eta)=0$, $\bar{\eta}=\phi(\psi)$. Thus in ${\operatorname{Def}}_{A,\eta_0}(R)$ we have $$(B,s,\psi)\cong ((A\otimes_k R,-\mu),-\operatorname{mod} m,\bar{\eta})$$ This proves essential surjectivity. Now we prove surjectivity on morphisms. Let $(\mu_1,\eta_1)$, $(\mu_2,\eta_2)\in
\operatorname{Ob}({\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R))$ and let $\phi$ be a unital algebra morphism $$(A\otimes_k R,-\mu_1){\longrightarrow}(A\otimes_k R,-\mu_2)$$ inducing the identity modulo $m$ and satisfying $\phi(\bar{\eta}_1)=
\bar{\eta}_2$.
It follows that $\phi(\eta_1)-\eta_2$ is a boundary in the negative cyclic complex of $(A\otimes_k R,-\mu_2)$. In other words there exists $\xi\in \overline{{\operatorname{CC}}}_{d+1}^{\,-}(A)\otimes_k R$ such that $$\phi(\eta_1)-\eta_2=(L_{\mu_2}+u\mathsf{B})(\xi)$$ We have to show that we may choose $\xi\in \overline{{\operatorname{CC}}}_{d+1}^{\,-}(A)\otimes_k m$
Since $\phi$ is the identity modulo $m$ we have $$\begin{aligned}
\phi(\eta_1)-
\eta_2 \mod m&=\eta_1-\eta_2\mod m\\
&=\hat{\eta}_0-\hat{\eta}_0\\
&=0\end{aligned}$$ It follows that $\phi(\eta_1)-\eta_2\in \overline{{\operatorname{CC}}}^{\,-}_d(A)\otimes_k m$ and hence $d\xi\mod m=0$. Since ${\operatorname{HC}}^-_{d+1}(A)=0$ by Proposition \[ref-5.7-25\] we see that there exists $\gamma \in \overline{{\operatorname{CC}}}^{\,-}_{d+2}(A)\otimes_k R$ such that $(L_{\mu_2}+u\mathsf{B})(\gamma)\cong \xi\mod m$. In other words $$\xi'=\xi-(L_{\mu_2}+u\mathsf{B})(\gamma)\in \overline{{\operatorname{CC}}}^{\,-}_d(A)\otimes_k m$$ Then the couple $(\phi,\xi')$ is a pre-image for $\phi$.
For completeness we state the following.
\[ref-6.3-31\] Let $\hat{\eta}_0'\in \overline{{\operatorname{CC}}}^{\,-}_0(A)$ be a different lift of $\eta_0$. Then ${\operatorname{Def}}^\flat_{A,\hat{\eta}'_0}(R)$ and ${\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R)$ are isomorphic.
We could easily prove this here directly, however we will postpone the proof till §\[ref-8-36\] where we reinterprete ${\operatorname{Def}}^\flat_{A,\hat{\eta}_0}(R)$ in terms of the Maurer-Cartan equation.
The Maurer-Cartan formalism {#ref-7-32}
===========================
In this section we briefly recall the construction of the deformation functor associated to a DG-Lie algebra.
Let ${\ensuremath{\mathfrak{h}}}^{\bullet}$ be a DG-Lie algebra over $k$. The set $${\operatorname{MC}}(\mathfrak{h}^\bullet)\overset{\text{def}}{=}\left\{y \in {\ensuremath{\mathfrak{h}}}^1 \, \left\vert
dy +\frac{1}{2}[y,y]=0 \right.\right\}$$ is the set of solutions to the *Maurer-Cartan equation* in ${\ensuremath{\mathfrak{h}}}^{\bullet}$. It has a natural structure of a groupoid which we now describe.
Assume that $\mathfrak{n}$ is a nilpotent Lie algebra and let $\widehat{U}(\mathfrak{n})$ be the enveloping algebra of $\mathfrak{n}$, completed at the augmentation ideal. Then the group $\exp(\mathfrak{n})$ is by definition the set of group like elements in $\widehat{U}(\mathfrak{n})$. It is well-known and easy to see that there is a bijection $$\exp:\mathfrak{n}{\longrightarrow}\exp(\mathfrak{n}):n\mapsto e^n$$ between the primitive and the group like elements in $\widehat{U}(\mathfrak{n})$.
Now assume that ${\ensuremath{\mathfrak{h}}}^{\bullet}$ is nilpotent. Then $\widehat{U}(\mathfrak{h}^0)$ acts on the graded Lie algebra $\mathfrak{h}^\bullet$ using the adjoint action and hence so does the *gauge group* $G(\mathfrak{h}^\bullet)\overset{\text{def}}{=}\exp(\mathfrak{h}^0)$. This action does not commute with the differential and in particular it does not preserve ${\operatorname{MC}}(\mathfrak{h}^\bullet)$. However the following modified *gauge action* does: $$\label{ref-7.1-33}
\begin{aligned}
\exp(x)\ast y&\overset{\text{def}}{=}e^{{\operatorname{ad}}x}(y)-\frac{e^{{\operatorname{ad}}x}-1}{{\operatorname{ad}}x}(dx)\\
&=e^{{\operatorname{ad}}x}(y)-\sum_{n=0}^\infty \frac{1}{(n+1)!} ({\operatorname{ad}}x)^n (dx)
\end{aligned}$$ where $x\in\mathfrak{h}^0$, $y\in \mathfrak{h}^1$ and $({\operatorname{ad}}x)(u)=[x,u]$.
An elegant derivation of this action is given by Manetti [@Manetti §V.4]. One first formally adjoins an element $\delta$ of degree one to $\mathfrak{h}^\bullet$ such that $dx=[\delta,x]$, $d\delta=0$ and $[\delta,\delta]=0$. Then can be rewritten as: $$\label{ref-7.2-34}
\exp(x)\ast y=e^{{\operatorname{ad}}x}(y+\delta)-\delta$$ This action preserves ${\operatorname{MC}}(\mathfrak{h}^\bullet)$ since for $y\in\mathfrak{h}^1$: $$y\in{\operatorname{MC}}(\mathfrak{h}^\bullet)\iff [y+\delta,y+\delta]=0$$ In the sequel we view ${\operatorname{MC}}(\mathfrak{h}^\bullet)$ as a groupoid through the $G(\mathfrak{h}^\bullet)$-action.
If $y\in {\operatorname{MC}}(\mathfrak{h}^\bullet)$ then by definition $\mathfrak{h}^\bullet_y$ is the DG-Lie algebra which is $\mathfrak{h}^\bullet$ as graded Lie algebra but which has the deformed differential $d_y=d+[y,-]$. Using one easily shows that for $x\in \mathfrak{h}^0$. $$\label{ref-7.3-35}
e^{{\operatorname{ad}}x}:\mathfrak{h}^\bullet_y{\longrightarrow}\mathfrak{h}^\bullet_{\exp(x)\ast y}$$ is an isomorphism of DG-Lie algebras
Assume $(R,m)\in{\operatorname{Nilp}}$. Given an arbitrary DG-Lie algebra $\mathfrak{g}^\bullet$ over $k$, the vector space ${\ensuremath{\mathfrak{g}}}^{\bullet}{\otimes}_k m$ becomes a nilpotent DG-Lie algebra. We define ${\operatorname{\mathcal{MC}}}(R)$ as ${\operatorname{MC}}(\mathfrak{g}^\bullet\otimes_k m)$ equipped with the groupoid structure introduced above. In this way we obtain a pseudo-functor ${\operatorname{\mathcal{MC}}}:{\operatorname{Nilp}}{\longrightarrow}{\operatorname{Gd}}$. This is the “deformation functor” associated to $\mathfrak{g}^\bullet$.
The DG-Lie algebra [ $\mathfrak{D}^\bullet(A,\eta)$]{} {#ref-8-36}
======================================================
Below $(A,\bar{\eta}_0)$ is a $d$-Calabi-Yau $k$-algebra where $\eta_0\in \overline{{\operatorname{CC}}}_d^-(A)$ satisfies $(L_{\mu_0}+u\mathsf{B})(\eta_0)=0$, with $-\mu_0\in\mathfrak{C}^1(A)$ being the multiplication on $A$. In this section we associate a DG-Lie algebra $\mathfrak{D}^\bullet(A,\eta_0)$ to $A$ and prove that its deformation functor (see §\[ref-7-32\]) is isomorphic to the functor ${\operatorname{Def}}^\flat_{A,\eta_0}$ introduced in §\[ref-6-27\].
If ${\ensuremath{\mathfrak{g}}}^{\bullet}$ is a DG-Lie algebra and $M^{\bullet}$ a ${\ensuremath{\mathfrak{g}}}^{\bullet}$-module then the direct sum complex ${\ensuremath{\mathfrak{g}}}^{\bullet}{\oplus}M^{\bullet}$ becomes a DG-Lie algebra when endowed with the following bracket: $$[(g,m),(g',m')]:=([g,g'], gm'-(-1)^{\vert g'\vert \vert m\vert} g'm)$$ The resulting DG-Lie algebra is called the *semi-direct product* of $\mathfrak{g}^\bullet$ and $M^\bullet$ and is denoted by ${\ensuremath{\mathfrak{g}}}^{\bullet}\ltimes M^{\bullet}$
By (see also §\[ref-4.6-16\]) we have a DG-Lie action $$\bar{{\ensuremath{\mathfrak{C}}}}^{\bullet}(A)\times \Sigma^{-d-1}\overline{{\operatorname{CC}}}_{\bullet}^-(A){\longrightarrow}\Sigma^{-d-1}\overline{{\operatorname{CC}}}_{\bullet}^-(A):(x,\eta)\mapsto L_x\eta$$ and we can form the corresponding semi-direct product $\mathfrak{D}^\bullet(A)^\sharp=\bar{{\ensuremath{\mathfrak{C}}}}^{\bullet}(A) \ltimes \Sigma^{-d-1}\overline{{\operatorname{CC}}}_{\bullet}^-(A)$.
The element $x=(0,s^{-d-1}\eta_0)\in \mathfrak{D}^\bullet(A)^\sharp$ satisfies $dx=0$ and $[x,x]=0$. So it satisfies the Maurer-Cartan equation. Put $\mathfrak{D}^\bullet(A,\eta_0)=\mathfrak{D}^\bullet(A)_x^\sharp$, with notation as in §\[ref-7-32\].
\[ref-8.1-37\] Let $(R,m)\in {\operatorname{Nilp}}$. There is an isomorphism of groupoids $$\Phi(R):{\operatorname{MC}}({\ensuremath{\mathfrak{D}}}^{\bullet}(A,\eta_0)\otimes_k m) {\longrightarrow}{\operatorname{Def}}^\flat_{A,\eta_0}(R)$$ which on objects is given by $$\label{ref-8.1-38}
(\mu,s^{-d-1}\eta)\mapsto(\mu_0+\mu,\eta_0+\eta)$$
\[ref-8.2-39\] There is a natural transformation of pseudo-functors $$\Phi:\mathcal{MC}({\ensuremath{\mathfrak{D}}}^{\bullet}(A,\eta_0)) {\longrightarrow}{\operatorname{Def}}^\flat_{A,\eta_0}$$ which, when evaluated on $R\in {\operatorname{Nilp}}$, is an isomorphism of groupoids.
We shall prove Theorem \[ref-8.1-37\] by combining some lemmas. Throughout we fix $(R,m)\in {\operatorname{Nilp}}$. The following lemma says that $\Phi(R)$ behaves correctly on objects.
\[ref-8.3-40\] Let $\mu \in \bar{\mathfrak{C}}^\bullet_1(A)\otimes_k m$ and $\eta \in \overline{{\operatorname{CC}}}^{\,-}_d(A)\otimes_k m$. The following are equivalent:
1. $(\mu,s^{-d-1}\eta) \in {\operatorname{MC}}({\ensuremath{\mathfrak{D}}}^{\bullet}(A,\eta_0)\otimes_k m)$;
2. $(\mu_0+\mu,\eta_0+\eta)\in {\operatorname{Def}}^\flat_{A,\eta_0}(R)$.
We will work out what it means for $(\mu,s^{-d-1}\eta) \in {\ensuremath{\mathfrak{D}}}^1(A,\eta_0)\otimes_k m$ to satisfy the Maurer-Cartan equation. To simplify the notations we write $\eta'_0=s^{-d-1}\eta_0$, $\eta'=s^{-d-1}\eta$. We compute $$\begin{aligned}
\frac{1}{2}[(\mu,\eta'),(\mu,\eta')]+d_{{\ensuremath{\mathfrak{D}}}}(\mu,\eta')
&= \frac{1}{2}([\mu,\mu], 2L_\mu(\eta'))+([\mu_0,\mu], (L_{\mu_0}+u\mathsf{B})(\eta'))+[(0,\eta_0'), (\mu,\eta')]\\
&=\frac{1}{2}([\mu,\mu], 2L_\mu(\eta'))+([\mu_0,\mu], (L_{\mu_0}+u\mathsf{B})(\eta'))+
[0,L_{\mu}(\eta_0')]\\
&= (\frac{1}{2}[\mu,\mu]+[\mu_0,\mu],L_{\mu}(\eta')+(L_{\mu_0}+u\mathsf{B})(\eta')+ L_{\mu}(\eta_0'))\\
&=([\mu_0+\mu,\mu_0+\mu], (L_{\mu+\mu_0}+u\mathsf{B})(\eta'+\eta'_0))\end{aligned}$$ where in the last line we have used $[\mu_0,\mu_0]=0$, $(L_{\mu_0}+u\mathsf{B})(\eta'_0)=0$. Thus if $(\mu_0+\mu,\eta_0+\eta)\in {\operatorname{Def}}^\flat_{A,\eta_0}(R)$ then $(\mu,s^{-d-1}\eta)\in {\operatorname{MC}}({\ensuremath{\mathfrak{D}}}^{\bullet}(A,\eta_0)\otimes_k m)$. To prove the converse the only thing we still need to check is that $-(\mu_0+\mu)$ defines a *unital* multiplication on $A\otimes_k R$. This follows immediately from the fact that $-\mu_0$ is unital and $\mu$ is normalized.
The next two lemmas will help us describing the gauge group action of $G(\mathfrak{D}^\bullet(A,\eta))$.
\[ref-8.4-41\] Let ${\ensuremath{\mathfrak{n}}}$ be a nilpotent Lie algebra over $k$ and let $M$ a representation of ${\ensuremath{\mathfrak{n}}}$. Then there is an isomorphism of groups $$\exp(\frak{n})\ltimes M{\longrightarrow}\exp(\frak{n}\ltimes M):
(\exp(n),m)\mapsto \exp(n,0)\exp(0,m)$$
This is a straightforward verification from the definition of “$\exp$” in §\[ref-7-32\] using the fact that $$U(\mathfrak{h}\ltimes M)\cong U(\mathfrak{h})\ltimes
\operatorname{Sym}(M)\qed$$
\[ref-8.5-42\] Let ${\ensuremath{\mathfrak{g}}}^{\bullet}$ be a nilpotent DG-Lie algebra over $k$ and $M^{\bullet}$ a nilpotent DG-module. Consider the DG-algebra ${\ensuremath{\mathfrak{h}}}^{\bullet}$ which is ${\ensuremath{\mathfrak{g}}}^{\bullet}\ltimes M^\bullet$ as graded Lie algebras and which is equipped with a deformed differential $(d_{\mathfrak{g}},d_M)+
d_0$ where $d_0: {\ensuremath{\mathfrak{g}}}^{\bullet}{\longrightarrow}M$ is of the form $g\mapsto (-1)^{|g|}gm_0$ for suitable $m_0\in M^1$. Then for $g\in\mathfrak{g}^0$, $m\in M^0$ and $(g_1, m_1)\in\mathfrak{h}^1$ we have $$\begin{aligned}
\exp(g,0)\ast(g_1,m_1)&=(\exp(g)\ast g_1,e^g(m_1-m_0)+m_0)\\
\exp(0,m)\ast(g_1,m_1)&=(g_1,m_1-(g_1+d_M)m)\end{aligned}$$
We compute $$\begin{aligned}
\exp(g,0)*(g_1,m_1)&= e^{{\operatorname{ad}}(g,0)}(g_1, m_1) -\sum_n \frac{1}{(n+1)!} {\operatorname{ad}}^n(g,0)(d_{\ensuremath{\mathfrak{h}}}(g,0))\\
&=(e^{{\operatorname{ad}}g}g_1, e^gm_1)-\sum_n \frac{1}{(n+1)!} {\operatorname{ad}}^n(g,0)(d_{\ensuremath{\mathfrak{g}}}g,d_0g)\\
&=(e^{{\operatorname{ad}}g}g_1,e^gm_1) -\sum_n \frac{1}{(n+1)!}({\operatorname{ad}}^n(g)(d_{\ensuremath{\mathfrak{g}}}g),g^{n+1}m_0)\\
&=(e^g*g_1,e^g(m_1-m_0)+m_0)\end{aligned}$$ Similarly: $$\begin{aligned}
\exp(0, m)*(g_1, m_1)&=e^{{\operatorname{ad}}(0, m)}(g_1, m_1) -\sum \frac{1}{(n+1)!} {\operatorname{ad}}^n(0, m)(d_{\ensuremath{\mathfrak{h}}}(0,m))\\
&=(g_1,m_1)-(0,g_1m)-(0,d_Mm)\\
&=(g_1, m_1-(g_1+d_M)m)\qed\end{aligned}$$
We will also use the following variant of
\[ref-8.6-43\] Let ${\ensuremath{\mathfrak{h}}}^{\bullet}$ be a nilpotent DG-Lie algebra with inner differential $d=[\mu_0,-]$. Then for $x\in \mathfrak{h}^0$, $y\in\mathfrak{h}^1$ one has $$\exp(x)*y= e^{{\operatorname{ad}}x}(y+\mu_0)-\mu_0$$
Direct evaluation of the righthand side yields the formula for $\exp(x)\ast y$.
We start by verifying that yields indeed a map of groupoids. To this end we have to define $\Phi(R)$ on maps. Note that by lemma \[ref-8.4-41\] each element of $\exp({\ensuremath{\mathfrak{D}}}^0(A, \mu_0, \eta_0)\otimes m)$ can be uniquely written as $\exp(0, s^{-d-1}\xi)\exp(f,0)$ for $f \in \bar{\mathfrak{C}}^0(A)\otimes_k m={\operatorname{Hom}}(A/k, A)
{\otimes}_k m\subset {\operatorname{Hom}}(A,A)\otimes_k m$ and $\xi \in \overline{{\operatorname{CC}}}^{\,-}_{d+1}(A){\otimes}_k m$. We put $
\phi=e^f
$. Then $\phi\in {\operatorname{Hom}}(A,A)\otimes_k R$ is such that $\phi\mod m={\operatorname{Id}}_A$.
Assume that $$\label{ref-8.2-44}
\exp(0,s^{-d-1} \xi)\ast\exp(f,0)\ast(\mu_1, s^{-d-1}\eta_1)= (\mu_2, s^{-d-1}\eta_2)$$ We define $\Phi(R)$ on maps as follows $$\label{ref-8.3-45}
\Phi(R)(\exp(0,s^{-d-1}\xi)\exp(f,0))=(e^f,(-1)^d\xi)$$ For this to be well defined we should have a morphism $$(\phi , (-1)^d\xi):(\mu_0+\mu_1,\eta_0+\eta_1){\longrightarrow}(\mu_0+\mu_2,\eta_0+\eta_2)$$ in ${\operatorname{Def}}^\flat_{A,\eta_0}(R)$. In other words:
1. $\phi :(A{\otimes}_k R, -(\mu_0+\mu_1)) {\longrightarrow}(A{\otimes}_k R, -(\mu_0+\mu_2))$ is an $R$-algebra morphism;
2. $\phi(\eta_0+\eta_1)= \eta_0+\eta_2+(-1)^d(L_{\mu_0+\mu_2}+u\mathsf{B})(\xi)$.
Put $\eta'_i=s^{-d-1}\eta_i$ for $i=0,1,2$, $\xi'=s^{-d-1}\xi$. We invoke Lemma \[ref-8.5-42\] with $m_0=-\eta'_0$. Then yields $$\label{ref-8.4-46}
(\mu_2, \eta'_2)=(\exp(f)\ast\mu_1,e^f(\eta'_0+\eta'_1)-\eta'_0-
L_{\exp(f)\ast\mu_1}(\xi')-(L_{\mu_0}+u\mathsf{B})(\xi'))$$ We may compute $\exp(f)\ast\mu_1$ inside unnormalized cochains ${\operatorname{C}}^\bullet(A)$ and then we may invoke lemma \[ref-8.6-43\]. We find $$\exp(f)\ast\mu_1=e^{{\operatorname{ad}}f}(\mu_0+\mu_1)-\mu_0$$ Furthermore a direct computation shows that $$\begin{aligned}
e^{{\operatorname{ad}}f}(\mu_0+\mu_1)&=e^f\circ (\mu_0+\mu_1)\circ (e^{-f},e^{-f})\\
&=\phi\circ (\mu_0+\mu_1)\circ (\phi^{-1},\phi^{-1})\end{aligned}$$ Hence translates into $$\begin{aligned}
\mu_0+\mu_2&=\phi\circ (\mu_0+\mu_1)\circ (\phi^{-1},\phi^{-1})\\
\eta_0'+\eta'_2&=\phi(\eta_0'+\eta'_1)-(L_{\mu_0+\mu_2}+u\mathsf{B})(\xi')\end{aligned}$$ The first of these equations yields (a). The second yields (b) taking into account that $L_{\mu_0+\mu_2}+u\mathsf{B}$ has degree one, which induces a sign change.
It remains to show that our assignment respects compositions. By Lemma \[ref-8.4-41\] we have for $f,g,h\in \bar{\mathfrak{C}}^0(A)\otimes_k m$ such that $\exp(h)=\exp(g)\exp(f)$, $\nu,\xi
\in \overline{{\operatorname{CC}}}_{-d-1}^{\,-}(A)\otimes_k m$: $$\begin{aligned}
\Phi(R)(\exp(0, s^{-d-1}\nu)&\exp(g,0)\circ \exp(0, s^{-d-1}\xi)\exp(f,0))\\
=& \Phi(R)(\exp(0, s^{-d-1}\nu)\exp(0, s^{-d-1} e^g\xi)\exp(g,0)\exp(f,0))\\
=&\Phi(R)( \exp(0,s^{-d-1}(\nu+e^g\xi))\exp(h,0))\\
=&(e^h, (-1)^d(\nu+e^g\xi))\\
=&(e^ge^f, (-1)^d(\nu+e^g\xi))\end{aligned}$$ and $$\begin{aligned}
\Phi(R)(\exp(0, s^{-d-1}\nu)\exp(g,0))&\circ \Phi(R)(\exp(0, s^{-d-1}\xi)\exp(f,0))\\
&=(e^g,(-1)^d\nu)(e^f,(-1)^d\xi)\\
&=(e^ge^f, (-1)^d(\nu+e^g\xi))\end{aligned}$$ by . We conclude that $\Phi(R)$ is indeed a map of groupoids. By Lemma \[ref-8.3-40\] it is bijective on objects, and running the above computation backwards, starting from , we see that it also bijective on maps. Thus $\Phi(R)$ is an isomorphism of groupoids.
The following result implies Propositon \[ref-6.3-31\].
Assume that $\eta_0,\eta'_0\in \overline{{\operatorname{CC}}}^-_{d}(A)$ induce the same element in ${\operatorname{HC}}^-_{d}(A)$. Then $\mathfrak{D}^\bullet(A,\eta_0)\cong
\mathfrak{D}^\bullet(A,\eta_0')$.
From one sees that it is sufficient to show that $(0,s^{-s-1}\eta_0),(0,s^{-d-1}\eta'_0)$ are in the same $G(\mathfrak{D}^\bullet(A)^\sharp)$ orbit. Pick $\xi\in \overline{{\operatorname{CC}}}^{\,-}_{d+1}(A)$ such that $\eta_0'=\eta_0+(-1)^d(L_{\mu_0}+u\mathsf{B})\xi$. We compute using $$\begin{aligned}
\exp(0,s^{-d-1}\xi)\ast (0,s^{-d-1}\eta_0)&=(0,s^{-d-1}\eta_0)-
(0,(L_{\mu_0}+u\mathsf{B})(s^{-d-1}\xi))\\
&=(0,s^{-d-1}\eta_0')\qed\end{aligned}$$
Relation with Hochschild cohomology {#ref-9-47}
===================================
Let $(A,\bar{\eta}_0)$ be a $d$-Calabi-Yau $k$-algebra and let $-\mu_0$ be the multiplication of $A$. Let $(R,m)\in {\operatorname{Nilp}}$. We may define pseudo-functors ${\operatorname{Def}}_A$, ${\operatorname{Def}}^\flat_{A}:{\operatorname{Nilp}}{\longrightarrow}{\operatorname{Gd}}$ in the same way as ${\operatorname{Def}}_{A,\eta_0}$, ${\operatorname{Def}}^\flat_{A,\eta_0}$, ignoring $\eta_0$. The induced morphism $${\operatorname{Def}}^\flat_A(R){\longrightarrow}{\operatorname{Def}}_A(R)$$ is essentially surjective on objects and surjective on morphisms. Furthermore there is an isomorphism of groupoids $$\Phi(R):{\operatorname{MC}}(\bar{\mathfrak{C}}^\bullet(A)\otimes_k m){\longrightarrow}{\operatorname{Def}}^\flat_A(R):\mu\mapsto \mu_0+\mu$$ The obvious morphism of DG-Lie algebras $$\phi:\mathfrak{D}^\bullet(A,\eta_0){\longrightarrow}\bar{\mathfrak{C}}^\bullet(A):(\mu,\eta)\mapsto \mu$$ makes the following diagram commutative: $$\xymatrix{
{\operatorname{\mathcal{MC}}}(\mathfrak{D}^\bullet(A,\eta_0))\ar[d]_{\Phi}\ar[r]^\phi& {\operatorname{\mathcal{MC}}}(\bar{\mathfrak{C}}^\bullet(A))\ar[d]^{\Phi}\\
{\operatorname{Def}}_{A,\eta} \ar[r]_{\text{forget $\eta$}}& {\operatorname{Def}}_A
}$$
Homology of [ $\mathfrak{D}^\bullet(A,\eta)$]{}
===============================================
Let $(A,\bar{\eta}_0)$ be a $d$-Calabi-Yau algebra as before with multiplication $-\mu_0$. In this section we prove that the homology of ${\ensuremath{\mathfrak{D}}}^\bullet(A, \eta_0)$ is isomorphic to ${\operatorname{HC}}^-_{-{\bullet}+d-1}(A)$. Furthermore we show that the induced Lie bracket on ${\operatorname{HC}}^-_{-{\bullet}+d-1}(A)$ is given Menichi’s string topology bracket [@Menichi].
In our statements and computations we will use the following conventions:
- Taking homology classes is indicated by overlining.
- Depending on context $\cong$ will mean either “up to homotopy” (when discussing maps) or “up to addition of a coboundary” (when discussing elements).
\[ref-10.1-48\] The map $$\Psi: {\ensuremath{\mathfrak{D}}}^{\bullet}(A,\eta_0) {\longrightarrow}\Sigma^{-d+1}
\overline{\mathfrak{{\operatorname{CC}}}}^{\,-}_{\bullet}(A): (\mu, s^{-d-1}\eta)
{\mapsto}(-1)^{\vert \mu\vert -1}(i_\mu+uS_\mu)(s^{-d+1}\eta_0)+us^{-d+1}\eta$$ is a quasi-isomorphism of complexes.
To simplify the notation we put $$I_\mu=i_\mu+uS_\mu$$ We first check that $\Psi$ does indeed commute with differentials. Write $\eta_0'=s^{-d-1}\eta_0$, $\eta'=s^{-d-1}\eta$. Then $$\label{ref-10.1-49}
\Psi(\mu,\eta')=s^2((-1)^{\vert \mu\vert -1}I_\mu\eta'_0+u\eta')$$ and hence $$\begin{aligned}
(d\circ \Psi)(\mu,\eta')&=
(\mathsf{b}+u\mathsf{B})s^2((-1)^{\vert \mu\vert -1}I_\mu \eta'_0+u\eta')\\
&= s^2((-1)^{\vert \mu \vert-1}(\mathsf{b}+u\mathsf{B}) I_\mu\eta'_0+u(\mathsf{b}+u\mathsf{B})\eta')\\
&= s^{2}((-1)^{\vert \mu \vert-1}[\mathsf{b}+u\mathsf{B}, I_\mu](\eta'_0)+u(\mathsf{b}+u\mathsf{B})\eta')&&(\text{since $(\mathsf{b}+u\mathsf{B})\eta'_0=0$)}\\
&= s^{2}((-1)^{\vert \mu \vert-1}(uL_{\mu}-I_{d\mu})\eta'_0+u(\mathsf{b}+
u\mathsf{B})\eta')&& (\text{by \eqref{ref-4.8-13}})\\
&= s^{2}((-1)^{\vert \mu \vert}I_{d\mu}\eta'_0+
u((\mathsf{b}+u\mathsf{B})\eta'-(-1)^{\vert \mu \vert}L_\mu \eta'_0))\\
&=\Psi(d\mu, (\mathsf{b}+u\mathsf{B})\eta'-(-1)^{\vert \mu \vert}L_\mu \eta'_0)
&&(\text{by \eqref{ref-10.1-49}})
\\
&=(\Psi\circ d)(\mu, \eta')\end{aligned}$$ To see that $\Psi$ is indeed a quasi-isomorphism, consider the following commutative diagram $$\xymatrix{
0\ar[r] & \Sigma^{-d-1}\overline{{\operatorname{CC}}}_{\bullet}^{\,-}(A)\ar[r] \ar[d]_\Psi& {\ensuremath{\mathfrak{D}}}^\bullet(A,\eta_0) \ar[r]\ar[d]\ar[d]_\Psi& \bar{{\ensuremath{\mathfrak{C}}}}^{\bullet}(A) \ar[r]\ar[d]^{\overline{\Psi}}&0\\
0\ar[r] & u\Sigma^{-d+1}\overline{{\operatorname{CC}}}_{\bullet}^{\,-}(A)\ar[r] & \Sigma^{-d+1}
\overline{{\operatorname{CC}}}_{\bullet}^{\,-}(A) \ar[r]& \Sigma^{-d+1}\bar{{\operatorname{C}}}_{\bullet}(A) \ar[r]&0
}$$ The map on the left is multiplication by $u$ which is an isomorphism. The map $\overline{\Psi}$ is given on cohomology by $$\overline{\mu}\mapsto\pm\overline{I_\mu\eta_0\,\operatorname{mod} u}= \pm\overline{i_{\mu} \pi(\eta_0)}$$ where $\pi$ is as in Proposition \[ref-5.7-25\]. Hence $\overline{\Psi}$ is an isomorphism by Proposition \[ref-5.5-20\]. From the five lemma we conclude that the middle arrow is an isomorphism on cohomology as well.
We now describe the Lie bracket on ${\operatorname{HC}}^-_\bullet(A)$ induced by the quasi-isomorphism $\Psi$. As already used in the above proof the map $$-\cap \pi(\bar{\eta}_0): {\operatorname{HH}}^i(A)
{\longrightarrow}{\operatorname{HH}}_{d-i}(A)$$ is invertible by Proposition \[ref-5.5-20\]. Let us denote its inverse by $j$. Using $j$, one can transport the cup product on ${\operatorname{HH}}^{\bullet}(A)$ to a product on ${\operatorname{HH}}_{\bullet}(A)$ $$\cdot:{\operatorname{HH}}_i(A)\times {\operatorname{HH}}_j(A){\longrightarrow}{\operatorname{HH}}_{i+j-d}(A)$$ with explicit formula $$a\cdot b=(j(a)\cup j(b))\cap \pi(\bar{\eta}_0)$$ or in a form more suitable for us below $$\label{ref-10.2-50}
i_{\mu_1}\pi(\bar{\eta}_0)\cdot i_{\mu_2}\pi(\bar{\eta}_0)=i_{\mu_1\cup \mu_2}\pi(\bar{\eta}_0)$$
\[ref-10.2-51\] The Lie bracket induced on $$H^\bullet(\Sigma^{-d+1}\overline{{\operatorname{CC}}}^{\,-}_\bullet(A))=
{\operatorname{HC}}_{\bullet+d-1}(A)$$ by the quasi-isomorphism $\Psi$ is given by $$[-,-]:
{\operatorname{HC}}^-_n(A) \times {\operatorname{HC}}^-_{m}(A) {\longrightarrow}{\operatorname{HC}}^-_{n+m-d+1}(A): (\eta_1, \eta_2)
{\mapsto}(-1)^{|\eta_1|+d}\mathsf{B}(\pi(\eta_1)\,\cdot\,\pi(\eta_2))$$ where $\mathsf{B}$ is given by $$\mathsf{B}:{\operatorname{HH}}_q(A){\longrightarrow}{\operatorname{HC}}_{q+1}^-(A):\bar{\nu}\mapsto \overline{\mathsf{B}\nu}$$
We first need the following technical lemma.
\[ref-10.3-52\] Let $\mu \in \bar{{\ensuremath{\mathfrak{C}}}}^{\bullet}(A)$ and $\eta \in \overline{{\operatorname{CC}}}_{\bullet}^{\,-}(A)$ be cocycles. Then $L_\mu \eta$ and $\mathsf{B}i_\mu \pi(\eta)$ are both cocycles in $\overline{{\operatorname{CC}}}_{\bullet}^{\,-}(A)$ and $\overline{\mathsf{B}i_\mu \pi(\eta})=\overline{L_\mu \eta}$ in ${\operatorname{HC}}^-_\bullet(A)$.
$L_\mu \eta$ is a cocycle by . $\mathsf{B}i_\mu \pi(\eta)$ is a cocycle since $\pi(\eta)$ is a cocycle in $\bar{{\operatorname{C}}}_\bullet(A)$ and $$(\mathsf{b}+u\mathsf{B})(\mathsf{B}i_\mu \pi(\eta))=
\mathsf{b}\mathsf{B}i_\mu \pi(\eta)= -\mathsf{B} \mathsf{b}i_\mu \pi(\eta)=0$$ where the last equality follows from .
For the second claim, we first multiply by $u$: $$\begin{aligned}
u(L_\mu\eta-\mathsf{B}i_\mu\pi(\eta))&=[\mathsf{b}+u\mathsf{B},I_\mu]\eta-u\mathsf{B}i_\mu\pi(\eta)&&(\text{by \eqref{ref-4.8-13}})\\
&=(\mathsf{b}+u\mathsf{B})I_\mu\eta-u\mathsf{B}i_\mu\pi(\eta)&&(\text{since $(\mathsf{b}+u\mathsf{B})\eta=0$})\\
&=(\mathsf{b}+u\mathsf{B})(I_\mu\eta-i_\mu \pi(\eta))&& (\text{since $\mathsf{b}i_\mu\pi(\eta)=0$})\end{aligned}$$
Now, $\pi(I_\mu\eta-i_\mu(\pi(\eta))= i_\mu\pi(\eta)-i_\mu\pi(\eta)=0$, which means that $I_\mu\eta-i_\mu\pi(\eta)$ is divisible by $u$. Thus it follows that $$L_\mu\eta-\mathsf{B}i_\mu\pi(\eta)=(\mathsf{b}+u\mathsf{B})(u^{-1}(I_\mu\eta-i_\mu\pi(\eta)))$$ hence the claim.
Let $(\mu_1, s^{-d-1}\eta_1)$ and $(\mu_2,s^{-d-1}\eta_2)$ be two cocycles in ${\ensuremath{\mathfrak{D}}}^{\bullet}(A, \eta_0)$. We must prove for $\eta_i'=s^{-d-1}\eta_i$ $$\label{ref-10.3-53}
\overline{s^{d-1}\Psi([(\mu_1, \eta'_1),(\mu_2,\eta'_2)])}=[\overline{s^{d-1}\Psi(\mu_1,\eta'_1)},\overline{s^{d-1}\Psi(\mu_2,\eta'_2)}]$$ We will first compute the lefthand side of . Writing out the differential in $\mathfrak{D}^\bullet(A,\eta'_0)$ explicitly, the fact that $({\mu_1}, {\eta'_1})$, $(\mu_2,{\eta'_2})$ are cocycles implies $$\label{ref-10.4-54}
\begin{gathered}
d{\mu_1}=d\mu_2=0\\
(\mathsf{b}+u\mathsf{B}){\eta'_1}-(-1)^{\vert {\mu_1} \vert}L_{\mu_1} {\eta'_0}=(\mathsf{b}+u\mathsf{B}){\eta'_2}-(-1)^{\vert \mu_2\vert }L_{\mu_2}{\eta'_0}=0
\end{gathered}$$ where ${\eta'_0}=s^{-d-1}\eta'_0$. We compute $$\begin{aligned}
x&\overset{\text{def}}{=}s^{d-1}\Psi([({\mu_1},{\eta'_1}),(\mu_2{\eta'_2})])\\
&= s^{d-1}\Psi([{\mu_1},\mu_2], L_{\mu_1}{\eta'_2}-(-1)^{\vert {\mu_1}\vert \vert {\eta'_2} \vert} L_{\mu_2} {\eta'_1})\\
&=s^{d+1}((-1)^{|{\mu_1}|+|\mu_2| -1}I_{[{\mu_1},\mu_2]} {\eta'_0}+u( L_{\mu_1}{\eta'_2}-(-1)^{\vert {\mu_1}\vert \vert \mu_2 \vert} L_{\mu_2} {\eta'_1})) \label{ref-10.5-55}
\end{aligned}$$ where we have used and the fact that $|{\eta'_2}|=|\mu_2|$.
We now consider the boundary element $(\mathsf{b}+u\mathsf{B})I_{\mu_1} {\eta'_2}$. By , we have $$(\mathsf{b}+u\mathsf{B})I_{\mu_1} {\eta'_2}-(-1)^{\vert{\mu_1} \vert+1} I_{\mu_1} (\mathsf{b}+u\mathsf{B}) {\eta'_2}+I_{d{\mu_1}}{\eta'_2}=uL_{\mu_1} {\eta'_2}$$ Taking into account this becomes $$\begin{aligned}
(\mathsf{b}+u\mathsf{B})I_{\mu_1} {\eta'_2}
&= (-1)^{\vert {\mu_1}\vert-1 } I_{\mu_1}(\mathsf{b}+u\mathsf{B}) {\eta'_2}+uL_{\mu_1} {\eta'_2}\\
&=(-1)^{\vert {\mu_1}\vert-1 +\vert \mu_2 \vert} I_{\mu_1} L_{\mu_2}{\eta'_0}+uL_{\mu_1} {\eta'_2}\end{aligned}$$ and similarly $$(\mathsf{b}+u\mathsf{B})I_{\mu_2}{\eta'_1}=(-1)^{\vert \mu_2\vert -1 + \vert {\mu_1} \vert } I_{\mu_2} L_{{\mu_1}}{\eta'_0}+uL_ {\mu_2} {\eta'_1}$$
We now subtract both boundaries with appropriate sign from to obtain the following homologous cocycle $$\begin{aligned}
\label{ref-10.6-56}
x&\cong
(-1)^{\vert {\mu_1}\vert+\vert \mu_2\vert
-1}
s^{d+1}(I_{[{\mu_1},\mu_2]}{\eta'_0}-
I_{\mu_1} L_{\mu_2}{\eta'_0}+ (-1)^{|{\mu_1}||\mu_2|}I_{\mu_2} L_{{\mu_1}}
{\eta'_0})\\
&=(-1)^{\vert {\mu_1}\vert+\vert \mu_2\vert
-1}
s^{d+1}
(
I_{[{\mu_1},\mu_2]}-
I_{\mu_1} L_{\mu_2}+ (-1)^{|{\mu_1}||\mu_2|}I_{\mu_2} L_{{\mu_1}}
){\eta'_0}
\end{aligned}$$ By Lemma \[ref-4.2-14\] and : $$[L_{\mu_1}, I_{\mu_2}]-(-1)^{\vert {\mu_1} \vert} I_{[{\mu_1}, \mu_2]}\cong 0$$ Thus $$\begin{aligned}
I_{[{\mu_1}, \mu_2]}&\cong (-1)^{\vert {\mu_1}\vert}(L_{\mu_1} I_{\mu_2}-(-1)^{\vert L_{\mu_1} \vert \vert I_{\mu_2}\vert }I_{\mu_2} L_{{\mu_1}})\\
&=(-1)^{\vert {\mu_1} \vert }( L_{\mu_1} I_{\mu_2}-(-1)^{\vert {\mu_1} \vert (\vert \mu_2 \vert +1)}I_{\mu_2} L_{\mu_1})\\
&=(-1)^{\vert {\mu_1} \vert } L_{\mu_1} I_{\mu_2}-(-1)^{\vert {\mu_1} \vert \vert \mu_2 \vert}I_{\mu_2} L_{\mu_1}\end{aligned}$$ Substituting this in we obtain $$\label{ref-10.7-57}
\nonumber x\cong
(-1)^{\vert {\mu_1}\vert+\vert \mu_2\vert
-1}
s^{d+1}((-1)^{\vert {\mu_1} \vert } L_{\mu_1} I_{\mu_2}\eta'_0
-
I_{\mu_1} L_{\mu_2}{\eta'_0})$$ Next we observe, using $$\label{ref-10.7-58}
\begin{split}
[\mathsf{b}+u\mathsf{B},I_{\mu_1} I_{\mu_2}-&(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}I_{\mu_2\cup{\mu_1}}]\\
&=[\mathsf{b}+u\mathsf{B},I_{\mu_1}]I_{\mu_2}
+(-1)^{|{\mu_1}|+1}I_{\mu_1}[\mathsf{b}+u\mathsf{B},I_{\mu_2}]-(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}[\mathsf{b}+u\mathsf{B},I_{{\mu_2}\cup \mu_1}]\\
&=u(L_{\mu_1} I_{\mu_2}+(-1)^{|{\mu_1}|+1}I_{\mu_1} L_{\mu_2}-(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}L_{\mu_2\cup {\mu_1}})
\end{split}$$ and also using : $$\begin{aligned}
I_{\mu_1} I_{\mu_2}-(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}I_{\mu_2\cup{\mu_1}}\mod u&=
i_{\mu_1} i_{\mu_2}-(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}i_{\mu_2\cup{\mu_1}}\\
&=0\end{aligned}$$ In other words $I_{\mu_1} I_{\mu_2}-(-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}I_{\mu_2\cup{\mu_1}}$ is divisible by $u$ and we obtain from $$L_{\mu_1} I_{\mu_2}+(-1)^{|{\mu_1}|+1}I_{\mu_1} L_{\mu_2}\cong (-1)^{(|{\mu_1}|+1)(|\mu_2|+1)}L_{\mu_2\cup {\mu_1}}$$ Substituting this back in we find $$\begin{aligned}
x&\cong (-1)^{|{\mu_1}|(|\mu_2|+1)}
s^{d+1}
L_{\mu_2\cup {\mu_1}}
{\eta'_0}\\
&= (-1)^{|{\mu_1}|(|\mu_2|+1)}s^{d+1}\mathsf{B}i_{\mu_2\cup{\mu_1}} \pi({\eta'_0})
&& (\text{by Lemma \ref{ref-10.3-52}})\\
&\cong(-1)^{|\mu_2|+1}s^{d+1}\mathsf{B}i_{\mu_1\cup{\mu_2}} \pi({\eta'_0})&&(\text{by \S\ref{ref-4.2-4}})\\
&\cong (-1)^{|\mu_2|+1} (-1)^{(|\mu_1|+|\mu_2|+1)(d+1)}\mathsf{B}i_{\mu_1\cup{\mu_2}} \pi({\eta_0})\end{aligned}$$ and hence by $$\bar{x}=(-1)^{|\mu_2|+1} (-1)^{(|\mu_1|+|\mu_2|+1)(d+1)}\mathsf{B}(i_{\mu_1}\pi(\bar{\eta}_0)\cdot i_{\mu_2}\pi(\bar{\eta}_0))$$
To compute the righthand side of we note $$\begin{aligned}
\pi(s^{d-1}\Psi(\mu_i,\eta_i'))&=\pi(s^{d+1}((-1)^{|\mu_i|-1}I_{\mu_i}\eta'_0+u\eta'_i)) &&(\text{by \eqref{ref-10.1-49}})\\
&=(-1)^{|\mu_i|-1}(-1)^{(|\mu_i|+1)(d+1)}i_{\mu_i}\pi(\eta_0)\end{aligned}$$ so that $$\begin{aligned}
[\overline{s^{d-1}\Psi(\mu_1,\eta'_1)},\overline{s^{d-1}\Psi(\mu_2,\eta'_2)}]&=(-1)^{|\mu_1|+d}\mathsf{B}
(\overline{s^{d-1}\Psi(\mu_1,\eta'_1)}\cdot \overline{s^{d-1}\Psi(\mu_1,\eta'_1)})\\
&=(-1)^{|\mu_1|+d+|\mu_1|+|\mu_2|}(-1)^{(|\mu_1|+|\mu_2|)(d+1)}\mathsf{B}(\overline{i_{\mu_1}\pi(\eta_0)}\cdot \overline{i_{\mu_2}\pi(\eta_0)})\\
&=(-1)^{|\mu_2|+1} (-1)^{(|\mu_1|+|\mu_2|+1)(d+1)}\mathsf{B}(i_{\mu_1}\pi(\bar{\eta}_0)\cdot i_{\mu_2}\pi(\bar{\eta}_0))\end{aligned}$$ finishing the proof.
The commutative case {#ref-11-59}
====================
In this section we will use formality results from [@Dolgushev; @Ko3; @Shoikhet2; @Tsygan; @Willwacher1] so we will assume that the ground field $k$ contains $\mathbb{R}$. It is likely that this condition can be removed by using the methods from [@dtt2; @dtt] but we have not checked it.
Let $A=\mathcal{O}(X)$ where $X$ is a smooth affine $d$-dimensional Calabi-Yau variety over $k$. Let $T^{{\operatorname{poly}},\bullet}(A)$ be the Lie algebra of poly-vector fields on $X$. We assume that $T^{{\operatorname{poly}},\bullet}(A)$ is shifted in such a way that the Lie bracket has degree zero. Similarly let $\Omega^\bullet(A)$ be the differential forms on $X$ (not shifted).
We fix a volume form $\eta\in \Omega^d(A)$. The Hochschild-Kostant-Rosenberg-map furnishes an isomorphism ${\operatorname{HH}}_d(A)\cong \Omega^d(A)$. So we may consider $\eta$ as an element in ${\operatorname{HH}}_d(A)$ and hence by Proposition \[ref-5.7-25\] as a cycle (still denoted by $\eta$) in ${\operatorname{CC}}^-_d(A)$. It is well-known and easy to see that $(A,\eta)$ is a Calabi-Yau algebra in the sense of Ginzburg. Let $${\operatorname{div}}:T^{{\operatorname{poly}},\bullet}(A){\longrightarrow}T^{{\operatorname{poly}},\bullet-1}(A)$$ be the divergence operator corresponding to $\eta$ (see §\[ref-11.4-68\] below). The divergence is a differential which acts as a derivation with respect to the Lie bracket on $T^{{\operatorname{poly}},\bullet}(A)$.
In this section we will prove the following result
\[ref-11.1-60\] There exists an isomorphism $$(T^{{\operatorname{poly}},\bullet}(A)[[u]],-u{\operatorname{div}})\cong \mathcal{D}^\bullet(A,\eta)$$ in the homotopy category of DG-Lie algebras which fits in a diagram like .
Recall that the homotopy category of DG-Lie algebras is the category of DG-Lie algebras with quasi-isomorphisms formally inverted.
Semi-direct products for $L_\infty$-algebras {#ref-11.1-61}
--------------------------------------------
We remind the reader of a few basic definition regarding $L_\infty$-algebras and modules. Let $\mathfrak{h}^\bullet$ be a graded $k$-vector space. Recall that an $L_\infty$-structure on $\mathfrak{h}^\bullet$ is a square zero, degree one coderivation $Q$ on the symmetric coalgebra $S^c(\Sigma
\mathfrak{h}^\bullet)$. Such an $L_\infty$-structure is determined by its Taylor coefficients $\partial^n Q$ which are maps $S^n(\Sigma
\mathfrak{h}^\bullet){\longrightarrow}\Sigma \mathfrak{h}^\bullet$. *Here and in related situations below we always assume that zeroth order Taylor coefficient are zero.*
A DG-Lie algebra can be made into an $L_\infty$-algebra by putting $\partial^1Q(sg)=-sdg$, $\partial^2Q(sg,sh)=(-1)^{|g|}s[g,h]$, $\partial^nQ=0$ for $n\ge 3$.
A morphism of $L_\infty$-algebras $\psi:(\mathfrak{g}^\bullet,Q)\rightarrow(\mathfrak{h}^\bullet,Q)$ is a coalgebra morphism $\psi:S^c(\Sigma\mathfrak{g}^\bullet){\longrightarrow}S^c(\Sigma\mathfrak{h}^\bullet)$ commuting with $Q$. It is also determined by its Taylor coefficients $\partial^n\psi:S^n(\Sigma\mathfrak{g}^\bullet){\longrightarrow}\Sigma
\mathfrak{h}^\bullet$.
If $V^\bullet$ is a graded $k$-vector space then an $L_\infty$-$\mathfrak{h}^\bullet$-module structure on $V^\bullet$ is a square zero, degree one differential $R:S^c(\Sigma \mathfrak{h}^\bullet)\otimes V^\bullet
\rightarrow S^c(\Sigma \mathfrak{h}^\bullet)\otimes V^\bullet$ satisfying $$(Q\otimes {\operatorname{Id}}_{S^c\mathfrak{h}}\otimes {\operatorname{Id}}_V+{\operatorname{Id}}_{S^c\mathfrak{h}}\otimes R)\circ(\Delta\otimes {\operatorname{Id}}_V)= (\Delta\otimes {\operatorname{Id}}_V)\circ R$$ as maps from $S^c(\Sigma\mathfrak{h}^\bullet)\otimes V^\bullet$ to $S^c(\Sigma\mathfrak{h}^\bullet)\otimes S^c(\Sigma\mathfrak{h}^\bullet) \otimes V^\bullet$ An $L_\infty$-$\mathfrak{h}^\bullet$-module structure $R$ on $V^\bullet$ is entirely determined by the maps $\partial^{n+1}R:S^n(\Sigma\mathfrak{h}^\bullet)\otimes V^\bullet{\longrightarrow}V^\bullet$. If $\mathfrak{h}^\bullet$ is a DG-Lie algebra and $V^\bullet$ is a DG-module over it then $V^\bullet$ can be made into an $L_\infty$-module over $\mathfrak{h}$ by putting $\partial^1R(v)=dv$, $\partial^2R(sg,v)=g\cdot v$, $\partial^nR=0$ for $n\ge 3$.
If $V^\bullet$ is an $L_\infty$-$\mathfrak{h}^\bullet$-module then so are $\Sigma^m V^\bullet$ for all $m$ using the obvious sign convention $\partial^{n+1}R(sg_1,\ldots,sg_n,s^mv)=(-1)^{m(n+|g_1|+\cdots+|g_n|)}\partial^{n+1}R(sg_1,\ldots,sg_n,v)$. We may combine the $L_\infty$-structures on $\mathfrak{h}^\bullet$ and $\Sigma V^\bullet$ to make the direct sum $\mathfrak{h}^\bullet\oplus
V^\bullet$ into an $L_\infty$-algebra. We will denote the resulting $L_\infty$-algebra by $\mathfrak{h}^\bullet\ltimes V^\bullet$ and call it the *semi-direct product* of $\mathfrak{h}^\bullet$. This is an obvious generalization of the semi-direct product of a DG-Lie algebra with a DG-module which was used in §\[ref-8-36\].
Assume that $(V^\bullet,R)$, $(W^\bullet,R)$ are $L_\infty$-$\frak{h}^\bullet$-modules. An $L_\infty$ morphism $\mu:V^\bullet{\longrightarrow}W^\bullet$ is a comodule map $\mu: S^c(\Sigma \mathfrak{g})\otimes V^\bullet
{\longrightarrow}S^c(\Sigma \mathfrak{g})\otimes W^\bullet$. commuting with $R$. It is determined by its Taylor coefficients $\partial^n \mu: S^n(\Sigma \mathfrak{h}^\bullet)
\otimes V^\bullet{\longrightarrow}W^\bullet$.
Given in addition an $L_\infty$-morphism $\psi:\mathfrak{g}^\bullet{\longrightarrow}\mathfrak{h}^\bullet$ the pullback $V_\psi^\bullet$ of $V^\bullet$ is defined as follows: $$\begin{gathered}
\partial^{n+1} R_{\psi}(sg_1,\ldots,sg_n,v)
=\\\sum_{t,1\le m_1<\cdots < m_{t-1}<n}
\pm \partial^{t+1}R(\partial^{m_1}\psi(sg_{i_1},\ldots,sg_{i_{m_1}}),
\partial^{m_2-m_1}\psi(sg_{i_{m_1+1}},\ldots,sg_{i_{m_2}}),\\
\ldots,
\partial^{n-m_{t-1}}\psi(sg_{i_{m_{t-1}+1}},\ldots,sg_{n}),v)\end{gathered}$$ where for all $j$: $i_{m_j+1}<\cdots<i_{m_{j+1}}$ and the sign is the Koszul sign of the corresponding shuffle of the $(sg_i)_i$. By construction we have a canonical $L_\infty$-morphism $$\psi_V:\mathfrak{g}^\bullet\ltimes V^\bullet_\psi {\longrightarrow}\mathfrak{h}^\bullet\ltimes V^\bullet$$ which restricted to $S^n(\Sigma \mathfrak{g})$ coincides with $\partial^n\psi$.
Twisting {#ref-11.2-62}
--------
Assume that $\psi:\frak{g}^\bullet{\longrightarrow}\frak{h}^\bullet$ is a $L_\infty$-morphism between $L_\infty$-algebras equipped with some type of topology and let $\omega\in \frak{g}^1$ be a Maurer-Cartan element in $\mathfrak{g}^1$, i.e. a solution of the $L_\infty$-Maurer-Cartan equation $$\sum_{i\ge 1} \frac{1}{i!} (\partial^i Q)(\underbrace{\omega\cdots\omega}_i)=0$$ One has to assume that one is in a situation where all occurring series are convergent and standard series manipulations are allowed. In our application below the series are in fact finite.
Define $Q_\omega$, $\psi_\omega$ and $\omega'$ by [@ye3] $$\begin{aligned}
\label{ref-11.1-63}
(\partial^i Q_\omega)(\gamma)&=\sum_{j\ge 0} \frac{1}{j!} (\partial^{i+j} Q)
(\underbrace{\omega\cdots\omega}_j \gamma)\qquad \text{(for $i>0$)}\\
\label{ref-11.2-64}
(\partial^i \psi_\omega)(\gamma)&=\sum_{j\ge 0} \frac{1}{j!} (\partial^{i+j} \psi)
(\underbrace{\omega\cdots \omega}_j \gamma)\qquad \text{(for $i>0$)}\\
\label{ref-11.3-65}
\omega'&=\sum_{j\ge 1} \frac{1}{j!} (\partial^{j} \psi)
(\underbrace{\omega\cdots\omega}_j )\end{aligned}$$ for $\gamma\in S^i(\Sigma\frak{g}^\bullet)$. Then e.g. by [@ye3] $\omega'$ is a solution of the Maurer-Cartan equation on $\frak{h}^\bullet$ and furthermore $\frak{g}^\bullet$, $\frak{h}^\bullet$, when equipped with $Q_\omega$, $Q_{\omega'}$ are again $L_\infty$-algebras. Let us denote these by $\frak{g}^\bullet_\omega$ and $\frak{h}^\bullet_{\omega'}$. Finally $\psi_\omega$ is an $L_\infty$ map $\frak{g}^\bullet_\omega{\longrightarrow}\frak{h}_{\omega'}^\bullet$.
Applying formality to [ $\mathfrak{D}(A,\eta)$]{}
-------------------------------------------------
By [@CF; @Ko3; @ye3] there is an $L_\infty$-quasi-isomorphism $$\mathfrak{U}:T^{{\operatorname{poly}},\bullet}(A)\rightarrow \bar{\mathfrak{C}}^\bullet(A)$$ such that $\partial^1 \mathfrak{U}$ is the standard Hochschild-Kostant-Rosenberg quasi-isomorphism.
View $(\overline{{\operatorname{CC}}}^{\,-}_\bullet(A),\mathsf{b}+u\mathsf{B})$ as an $L_\infty$-module over $T^{{\operatorname{poly}},\bullet}(A)$ via $\mathfrak{U}$ as in §\[ref-11.1-61\]. We also view $(\Omega^\bullet(A)[[u]],ud)$ as a DG-Lie module over $T^{{\operatorname{poly}},\bullet}(A)$ via the Lie derivative. Then by [@Dolgushev; @Shoikhet2; @Tsygan; @Willwacher1] there is an $L_\infty$-quasi-morphism of $L_\infty$-modules over $T^{{\operatorname{poly}},\bullet}(A)$ $$\mathfrak{S}:(\overline{{\operatorname{CC}}}^{\,-}_\bullet(A),\mathsf{b}+u\mathsf{B}){\longrightarrow}(\Omega^\bullet(A)[[u]],ud)$$ where $\partial^1\mathfrak{S}$ is again the HKR quasi-isomorphism. Thus we get a roof of $L_\infty$-quasi-morphisms of graded DG-Lie algebras $${\tiny
\label{ref-11.4-66}
\xymatrix{& T^{{\operatorname{poly}},\bullet}(A)\ltimes \Sigma^{-d-1}\overline{{\operatorname{CC}}}^{\,-}_\bullet(A)\ar[dl]_{\mathfrak{S}}\ar[dr]^{\mathfrak{U}}&\\
T^{{\operatorname{poly}},\bullet}(A)\ltimes \Sigma^{-d-1}\Omega^\bullet(A)[[u]] &&
\bar{\mathfrak{C}}^{\bullet}(A)\ltimes \Sigma^{-d-1}\overline{{\operatorname{CC}}}^{\,-}_\bullet(A)
}
}$$
We obtain a new roof by twisting with $(0,\eta')$ where $\eta'=s^{-d-1}\eta$. $$\label{ref-11.5-67}
{\tiny
\xymatrix{& (T^{{\operatorname{poly}},\bullet}(A)\ltimes \Sigma^{-d-1}\overline{{\operatorname{CC}}}^{\,-}_\bullet(A)\ar[dl]_{\mathfrak{S}_{(0,\eta')}}\ar[dr]^{\mathfrak{U}_{(0,\eta')}})_{(0,\eta')}&\\
\mathfrak{T}^\bullet(A,\eta)
&&
\mathfrak{D}^\bullet(A,\eta)
}
}$$ where $$\mathfrak{T}^\bullet(A,\eta)=(T^{{\operatorname{poly}},\bullet}(A)\ltimes \Sigma^{-d-1}\Omega^\bullet(A)[[u]])_{(0,\eta')}$$ The complexes here are are $2$-step filtered. The arrows are quasi-isomorphisms since if we take the associated graded complexes for the $2$-step filtrations we find the same arrows as in .
Divergence etc… {#ref-11.4-68}
---------------
The divergence operator is defined by $${\operatorname{div}}: T^{\bullet,{\operatorname{poly}}}(A){\longrightarrow}T^{\bullet-1,{\operatorname{poly}}}(A)$$ via the following identity $$d(\gamma \cap \eta)={\operatorname{div}}\gamma \cap \eta$$ We conclude immediately $${\operatorname{div}}^2=0$$ and furthermore the following is true [@Schechtman]: $$(-1)^{|\gamma_1|}[\gamma_1,\gamma_2]={\operatorname{div}}(\gamma_1\gamma_2)-{\operatorname{div}}(\gamma_1)\gamma_2-(-1)^{|\gamma_1|+1}\gamma_1{\operatorname{div}}\gamma_2$$ So $(T^{{\operatorname{poly}},\bullet}(A),-{\operatorname{div}},\cup)$ is a BV-algebra (see App. \[ref-A-76\]).
\[ref-11.2-69\] There is an $L_\infty$-isomorphism of DG-Lie algebras $$\delta:(T^{{\operatorname{poly}},\bullet}(A)[[u]],-u{\operatorname{div}}){\longrightarrow}\mathfrak{T}^\bullet(A)$$
According to Proposition \[ref-A.4-81\] there exists an $L_\infty$-isomorphism $$(T^{{\operatorname{poly}},\bullet}(A)[[u]],-u{\operatorname{div}}){\longrightarrow}(T^{{\operatorname{poly}},\bullet}(A)\ltimes \mathfrak{a},-u{\operatorname{div}})$$ where $\mathfrak{a}$ is the abelian graded Lie algebra on the vector space $u T^{{\operatorname{poly}},\bullet}(A)[[u]]$. The action of $T^{{\operatorname{poly}},\bullet}(A)$ on $\mathfrak{a}$ is given by $$\gamma\star a=[\gamma,a]+(-1)^{|\gamma|}{\operatorname{div}}\gamma\cup a$$ To finish the proof it is sufficient to show that the following map $$\delta':(T^{{\operatorname{poly}},\bullet}(A)\ltimes \mathfrak{a},-u{\operatorname{div}}){\longrightarrow}\mathfrak{T}^\bullet(A,\eta)=(T^{{\operatorname{poly}},\bullet}(A)\ltimes \Sigma^{-d-1}\Omega^\bullet(A)[[u]])_{(0,\eta')}
:(\gamma,a)\mapsto (\gamma,(-1)^{|a|}u^{-1}a\cap \eta')$$ is an isomorphism of DG-Lie algebras. First we show that $$\delta': \mathfrak{a}{\longrightarrow}\Sigma^{-d-1}\Omega^\bullet(A)[[u]]:a\mapsto (-1)^{|a|}u^{-1}(a\cap \eta')$$ is compatible with the action of $T^{{\operatorname{poly}},\bullet}(A)$. We compute for $\gamma\in T^{{\operatorname{poly}},\bullet}(A)$ and $a\in \mathfrak{a}$. $$\begin{aligned}
\delta'(\gamma\star {a})&=\delta'([\gamma,{a}]+(-1)^{|\gamma|}{\operatorname{div}}\gamma\cup {a})\\
&=(-1)^{|\gamma|+|{a}|}u^{-1}([\gamma,{a}]+(-1)^{|\gamma|}{\operatorname{div}}\gamma\cup {a})\cap \eta'\\
&=(-1)^{|\gamma|+|{a}|}u^{-1}((-1)^{|\gamma|}{\operatorname{div}}(\gamma\cup {a})-(-1)^{|\gamma|}{\operatorname{div}}(\gamma)\cup
{a}+\gamma\cup
{\operatorname{div}}({a})+(-1)^{|\gamma|}{\operatorname{div}}(\gamma)\cup {a})\cap \eta'
\\
&=(-1)^{|\gamma|+|{a}|}u^{-1}((-1)^{|\gamma|}{\operatorname{div}}(\gamma\cup {a})+\gamma\cup {\operatorname{div}}{a})\cap \eta'\\
&=(-1)^{|\gamma|+|{a}|}u^{-1} ((-1)^{|\gamma|}d(\gamma\cap({a}\cap \eta'))+\gamma \cap d({a}\cap \eta'))\\
&=L_\gamma(\delta'({a}))\end{aligned}$$ Now we check compatibility with the differential of $\delta'$ on an element $a\in\mathfrak{a}$. $$\begin{aligned}
\delta'(-u{\operatorname{div}}a)&=-(-1)^{|a|+1}{\operatorname{div}}a\cap \eta'\\
&=(-1)^{|a|}d(a\cap \eta')\\
&=d(\delta'(a))\end{aligned}$$ Finally we check compatibility with the differential of $\delta'$ on $\gamma\in T^{{\operatorname{poly}},\bullet}(A)$. $$\begin{aligned}
\delta'(-u{\operatorname{div}}\gamma)&=-(-1)^{|\gamma|+1}{\operatorname{div}}\gamma\cap \eta'\\
&=(-1)^{|\gamma|}d(\gamma\cap \eta')\\
&=(-1)^{|\gamma|}L_a\eta'\\
&=[(0,\eta'),(\gamma,0)]\qed\end{aligned}$$
It suffices to combine diagram with Proposition \[ref-11.2-69\], taking into account that an $L_\infty$-quasi-isomorphism yields an isomorphism in the homotopy category of DG-Lie algebras via the bar cobar construction.
Obstructions
============
Let $\mathfrak{g}^\bullet$ be a DG-Lie algebra and let $(S,n)\rightarrow (R,m)$ be a surjective morphism in ${\operatorname{Nilp}}$ with one-dimensional kernel $ks\subset n$. Let $x\in \mathfrak{g}^1\otimes m$ be a solution to the Maurer-Cartan equation. Lift $x$ to an arbitrary element $\hat{x}$ of $\mathfrak{g}^1\otimes n$ and let $p(\hat{x})\in
\mathfrak{g}^2$ be such that $p(\hat{x})s=d\hat{x}+\frac{1}{2}[\hat{x},\hat{x}]$. Then clearly $dp(\hat{x})=0$ and furthermore the cohomology class $o(x)\overset{\text{def}}{=}\overline{p(\hat{x})}\in
H^1(\mathfrak{g}^\bullet)$ does not depend on the chosen lift $\hat{x}$ of $x$. It is easy to see that $o(x)=0$ if and only if $x$ can be lifted to an element of ${\operatorname{MC}}(\mathfrak{g}^\bullet\otimes n)$. Consequently $o(x)$ is called the *obstruction class* of $x$.
The *obstruction space* $O(\mathfrak{g}^\bullet)$ is the linear span in $\mathfrak{g}^2$ of all $o(x)$ for all morphisms $(S,n)\rightarrow (R,m)$ with one-dimensional kernel and all $x\in {\operatorname{MC}}(\mathfrak{g}^\bullet\otimes m)$ as above.
Clearly $o(x)$ and hence $O(\mathfrak{g}^\bullet)$ is functorial under DG-Lie algebra morphisms. It is well-known and easy to see that this functoriality extends to $L_\infty$-morphisms.
Recall that the *periodic cyclic complex* ${\operatorname{CC}}_\bullet^{\text{per}}(A)$ of a $k$-algebra $A$ is obtained by inverting $u$ in ${\operatorname{CC}}^-_\bullet(A)$. Its homology will be denoted by ${\operatorname{HC}}_\bullet^{\text{per}}(A)$. The following is the main result of this section.
\[ref-12.1-70\] Let $(A,\bar{\eta})$ be a $d$-Calabi-Yau algebra. Then the composition $$O(\mathfrak{D}^\bullet(A,\eta))\hookrightarrow H^2(\mathfrak{D}^\bullet(A,\eta))
\overset{\text{Thm \ref{ref-10.1-48}}}{\cong} {\operatorname{HC}}_{d-3}^-(A)\rightarrow {\operatorname{HC}}_{d-3}^{\text{per}}(A)$$ is zero.
The proof depends on the following beautiful result by Tsygan and Daletskii [@Tsygan1 Thm 1] (see also [@DaTs]).
The Lie action of $\mathfrak{C}^\bullet(A)$ on ${\operatorname{CC}}^-_\bullet(A)$ can be extended to a $u$-linear $L_\infty$-action of the DG-Lie algebra $(\mathfrak{C}^\bullet(A)[u,\epsilon],d+u\partial/\partial\epsilon)$, with $|\epsilon|=1$, $\epsilon^2=0$ and such that $$\begin{aligned}
\partial^1R(\gamma)&=d\gamma\\
\partial^2R(s\sigma,\gamma)&=L_\sigma \gamma\\
\partial^2R(s(\epsilon \sigma),\gamma)&=I_\sigma \gamma\end{aligned}$$ for $\sigma\in \mathfrak{C}^\bullet(A)$, $\gamma\in{\operatorname{CC}}^-_\bullet(A)$.
The statement about $\partial^2R(s(\epsilon \sigma),\gamma)$ does not occur in [@Tsygan Thm 1] but it follows easily from the proof.
In the rest of this section $(A,\bar{\eta})$ is a $d$-Calabi-Yau algebra.
There is a commutative diagram of complexes $$\label{ref-12.1-71}
\xymatrix{
(\mathfrak{C}^\bullet(A)\ltimes \Sigma^{-d-1}{\operatorname{CC}}^-_\bullet(A))_{(0,\eta')}
\ar[rr]\ar[dr]_{\Psi} && (\mathfrak{C}^\bullet(A)[u,\epsilon]\ltimes \Sigma^{-d-1}{\operatorname{CC}}^-_\bullet(A))_{(0,\eta')}\ar[dl]^{\Psi'}\\
&
\Sigma^{-d+1}{\operatorname{CC}}^-_\bullet(A))&
}$$ where
- $\Psi$ was introduced in Theorem \[ref-10.1-48\];
- $\eta=s^{-d-1}\eta'$;
- the horizontal map is a twist (see §\[ref-11.2-62\]) of the map obtained from the obvious inclusion of DG-Lie algebras $(\mathfrak{C}^\bullet(A),d)\hookrightarrow (\mathfrak{C}^\bullet(A)[u,\epsilon],d+\partial/\partial\epsilon)$.
- $\Psi'$ restricted to $\mathfrak{C}^\bullet(A)[u,\epsilon]$ is $u$-linear and satisfies $$\label{ref-12.2-72}
\begin{aligned}
\Psi'(\sigma)&=(-1)^{|\sigma|+1}I_\sigma \eta'\\
\Psi'(\epsilon\sigma)&=0
\end{aligned}$$ for $\sigma\in \mathfrak{C}^\bullet(A)$.
- $\Psi'$ restricted to $\Sigma^{-d-1}{\operatorname{CC}}_\bullet^-(A)$ is multiplication by $u$.
The commutativity of the diagram is clear. We only have to show that $\Psi'$ commutes with the differential. For $\Psi'$ restricted to $\Sigma^{-d-1}{\operatorname{CC}}_\bullet^-(A)$ this is obvious. As far as the restriction of $\Psi'$ to $\mathfrak{C}^\bullet(A)[u,\epsilon]$ is concerned: the only non-trivial case (given that $\Psi$ already commutes with the differential) is the evaluation on an element of $\epsilon\mathfrak{g}$.
Using we find for $\sigma \in \mathfrak{C}^\bullet(A)$ $$d_{(0,\eta')}(\epsilon\sigma)=(d(\epsilon \sigma),(-1)^{|g|}I_\sigma\eta')$$ Given we have to show $$\Psi'(d_{(0,\eta')}(\epsilon\sigma)=0$$ We compute $$\begin{aligned}
\Psi' (d_{(0,\eta')}(\epsilon \sigma))&=\Psi'(d(\epsilon \sigma),(-1)^{|\sigma|}I_\sigma\eta')\\
&=\Psi'(-\epsilon d\sigma+u\sigma,(-1)^{|\sigma|}I_\sigma\eta')\\
&=(-1)^{|\sigma|+1}uI_\sigma\eta'+(-1)^{|\sigma|}uI_\sigma\eta'\\
&=0\qed\end{aligned}$$
\[ref-12.4-73\] Consider $\Sigma^{-d+1}{\operatorname{CC}}^{\text{per}}_\bullet(A)$ as an abelian DG-Lie algeba. Then there exists an $L_\infty$-morphism $$\Delta:\mathfrak{D}^\bullet(A,\eta)\rightarrow \Sigma^{-d+1}{\operatorname{CC}}^{\text{per}}_\bullet(A)$$ such that the following diagram is commutative $$\xymatrix{
H^\bullet(\mathfrak{D}^\bullet(A,\eta))\ar[drr]_{H^\bullet(\Delta)}\ar[rr]^-{H^\bullet(\Psi)}_-{}&&
H^\bullet(\Sigma^{-d+1}{\operatorname{CC}}^{-}_\bullet(A))\ar[d]^{\text{canonical}}\\
&&H^\bullet(\Sigma^{-d+1}{\operatorname{CC}}^{\text{per}}_\bullet(A))
}$$
To simplify the notations put $\mathfrak{g}^\bullet=\mathfrak{C}^\bullet(A)$, $V^-=\Sigma^{-d-1}{\operatorname{CC}}_\bullet^-(A)$, $V^{\text{per}}=\Sigma^{-d-1}{\operatorname{CC}}_\bullet^-(A)$. Thus we get $L_\infty$-morphisms (see §\[ref-11.1-61\],§\[ref-11.2-62\]) $$\begin{gathered}
\label{ref-12.3-74}
(\mathfrak{g}^\bullet\ltimes V^-)_{(0,\eta')}\rightarrow
(\mathfrak{g}^\bullet[u,\epsilon]\ltimes V^-)_{(0,\eta')}
\rightarrow
(\mathfrak{g}^\bullet[u,u^{-1},\epsilon]\ltimes V^{\text{per}})_{(0,\eta')}
\xleftarrow{c} (0\ltimes V^{\text{per}})_{(0,\eta')}
\\\cong
V^{\text{per}}
\overset{\times u}\cong \Sigma^2 V^{\text{per}}\end{gathered}$$ Here $c$ goes in the wrong direction but it is easy to see that $(\mathfrak{g}^\bullet[u,u^{-1},\epsilon],d+u\partial/\partial\epsilon)$ is acyclic. Hence $c$ is an quasi-isomorphism. This means that there is an $L_\infty$-quasi-isomorphism $c'$ which goes in the opposite direction and which inverts $c$ on the level of cohomology. Taking the composition of everything we obtain an $L_\infty$-morphism $$(\mathfrak{g}^\bullet\ltimes V^-)_{(0,\eta')}{\longrightarrow}\Sigma^2V^{\text{per}}$$ which is the desired $\Delta$.
It remains to show that $\Delta$ and $\Psi$ are compatible on the level of cohomology. This follows from the following commutative diagram whose upper row is a compressed version of and whose lower row we obtain from . $$\xymatrix{
(\mathfrak{g}^\bullet\ltimes V^-)_{(0,\eta')}\ar@/^2em/[rrrr]^{\Delta}\ar[r]\ar[d]_{\Psi}&(\mathfrak{g}^\bullet[u,u^{-1},\epsilon]\ltimes V^{\text{per}})_{(0,\eta')}\ar[d]^{\Psi'}\ar@<1ex>[rr]^-{c'}&\cong&V^{\text{per}}\ar@{=}[d]\ar@<1ex>[ll]^-{c}\ar[r]_{\times u}&\Sigma^2 V^{\text{per}}\ar@{=}[d]\\
\Sigma^2 V^-\ar[r]\ar@/_2em/[rrrr]_{\text{canonical}}&\Sigma^2 V^{\text{per}}&&V^{\text{per}}\ar[ll]^{\times u}\ar[r]^{\times u}&\Sigma^2 V^{\text{per}}
}\qed$$
The theorem follows from Lemma \[ref-12.4-73\] together with the functoriality of obstruction spaces under $L_\infty$-morphisms and the fact that the obstruction space of an abelian Lie algebra is trivial.
\[ref-12.5-74\] If the map ${\operatorname{HC}}^{-}_{d-3}(A){\longrightarrow}{\operatorname{HC}}^{\text{per}}_{d-3}(A)$ is injective then the deformation theory of $A$ is unobstructed.
This corollary applies for example in the case $d\le 3$ by the following well-known lemma.
\[ref-12.6-75\] ${\operatorname{HC}}^{-}_{n}(A){\longrightarrow}{\operatorname{HC}}^{\text{per}}_{n}(A)$ is an isomorphism for $n\le 0$.
There is an exact sequence $${\operatorname{HC}}_{n-1}(A) {\longrightarrow}{\operatorname{HC}}^{-}_{n}(A){\longrightarrow}{\operatorname{HC}}^{\text{per}}_{n}(A){\longrightarrow}{\operatorname{HC}}_{n-2}(A)$$ (e.g. [@Loday1 Prop. 5.1.5]) where ${\operatorname{HC}}_\bullet(A)$ denotes ordinary cyclic homology. The complex computing ordinary cyclic homology is concentrated in homological degrees $\ge 0$. Hence ${\operatorname{HC}}_n(A)=0$ for $n<0$. This finishes the proof.
Many 3-dimensional Calabi-Yau algebras are obtained from superpotentials (see [@Bocklandt; @VdBsuper]). For those it is is not very surprising that the deformation theory is unobstructed (the deformations come from deforming the superpotential). However there are examples of 3-dimensional Calabi-Yau algebras which are not obtained from superpotentials. See e.g. [@davison]. Simple examples are given by 3-dimensional smooth commutative Calabi-Yau algebras with no exact volume form.
A technical result on BV-algebras {#ref-A-76}
=================================
Recall that a DG-BV-algebra is a quadruple $(\mathfrak{g}^\bullet,d,\Delta,\cup)$ where $(\mathfrak{g}^\bullet,d)$ is a complex, $\cup$ is a commutative, associative product of degree[^4] $1$ on $\mathfrak{g}^\bullet$ compatible with $d$, $\Delta$ is a differential of degree $-1$, $(\mathfrak{g}^\bullet,d,[-,-])$ is a DG-Lie algebra with $[-,-]$ defined by: $$[g,h]=(-1)^{|g|+1}(\Delta(g\cup h)-\Delta g\cup h-(-1)^{|g|+1}g\cup \Delta h)$$ and $\cup$, $[-,-]$ are related by the Leibniz rule: $$[g,h_1\cup h_2]=[g,h_1]\cup h_2+(-1)^{|g|(|h_1|+1)}h_1\cup [g,h_2]$$ It is shown in [@KKP; @Terilla] that if $\mathfrak{h}^\bullet$ is a DG-BV-algebra then $(\mathfrak{h}^\bullet((u)),d+u\Delta)$ is homotopy abelian. The same proof works for $u\mathfrak{h}^\bullet[[u]],d+u\Delta)$ but not for $(\mathfrak{h}^\bullet[[u]],d+u\Delta)$. Our aim in this section is to make $(\mathfrak{h}^\bullet[[u]],d+u\Delta)$ as “commutative as possible” (see Proposition \[ref-A.4-81\] below) by making at least its sub-DG-Lie algebra $(u\mathfrak{h}^\bullet[[u]],d+u\Delta)$ abelian. This is not completely straightforward since in order to do this we have to twist the action of $\mathfrak{h}^\bullet$ on $u\mathfrak{h}^\bullet[[u]]$.
The fact that $(\mathfrak{h}^\bullet((u)),d+u\Delta)$ and $(u\mathfrak{h}^\bullet[[u]],d+u\Delta)$ are homotopy abelian is in fact a special case of a general result in [@ShTa]. For the benefit of the reader we repeat the proof of this result. Afterwards we will reuse the proof to treat $(\mathfrak{h}^\bullet[[u]],d+u\Delta)$.
It is convenient to use the following adhoc definition.
A $\BVm$ algebra is a DG-Lie algebra $\mathfrak{g}^\bullet$ equipped with a commutative, associative product $\cup$ of degree $-1$, compatible with $d$, such that $$\label{ref-A.1-77}
[g,h]=(-1)^{|g|+1}(d(g\cup h)-dg\cup h-(-1)^{|g|+1}g\cup dh)$$ and $$\label{ref-A.2-78}
[g,h_1\cup h_2]=[g,h_1]\cup h_2+(-1)^{|g|(|h_1|+1)}h_1\cup [g,h_2]$$
[@ShTa] \[ref-A.2-79\] Let ${\mathfrak{g}^\bullet}$ be a $\BVm$-algebra and let ${\mathfrak{a}^\bullet}$ be the same as ${\mathfrak{g}^\bullet}$ but with the Lie bracket set to zero. Then there is a $L_\infty$-morphism $\psi:{\mathfrak{g}^\bullet}{\longrightarrow}{\mathfrak{a}^\bullet}$ such $\partial^1\psi$ is the identity. In other words ${\mathfrak{g}^\bullet}$ is homotopy abelian.
Let $({\mathfrak{h}^\bullet},d,\Delta,\cup)$ be a DG-BV-algebra. Then $(u{\mathfrak{h}^\bullet}[[u]],d+u\Delta,[-,-],u^{-1}(-\cup-))$ is a $\BVm$-algebra and hence by the previous lemma $(u{\mathfrak{h}^\bullet}[[u]],d+u\Delta)$ is homotopy abelian. The same reasoning applies to $(\mathfrak{h}^\bullet((u)),d+u\Delta)$.
Put $V^\bullet=\Sigma {\frak{g}^\bullet}$. The coderivation $Q$ on $S^cV^\bullet$ corresponding to the DG-Lie structure is given by $$\partial^1 Q:V^\bullet{\longrightarrow}V^\bullet:sg\mapsto -s\,dg$$ $$\partial^2 Q:S^2 V^\bullet{\longrightarrow}V^\bullet:(sg,sh)\mapsto (-1)^{|g|}s[g,h]$$ and all other $\partial^n Q$ are zero.
For simplicity of notation we put $$sg_1\cdot sg_2 \cdots sg_n=s(g_1\cup\cdots \cup g_n)$$ From we obtain: $$\label{ref-A.3-80}
\begin{aligned}
\partial^1 Q(v_1\cdot v_2\cdots v_n)&=
\sum_i \epsilon_i \partial^1 Q(v_i)v_1\cdots \hat{v}_i\cdots v_n
+\sum_{i<j} \epsilon_{i,j}\partial^2 Q(v_i,v_j)v_1\cdots \hat{v}_i\cdots \hat{v}_j\cdots v_n
\end{aligned}$$ where the signs are determined by $$\begin{aligned}
v_1\cdot v_2\cdots v_n&=\epsilon_i v_i\cdot v_1\cdots \hat{v}_i\cdots v_n\\
&=\epsilon_{i,j}v_i\cdot v_j\cdot v_1\cdots \hat{v}_i\cdots \hat{v}_j\cdots v_n\end{aligned}$$ Consider $\partial^1 Q$ as a coderivation of $S^cV^\bullet$ and let $\psi:S^cV^\bullet{\longrightarrow}S^cV^\bullet$ be the coalgebra automorphism determined by $$\partial^n\psi(v_1,\ldots,v_n)=v_1\cdot v_2\cdots v_n$$ Then becomes $$\partial^1 Q\circ \psi=\psi\circ Q$$ which finishes the proof.
\[ref-A.4-81\] Let $({\mathfrak{h}^\bullet},d,\Delta,\cup)$ be a DG-BV-algebra. Let ${\mathfrak{a}^\bullet}$ be the graded vector space $u{\mathfrak{h}^\bullet}[[u]]$. The following operation $$\label{ref-A.4-82}
h\star a=[h,a]+(-1)^{|h|+1}\Delta(h)\cup a$$ for $h\in {\mathfrak{h}^\bullet}$, $a\in {\mathfrak{a}^\bullet}$ makes ${\mathfrak{a}^\bullet}$ into a graded ${\mathfrak{h}^\bullet}$-representation. Furthermore $d+u\Delta$ defines a derivation on the Lie algebra ${\mathfrak{h}^\bullet}\ltimes {\mathfrak{a}^\bullet}$ and finally there is an $L_\infty$-isomorphism $$\phi:{\mathfrak{h}^\bullet}[[u]]{\longrightarrow}({\mathfrak{h}^\bullet}\ltimes {\mathfrak{a}^\bullet},d+u\Delta)$$ such that $\partial^1\phi$ is the identity.
In the proof below we identify the underlying vector spaces of $\mathfrak{h}^\bullet[[u]]$ and $\mathfrak{h}^\bullet\ltimes\mathfrak{a}$ in the obvious way. The fact that defines indeed a representation as well as compatibility with differentials is an easy direct verification: Now put $V^\bullet=\Sigma {\mathfrak{a}^\bullet}$, $W^\bullet=\Sigma {\mathfrak{h}^\bullet}$. Let $Q$ be the coderivation on $S^c(W^\bullet\oplus V^\bullet)$ corresponding to ${\mathfrak{h}^\bullet}[[u]]$. We observe that $\partial^1Q{\mid} W^\bullet=\partial^1 Q_1+\partial^2 Q_2$ where $\partial^1 Q_1=-d$ and $\partial^1 Q_2=-u\Delta$. Let $Q'$ be the coderivation on $S^c(W^\bullet\oplus V^\bullet)$ corresponding to $({\mathfrak{h}^\bullet}\ltimes {\mathfrak{a}^\bullet},d+u\Delta)$. We have $\partial^1Q'=\partial^1 Q$. Furthermore $$\begin{aligned}
\partial^2 Q'(w_1,w_2)&=\partial^2 Q(w_1,w_2)&&\text{for $w_1,w_2\in W^\bullet$}\\
\partial^2 Q'(v_1,v_2)&=0&&\text{for $v_1,v_2\in V^\bullet$}\end{aligned}$$ and for $h\in {\mathfrak{h}^\bullet}$, $a\in {\mathfrak{a}^\bullet}$ $$\begin{aligned}
\partial^2 Q'(sh,sa)&=(-1)^{|h|}s(h\star a)\\
&=(-1)^{|h|}s[h,a]-s(\Delta h\cup a)\\
&=\partial^2 Q(sh,sa)+\partial^1 Q_2(sh)\cdot sa\end{aligned}$$ where as above $x\cdot y=u^{-1}(x\cup y)$. In other words $$\begin{aligned}
\label{ref-A.5-83}
\partial^2 Q'(w,v)&=\partial^2 Q(w,v)+\partial^1 Q_2(w)\cdot v
&&\text{for $w\in W^\bullet$, $v\in V^\bullet$}\end{aligned}$$ We now construct the desired $L_\infty$-morphism. By definition $\partial^n \psi={\operatorname{Id}}$ for $n=1$. For $n>1$, $i\ge 1$, $w_1,\ldots,w_i\in W^\bullet$, $v_1,\ldots,v_j\in V^\bullet$ we put $$\partial^n\psi(w_1,\ldots,w_i,v_1,\ldots,v_j)=0$$ and $$\partial^n\psi(v_1,\ldots,v_j)=v_1\cdot v_2\cdots v_n$$ We now verify $$\psi \circ Q=Q'\circ \psi$$ We must evaluate both sides on $S^i W^\bullet\otimes S^j V^\bullet$. If $i=0$ then the desired equality follows from the proof of Lemma \[ref-A.2-79\]. If $i> 2$ then both sides are zero so this case is trivial as well. If $i=2$ then both sides are zero unless $j=0$ in which case we reduce to $\partial^2 Q{|} S^2W^\bullet=\partial^2 Q'{|} S^2W^\bullet$.
We concentrate on the case $i=1$. We find $$(Q'\circ \psi)(w_1,v_1,\ldots,v_j)=\partial^2Q'(w_1,v_1\cdot v_2\cdots v_n)$$ and $$\begin{aligned}
(\psi\circ Q)(w_1,v_1,\ldots,v_j)&=\partial^1Q_2(w_1)\cdot v_1\cdots v_j+\sum_l \pm\partial^2Q(w_1,v_l)\cdot v_1
\cdots \hat{v}_l\cdots v_j\\
&=\partial^1Q_2(w_1)\cdot v_1\cdots v_j+\partial^2Q(w_1,v_1\cdot v_2\cdots v_j)\end{aligned}$$ We conclude by .
[10]{}
R. Bocklandt, [*Graded [C]{}alabi [Y]{}au algebras of dimension 3*]{}, J. Pure Appl. Algebra [**212**]{} (2008), no. 1, 14–32.
D. Calaque and C. A. Rossi, [*Compatibility with cap-products in [T]{}sygan’s formality and homological [D]{}uflo isomorphism*]{}, Lett. Math. Phys. [**95**]{} (2011), no. 2, 135–209.
D. Calaque, C. A. Rossi, and M. Van den Bergh, [*Hochschild (co)homology for [L]{}ie algebroids*]{}, Int. Math. Res. Not. IMRN (2010), no. 21, 4098–4136.
A. S. Cattaneo, G. Felder, and L. Tomassini, [*From local to global deformation quantization of [P]{}oisson manifolds*]{}, Duke Math. J. [**115**]{} (2002), no. 2, 329–352.
Y. L. Daletskii and B. L. Tsygan, [*Operations on [H]{}ochschild and cyclic complexes*]{}, Methods Funct. Anal. Topology [**5**]{} (1999), no. 4, 62–86.
B. Davison, [*Superpotential algebras and manifolds*]{}, arXiv:1010.3564v1.
V. Dolgushev, [*A formality theorem for [H]{}ochschild chains*]{}, Adv. Math. [**200**]{} (2006), no. 1, 51–101.
[to3em]{}, [*The [V]{}an den [B]{}ergh duality and the modular symmetry of a [P]{}oisson variety*]{}, Selecta Math. (N.S.) [**14**]{} (2009), no. 2, 199–228.
V. Dolgushev, D. Tamarkin, and B. Tsygan, [*Formality of the homotopy calculus algebra of hochschild (co)chains*]{}, arXiv:0807.5117v1 \[math.KT\].
[to3em]{}, [*The homotopy [G]{}erstenhaber algebra of [H]{}ochschild cochains of a regular algebra is formal*]{}, J. Noncommut. Geom. [**1**]{} (2007), no. 1, 1–25.
M. Gerstenhaber and A. A. Voronov, [*Higher-order operations on the [H]{}ochschild complex*]{}, Funktsional. Anal. i Prilozhen. [**29**]{} (1995), no. 1, 1–6, 96.
E. Getzler, [*Cartan homotopy formulas and the [G]{}auss-[M]{}anin connection in cyclic homology*]{}, Quantum deformations of algebras and their representations ([R]{}amat-[G]{}an, 1991/1992; [R]{}ehovot, 1991/1992) (Ramat Gan), Israel Math. Conf. Proc., vol. 7, Bar-Ilan Univ., Ramat Gan, 1993, pp. 65–78.
V. Ginzburg, [*Calabi-[Y]{}au algebras*]{}, arXiv:math/0612139.
L. Katzarkov, M. Kontsevich, and T. Pantev, [*Hodge theoretic aspects of mirror symmetry*]{}, From [H]{}odge theory to integrability and [TQFT]{} tt\*-geometry (Providence, RI), Proc. Sympos. Pure Math., vol. 78, Amer. Math. Soc., Providence, RI, 2008, pp. 87–174.
B. Keller, 2006, private communication.
M. Kontsevich, [*Deformation quantization of [P]{}oisson manifolds*]{}, Lett. Math. Phys. [**66**]{} (2003), no. 3, 157–216.
M. Kontsevich and Y. Soibelman, [*Notes on [$A_\infty$]{}-algebras, [$A_\infty$]{}-categories and non-commutative geometry*]{}, Homological mirror symmetry (Berlin), Lecture Notes in Phys., vol. 757, Springer, Berlin, 2009, pp. 153–219.
J.-L. Loday, [*Cyclic homology*]{}, second ed., Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\], vol. 301, Springer-Verlag, Berlin, 1998, Appendix E by María O. Ronco, Chapter 13 by the author in collaboration with Teimuraz Pirashvili.
M. Manetti, [*Lectures on deformations of complex manifolds (deformations from differential graded viewpoint)*]{}, Rend. Mat. Appl. (7) [**24**]{} (2004), no. 1, 1–183.
L. Menichi, [*Batalin-[V]{}ilkovisky algebra structures on [H]{}ochschild cohomology*]{}, Bull. Soc. Math. France [**137**]{} (2009), no. 2, 277–295.
V. Schechtman, [*Remarks on formal deformations and [B]{}atalin-[V]{}ilkovisky algebras*]{}, arXiv:math/9802006v2 \[math.AG\], 1998.
G. Sharygin and D. Talalaev, [*On the [L]{}ie-formality of [P]{}oisson manifolds*]{}, J. K-Theory [**2**]{} (2008), no. 2, Special issue in memory of Yurii Petrovich Solovyev. Part 1, 361–384.
B. Shoikhet, [*A proof of the [T]{}sygan formality conjecture for chains*]{}, Adv. Math. [**179**]{} (2003), no. 1, 7–37.
D. Tamarkin and B. Tsygan, [*The ring of differential operators on forms in noncommutative calculus*]{}, Graphs and patterns in mathematics and theoretical physics (Providence, RI), Proc. Sympos. Pure Math., vol. 73, Amer. Math. Soc., Providence, RI, 2005, pp. 105–131.
J. Terilla, [*Smoothness theorem for differential [BV]{} algebras*]{}, J. Topol. [**1**]{} (2008), no. 3, 693–702.
J. Terilla and T. Tradler, [*Deformations of associative algebras with inner products*]{}, Homology, Homotopy Appl. [**8**]{} (2006), no. 2, 115–131.
B. Tsygan, [*Formality conjectures for chains*]{}, Differential topology, infinite-dimensional [L]{}ie algebras, and applications (Providence, RI), Amer. Math. Soc. Transl. Ser. 2, vol. 194, Amer. Math. Soc., Providence, RI, 1999, pp. 261–274.
[to3em]{}, [*On the [G]{}auss-[M]{}anin connection in cyclic homology*]{}, Methods Funct. Anal. Topology [**13**]{} (2007), no. 1, 83–94.
M. Van den Bergh, [*Calabi-[Y]{}au algebras and superpotentials*]{}, arXiv:1008.0599v1 \[math.KT\].
[to3em]{}, [*A relation between [H]{}ochschild homology and cohomology for [G]{}orenstein rings*]{}, Proc. Amer. Math. Soc. [**126**]{} (1998), no. 5, 1345–134, Erratum : \[[P]{}roc. [A]{}mer. [M]{}ath. [S]{}oc. [**130**]{} (2002), 2809–2810\].
C. A. Weibel, [*An introduction to homological algebra*]{}, Cambridge Studies in Advanced Mathematics, vol. 38, Cambridge University Press, Cambridge, 1994.
T. Willwacher, [*Formality of cyclic chains*]{}, arXiv:0804.3887v1 \[math.QA\], 2008.
A. Yekutieli, [*Deformation quantization in algebraic geometry*]{}, Adv. Math. [**198**]{} (2005), no. 1, 383–432.
[^1]: This isomorphism is sometimes required to satisfy a certain symmetry condition but this happens to be automatically satisfied. See [@VdBsuper App. C].
[^2]: In fact this is slightly imprecise as $\mathfrak{D}^\bullet(A,\eta)$ is only determined up to a non-unique isomorphism. The actual definition of $\mathfrak{D}^\bullet(A,\eta)$ depends on the lift of $\eta$ to an explicit cycle in a suitable complex but we will ignore this subtlety in the introduction.
[^3]: This is a special case Dolgushev’s result. Dolgushev does not restrict to the Calabi-Yau case.
[^4]: As always our grading conventions are such that Lie brackets have degree zero.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Directed sandpile models with different toppling rules are studied by means of numerical simulations in two dimensions, with the purpose of determining the different universality classes. It is concluded that the random-threshold directed model is in the same universality of the Manna directed model where multiple toppling events plays a determinant role. The BTW model with uniform and noisy driving are found to display the same critical behavior. Moreover it is observed that the Zhang model does not satisfy simple finite size scaling.'
address:
- |
$^1$ Abdus Salam International Center for Theoretical Physics\
Strada Costiera 11, P.O. Box 586, 34100 Trieste, Italy
- |
$^2$ Department of Theoretical Physics, Havana University\
San Lázaro y L, Havana 10400, Cuba
author:
- 'Alexei Vázquez$^{1,2}$'
title: Numerical simulations of directed sandpile models
---
[2]{}
Introduction
============
The classification of sandpile models in their different universality class has been a topic of intensive research in the last decade [@ns]. However, most of the work has been devoted to models with undirected toppling while their corresponding directed variants has been less studied. From the theoretical side we count with the exact solution for the directed BTW model [@bak87] obtained by Dhar and Ramaswamy [@dhar89], which can be taken as a reference for numerical simulations. On the other hand, Pastor-Satorras and Vespignani [@pastor99] has recently reported numerical simulations for the Manna model [@manna91], which puts clear evidence about the classification of the BTW and Manna directed models in different universality classes. Other studies include a directed model with a probabilistic toppling [@tadic97] and a recent report concerning the effect of local dissipation on the SOC state [@manna99].
The analysis of other directed models, the non Abelian Zhang model for instance, is of great interested and may give light to the study of their corresponding undirected variants. With this scope three different directed models are investigated by means of numerical simulations. This includes the well known Zhang model [@zhang89], the random threshold model (RT) [@christensen96], and the BTW model under an uniform driving. The influence of a uniform driving in the Zhang model has been already discussed in the literature [@olami97] although some aspects are still not clear. However, the same analysis has not been made for models with a discrete toppling rule, like the BTW model, where the energy transfer always takes place in discrete units. In this direction Narayan and Middleton [@narayan96] suggested that the BTW model under noisy and uniform driving has the same critical behavior.
From the numerical data it is concluded that the Manna and RT directed models belong to the same universality class, in agreement with the general believe for the corresponding undirected variants [@vazquez99; @lauritsen99]. Moreover, the data obtained for the Zhang model does not satisfy finite-size scaling which suggest that this model is in different universality class from that of the models mentioned above.
In the case of the BTW model under uniform driving it is shown, after some algebra with the toppling operators [@dhar89], that its evolution is periodic in time with a period scaling linearly with system size. In spite of this periodicity, which is not present in the original model with noisy driving [@dhar89], the statistics of the avalanches is found to be practically identical to its noisy driving counterpart.
Models and simulations
======================
Models with noisy driving
-------------------------
Consider a square lattice of $L^2$ sites labeled by index $(i,j)$ ($i,j=1,\ldots,N$) and assign a variable $z_{ij}$ to each of them. $z_{ij}$ can be continuous or discrete and may have different interpretations depending on the system one is modeling. It will be referred here as the energy storage at the corresponding site. The geometry used is shown in fig. \[fig:1\], in which a site can transfer energy only to its three downward nearest neighbors (nn). In the horizontal direction periodic boundary conditions are considered while the downward boundary is taken open. This geometry allows a natural implementation of the Manna toppling rule [@pastor99].
One motivation for the use of this geometry was given in [@pastor99], it is just introduced to allow the implementation of the Manna toppling rule.
To completely define a model one should specify the initial condition and the evolution rules (addition of energy and toppling) of the sandpile cellular automaton model. A threshold $z_c$ is considered in such a way that sites with $z<z_c$ are say to be stable and their energy remains constant, while those with $z\geq z_c$ are say to be active and topple transferring energy to their downward nn. First the usual noisy addition of energy is considered. In this case, if all sites are stable a unit of energy is added to a site selected at random. Then the system is updated in parallel using the toppling rule until all sites are stable. The number of toppling events required to drive the system to an stable configuration is the size of the avalanche and is denoted by $s$. On the other hand, the number of steps (parallel updates) required is its duration and is denoted by $T$. Since the driving acts at random after some avalanches the system “forget” its initial condition and reaches a stationary state. In other words the initial condition is irrelevant.
Models will differ from each other depending on the specific toppling rule one implements. Here the following toppling rules are considered,
- [BTW:]{} $z_c=3$, $z_{ij}\rightarrow z_{ij}-3$ and $z_{kj-1}\rightarrow
z_{kj-1}+1$ ($k=i-1,i,i+1$);
- [Manna:]{} $z_c=2$, $z_{ij}\rightarrow z_{ij}-2$ and $z_{kj-1}\rightarrow
z_{kj-1}+\delta_k$ ($k=i-1,i,i+1$), where $\delta_k$ can take the values $0,1,2$ at random but with the constraint of conservation $\sum_k\delta_k=2$;
- [Zhang:]{} $z_c=1$, $z_{ij}\rightarrow 0$ and $z_{kj-1}\rightarrow
z_{kj-1}+z_{ij}/3$ ($k=i-1,i,i+1$);
- [RT:]{} $z_c$ takes the values 3 and 4 at random after each toppling, $z_{ij}\rightarrow z_{ij}-3$ and $z_{kj-1}\rightarrow z_{kj-1}+1$ ($k=i-1,i+1$).
BTW model under uniform driving
-------------------------------
In the noisy driving described above the addition of energy takes place at one site selected at random. However, there are many situations where a uniform driving in which the energy at all sites increases in the same amount becomes more realistic. Examples can be found in earthquake dynamics [@olami97], interface depinning [@narayan96] and also in some experimental setups to for granular materials [@granular]. This type of driving has been investigated in models with a toppling rule similar to that of the Zhang models but with local dissipation [@olami97].
In the case of the BTW model we should be careful when introducing a global driving. If as usual $z_{ij}$ is an integer variable and the energy at all sites is increased at a constant rate $c$ then many sizes will reach the threshold energy at the same time and, therefore, many avalanches will start at different points of the lattice leading to the superposition of avalanches.
This problem can be solved considering a continuum energy profile. This is still not enough because if all sites start with a discrete energy it will remain discrete forever. We are thus forced to consider a continuum initial profile $z_{ij}(0)$. Then as it was already shown by Narayan and Middleton [@narayan96] the continuum addition of energy can be replaced by a sequential addition of energy. For simplicity consider the low disorder regime where $z_{ij}(0)<1$ for all sites. For the analysis below is irrelevant if the initial energy profile is displaced uniformly at all sites.
Now, suppose that the energy increases at rate $c$ at all sites. An example is shown in fig. \[fig:2\] for a lattice made of a horizontal line of three sites. Notice that in this case one has only input of energy coming form the driving field and output dissipation under toppling, a simplification considered for illustrative purposes.
In the continuum time scale the energy increases linearly until it reaches the threshold where the site topples and its energy decreases by 3 (BTW toppling rule). But the system can be also monitored in a discrete time and energy scales. In these scales, at step $t=0$ all sites have energy 0. In step 1, 2 and 3 sites 1, 2 and 3 receives one unit of energy. Then in subsequent steps the same sequence of addition is repeated. The order in the sequence of addition is clearly determined by the initial condition and all sites should receive a unit of energy before the first site of the second receives the second grain.
Now consider an square lattice, where sites can also receive energy from nearest neighbors in the layer above. The picture will not change in relation to the addition of energy from the external field. In the BTW model the energy is transferred in discrete units and, therefore, the toppling only modifies the integer part of $z$ with no modification of the sequence of addition. This is a fundamental difference with the Zhang toppling rule which not only involve the integer par of the energy but on toppling all the energy at the active site is transferred. The consequences derived from this periodic sequential driving is investigated below, using the formalism introduced by Dhar [*et al*]{} [@dhar89].
Let be $a_{ij}$ the operator which add a particle at site $(i,j)$ and lead the system relax to an equilibrium position [@dhar89]. After $N$ steps all sites receives one, and only one, unit of energy in certain order determined at $t=0$. Thus, if at time $t$ we have a configuration ${\cal C}(t)$ then at time $t+N$ we will obtain the configuration $${\cal C}(t+N)=\prod_{i=1}^L\prod_{j=1}^L a_{ij}{\cal C}(t).
\label{eq:b2}$$ The order in which the string of operators appears in this equation is irrelevant because the operators $a_{ij}$ commute among them self.
Applying these string three times it results that $${\cal C}(t+3N)=\prod_{i=1}^L\prod_{j=1}^L a_{ij}^3{\cal C}(t).
\label{eq:b3}$$ This expression can be simplified using the following property of the toppling operators [@dhar89] $$a_{ij}^3=
\left\{
\begin{array}{ll}
a_{i-1j-1}a_{ij-1}a_{i+1j-1}, & \text{for}\ j<L,\\
1, & \text{for}\ j=L.
\end{array}
\right.
\label{eq:b4}$$ The first equality express the fact that the addition of three grains at a site $(i,j)$, with $j<L$, makes this site active transferring one grain to each of its downward nn. The second one applies for the boundary sites which after receiving three grains become active dissipating these three grains through the boundary and, therefore, leaving the energy configuration invariant.
Starting at layer $j=1$ all the operators $a_{i1}$ are eliminated using eq. (\[eq:b4\]). This will increase the power of operators $a_{i2}$ in eq. (\[eq:b3\]) by 3. The same procedure is applied to the second, third ... $L-1$ layer finally resulting $${\cal C}(t+3N)=\prod_{i=1}^La_{iL}^{3L}{\cal C}(t).
\label{eq:b5}$$ The application of the operator $a_{iL}$ three consecutive times lead its energy invariant and, therefore, eq. (\[eq:b5\]) is reduced to $${\cal C}(t+3N)={\cal C}(t).
\label{eq:b6}$$ Hence the evolution of the energy profile is periodic with period 3N.
This property is not observed in the noisy driving case where the randomness introduced by the driving field makes the dynamics Markovian [@dhar89]. Nevertheless, as it is shown in the next section, the statistical properties of the avalanches in the BTW directed model are independent of the driving mechanism.
Numerical simulations and discussion
------------------------------------
Numerical simulations of the BTW, Manna, Zhang and RT models with directed toppling were performed. In all cases lattice sizes ranging from $L=64$ to $L=2048$ where used. The numerical results obtained for the BTW and Manna models are taken only as a reference because we count with the largest scale simulations reported in [@pastor99] (up to $L=6400$).
[*Noisy driving*]{}: Starting from an initial flat profile all systems were updated until they reach the stationary state. After that statistics over $10^8$ avalanches was taken, recording the avalanche sizes and durations.
[*Uniform driving*]{}: the evolution in time of the energy profile is periodic and, therefore, average was taken over the period $3N$. Different initial conditions where simulated using different permutations of the sequence of addition of energy.
To extract the scaling exponents we use the moment analysis technique [@stella]. The $q$ moment of the probability density $p_x(x)$ of a magnitude $x$ is defined by $$\langle x^q\rangle=\int dx p_x(x)x^q.
\label{eq:c1}$$ where $x=s,T$. As defined above $s$ and $T$ are the avalanche size and duration, i.e. the number of toppling events and parallel updates, respectively, required to drive the system to an stable configuration.
If the hypothesis of finite-size scaling is satisfied, that is the distribution of avalanche size and duration can be written in the form $p_x(x)=x^{-\tau_x}f_x(x/L^{\beta_x})$, then the $q$ moment scales with system size according to the power law $$\langle x^q\rangle\sim L^{\sigma_x(q)},
\label{eq:c2}$$ with $$\sigma_x(q)=\beta_x(1-\tau_x)+\beta_xq,
\label{eq:c3}$$ where $\beta_s=D$ and $\beta_T=z$ are effective dimensions which characterize how the cutoffs of the distribution of avalanche sizes and durations, respectively, scales with system size. On the other hand $\tau_x$ is the power low exponent which can be measured in the scaling region before the finite-size cutoffs.
The plot of $\sigma_s(q)$ and $\sigma_T(q)$ vs. $q$ is shown in figs. \[fig:3\] for different directed models. and \[fig:4\], respectively. If two models belong to the same universality class then the linear part of the plot should overlap. Based on this argument it is then concluded that the RT model belong to the same universality class of the Manna model. A more quantitative comparison can be seen in table \[tab:1\] where the exponents computed here for the RT model are compared with those reported in [@pastor99] for the Manna model. The scaling exponents are found in very good agreement within the numerical error.
If the hypothesis of finite size scaling is valid then one can take the scaling exponents obtained from the moment analysis and plot the different distributions on rescaled variables in such a way that curve for different system sizes overlap. This is done in figs. \[fig:5\] and \[fig:6\] for the RT model resulting in a very good data collapse, as it has been also observed for the directed Manna model [@pastor99].
On the other hand, one cannot distinguish between the curve for the BTW model with noisy or uniform driving, leading to the same scaling exponents. Thus, the periodicity introduced by the uniform driving carry no consequence for the critical behavior of the BTW model. Hence, the noisy driving can be substituted by a uniform driving together with an initial random energy profile. This will correspond in an interface depinning description, the number of toppling events playing the role of the interface height, to a columnar disorder. A similar conclusion was obtained by Lauritsen and Alava using a different argument [@lauritsen99].
The things becomes less clear when analyzing the Zhang model. In this case from the moment technique it results that $D\approx1.55$, $z\approx1.03$, $\tau_s\approx1.31$ and $\tau_t\approx1.53$. These exponents define by them self a new universality class. However, the moment analysis technique is based on the hypothesis of finite size scaling which in this case is not satisfied. This fact becomes clear in figs. \[fig:7\] and \[fig:8\], where the data collapse is shown, revealing that in this case the finite-size scaling is not satisfied. Deviations are observed not only for the smallest avalanches but also for the largest avalanches where the finite size scaling is expected to be better.
The anomalies observed for the Zhang model are associated with the existence of huge avalanches which practically empties the system. After one of these huge avalanches the system needs some time to reach again the critical state. This means that the mean energy of the system displays strong fluctuations and, therefore, the overall avalanche statistics is given by the small avalanches taking place during the accumulation of new grains and these huge avalanches. This picture is illustrated in fig. \[fig:9\] where the fraction of avalanches of size $s$ is plotted. It is characterized by a rounded peak at the largest avalanche sizes which shifts with lattice size. On the other hand, the other part of the distribution cannot be fitted by a single power law.
The classification of the Manna and RT directed models in the same universality class is in agreement with a similar report for the corresponding undirected variants [@vazquez99]. Thus, there should be some common element in these models, which is off course not present in the BTW model. A clue was given in [@pastor99] related with the possibility of multiple toppling events. In this final part of this section we discuss this statement in more detail.
In the directed BTW model the cluster of sites which topples within an avalanche is compact and these sites topple only once. On the contrary, Pastor-Satorras and Vespignani [@pastor99] observed that in the directed Manna model the cluster of sites touched by the avalanche is still compact but each site participating in the avalanche can topple more than once. If the existence of multiple toppling events is the property that puts the Manna model in a different universality class then a similar behavior should be observed in the present simulations of the directed RT model.
A decomposition of the sites participating in an avalanche based on the number of toppling events performed at these sites is shown in fig. 9, for the case of the directed random-threshold model. In this particular realization the cluster of sites touched by the avalanche is decomposed in three sub clusters where sites have toppled one, two and three times. The fraction of sites toppling three times is small but the one with two toppling is comparable with that of one. In general it was observed that in large avalanches the fraction of sites which topple one and two times are of the same order and, therefore, multiple toppling events are relevant.
In the case of undirected models it is known that multiple toppling events are present even in the BTW model, which leads to the decomposition of the avalanches in waves [@priezzhev96]. However, their origin is different than in directed models. In the undirected BTW model a site may topple more than once because after a first toppling (let say at step $t$) it is possible that all its neighbors become active and topple (at step $t+1$) and, therefore, the site will again be active and topple (at step $t+2$). In the decomposition of waves one apply the toppling rule to all sites until they are stable before toppling the initial active site a second, third, ... time, generating in this way the first, second, ... wave.
One may think in applying a similar approach for the avalanches in the Manna and RT directed models, decomposing the avalanche as a superposition of waves. A fundamental property of the waves is that within it sites can toppling only once, otherwise the concept is useless. Below its is shown that such a decomposition is not possible in the Manna and RT directed models, a least not in such a simple way.
Let us analyze in detail how a multiple toppling event can be generated in the RT directed model. Suppose the lattice has a configuration where a site has height $3$ and threshold $4$ and its three upward nearest neighbors are active. Then in the next step the site will receive three grains, one per active neighbor, taking an energy $3+3=6>4$, becoming active. After toppling the energy will decrease to $6-3=3$ and a new threshold is assigned. But the new threshold can be either 3 or 4. If it is 4 the site will be stable but if it is 3 it will be still active and topple in the next step. Since in the particular model considered here the two threshold are selected with equal probability then the multiple toppling can take place with the same probability than the single one, which explains the previous observation that in large avalanche the fraction of two-toppling events at the same site are of the order of the one-toppling one.
During the evolution of an avalanche which started at layer $j_0$ is possible that a site at a layer below $j_1>j_0$ needs two consecutive toppling to be stable. Thinking in a decomposition in waves one can delay the second toppling until all the sites below are stable (first wave) and then topple the site the second time generating the second wave. However, during the first wave is possible that a site at a deeper layer $j_2>j_1>j_0$ also needs two toppling to become stable and, therefore, the first wave has to be decomposed in sub-waves where sites topple only once. The same process may occur even at deeper layer thus generating a hierarchical structure of sub-waves. Hence, the decomposition of avalanche in waves in these models lead to a more complex structure which nevertheless may be exploited to obtain some estimate of the scaling exponents. This is nevertheless out of the scope of this work.
Summary and conclusions
=======================
Directed sandpile models with different toppling rule has been studied by means of numerical simulations, with the purpose of determining the different universality classes. To extract the scaling exponents the moment analysis technique was used and the resulted exponents were latter corroborated by finite size scaling of the distribution of avalanche size and duration.
The numerical analysis reveals that the introduction of a uniform driving in the BTW directed model does not change the critical properties. The evolution in time of the energy profile is in this case periodic with a period which scales linearly with the system size. In spite of this periodicity the avalanche distributions are practically identical to that obtained for the same model but with the usual noisy driving.
It is concluded that the Manna and RT models are in the same universality, where multiple toppling events appear to be a fundamental property. The existence of multiple toppling events leads to a decomposition of the avalanche in a hierarchical structure of waves which my be a starting point for future research.
Finally, it is observed that the avalanches in the directed Zhang model displays a complex structure which does not satisfy the finite-size scaling hypothesis. It is given by the superposition of huge avalanches involving a large dissipation of energy through the boundary a small avalanches taking place during the accumulation of energy.
Acknowledgements {#acknowledgements .unnumbered}
================
I thanks R. Pastor Satorras and A. Vespignani for useful comments and discussion during the elaboration of this manuscript. The numerical simulations where performed using the computing facilities at the ICTP.
[99]{}
A. Ben-hur and O. Biham, Phys. Rev. E [**53**]{}, R1317 (1996); S. Lübeck and K. D. Usadel, Phys. Rev. E [**55**]{}, 4095 (1997); ibid. [**56**]{}, 5138 (1997); A. Chessa, H. E. Stanley, A. Vespignani and S. Zapperi, Phys. Rev. E [**59**]{}, R12 (1999); C. Tebaldi, M. De Menech, and A. L. Stella, Phys. Rev. Lett. [**83**]{}, 3952 (1999); S. Lübeck, Phys. Rev. E [**61**]{}, 204 (2000).
P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. [**59**]{}, 381 (1987); Phys. Rev. A [**38**]{}, 364 (1988).
D. Dhar and R. Rammaswamy, Phys. Rev. Lett. [**63**]{}, 1659 (1989); for a review see cond-mat/9909009.
R. Pastor-Satorras and A. Vespignani, J. Phys. A [**33**]{}, L33 (2000).
S. S. Manna, J. Phys. A [**24**]{}, L363 (1991).
B. Tadić and D. Dhar, Phys. Rev. Lett. [**79**]{}, 1519 (1997).
S. S. Manna, A. D. Chakrabarti, and R. Cafiero, Phys. Rev. E [**60**]{}, R5005 (1999).
Y.-C. Zhang, Phys. Rev. Lett. [**63**]{}, 470 (1989).
K. Christensen, A. Corral, V. Frette, J. Feder, and T. Tossang, Phys. Rev. Lett. [**77**]{}, 107 (1996).
Z. Olami, H. J. S. Feder, and K. Christensen, Phys. Rev. Lett [**68**]{}, 1244 (1992); PRE
O. Narayan and A. A. Middleton, Phys. Rev. B [**49**]{}, 244 (1994).
A. Vázquez and O. Sotolongo-Costa, cond-mat/9811417.
K. B. Lauritsen and M. J. Alava, cond-mat/9903346.
H. M. Jaeger and S. R. Nagel, Science [**255**]{}, 1522 (1992).
M. De Menech, A. L. Stella, and C. Tebaldi, Phys. Rev. E [**58**]{}, R2677 (1998).
V. B. Priezzhev, D. V. Ktitarev, and E. V. Ivashkevich, Phys. Rev. Lett. [**76**]{}, 2093 (1996); D. V. Ktitarev and V. B. Priezzhev, Phys. Rev. E [**58**]{}, 2883 (1998).
Model $D$ $z$ $\tau_s$ $\tau_t$
------------------ --------- --------- ---------- ----------
BTW[@dhar89] $3/2$ 1 $4/3$ $3/2$
Manna[@pastor99] 1.75(1) 0.99(1) 1.43(1) 1.74(4)
RT 1.73(2) 0.99(3) 1.44(2) 1.70(4)
: The scaling exponents for the BTW, Manna and RT directed models. Those for the BTW model are exact while the others are numerical estimates.[]{data-label="tab:1"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'State-of-the-art simulation tools for non-equilibrium quantum transport systems typically take the current-carrier occupations to be described in terms of equilibrium distribution functions characterised by two different electro-chemical potentials, while for the description of electronic exchange and correlation, the local density approximation (LDA) to density functional theory (DFT) is generally used. However this involves an inconsistency because the LDA is based on the homogeneous electron gas *in equilibrium*, while the system is not in equilibrium and may be far from it. In this paper we analyze this inconsistency by studying the interplay between non-equilibrium occupancies obtained from a maximum entropy approach and the Hartree-Fock exchange energy, single-particle spectrum and exchange hole, for the case of a two-dimensional homogeneous electron gas. The current-dependence of the local exchange potential is also discussed. It is found that the single-particle spectrum and exchange hole have a significant dependence on the current which has not been taken into account in practical calculations. The exchange energy and the local exchange potential, however, are shown to change very little with respect to their equilibrium counterparts. The weak dependence of these quantities on the current is explained in terms of the symmetries of the exchange hole.'
author:
- 'H. Mera'
- 'P. Bokes'
- 'R.W. Godby'
title: 'Hartree-Fock theory of a current-carrying electron gas'
---
Introduction
============
One of the uncontrollable approximations introduced in [*ab initio*]{} calculations of the transport properties of nano-scale conductors consists in the application of DFT, a ground state theory, outside the equilibrium regime. An immediate consequence of this approximation is that these properties are typically calculated at the level of the LDA, which is derived from the case of a homogeneous electron gas in equilibrium. The extent to which these approximations might affect the calculated electronic structure of the non-equilibrium systems remains largely unknown and thus a comparison between electronic properties calculated *exactly* for an admittedly highly idealised non-equilibrium system and those of the same system in equilibrium constitutes a particularly simple way of approaching and illustrating this problem.
In order to put these ideas into practice we will consider a two-dimensional electron gas in equilibrium and in a *model* non-equilibrium state. To model a homogeneous electron gas outside equilibrium we will assume that the non-equilibrium steady-state of the two dimensional electron gas can be characterized by the average total energy of the electron gas and by *different* average numbers of left- and right-moving electrons and that the non-equilibrium steady-state is given by the density matrix that maximises the entropy of the electron gas with constraints on the above mentioned averages.
Such an assumption leads in the non-interacting case to a momentum distribution characterized by *two Fermi hemispheres of different radii*; we take a pragmatic approach here and ignore the problems associated with the discontinuous character of this momentum distribution for the time being since we are interested in the question of how these current-inducing constraints affect the electronic properties of the two dimensional electron gas. Note that this type of momentum distribution is precisely of the form used in Landauer-Büttiker-type of approaches and thus familiar to the [*ab initio*]{} quantum transport community [@dattabook; @JauhoBook; @Taylor1; @Taylor2] which constantly makes use of it. Similar momentum distributions are predicted by semi-classical transport theories in two dimensional quantum point contacts [@Jansen1980]. Alternatively, and perhaps also more physically, a current-constraint may be used instead of the above-mentioned constraint to search for the non-equilibrium maximum entropy density matrix [@Ng92; @Heinonen93; @Bokes03; @Bokes05a].
To summarise, we will maximize the entropy of a two-dimensional homogeneous electron gas with constraints on the average numbers of left- and right-moving electrons to obtain a description of a steady-state at the Hartree-Fock level of approximation, which can then be used to obtain the electronic structure of the gas in the presence of a current and to compare it with the usual approximations. The rest of the paper is organized as follows: in the next section we discuss our theoretical approach to the problem and its numerical implementation; in Section III we discuss the current-dependence of the Hartree-Fock pair probability distribution, single particle spectrum, total energy and local exchange potential. We conclude with a discussion of the relevance of our work for practical calculations.
Theory
======
In order to proceed let us consider the entropy per unit area of the two dimensional electron gas to be a functional of the electronic occupancies given by [@Todorov00]: $$\label{eq:entropy}
S\left[ f\right(\mathbf{k}\left)\right]
=-
\int_{\Re^2} \frac{d^2\mathbf{k}}{2\pi^2} \, \left[
f(\mathbf{k})\ln f(\mathbf{k})+\left(1-f(\mathbf{k}) \right) \ln \left( 1-f(\mathbf{k})
\right)
\right] .$$ The electronic occupancies are written as: $$f(\mathbf{k})= \left\{\begin{array}{ll}
f_L(\mathbf{k}) & \textrm{if $k_x < 0$} \\
f_R(\mathbf{k}) & \textrm{if $k_x > 0$}
\end{array}\right.$$ where $\mathbf{k}=\left(k_x , k_y \right)$ and $f_{L/R}$ are the occupation functions to be varied independently in order to maximise Eq. (\[eq:entropy\]) with constraints on the average total energy per unit area and different average numbers of left- and right-going particles per unit area. In the finite-temperature Hartree-Fock approximation the average total energy is given by: $$\label{eq:hfetot}
\langle E \rangle=2\int_{\Re^2}\frac{d^2\mathbf{k}}{\left( 2\pi \right)^2}\,f(\mathbf{k})\frac{k^2}{2}
-\int_{\Re^2}\frac{d^2\mathbf{k}^\prime}{\left( 2\pi \right)^2}\int_{\Re^2}\frac{d^2\mathbf{k}}{\left( 2\pi \right)^2}
f(\mathbf{k})f(\mathbf{k}^\prime)v(\mathbf{k},\mathbf{k}^\prime)$$ where $v(\mathbf{k},\mathbf{k}^\prime)=2\pi/\left|\mathbf{k}-\mathbf{k}^\prime \right|$ is the Fourier transform of the Coulomb interaction in two dimensions. The number of left- and right-going electrons per unit area can be written as: $$\label{eq:densdes}
n_{L(R)}=\frac{2}{\left( 2\pi\right)^2}\int_{k_x<(>)0} d^2\mathbf{k} \, f_{L(R)}(\mathbf{k})$$ In order to maximise the entropy functional with respect to $f_{L/R}$ subject to the above-mentioned constraints we use the method of Lagrange multipliers and consider the auxiliary functional $$\mathcal{L}\left[ f\right(\mathbf{k}\left)\right]=S-\beta \left(
\langle E \rangle-\mu_L n_L -\mu_R n_R
\right)\, ,$$ together with the extremal condition $$\frac{\delta \mathcal{L}}{\delta f_{L/R}}=0 .$$ A straightforward calculation shows that the occupation functions that maximise the entropy functional with constraints in the above-mentioned averages are given by: $$f_{L,R}(\mathbf{k})=\frac{1}{1+\exp \left[ \beta \left(
k^2/2+\epsilon_x(\mathbf{k})-\mu_{L,R}
\right) \right]}$$ where $$\epsilon_x(\mathbf{k})= -\frac{1}{(2\pi)^2}\left(\int_{k_x<0} d^2\mathbf{k}^\prime \, f_L(\mathbf{k}^\prime) v(\mathbf{k},\mathbf{k}^\prime)
+\int_{k_x>0} d^2\mathbf{k}^\prime \, f_R(\mathbf{k}^\prime) v(\mathbf{k},\mathbf{k}^\prime)
\right)
\label{eq:xspec}$$ i.e., the occupations that maximise the entropy are similar to the ones of the Landauer-Büttiker approach but with a modified exchange part of the spectrum. In the calculation we fix the ratio $n_L/n_R$, that together with the charge neutrality condition $n_L+n_R=1/(\pi r_s^2)$ completely determines both $n_L$ and $n_R$. With the equilibrium spectrum as a trial $\epsilon_x(\mathbf{k})$ we solve Eqs. (\[eq:densdes\]) for $\mu_L$ and $\mu_R$. With these values of $\mu_{L,R}$ a new spectrum is constructed using Eq. (\[eq:xspec\]) and the iteration is completed and subsequently repeated until the input and output spectra are identical to each other within the desired tolerance. All the results presented here are obtained in the $\beta \rightarrow
\infty$-limit, where our approach is equivalent to that of Hershfield [@Hershfield93] in the Hartree-Fock approximation [^1]. Once the self-consistent spectrum and occupation factors are obtained, other quantities like the exchange-energy and exchange hole can be easily obtained. From these we can study how the local exchange potential of the electron gas depends on the current density [^2].
Results
=======
Hartree-Fock pair distribution function
---------------------------------------
Let us begin by discussing the current dependence of the Hartree-Fock pair distribution function for spin-like electrons, which is given by: $$g(\mathbf{r},\mathbf{r^\prime})=1-\left| \frac{1}{n}\int \frac{d^2\mathbf{k}}{(2\pi)^2}
\exp\left[-i\mathbf{k}\cdot (\mathbf{r-r^\prime})\right]f(\mathbf{k})\right|^2$$ and shown for $n_L/n_R=0.5$ in Figure \[fig:hole\]-(a). For $n_L=n_R$, $g$ is spherically symmetric while for $n_L \neq n_R$ is elongated in the direction of the current. Similar phenomenology has been reported previously by Skudlarski and Vignale for the three dimensional electron gas in the presence of a magnetic field [@sk93], where the exchange hole is elongated in the direction of the field. In Ref. the elongation arises from the change of occupancies associated with the Zeeman splitting due to the externally applied magnetic field. In the present case the elongation of the hole can be understood in terms of the change in the electronic occupancies that result from our constrained maximization of the entropy functional. In both cases the elongation of the hole is the result of the change in the polarizability induced by the change in the occupancies [@sk93].
Note that the difference between the equilibrium and non-equilibrium exchange holes, $\Delta g=g_{eq}-g_{neq}$, shown in Figure \[fig:hole\]-(b), has a strong antisymmetric character, i.e., defining $\mathbf{R}=\mathbf{r}-\mathbf{r}^\prime=(X,Y)$ then $\Delta g (X,Y)\sim -\Delta g
(Y,X)$. We shall return to this point later in the text when discussing the weak dependence of the exchange energy on the current density.
Single-Particle spectrum
------------------------
Figure \[fig:scfspec\] shows the self-consistent single-particle energy spectrum. Fig.\[fig:scfspec\]-(a) shows the total (kinetic + exchange) spectrum while in Fig.\[fig:scfspec\]-(b) we plotted only its exchange part on the $k_y=0$ line as given by Eq. (\[eq:xspec\]) for $n_L=n_R$ and $n_L/n_R=0.5$.
The combined effect of the constraints and the exchange interaction shifts the spectrum towards higher values of $k_x$. Note also that, when compared to the equilibrium spectrum, the minimum of the non-equilibrium spectrum is less negative. Hence we expect the total non-equilibrium exchange energy to increase with respect to the equilibrium one. Note that the constraints alter the total kinetic energy of the system but do not change the kinetic contribution to the single particle spectrum, since this contribution does not depend on the electronic occupancies. Hence the changes in the single particle spectrum are entirely due to the exchange interaction, which raises (lowers) the single particle energy of electrons with $k_x<0$ ($k_x>0$). The anomalous behaviour in the $k_x=0$ plane inherited from the discontinuous character of the maximum entropy momentum distribution can be seen clearly in Fig.\[fig:scfspec\]-(a), between $\mu_L$ and $\mu_R$.
 
The interplay between non-equilibrium occupancies and the single-particle spectrum observed here is just a consequence of the orbital dependence of the Fock operator and will also be seen in any practical calculation that combines a non-equilibrium theory such as the Landauer-Büttiker approach or the Keldysh-NEGF formalism, with an *orbital-dependent* description of the interactions between the electrons, such as the Hartree-Fock approximation. We would like to point out that practical implementations of NEGF formalism typically take the electronic structure of the leads to be that of the equilibrium system (see Ref. and references therein), and hence the dependence of the single-particle spectrum on the non-equilibrium current (and vice-versa) is commonly ignored. The validity of this approximation is geometry dependent: it works in quantum point contact geometries while it does not in planar electrode geometries at high currents. As a consequence under the “non-interacting equilibrium lead approximation” the distribution of incoming electrons would be current-independent, while, as this example shows[^3], the unavoidable presence of interactions in the leads induces a current dependence in the non-equilibrium occupancies through the exchange part of the single particle spectrum. Unless the geometry is adequately chosen the distribution of incoming electrons will be that of a *non-equilibrium* lead such as ours.
Total energy
------------
Once the self-consistent single-particle spectrum is calculated the total exchange energy, $E_x$, can be obtained from the second term in the right hand side of Eq. (\[eq:hfetot\]). Figure \[fig:exvsj\] shows the dependence of the $r_s$-invariant quantity $-E_x/E_x^{eq}$ on $(1-n_L/n_R)$. For $n_L/n_R=0.25$, the exchange energy deviates by about $1-2\%$ from its equilibrium value. We also see that, even though the non-self-consistent results provide a good estimate to the self-consistent ones, full self-consistency is needed in the non-equilibrium case, even for a homogeneous gas. The error bars in the self-consistent results are estimated by comparing the exact exchange energy in equilibrium with the exchange energy obtained from our code for $n_L=n_R$ and different values of $r_s$. Therefore, the exchange-energy depends on the current-density, but this dependence is extremely weak in our model system. One could now proceed to calculate this current density explicitly and work out a current dependent local density approximation, from the dependence of $E_x$ on the current density. However, the weak dependence of the exchange energy on the current density deduced from Fig. \[fig:exvsj\] means that the current dependence of the local exchange functional is also very weak, and the changes it will induce in the associated LDA-Kohn-Sham effective potential will be well within the error bar of the LDA itself.
![(Color online). Exchange energy (in units of the equilibrium exchange energy) versus $1-n_L/n_R$. In equilibrium $1-n_L/n_R=0$. The dashed line shows the self-consistent results with estimated error bars. The non-self-consistent results are also shown with points calculated for different values of $r_s$ showing that the exchange energy scales with $r_s$ as $1/r_s$. The lines are fits to parabolic functions.[]{data-label="fig:exvsj"}](mainresult.eps)
Local Exchange Potential
------------------------
The weak dependence of the local exchange potential on the current density can be seen clearly in terms of the symmetries of the exchange hole. Consider the expression for Slater’s exchange potential, $v^s_x$, in terms of the Hartree-Fock pair distribution function: $$v^s_x(\mathbf{r})=\int d^2 \mathbf{r'} \frac{[g(\mathbf{r}-\mathbf{r'})-1]}{|\mathbf{r}-\mathbf{r'}|}
n(\mathbf{r'})$$ where $n(\mathbf{r'})$ is the electron density and $g(\mathbf{r}-\mathbf{r'})$ is the exchange hole. Then, the difference between equilibrium and non-equilibrium exchange potentials is, for our homogeneous system, given by: $$\Delta v^s_x =n \int d^2 \mathbf{R} \frac{\Delta g(\mathbf{R})}{|\mathbf{R}|}
\label{eq:vxandhole}$$ where $\mathbf{R}$ and $\Delta g$ are defined as above. From Eq. (\[eq:vxandhole\]) follows that: $$\Delta g(X,Y)=-\Delta g(Y,X) \Rightarrow \Delta v^s_x=0$$ and hence only the symmetric part of $\Delta g(X,Y)$ contributes to the deviation of exchange potential with respect to its equilibrium value. Note that $\Delta g(X,Y)$ is an oscillatory function that integrates to zero which also has a marked antisymmetric character shown in Fig. \[fig:hole\]-(b). This explains the weak dependence of $E_x$ and $v_x$ on the current-density.
Conclusions
===========
In conclusion we have maximised the entropy of a two-dimensional homogeneous electron gas with constraints on the average total energy and average numbers of left- and right-going electrons to obtain a simplified description of the steady-state within the Hartree-Fock approximation. Our results show that both the single-particle spectrum and the exchange hole depend significantly on the current density while averaged quantities like the local exchange potential or the exchange energy do not.
The authors gratefully acknowledge useful discussions with J.J. Palacios and J.Fernández-Rossier. We are grateful to Matthieu Verstraete for useful comments on the manuscript. This work was supported by the EU’s 6th Framework Programme through the NANOQUANTA Network of Excellence (NMP4-CT-2004-500198), ERG programme of the European Union QuaTraFo (contract MERG-CT-2004-510615), the Slovak grant agency VEGA (project No. 1/2020/05) and the NATO Security Through Science Programme (EAP.RIG.981521).
[13]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, ).
, ** (, ).
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
[^1]: i.e., to find the Slater determinant that minimizes the expectation value of the effective Hamiltonian $\widehat{F}=\widehat{H}_{HF}-\mu_L \widehat{N}_R-\mu_L
\widehat{N}_R$.
[^2]: For two Fermi-hemispheres of radii $k_L$ and $k_R$ the non-interacting electronic current is related to these densities by the expression $j=\frac{2}{3\sqrt{2\pi}}(n_L+n_R)(n_R^3-n_L^3)$
[^3]: we can see our two-dimensional electron gas as a rough model of one of the leads to which the nanoscale conductor is attached
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the online saddle point problem, an online learning problem where at each iteration a pair of actions need to be chosen without knowledge of the current and future (convex-concave) payoff functions. The objective is to minimize the gap between the cumulative payoffs and the saddle point value of the aggregate payoff function, which we measure using a metric called “SP-Regret”. The problem generalizes the online convex optimization framework but here we must ensure both players incur cumulative payoffs close to that of the Nash equilibrium of the sum of the games. We propose an algorithm that achieves SP-Regret proportional to $\sqrt{\ln(T)T}$ in the general case, and $\log(T)$ SP-Regret for the strongly convex-concave case. We also consider the special case where the payoff functions are bilinear and the decision sets are the probability simplex. In this setting we are able to design algorithms that reduce the bounds on SP-Regret from a linear dependence in the dimension of the problem to a *logarithmic* one. We also study the problem under bandit feedback and provide an algorithm that achieves sublinear SP-Regret. We then consider an online convex optimization with knapsacks problem motivated by a wide variety of applications such as: dynamic pricing, auctions, and crowdsourcing. We relate this problem to the online saddle point problem and establish $O(\sqrt{T})$ regret using a primal-dual algorithm.'
author:
- |
Adrian Rivera Cardoso\
School of Industrial and Systems Engineering\
Georgia Institute of Technology\
Atlanta, GA 30332\
`[email protected]`\
He Wang\
School of Industrial and Systems Engineering\
Georgia Institute of Technology\
Atlanta, GA 30332\
`[email protected]`\
Huan Xu\
School of Industrial and Systems Engineering\
Georgia Institute of Technology\
Atlanta, GA 30332\
`[email protected]`\
bibliography:
- 'mybib.bib'
title: The Online Saddle Point Problem and Online Convex Optimization with Knapsacks
---
Introduction {#Intro}
============
In this paper, we study the *online saddle point* (OSP) problem. The OSP problem involves a sequence of two-player zero-sum convex-concave games which are selected arbitrarily by Nature. In each iteration, player 1 chooses an action to minimize its payoffs, while player 2 chooses an action to maximize its payoffs. Both players choose actions without knowledge of the current and future payoff functions. Our goal is to *jointly* choose a pair of actions for both players at each iteration, such that each player’s cumulative payoff at the end is as close as possible to that of the Nash equilibrium (i.e. saddle point) of the aggregate game.
More formally, we define the OSP problem as follows. There is a sequence of unknown functions $\{{\mathcal{L}}_t(x, y)\}_{t=1}^T$ that are convex in $x \in X$ and concave in $y \in Y$. Here, $X$ and $Y$ are compact convex sets in Euclidean space. As a result, there exists a saddle point $(x^*, y^*) \in X \times Y$ such that $${\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) = \sum_{t=1}^T {\mathcal{L}}_t(x^*,y^*) = \max_{y\in Y} \min_{x\in X} \sum_{t=1}^T {\mathcal{L}}_t(x,y) .$$ At each iteration $t$, the decision makers jointly choose a pair of actions $(x_t , y_t) \in X \times Y$, and then the function ${\mathcal{L}}_t$ is revealed. The goal is to design an algorithm to minimize the cumulative *saddle-point regret* (SP-Regret), defined as $$\label{eq:sp-reg}
{\mathsf{SP}\text{-}\mathsf{Regret}}(T) = \left|\sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) -{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y)\right|.$$ In other words, we would like to obtain a cumulative payoff that is as close as possible to the saddle-point value if we had known all the functions $\{{\mathcal{L}}_t\}_{t=1}^T$ in advance.
We would like to emphasize an important distinction between the OSP problem and the standard Online Convex Optimization (OCO) problem [@hazan2016introduction]. In the OCO problem, Nature selects an arbitrary sequence of convex functions $\{f_t(\cdot)\}_{t=1}^{T}$, and the decision maker chooses an action $x_t \in X$ before each function $f_t(\cdot)$ is revealed. The objective is to minimize the regret defined as $$\sum_{t=1}^T f_t(x_t) - \min_{x\in X} \sum_{t=1}^T f_t(x).$$ [ The objective in the OSP problem is to choose the actions of two players *jointly* such that the aggregate payoffs of both players are close to the Nash equilibrium payoff.]{} In contrast, OCO involves only an individual player against Nature. The OCO framework can be viewed as a [special case]{} of the OSP problem where the action set of the second player $Y$ is a singleton. Moreover, the standard OCO setting is applicable to the OSP problem when only one of the players’ payoff is optimized at a time. To be specific, we define the *individual-regret* of players 1 and 2 as
\[eq:indiv-reg\] $$\begin{aligned}
{\mathsf{Ind}\text{-}\mathsf{Regret}}_x (T) = \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) -\min_{x\in X}\sum_{t=1}^T {\mathcal{L}}_t(x,y_t), \label{eq:indiv-reg-x} \\
{\mathsf{Ind}\text{-}\mathsf{Regret}}_y (T) = \max_{y\in Y}\sum_{t=1}^T {\mathcal{L}}_t(x_t,y)- \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t). \label{eq:indiv-reg-y}\end{aligned}$$
The individual-regret measures each player’s own regret while fixing the other player’s actions. It is easy to see that minimizing individual-regret or can be cast as a standard OCO problem.
However, we will show that SP-Regret and individual-regret do not imply one another, so existing OCO algorithms cannot be directly applied to the OSP problem. More surprisingly, we show that any OCO algorithm with a sublinear ($o(T)$) individual-regret will inevitably have a linear ($\Omega(T)$) SP-Regret in the general OSP problem (see details in §\[choose\_one\]).
In addition to establishing general results for the OSP problem, we focus on one of its prominent applications: the *online convex optimization with knapsacks* (OCOwK) problem. Several variants of the OCOwK problem have recently received a lot of attention in recent literature, but we found its connection to the OSP problem has not been well exploited. We show that the OCOwK problem is closely related to the OSP problem through Lagrangian duality; thus, we are able to apply our results for the OSP problem to the OCOwK problem.
In the OCOwK problem, a decision maker is endowed with a fixed budget of resource at the beginning of $T$ periods. In each period $t=1,\ldots,T$, the decision maker chooses an action $x_t \in X$, and then Nature reveals a reward function $r_t$ and a budget consumption function $c_t$. The objective is to maximize total reward $\sum_{t=1}^T r_t(x_t)$ while keeping the total consumption $\sum_{t=1}^T c_t(x_t)$ within the given budget.
The OCOwK model also generalizes the standard OCO problem by having an additional budget constraint. Additionally, it also has a wide range of practical applications (see more discussion in [@badanidiyuru2018bandits]), some notable examples include:
- Dynamic pricing: a retailer is selling a fixed amount of goods in a finite horizon. The actions correspond to pricing decisions, the reward is the retailer’s revenue, and the budget represents finite item inventory. The reward functions are unknown initially due to high uncertainty in customer demand.
- Online ad auction: a firm is bidding for advertising on a platform (e.g. Google) with limited daily budget. The actions refer to auction bids, and the reward represents impressions received from displayed ads. The reward function is unknown because the firm is unaware of other firms’ bidding strategies.
- Crowdsourcing: suppose an organization is purchasing labor tasks on a crowdsourcing platform (e.g. Amazon Mechanical Turk). The actions correspond to prices offered for each micro-task, and the budget corresponds to the maximum amount of money to be spent on acquiring these tasks. The reward functions are unknown a priori because of uncertainty in the crowd’s abilities.
Main Contributions
------------------
[ We first propose an algorithm called (Saddle-Point Follow-the-Leader) for the online saddle point problem when the payoff function ${\mathcal{L}}_t(x, y)$ is Lipschitz continuous and strongly-convex in $x$ and strongly-concave in $y$, the algorithm has a SP-regret that scales as $\ln(T)$, which is optimal. When the payoff functions are convex-concave we show that a variant of attains a SP-Regret that scales as $\sqrt{\ln(T) T}$, which matches the lower bound of $\Omega(\sqrt{T})$ up to a logarithmic factor. In the special case where the payoff functions are bilinear and the decision sets are the probability simplex, a setup which we call Online Matrix Games, we show that a variant of can attain SP-Regret that scales almost optimally with $T$ and *logarithmically* with the dimension of the problem. This is in contrast to the general convex-concave case where the SP-Regret scales *linearly* with the dimension of the problem. We also study Online Matrix Games under bandit feedback. Here the players only observe the loss function evaluated at their decisions instead of observing the whole payoff function, this makes the problem significantly more challenging. For this setting we derive an algorithm that attains sublinear SP-Regret. ]{}
In addition, we show that no algorithm can simultaneously achieve sublinear (i.e. $o(T)$) SP-Regret and sublinear individual-regrets (defined in and ) in the general OSP problem. This impossibility result further illustrates the contribution of the algorithm, as existing OCO algorithms designed to achieve sublinear individual-regret are not able to achieve sublinear SP-Regret.
Then, we consider the OCOwK problem. We show that this problem is related to the OSP problem by Lagrangian duality, and a sufficient condition to achieve a sublinear regret for OCOwK is that an algorithm must have *both* sublinear SP-Regret and sublinear individual-regret for the OSP problem. In light of the previous impossibility result, we consider the OCOwK problem in a stochastic setting where the reward and consumption functions are sampled i.i.d. from some unknown distribution. By applying the algorithm and exploiting the connection between OCOwK and OSP problems, we obtain a $\tilde{O}(T^{5/6})$ regret. We then propose a new algorithm called (Primal-Dual Regularized-Follow-the-Leader) that achieves an $O(\sqrt{T})$ regret bound. The result matches the lower bound $\Omega(\sqrt{T})$ for the OCOwK problem in the stochastic setting. We then provide numerical experiments to compare the empirical performances of and .
Literature Review
=================
Saddle point problems emerge from a variety of fields such as machine learning, statistics, computer science, and economics. Some applications of the saddle point problem include: minimizing the maximum of smooth convex functions, minimizing the maximal eigenvalue, $l_1$-minimization (an important tool in sparsity-oriented Signal processing), nuclear norm minimization, robust learning problems, and two-player zero-sum games [@presentation_arkadi; @lu2007large; @cox2017decomposition; @ahmadinejad2016duels; @myerson1993incentives; @laslier2002distributive; @chowdhury2013experimental; @kovenock2012coalitional].
[ We now discuss some related works that focus on learning in games. [@singh2000nash] study a two player, two-action general sum static game. They show that if both players use Infinitesimal Gradient Ascent, either the strategy pair will converge to a Nash equilibrium (NE), or even if they do not, then the average payoffs are close to that of the NE. A result of similar flavor was derived in [@cesa2007improved] for any zero-sum convex-concave game. Given a payoff function ${\mathcal{L}}(x,y)$, they show that if both players minimize their individual-regrets, then the average of actions $(\bar{x},\bar{y})$ will satisfy $|{\mathcal{L}}(\bar{x},\bar{y})-{\mathcal{L}}(x^*,y^*)|\to 0$ as $T \to \infty$, where $(x^*,y^*)$ is a NE. [@bowling2001convergence] improve upon the result of [@singh2000nash] by proposing an algorithm called WoLF (Win or Learn Fast), which is a modification of gradient ascent; they show that the iterates of their algorithm indeed converge to a NE. [@conitzer2007awesome] further improve the results in [@singh2000nash] and [@bowling2005convergence] by developing an algorithm called GIGA-WoLF for multi-player nonzero sum static games. Their algorithm learns to play optimally against stationary opponents; when used in self-play, the actions chosen by the algorithm converge to a NE. More recently, [@balduzzi2018mechanics] studied general multi-player static games and show that by decomposing and classifying the second order dynamics of these games, one can prevent cycling behavior to find NE. We note that unlike our paper, all of the papers above consider repeated games with a static payoff matrix, whereas we allow the payoff matrix to change arbitrarily. An exception is the work by [@ho2016role], who consider the same setting as our OMG problem; however their paper only shows that the sum of the individual regrets of both players is sublinear and does not study SP-Regret. Related to the OMG problem with bandit feedback is the seminal work of [@flaxman2005online]. They provide the first sublinear regret bound for Online Convex Optimization with bandit feedback, using a one-point estimate of the gradient. The one-point gradient estimate used in [@flaxman2005online] is similar to those independently proposed in [@spall1997one]. The regret bound provided in [@flaxman2005online] is $O(T^{3/4})$, which is suboptimal. In [@abernethy2009competing], the authors give the first $O(\sqrt{T})$ bound for the special case when the functions are linear. More recently, [@hazan2016optimal] and [@bubeck2016kernel] designed the first efficient algorithms with $\tilde{O}(poly(d)\sqrt{T})$ regret for the general online convex optimization case; unfortunately, the dependence on the dimension $d$ in the regret rate is a very large polynomial. Our one-point matrix estimate is most closely related to the random estimator in [@auer1995gambling] for linear functions. It is possible to use the more sophisticated techniques from [@abernethy2009competing; @hazan2016optimal; @bubeck2016kernel] to improve our SP-Regret bound in Section \[section:omg\_bandit\]; however, the result does not seem to be immediate and we leave this as future work. ]{}
The Online Convex Optimization with Knapsacks (OCOwK) problem studied in this paper is related to several previous works on constrained multi-armed bandit problems, online linear programming, and online convex programming. We next give an overview of the work related to OCOwK . Agrawal et al. [@agrawal2014dynamic] and Agrawal and Devanur [@agrawal2014fast] consider online linear/convex programming problems. A key difference between the online linear/convex programming problems and the OCOwK problem is that we assume the action must be chosen without knowledge of the function associated with the current iteration. In [@agrawal2014dynamic; @agrawal2014fast], it is assumed that these functions are revealed *before* the action is chosen. Related work is that of Buchbinder and Naor [@buchbinder2009online], where they study an online fractional covering/packing problem, and that of Gupta and Molinaro [@gupta2016experts] where they consider a packing/covering multiple choice LP problem in a random permutation model. Another relevant paper is [@lan2016algorithms] where the authors provide an algorithm to solve convex problems with expectation constraints, such as the benchmark in Section \[OCOwK\_setup\]. However it is unclear if their optimization algorithm has any sublinear regret properties.
Mannor et al. [@mannor2009online] consider a variant of the online convex optimization (OCO) problem where the adversary may choose extra constraints that must be satisfied. They construct an example such that no algorithm can attain an $\epsilon$-approximation to the offline problem. In view of such result, several papers [@mahdavi2012online; @neely2017online; @yu2017online] study problems similar to [@mannor2009online] with further restrictions on how constraints are selected by the adversary. The objective in this line of work is to choose a sequence of decisions to achieve the offline optimum while making sure the constraints are (almost) satisfied. In this line of research, the most relevant work to ours is that of [@neely2017online]. They study OCO with time-varying constraints, the model is similar to that of [@mannor2009online], however in view of the existing 3 negative results they consider three different settings. In the first one, both the cost functions and the constraints are arbitrary sequences of convex functions, however in view of the negative result from [@mannor2009online], the constraints must all be non positive over a common subset of $\mathbb{R}^n$. In the second setting the sequences of loss functions remain adversarially chosen however the constraints are sampled i.i.d. from some unknown distribution. Finally, in the third setting both the sequences of loss functions and constraints are sampled i.i.d. from some unknown distribution. They develop algorithms for all the three different settings that ensure the total loss incurred by the algorithm is not too far from the offline optimum and such that the constraints are almost satisfied. The setup and results are different than ours because they only require the cumulative constraint violation to be sublinear whereas in OCOwK, once the player exceeds the budget it can no longer collect rewards. Closely related is the problem of “Online Convex Optimization with Long Term Constraints”. The setup is similar to that of OCO where the functions are chosen adversarially with the difference that it is not required that the decisions the player makes at each step belong to the set. Instead, it is required that the average decision lies in the set (which is fully known in advance). As the authors explain, this problem is useful to avoid the projection step of online gradient descent (OGD) and it allows to solve problems such as multi-objective online classification [@bernstein2010online], and for using the popular online-to-batch conversion. The algorithms they develop consist of simultaneously running two copies of variants of OGD on convex-concave functions. Better rates and slightly different guarantees were obtained for the same problem in [@yuan2018online; @jenatton2015adaptive; @yu2016low]. In [@paternain2015online], the authors study a continuous time version of a problem similar to that of [@mahdavi2012online] and show that a continuous time version of primal-dual online gradient descent in continuous time guarantees small regret. Motivated by an application in low-latency fog computing, [@chen2018harnessing] consider a problem similar to that in [@mahdavi2012trading] however there is bandit feedback in the loss function. The algorithm proposed in [@chen2018harnessing] is primal-dual online gradient descent that combines ideas from [@flaxman2005online] to deal with bandit feedback.
Most closely related to our model is the Bandits with Knapsacks problem studied by Badanidiyuru et al. [@badanidiyuru2018bandits] and Wu et al. [@wu2015algorithms]. In this problem, there is a finite set of arms, and each arm yields a random reward and consumes resources when it is pulled. The goal is to maximize total reward without exceeding a total budget. The Bandits with Knapsacks problem can be viewed as a special case of the OCOwK problem, where the reward and consumption functions are both linear. Agrawal and Devanur [@agrawal2014bandits] study a generalization of bandits with concave rewards and convex knapsack constraints. Similar problems have also been studied in specific application contexts, such as online ad auction [@balseiro2017learning] and dynamic pricing [@besbes2012blind; @ferreira2017online]. [ Recently, [@immorlica2019adversarial] study an adversarial version of the multi-armed bandits with knapsack problem. A key part of their algorithm uses a primal-dual approach similar to the one we propose in Section \[OCOwK\_setup\].]{}
Preliminaries
=============
We introduce some notation and definitions that will be used in later sections. By default, all vectors are column vectors. A vector with entries $x_1,...,x_n$ is written as $x = [x_1;...;x_n] = [x_1,...,x_n]^\top$, where $\top$ denotes the transpose. [Let $\Vert\cdot\Vert$ be any norm of a vector; the ones we will frequently use are $\Vert\cdot\Vert_2, \Vert\cdot\Vert_1, \Vert\cdot\Vert_\infty$.]{}
[We say a function ${\mathcal{L}}(x,y)$ is *convex-concave* if it is convex in $x \in X$, for every fixed $y\in Y$, and concave in $y \in Y$, for every fixed $x\in X$.]{} A pair $(x^*,y^*)$ is called a saddle point for ${\mathcal{L}}$ if for any $x\in X$ and any $y \in Y$, we have $$\label{def_sp}
{\mathcal{L}}(x^*,y) \leq {\mathcal{L}}(x^*,y^*) \leq {\mathcal{L}}(x,y^*).$$ It is well known that if ${\mathcal{L}}$ is convex-concave, and $X$ and $Y$ are convex compact sets, there always exists at least one saddle point (see e.g. [@boyd2004convex]).
We say that a function $f:X\rightarrow \mathbb{R}$ is $H$-strongly convex if for any $x_1, x_2 \in X$, it holds that $$\begin{aligned}
f(x_1) \geq f(x_2) + \nabla f(x_2)^\top(x_1-x_2) + \frac{H}{2}\Vert x_1-x_2 \Vert^2.\end{aligned}$$ Here, $\nabla f(x)$ denotes a subgradient of $f$ at $x$. Strong convexity implies that the problem $\min_{x \in X} f(x)$ has a unique solution. We say a function $g$ is $H$-strongly concave if $-g$ is $H$-strongly convex.
Furthermore, we say a function ${\mathcal{L}}(x,y)$ is $H$-strongly convex-concave if for any fixed $y_0 \in Y$, the function ${\mathcal{L}}(x,y_0)$ is $H$-strongly convex in $x$, and for any fixed $x_0\in X$, the function ${\mathcal{L}}(x_0,y)$ is $H$-strongly concave in $y$. If ${\mathcal{L}}$ is $H$-strongly convex-concave, then there exists a unique saddle point.
[ We say a function ${\mathcal{L}}(x,y)$ is $G$-Lipschitz continuous with respect to norm $\Vert\cdot\Vert$, if $$\begin{aligned}
|{\mathcal{L}}(x_1, y_1)- {\mathcal{L}}(x_2, y_2)| \leq G \Vert [x_1;y_1] - [x_2;y_2] \Vert.\end{aligned}$$ It is well known that the previous inequality holds if and only if $$\begin{aligned}
\Vert [\nabla_x {\mathcal{L}}(x,y); \nabla_y {\mathcal{L}}(x,y)] \Vert_* \leq G \end{aligned}$$ for all $x\in X$, $y \in Y$ where $\Vert\cdot\Vert_*$ is the norm dual to $\Vert\cdot\Vert$ (see Lemma 2.6 in [@shalev2012online]). ]{}
Throughout the paper we will use the big $O$ notation to hide constant factors. For two functions $f(T)$ and $g(T)>0$, we write $f(T)=O(g(T))$ if there exists a constant $M_1$ and a constant $T_1$ such that $f(T) \leq M_1 g(T)$ for all $T \geq T_1$; we write $f(T)=\Omega(g(T))$ if there exists a constant $M_2$ and a constant $T_2$ such that $f(T) \geq M_2 g(T)$ for all $T \geq T_2$. We use the $\tilde{O}$ notation to hide constant factors and poly-logarithmic factors. More specifically, for two functions $f(T)$ and $g(T)>0$, we write $f(T)=\tilde{O}(g(T))$ if there exists constants $M_3$, $T_3$ and an integer $k\geq 0$ such that $f(T) \leq M_3 g(T)\log^k (g(T))$ for all $T \geq T_3$.
The Online Saddle Point Problem {#setup_problem}
===============================
The Strongly Convex-Concave Case
--------------------------------
We now present algorithms for the OSP problem with guaranteed sublinear SP-Regret. Recall that the SP-regret defined in measures the gap between the cumulative value achieved by an online algorithm and the value of the game under the Nash equilibrium if all functions are known in hindsight.
For simplicity we assume $T$ is known in advance (this assumption can be relaxed using the well known doubling trick from [@cesa2006prediction; @shalev2012online]). We first consider the case where the functions $\{{\mathcal{L}}_t\}_{t=1}^T$ are strongly convex-concave. We show that the following simple algorithm Saddle-Point Follow-the-Leader (), which is a variant of the Follow-the-Leader () algorithm by Kalai and Vempala [@kalai2002], attains sublinear SP-Regret.
$x_1 \in X$, $y_1 \in Y$ Choose actions $(x_t, y_t)$ Observe function ${\mathcal{L}}_t$ Set $x_{t+1} \leftarrow \arg \min_{x\in X} \max_{y\in Y} \sum_{\tau=1}^t {\mathcal{L}}_{\tau}(x,y)$ Set $y_{t+1} \leftarrow \arg \max_{y\in Y} \min_{x\in X} \sum_{\tau=1}^t {\mathcal{L}}_{\tau}(x,y)$
The main difference between and is that in both players update jointly and play the (unique) saddle point of the sum of the games observed so far. In contrast, the updates for Follow-the-Leader would be $x^{FTL}_{t+1}\leftarrow \arg \min_{x\in X} \sum_{\tau=1}^t {\mathcal{L}}_\tau(x, y^{FTL}_\tau)$ and $y^{FTL}_{t+1}\leftarrow \arg \max_{y\in Y} \sum_{\tau=1}^t {\mathcal{L}}_\tau(x^{FTL}_\tau, y)$ for $t=2,...,T$ and $x^{FTL}_1$, $y^{FTL}_1$ are arbitrarily chosen from their respective sets $X$ and $Y$. It is easy to see that the sequence of iterates is in general not the same. In fact, in view of Theorem \[thm:impossible\] we will see that can not achieve sublinear ${\mathsf{SP}\text{-}\mathsf{Regret}}$ when the sequence of functions is chosen arbitrarily.
\[thm:sp\_regret\_str\] Let $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$ be any sequence of $H$-strongly convex-concave, $G$-Lipschitz functions. Then, the $\mathsf{SP}$-$\mathsf{FTL}$ algorithm guarantees $$\begin{aligned}
{\mathsf{SP}\text{-}\mathsf{Regret}}(T) =
\Bigl| \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) -{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \Bigr| \leq \frac{8G^2}{H} (1+ \log T).\end{aligned}$$
[We remark that since Theorem \[thm:sp\_regret\_str\] holds against *all* sequences of functions $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$, this means that the sequence $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$ can even be chosen by an adaptive adversary. In particular, this means that ${\mathcal{L}}_t(x,y)$ can be a function of all the previous iterates $\{(x_\tau,y_\tau)\}_{\tau=1}^{t-1}$. ]{}
The proof of Theorem \[thm:sp\_regret\_str\] is based on the following two lemmas. We first analyze a quantity that is similar to SP-Regret, but with actions $(x_t, y_t)$ replaced by $(x_{t+1},y_{t+1})$ (Lemma \[loss\_BTL\]). This analysis framework is known as the Follow-the-Leader vs. Be-the-Leader scheme [@kalai2002]. We then show that consecutive iterates of have distances diminishing proportionally to $1/t$. The proof heavily utilizes the KKT conditions associated with points $(x_t, y_t)$ and $(x_{t+1},y_{t+1})$ (Lemma \[FTL\_close\_iterates\]).
\[loss\_BTL\] [ Let $\{{\mathcal{L}}_t\}_{t=1}^T$ be an arbitrary sequence of convex-concave functions that are $G$-Lipschitz with respect to norm $\Vert \cdot \Vert$. Here, ${\mathcal{L}}_t:X \times Y \rightarrow \mathbb{R}$, where $X\subset \mathbb{R}^{d_1}$ and $Y\subset \mathbb{R}^{d_2}$ are convex compact sets. Let $\{(x_t,y_t)\}_{t=1}^T$ be the iterates of $\mathsf{SP}$-$\mathsf{FTL}$ when run on the sequence of functions $\{{\mathcal{L}}_t\}_{t=1}^T$. It holds that $$-G \sum_{t=1}^T \Vert x_t-x_{t+1} \Vert \leq \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \leq G \sum_{t=1}^T \Vert y_t-y_{t+1} \Vert.$$ ]{}
We first prove the second inequality, namely $$\sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \leq G \sum_{t=1}^T \Vert y_t-y_{t+1} \Vert.$$ We proceed by induction. The base case $t=1$ holds by definition of $(x_2,y_2)$: $${\mathcal{L}}_1(x_2,y_2) - {\min_{x\in X}\max_{y\in Y}}{\mathcal{L}}_1(x,y) = {\mathcal{L}}_1(x_2,y_2) - {\mathcal{L}}_1(x_2,y_2) = 0 \leq G \Vert y_1-y_2\Vert.$$ We now assume the following claim holds for $T-1$, $$\label{ind_hypo}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^{T-1} {\mathcal{L}}_t(x,y) \geq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{t+1},y_{t+1}) - G \sum_{t=1}^{T-1} \Vert y_t-y_{t+1}\Vert,$$ and show it holds for $T$. By definition of $(x_{T+1},y_{T+1})$, we have $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) = & \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{T+1},y_{T+1}) + {\mathcal{L}}_T(x_{T+1},y_{T+1})\\
\geq & \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{T+1},y_T) + {\mathcal{L}}_{T}(x_{T+1},y_T)\\
\geq &\sum_{t=1}^{T-1} {\mathcal{L}}_t(x_T,y_T) + {\mathcal{L}}_{T}(x_{T+1},y_T).\end{aligned}$$ The first inequality holds because $(x_{T+1},y_{T+1})$ is the saddle point of $\sum_{t=1}^T{\mathcal{L}}_t(x,y)$, so $y_{T+1}$ is a maximizer for function $\sum_{t=1}^T{\mathcal{L}}_t(x_{T+1},y)$, see Equation . Similarly, the second inequality follows since $(x_T,y_T)$ is the saddle point of $\sum_{t=1}^{T-1} {\mathcal{L}}_t (x,y)$. By the induction hypothesis, see Equation , and the definition of $(x_T,y_T)$ we have $$\begin{aligned}
& \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_T,y_T) + {\mathcal{L}}_{T}(x_{T+1},y_T)\\
\geq & \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{t+1},y_{t+1}) - G \sum_{t=1}^{T-1} \Vert y_t-y_{t+1}\Vert + {\mathcal{L}}_{T}(x_{T+1},y_T)\\
= & \sum_{t=1}^{T} {\mathcal{L}}_t(x_{t+1},y_{t+1})- G \sum_{t=1}^{T-1} \Vert y_t-y_{t+1}\Vert+ {\mathcal{L}}_{T} (x_{T+1},y_T) - {\mathcal{L}}_{T}(x_{T+1},y_{T+1})\\
\geq & \sum_{t=1}^{T} {\mathcal{L}}_t(x_{t+1},y_{t+1})- G \sum_{t=1}^{T-1} \Vert y_t-y_{t+1}\Vert - G \Vert y_T - y_{T+1}\Vert \qquad \qquad \text{since ${\mathcal{L}}_T$ is $G$-Lipschitz}\\
= & \sum_{t=1}^{T} {\mathcal{L}}_t(x_{t+1},y_{t+1})- G \sum_{t=1}^{T} \Vert y_t-y_{t+1}\Vert.\end{aligned}$$ This proves the second inequality in the lemma.
Using a similar argument, we now show by induction the first inequality in the statement of the lemma, namely that $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \leq \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^T \Vert x_t-x_{t+1}\Vert. \end{aligned}$$ Indeed, $t=1$ follows from the definition of $(x_2,y_2)$: $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}{\mathcal{L}}_1(x,y) - {\mathcal{L}}_1(x_2,y_2) = {\mathcal{L}}_1(x_2,y_2) - {\mathcal{L}}_1(x_2,y_2 = 0 \leq G \Vert x_1,x_2 \Vert.\end{aligned}$$ We now assume the following claim holds for $T-1$,
$$\label{ind_hypo_2}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^{T-1} {\mathcal{L}}_t(x,y) \leq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^{T-1} \Vert x_t-x_{t+1}\Vert$$
and prove it for $T$. By definition of $(x_{T+1},y_{T+1})$, we have $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) &= \sum_{t=1}^T {\mathcal{L}}_t(x_{T+1},y_{T+1})\\
&\leq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{T},y_{T+1}) + {\mathcal{L}}_T(x_T, y_{T+1})\\
&\leq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{T},y_{T}) + {\mathcal{L}}_T(x_T, y_{T+1}).
\end{aligned}$$
The first inequality holds because $(x_{T+1},y_{T+1})$ is the saddle point of $\sum_{t=1}^T {\mathcal{L}}_t(x,y)$, so $x_{T+1}$ is a minimizer of $\sum_{t=1}^T {\mathcal{L}}_t(x,y_{T+1})$, see Equation . Similarly, the second inequality follows since $(x_T,y_T)$ is the saddle point of $\sum_{t=1}^{T-1}{\mathcal{L}}_t(x,y)$ so $y_T$ is the maximizer of $\sum_{t=1}^{T-1}{\mathcal{L}}_t(x_T,y)$.
By the induction hypothesis (see Equation ) and the definition of $(x_T,y_T)$, we have $$\begin{aligned}
& \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{T},y_{T}) + {\mathcal{L}}_T(x_T, y_{T+1})\\
&\leq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^{T-1}\Vert
x_{t}-x_{t+1}\Vert + {\mathcal{L}}_T(x_T, y_{T+1})\\
& = \sum_{t=1}^{T} {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^{T-1}\Vert x_{t}-x_{t+1}\Vert + {\mathcal{L}}_T(x_T, y_{T+1}) - {\mathcal{L}}_T(x_{T+1},y_{T+1})\\
&\leq \sum_{t=1}^{T} {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^{T}\Vert x_{t}-x_{t+1}\Vert, \qquad \qquad \text{since ${\mathcal{L}}_T$ is $G$-Lipschitz}. \end{aligned}$$ This concludes the proof.
\[FTL\_close\_iterates\] Let $\{{\mathcal{L}}_t\}_{t=1}^T$ be an arbitrary sequence of $H$-strongly convex-concave functions (with respect to norm $\Vert \cdot \Vert$) which is also $G$-Lipschitz with respect to the same norm. Let $\{(x_t,y_t)\}_{t=1}^T$ be the iterates of $\mathsf{SP}$-$\mathsf{FTL}$ run on the sequence $\{{\mathcal{L}}_t\}_{t=1}^T$. It holds that $$\begin{aligned}
\Vert x_t - x_{t+1}\Vert + \Vert y_t - y_{t+1}\Vert \leq \frac{4G}{H t}.\end{aligned}$$
Consider a fixed period $t$. Define $$\begin{aligned}
J(x,y) \triangleq \sum_{\tau=1}^{t-1} {\mathcal{L}}_\tau (x,y) + {\mathcal{L}}_t(x,y)\end{aligned}$$ so that $(x_{t+1},y_{t+1})$ is a saddle point of $J$. Since $J$ is $Ht$-strongly convex it holds that for any $x\in X$ and any $y\in Y$ $$\begin{aligned}
J(x,y) \geq J(x_{t+1},y) + \nabla_x J(x_{t+1},y)^{\top}(x - x_{t+1}) + \frac{Ht}{2}\Vert x-x_{t+1}\Vert^2.\end{aligned}$$ Plugging in $y=y_{t+1}$ and recalling the KKT condition $\nabla_x J(x_{t+1},y_{t+1})^{\top}(x - x_{t+1}) \geq 0$ (see Chapter 2 in [@hazan2016introduction]), we have that for any $x\in X$ $$\label{J_KKT_convex}
\frac{2}{Ht}\big[ J(x,y_{t+1}) - J(x_{t+1},y_{t+1})\big] \geq \Vert x - x_{t+1}\Vert^2.$$ Similarly, since $J$ is $Ht$-strongly concave, for any $y\in Y$ we have $$\begin{aligned}
J(x_{t+1},y) \leq J(x_{t+1},y_{t+1}) + \nabla_y J(x_{t+1},y_{t+1})^{\top}(y-y_{t+1}) - \frac{Ht}{2} \Vert y-y_{t+1}\Vert^2.\end{aligned}$$ Together with the KKT condition $\nabla_y J(x_{t+1},y_{t+1})^{\top}(y-y_{t+1})\leq 0$, we get that for any $y\in Y$ $$\label{J_KKT_concave}
\frac{2}{Ht}\big[ J(x_{t+1},y_{t+1}) - J(x_{t+1},y)\big] \geq \Vert y-y_{t+1}\Vert^2.$$ Adding up Equations and , plugging $x=x_t$ and $y=y_t$, we get $$\begin{aligned}
& \frac{2}{Ht} \big[ J(x_t,y_{t+1})-J(x_{t+1},y_t) \big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$ Plugging in the definition of function $J(\cdot)$, we get $$\begin{aligned}
\frac{2}{Ht} \big[ \sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t+1}) + {\mathcal{L}}_t (x_t,y_{t+1}) - [ \sum_{\tau=1}^{t-1} {\mathcal{L}}_{\tau}(x_{t+1},y_t) + {\mathcal{L}}_t (x_{t+1},y_t)]\big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$ Since $(x_t,y_t)$ is the saddle point of $\sum_{\tau=1}^{t-1} {\mathcal{L}}_\tau(x,y)$, it holds that $\sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t+1}) \leq \sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t})$. Therefore, we have $$\begin{aligned}
&\frac{2}{Ht} \big[ \sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t}) + {\mathcal{L}}_t (x_t,y_{t+1}) - [ \sum_{\tau=1}^{t-1} {\mathcal{L}}_{\tau}(x_{t+1},y_t) + {\mathcal{L}}_t (x_{t+1},y_t)]\big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$ Additionally, since $(x_t,y_t)$ is the saddle point of $\sum_{\tau=1}^{t-1} {\mathcal{L}}_\tau(x,y)$, it holds that $- \sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t}) \geq -\sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_{t+1},y_{t})$. This implies $$\begin{aligned}
&\frac{2}{Ht} \big[ \sum_{\tau=1}^{t-1}{\mathcal{L}}_\tau (x_t,y_{t}) + {\mathcal{L}}_t (x_t,y_{t+1}) - \sum_{\tau=1}^{t-1} {\mathcal{L}}_{\tau}(x_{t},y_t) - {\mathcal{L}}_t (x_{t+1},y_t) \big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$ Notice the two summations cancel, thus $$\begin{aligned}
\frac{2}{Ht} \big[ {\mathcal{L}}_t (x_t,y_{t+1}) - {\mathcal{L}}_t (x_{t+1},y_t) \big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$
Since ${\mathcal{L}}_t$ is $G$-Lipschitz with respect to norm $\Vert \cdot \Vert$, it holds that $$\begin{aligned}
& \frac{2}{Ht} G \Vert[x_t;y_{t+1}] - [x_{t+1};y_t]\Vert \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2,\end{aligned}$$ which then implies that $$\begin{aligned}
\frac{2}{Ht} G\big[ \Vert x_t - x_{t+1}\Vert + \Vert y_t - y_{t+1}\Vert\big] \geq \Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2.\end{aligned}$$ Rearranging the terms of the inequality above, we get $$\begin{aligned}
\frac{2G}{Ht} \geq \frac{\Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2}{\Vert x_t - x_{t+1}\Vert + \Vert y_t - y_{t+1}\Vert}.\end{aligned}$$ For any $a,b \in \mathbb{R}$, we have $a^2 + b^2 \geq \frac{(a+b)^2}{2}$, which implies $$\begin{aligned}
\Vert x_t - x_{t+1}\Vert^2 + \Vert y_t-y_{t+1}\Vert^2 \geq \frac{\left(\Vert x_t - x_{t+1}\Vert + \Vert y_t-y_{t+1}\Vert \right)^2}{2}.\end{aligned}$$ Therefore, we have $$\begin{aligned}
\frac{2G}{Ht} \geq \frac{ \left(\Vert x_t - x_{t+1}\Vert + \Vert y_t-y_{t+1}\Vert \right)^2}{2 \left(\Vert x_t - x_{t+1}\Vert + \Vert y_t - y_{t+1}\Vert \right)}.\end{aligned}$$ Rearranging the terms, we get $$\begin{aligned}
\frac{4G}{Ht}\geq \Vert x_t - x_{t+1}\Vert + \Vert y_t - y_{t+1}\Vert. \end{aligned}$$ This concludes the proof.
Now we are ready to prove Theorem \[thm:sp\_regret\_str\].
We first prove one side of the inequality, $$\begin{aligned}
\sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) -{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \leq \frac{8G^2}{H} (1+ \ln T).\end{aligned}$$ We have $$\begin{aligned}
&\quad \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) -{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) \\
&\leq \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) - \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) + G \sum_{t=1}^T \Vert y_t-y_{t+1}\Vert & \text{by Lemma \ref{loss_BTL}}\\
& \leq \sum_{t=1}^T G \Vert [x_t;y_t] - [x_{t+1};y_{t+1}]\Vert + G \sum_{t=1}^T \Vert y_t-y_{t+1}\Vert & \text{since ${\mathcal{L}}_t$ is $G$-Lipschitz}\\
&\leq G\sum_{t=1}^T \Vert x_t - x_{t+1}\Vert + \Vert y_t- y_{t+1}\Vert + G \sum_{t=1}^T \Vert y_t-y_{t+1}\Vert \\
&\leq G\sum_{t=1}^T \frac{4G}{Ht} + G \sum_{t=1}^T \frac{4G}{Ht} & \text{by Lemma \ref{FTL_close_iterates}}\\
&\leq \frac{8G^2}{H} (1+ \int_{1}^T \frac{1}{t}dt)\\
&= \frac{8G^2}{H} (1+ \ln T).
\end{aligned}$$
[ We now prove the other side of the inequality, $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) \leq \frac{8G^2}{H} (1+ \ln T),\end{aligned}$$ using a similar argument. In particular, we have $$\begin{aligned}
& \quad {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) \\
& \leq \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) + G \sum_{t=1}^T \Vert x_t - x_{t+1} \Vert & \text{by Lemma \ref{loss_BTL}}\\
& \leq \sum_{t=1}^T G \Vert [x_t;y_t] - [x_{t+1};y_{t+1}]\Vert + G \sum_{t=1}^T \Vert x_t-x_{t+1}\Vert & \text{since ${\mathcal{L}}_t$ is $G$-Lipschitz}\\
&\leq G\sum_{t=1}^T \Vert x_t - x_{t+1}\Vert + \Vert y_t- y_{t+1}\Vert + G \sum_{t=1}^T \Vert x_t-x_{t+1}\Vert \\
&\leq G\sum_{t=1}^T \frac{4G}{Ht} + G \sum_{t=1}^T \frac{4G}{Ht} & \text{by Lemma \ref{FTL_close_iterates}}\\
&\leq \frac{8G^2}{H} (1+ \int_{1}^T \frac{1}{t}dt)\\
&= \frac{8G^2}{H} (1+ \ln T).\end{aligned}$$ This concludes the proof. ]{}
We note that the rate in Theorem \[thm:sp\_regret\_str\] is optimal with respect to $T$, since when $Y$ is a singleton, the problem reduces to the OCO problem with strongly convex loss functions. In that case, it is well known that no algorithm can achieve regret smaller than $\Omega(\frac{G^2}{H} \log (T))$ [@hazan2014beyond].
The General Convex-Concave Case
-------------------------------
In this section we propose an algorithm to solve the online Saddle Point Problem in the full information setting when the payoff functions are arbitrary convex-concave Lipschitz functions, and the action sets of Player 1 and Player 2 ($X\subset \mathbb{R}^{d_1}$ and $Y\subset \mathbb{R}^{d_2}$ respectively) to be arbitrary convex compact sets.
Let the sequence of convex-concave functions be $\{\bar{{\mathcal{L}}}_t(x,y)\}_{t=1}^T$, which are $G_{\bar{{\mathcal{L}}}}$-Lipschitz with respect to some norm $\Vert \cdot \Vert$. We propose an algorithm called Saddle Point Regularized Follow the Leader (), shown in Algorithm \[alg:SPRFTL\].
$x_1 \in X$, $y_1 \in Y$, parameters: $\eta>0$, strongly convex functions $R_X$, $R_Y$ Play $(x_t,y_t)$ Observe $\bar{{\mathcal{L}}}_t$ ${\mathcal{L}}_t(x,y) \gets \bar{{\mathcal{L}}}_t + \frac{1}{\eta}R_X(x)-\frac{1}{\eta}R_Y(y)$ $x_{t+1}\leftarrow \arg \min_{x\in X} \max_{y \in Y} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y)$ $y_{t+1}\leftarrow \arg \max_{y \in Y} \min_{x\in X} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y) $
The regularizers $R_X, R_Y$ are used as input for the algorithm. We will choose regularizers that are strongly convex with respect to norm $ \Vert \cdot \Vert $, and $G_{R_1}$ and $G_{R_2}$ Lipschitz with respect to norm $ \Vert \cdot \Vert $, which means that $\Vert \nabla R_X(x)\Vert_* \leq G_{R_1}$ for all $x\in X$, and $\Vert \nabla R_Y(y)\Vert_* \leq G_{R_2}$ for, all $y\in Y$. Finally, we assume $R_X(x)\geq 0 $ for all $x\in X$ and $R_Y(y)\geq 0 $ for all $y\in Y$.
We have the following guarantee for .
\[theorem:sp\_regret\_convex\_concave\] Let $X\subset \mathbb{R}^{d_1}$ and $Y\subset \mathbb{R}^{d_2}$ be convex and compact sets. Let $\{\bar{{\mathcal{L}}}_{t}(x,y)\}_{t=1}^T$ be any sequence of convex-concave functions. For $t=1,...,T$, let $\bar{{\mathcal{L}}}_{t}$ be $G_{\bar{{\mathcal{L}}}}$-Lipschitz with respect to norm $\Vert \cdot \Vert$. Let $R_X$, $R_Y$ be two strongly convex regularization functions with respect to the same norm, and let $G_{R_X},G_{R_Y}$ be the Lipschitz constants of $R_X$, $R_Y$. Let $\{(x_t,y_t)\}_{t=1}^T$ be the iterates generated by when run on the sequence $\{\bar{{\mathcal{L}}}_{t}(x,y)\}_{t=1}^T$. It holds that $$\begin{aligned}
& \left| \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) \right| \\
\leq & 8 \eta \left[G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})\right]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in Y} R_Y(y) + \frac{T}{\eta} \max_{x\in X} R_X(x),\end{aligned}$$ where $\eta>0$ is the parameter chosen in Algorithm \[alg:SPRFTL\].
As a corollary, we have the following result that shows our algorithm guarantees a sublinear SP-Regret.
\[first\_corollary\] Let $X\subset \mathbb{R}^{d_1}$ and $Y\subset \mathbb{R}^{d_2}$ be convex and compact sets containing the origin such that $\max_{x\in X}\Vert x \Vert_2, \max_{y\in Y}\Vert y \Vert_2 \leq D$ for some $0 < D \in \mathbb{R}$. Let $R_X(x) = \Vert x \Vert_2^2$ and $R_Y(y) = \Vert y \Vert_2^2$. For $t=1,...,T$, let $\bar{{\mathcal{L}}}_{t}$ be $G_{\bar{{\mathcal{L}}}}$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. Setting $\eta = \frac{D \sqrt{T}}{G_{\bar{{\mathcal{L}}}}\sqrt{\ln(T)}}$ in guarantees $$\begin{aligned}
\left| \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) \right| \leq O\left(G_{\bar{{\mathcal{L}}}} D \sqrt{\ln(T) T}\right),\end{aligned}$$ where the $O(\cdot)$ notation hides an absolute constant.
We will instantiate the result from Theorem \[theorem:sp\_regret\_convex\_concave\]. By our choice of regularizers, we have $\max_{x\in X} R_X(x), \max_{y\in Y} R_Y(y) \leq D^2$, as well as $\Vert \nabla R_X (x) \Vert_2 \leq 2 D$. Thus, we have $$\begin{aligned}
&\left| \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) \right| \leq 8 \eta \left[G_{\bar{{\mathcal{L}}}}+ \frac{2D}{\eta}\right]^2 ( 1 + \ln(T) )+ \frac{D^2 T}{\eta} + \frac{D^2 T}{\eta}.\end{aligned}$$ Since for any $a,b \in \mathbb{R}$ it holds that $(a+b)^2\leq 2a^2 + 2b^2$, we have $$\begin{aligned}
8 \eta \left[G_{\bar{{\mathcal{L}}}}+ \frac{2D}{\eta}\right]^2 ( 1 + \ln(T) )+ \frac{D^2 T}{\eta} + \frac{D^2 T}{\eta} & \leq 8\eta \left( 2 G_{\bar{{\mathcal{L}}}}^2 + \frac{8D^2}{\eta^2}\right) ( 1 + \ln(T) )+ \frac{2 D^2 T}{\eta}\\
& \leq O\left(G_{\bar{{\mathcal{L}}}} D \sqrt{\ln(T) T}\right),\end{aligned}$$ where the last inequality follows by the choice of $\eta$ and the $O(\cdot)$ hides an absolute constant.
We note that the bound in Corollary \[first\_corollary\] is optimal up to the $\sqrt{\ln(T)}$ factor. This is because our setup is a special case of Online Convex Optimization and there is a known lower bound $\Omega(d \sqrt{T})$, see Chapter 3 in [@hazan2016introduction]. In our setup if $\bar{{\mathcal{L}}}_t$ is bilinear, that is $\bar{{\mathcal{L}}}_t = x^{\top} A_t y$ for some $d_1 \times d_2$ matrix $A_t$ with bounded entries, and $X,Y$ are unit boxes of dimensions $d_1$ and $d_2$ respectively, we have $G_{\bar{{\mathcal{L}}}}, D = O(\max\{d_1,d_2\})$ and so our SP-Regret bound becomes $O(\max\{d_1,d_2\} \sqrt{\ln(T)\sqrt{T}})$.
Although the focus of our work is mainly concerned with showing sublinear rate of SP-Regret, it is worth discussing the computation complexity for each iteration of our algorithms. Notice that in each iteration we must solve a strongly convex strongly concave constrained saddle point problem. It is well known that by simultaneously playing two no Individual Regret algorithms for strongly convex functions (such as those in [@hazan2007logarithmic] which achieve Individual Regret $O(\log(K)))$, one can generate after $K$ rounds a solution to the problem that is $O(\log(K)/K)$ close to the Nash equilibrium (in terms of the value of the game) (See Theorem 9 in [@abernethy2018faster]). Recently [@abernethy2018faster] showed that with additional smoothness assumptions it is possible to obtain linear convergence rates for some static saddle point problems. It is also possible to solve the subproblem for each iteration using the (Stochastic Approximation) Mirror Descent algorithm from [@nemirovski2009robust]. All the previously discussed algorithms are variants of the seminal work of [@arrow1958studies].
In the rest of this subsection we will prove Theorem \[theorem:sp\_regret\_convex\_concave\]. Define ${\mathcal{L}}_t(x,y) \triangleq \bar{{\mathcal{L}}}_t(x,y) + \frac{1}{\eta}R_X(x) - \frac{1}{\eta}R_Y(y)$. Notice that it is $\frac{1}{\eta}$-strongly convex in $x$ with respect to norm $\Vert \cdot \Vert$ for all $y\in Y$ and $\frac{1}{\eta}$-strongly concave with respect to norm $\Vert \cdot \Vert$ for all $x \in X$. Additionally, notice that ${\mathcal{L}}_t$ is $G_{\mathcal{L}}\triangleq G_{\bar{{\mathcal{L}}}} + \frac{1}{\eta} (G_{R_X}+G_{R_Y})$-Lipschitz with respect to norm $ \Vert \cdot \Vert $. Finally, notice that by nonnegativity of $R_X$ and $R_Y$ for $t=1,...,T$, all $x \in X$ and all $y\in Y$ it holds that $$\label{eq:diff_bar_not_bar}
-\frac{1}{\eta}R_Y(y) \leq {\mathcal{L}}_t(x,y) - \bar{{\mathcal{L}}}_t(x,y) \leq \frac{1}{\eta} R_X(x).$$
The following lemma shows that the value of the convex-concave games defined by $\sum_{t=1}^T {\mathcal{L}}_t$ and $\sum_{t=1}^T \bar{{\mathcal{L}}}_t$ are not too far from each other.
\[lemma:mm\_bar\_not\_bar\] Let $X \subseteq \mathbb{R}^{d_1}$, $Y \subseteq \mathbb{R}^{d_2}$ be convex and compact sets. Let $\{{\mathcal{L}}_t\}_{t=1}^T$ be any sequence of convex-concave functions where ${\mathcal{L}}_t: X \times Y \rightarrow \mathbb{R}$ for all $t=1,...,T$. Let $$\begin{aligned}
\textstyle \bar{x}_{T+1} \in \arg \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T\bar{{\mathcal{L}}}_t(x,y), \\
\textstyle \bar{y}_{T+1} \in \arg \max_{y\in Y} \min_{x\in X} \sum_{t=1}^T\bar{{\mathcal{L}}}_t(x,y).\end{aligned}$$ It holds that $$\begin{aligned}
-\frac{T}{\eta} R_Y(\bar{y}_{T+1}) \leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t (x,y) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) \leq \frac{T}{\eta} R_X(\bar{x}_{T+1}).\end{aligned}$$
We will first show that $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) \leq \frac{T}{\eta}R_X(\bar{x}_{T+1}).\end{aligned}$$
Plugging in the definition of ${\mathcal{L}}_t$, we have $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) &= \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(x_{T+1},y_{T+1}) + \frac{1}{\eta}R_X(x_{T+1}) - \frac{1}{\eta} R_Y(y_{T+1}) ]\\
&\leq \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(\bar{x}_{T+1},y_{T+1}) + \frac{1}{\eta}R_X(\bar{x}_{T+1}) - \frac{1}{\eta} R_Y(y_{T+1}) ]\\
&\leq \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(\bar{x}_{T+1},\bar{y}_{T+1}) + \frac{1}{\eta}R_X(\bar{x}_{T+1}) - \frac{1}{\eta} R_Y(y_{T+1}) ],\end{aligned}$$ where the first inequality holds since $(x_{T+1},y_{T+1})$ is the saddle point of $\sum_{t=1}^T {\mathcal{L}}_t(x,y)$ and thus $x_{T+1}$ is the minimizer of $\sum_{t=1}^T {\mathcal{L}}_t(x,y_{T+1})$ see Equation . The second inequality holds since $(\bar{x}_{T+1},\bar{y}_{T+1})$ is a saddle point of $\sum_{t=1}^T \bar{{\mathcal{L}}}_{t}(x,y)$, and thus $\bar{y}_{T+1}$ is the maximizer of $\sum_{t=1}^T {\mathcal{L}}_t(\bar{x}_{T+1},y)$, see Equation . By definition of $(\bar{x}_{T+1},\bar{y}_{T+1})$, we have $$\begin{aligned}
&\sum_{t=1}^T[\bar{{\mathcal{L}}}_t(\bar{x}_{T+1},\bar{y}_{T+1}) + \frac{1}{\eta}R_X(\bar{x}_{T+1}) - \frac{1}{\eta} R_Y(y_{T+1}) ]\\
&= {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T[\bar{{\mathcal{L}}}_t(x,y) + \frac{T}{\eta}R_X(\bar{x}_{T+1}) - \frac{T}{\eta} R_Y(y_{T+1}) ]\\
& \leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) + \frac{T}{\eta}R_X(\bar{x}_{T+1}),\end{aligned}$$ where the inequality holds by nonnegativity of $R_Y$.
Using a similar argument we now show that $$\begin{aligned}
-\frac{T}{\eta}R_Y(\bar{y}_{T+1}) \leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y).\end{aligned}$$
Plugging in the definition of ${\mathcal{L}}_t$, we have $$\begin{aligned}
{\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) &= \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(x_{T+1},y_{T+1}) + \frac{1}{\eta}R_X(x_{T+1}) - \frac{1}{\eta} R_Y(y_{T+1}) ]\\
& \geq \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(x_{T+1},\bar{y}_{T+1}) + \frac{1}{\eta}R_X(x_{T+1}) - \frac{1}{\eta} R_Y(\bar{y}_{T+1}) ] \\
&\geq \sum_{t=1}^T[\bar{{\mathcal{L}}}_t(\bar{x}_{T+1},\bar{y}_{T+1}) + \frac{1}{\eta}R_X(x_{T+1}) - \frac{1}{\eta} R_Y(\bar{y}_{T+1}) ],\end{aligned}$$ where the first inequality holds since $(x_{T+1},y_{T+1})$ is a saddle point of $\sum_{t=1}^T {\mathcal{L}}_t(x,y)$ and thus $y_{T+1}$ is a maximizer of $\sum_{t=1}^T {\mathcal{L}}_t(x_{T+1},y)$, see Equation . The second inequality holds since $(\bar{x}_{T+1},\bar{y_{T+1}})$ is a saddle point of $\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y)$ thus $\bar{x}_{T+1}$ is a minimizer of $\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,\bar{y}_{T+1})$, see Equation .
By definition of $(\bar{x}_{T+1},\bar{y}_{T+1})$, we have $$\begin{aligned}
\sum_{t=1}^T[\bar{{\mathcal{L}}}_t(\bar{x}_{T+1},\bar{y}_{T+1}) + \frac{1}{\eta}R_X(x_{T+1}) - \frac{1}{\eta} R_Y(\bar{y}_{T+1}) ] &= {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T[\bar{{\mathcal{L}}}_t(x,y) + \frac{T}{\eta}R_X(x_{T+1}) - \frac{T}{\eta} R_Y(\bar{y}_{T+1}) ]\\
& \geq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) - \frac{T}{\eta}R_Y(\bar{y}_{T+1}),\end{aligned}$$ where the inequality holds by nonnegativity of $R_Y$. This concludes the proof.
To prove the SP-Regret bound, we note that is running on functions $\{{\mathcal{L}}_{t=1}^T\}$ so the proof will be similar to that of Theorem \[thm:sp\_regret\_str\].
We first prove one side of the inequality. $$\begin{aligned}
& \quad \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y)\\
& \leq \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) \quad \text{by Equation \eqref{eq:diff_bar_not_bar}}\\
& \leq \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1}) \quad \text{by Lemma \ref{lemma:mm_bar_not_bar}}\\
& \leq \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) - \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1}) + G_{{\mathcal{L}}} \sum_{t=1}^T \Vert y_t- y_{t+1}\Vert \quad \text{by Lemma \ref{loss_BTL}}\\
& \leq \sum_{t=1}^T G_{{\mathcal{L}}} (\Vert x_t - x_{t+1}\Vert + \Vert y_{t}-y_{t+1}\Vert) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1}) + G_{{\mathcal{L}}} \sum_{t=1}^T \Vert y_t- y_{t+1}\Vert \\
& \quad \text{since ${\mathcal{L}}_t$ is $G_{\mathcal{L}}$-Lipschitz} \\
& \leq 2\sum_{t=1}^T G_{{\mathcal{L}}} (\Vert x_t - x_{t+1}\Vert + \Vert y_{t}-y_{t+1}\Vert) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1}).\end{aligned}$$
Applying Lemma \[FTL\_close\_iterates\] using $H=\frac{1}{\eta}$ and $G_{\mathcal{L}}= G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})$, we have $$\begin{aligned}
&\quad 2\sum_{t=1}^T G_{{\mathcal{L}}} (\Vert x_t - x_{t+1}\Vert + \Vert y_{t}-y_{t+1}\Vert) + \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1})\\
& \leq 8 G_{{\mathcal{L}}} \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})] ( 1 + \int_{1}^T \frac{1}{t} dt )+ \sum_{t=1}^T \frac{1}{\eta} R_Y(y_t) + \frac{T}{\eta} R_X(x_{T+1})\\
& \leq 8 G_{{\mathcal{L}}} \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})] ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in Y} R_Y(y) + \frac{T}{\eta} \max_{x\in X} R_X(x)\\
& \leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in Y} R_Y(y) + \frac{T}{\eta} \max_{x\in X} R_X(x).\end{aligned}$$ This completes the proof for one side of the inequality. We now prove the other side of the inequality. $$\begin{aligned}
& \quad {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) - \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t)\\
& \leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) \quad \text{by Equation \eqref{eq:diff_bar_not_bar}}\\
& \leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1}) \quad \text{by Lemma \ref{lemma:mm_bar_not_bar}}\\
& \leq \sum_{t=1}^T {\mathcal{L}}_t(x_{t+1},y_{t+1}) - \sum_{t=1}^T {\mathcal{L}}_t(x_t,y_t) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1}) + G_{{\mathcal{L}}} \sum_{t=1}^T \Vert x_t- x_{t+1}\Vert \quad \text{by Lemma \ref{loss_BTL}}\\
& \leq \sum_{t=1}^T G_{{\mathcal{L}}} (\Vert x_t - x_{t+1}\Vert + \Vert y_{t}-y_{t+1}\Vert) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1}) + G_{{\mathcal{L}}} \sum_{t=1}^T \Vert x_t- x_{t+1}\Vert \\
& \quad \text{since ${\mathcal{L}}_t$ is $G_{\mathcal{L}}$-Lipschitz} \\
& \leq 2\sum_{t=1}^T G_{{\mathcal{L}}} (\Vert x_t - x_{t+1}\Vert + \Vert y_{t}-y_{t+1}\Vert) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1}) \\
& \leq 2\sum_{t=1}^T G_{{\mathcal{L}}} ( \frac{4 \eta }{t} [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1}).\end{aligned}$$
Applying Lemma \[FTL\_close\_iterates\] using $H=\frac{1}{\eta}$ and $G_{\mathcal{L}}= G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})$, we have $$\begin{aligned}
& \quad 2\sum_{t=1}^T G_{{\mathcal{L}}} ( \frac{4 \eta }{t} [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]) + \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1})\\
& \leq 8 G_{{\mathcal{L}}} \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})] ( 1 + \int_{1}^T \frac{1}{t} dt )+ \sum_{t=1}^T \frac{1}{\eta} R_X(x_t) + \frac{T}{\eta} R_Y(y_{T+1})\\
& \leq 8 G_{{\mathcal{L}}} \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})] ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in Y} R_Y(y) + \frac{T}{\eta} \max_{x\in X} R_X(x)\\
& \leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in Y} R_Y(y) + \frac{T}{\eta} \max_{x\in X} R_X(x).\end{aligned}$$
This concludes the proof.
Online Matrix Games
===================
In Section \[setup\_problem\], we analyzed the OSP problem by treating the payoff functions as general convex-concave functions and the action spaces as general convex compact sets. We explained that, in general, one should expect to achieve SP-Regret which depends *linearly* in the dimension of the problem (see discussion after Corollary \[first\_corollary\]).
In this section, we consider a special case of the OSP problem with bilinear payoff functions, which we call Online Matrix Games (OMG). In this setting, player 1 has $d_1$ available actions and player 2 has $d_2$ available actions. At each time step $t=1,...,T$, the payoff of the players will be given by a payoff matrix $A_t \in [-1,1]^{d_1\times d_2}$, where the $(i,j)$-th entry specifies the loss of player 1 and the reward of player 2 when they choose actions $i$ and $j$ respectively. We allow the players to choose probability distributions over their available actions. That is, the decision sets of player 1 and player 2 are the probability simplexes, $\Delta_X\subset \mathbb{R}^{d_1}$ and $\Delta_Y\subset \mathbb{R}^{d_2}$, respectively. Here $\Delta_X$ denotes the probability simplex over $d_1$ actions, that is $\Delta_X \triangleq \{x\in \mathbb{R}^{d_1}: x\geq 0, \Vert x \Vert_1\ = 1\}$, $\Delta_Y$ is defined similarly. Notice that this is a special case of the OSP problem studied in Section \[setup\_problem\], where the convex-concave function is defined as ${\mathcal{L}}_t(x,y) = x^\top A_t y$, which specifies the expected payoff for the players when they choose distributions $x\in \Delta_X,y\in \Delta_Y$.
Our goal in this section is to obtain sharper SP-Regret bounds that scale *logarithmically* in the dimensions $d_1$ and $d_2$. This will allow us to solve games that may have exponentially many actions, which often arise in combinatorial optimization settings.
To achieve this goal, we exploit the geometry of the probability simplexes $\Delta_X, \Delta_Y$ and the bilinear structure of the payoff functions. We use the negative entropy as a regularization function (which is strongly convex with respect to $\Vert \cdot \Vert_1$), that is $R_X(x) = \sum_{i=1}^{d_1} x_i\ln(x_i) + \ln(d_1)$ and $R_Y(y)= \sum_{i=1}^{d_2} y_i \ln(y_i)+\ln(d_2)$ where the extra logarithmic terms ensure $R_X, R_Y$ are nonnegative everywhere in their respective simplexes. Unfortunately, the negative entropy is not Lipschitz over the simplex, so we can not leverage our result from Theorem \[theorem:sp\_regret\_convex\_concave\]. To deal with this challenge, we will restrict the new algorithm to play over a restricted simplex:[^1] $$\Delta_\theta = \{z \in \mathbb{R}^d: \Vert z\Vert_1=1, z_i\geq \theta, i=1,...,d \}.$$ The tuning parameter $\theta \in [0,1/d]$ used for the algorithm will be defined later in the analysis. (Notice that when $\theta > {1}/{d}$, the set is empty). We have the following result.
\[lemma:negative\_entropy\_lipschitz\] The function $R(x)\triangleq \sum_{i=1}^d x_i \ln(x_i)$ is $G_R$-Lipschitz continuous with respect to $\Vert \cdot \Vert_1$ over $\Delta_\theta$ with $G_R = \max\{|\ln(\theta)|,1\}$.
We need to find $G_R>0$ such that $\Vert \nabla R(x)\Vert_\infty \leq G_R$ for all $x\in \Delta_\theta$. Notice that $[\nabla R(x)]_i = 1 + \ln(x_i)$ for $i=1,...d$. Moreover, since for every $i=1,...,d$ we have $\theta \leq x_i\leq 1$ the following sequence of inequalities hold: $\ln(\theta)\leq 1 +\ln(\theta) \leq 1+ \ln(x_i) \leq 1$. It follows that $G_R = \max\{|\ln(\theta)|,1\}$.
The algorithm is an instantiation of with a particular choice of regularization functions, which are nonegative and Lipschitz with respect to the $\Vert \cdot \Vert_1$ norm over the sets $\Delta_{X,\theta}$, $\Delta_{Y,\theta}$. With this, we can prove a SP-Regret bound for the OMG problem. For the remainder of the section, the regularization functions will be set as follows: $$R_X(x) \triangleq \textstyle \sum_{i=1}^{d_1} x_i \ln(x_i)+\ln(d_1), \quad
R_Y(y) \triangleq \textstyle \sum_{i=1}^{d_2} y_i\ln(y_i)+\ln(d_2).$$
$x_1 \in \Delta_{X,\theta}\subset \mathbb{R}^{d_1}$, $y_1 \in \Delta_{Y,\theta}\subset \mathbb{R}^{d_2}$, parameters: $\eta>0$, $\theta<\min \{\frac{1}{d_1}, \frac{1}{d_2}\}$. Play $(x_t,y_t)$, observe matrix $A_t$ $\bar{{\mathcal{L}}}_t \gets x^\top A_t y$ ${\mathcal{L}}_t(x,y) \gets \bar{{\mathcal{L}}}_t + \frac{1}{\eta}R_X(x)-\frac{1}{\eta}R_Y(y)$ $x_{t+1}\leftarrow \arg \min_{x\in \Delta_{X,\theta}} \max_{y \in \Delta_{Y, \theta}} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y)$ $y_{t+1}\leftarrow \arg \max_{y \in \Delta_{Y, \theta}} \min_{x\in \Delta_{X,\theta}} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y) $
We have the following guarantee for .
\[thm:omg\_rftl\_regret\] Let $\{A_t\}_{t=1}^T$ be an arbitrary sequence of matrices in $[-1, 1]^{d_1 \times d_2}$. Let $G_{\bar{{\mathcal{L}}}}$ be the Lipschitz constant (with respect to $\Vert \cdot \Vert_1$) of $\bar{{\mathcal{L}}}_t \triangleq x^\top A_t y$ for $t=1,...,T$. Let $\{(x_t,y_t)\}_{t=1}^T$ be the iterates of and choose $\theta = e^{-\eta G_{\bar{{\mathcal{L}}}}}\leq \min\{\frac{1}{d_1}, \frac{1}{d_2}\}$ such that $\frac{|\ln(\theta)|}{\eta} = G_{\bar{{\mathcal{L}}}}$. By setting $\eta = \frac{\sqrt{T}}{G_{\bar{{\mathcal{L}}}}}$ in Algorithm \[alg:OMG-RFTL\], it holds that $$\begin{aligned}
& \quad \left|\sum_{t=1}^T x_t^\top A_t y_t - {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^\top A_t y \right|\\
& \leq 32 G_{\bar{{\mathcal{L}}}} \sqrt{T} (1 + \ln(T)) + 2 \sqrt{T} \max\{\ln d_1 ,\ln d_2\} + 2 \max\{d_1,d_2\} G_{\bar{{\mathcal{L}}}} T e^{-\sqrt{T}}\\
& = O\left(\ln(T)\sqrt{T} + \sqrt{T} \max\{\ln d_1 ,\ln d_2\}\right) + o(1)\max\{d_1,d_2\}.\end{aligned}$$
To prove the theorem, we require a few intermediate results. Since Algorithm \[alg:OMG-RFTL\] selects actions over a restricted simplex, we must quantify the loss in the SP-Regret bound imposed by this restriction. The next two lemmas make this precise.
\[lemma:dist\_proj\_sp\] Let $z^* \in \Delta \subset \mathbb{R}^d$ define $z^*_p \triangleq \arg \min_{z\in \Delta_\theta} \Vert z - z^*\Vert_1$, with $\theta \leq \frac{1}{d}$. Notice $z^*_p$ is unique since it is a projection. It holds that $\Vert z^*_p - z^*\Vert_1 \leq 2\theta(d-1)$.
Choose $z^*=[1;0;0;...;0;0]$, it is easy to see that $z^*_p = [1-\theta (d-1); \theta; \theta; ...; \theta; \theta]$ and $\Vert z^* - z^*_p \Vert_1 = 2\theta(d-1).$
\[lemma:sp\_val\_error\_theta\]
Let $\{\bar{{\mathcal{L}}}_t(x,y)\}_{t=1}^T$ be an arbitrary sequence of convex-concave functions, $\bar{{\mathcal{L}}}_t:\Delta_X \times \Delta_Y \rightarrow \mathbb{R}$, that are $G_{\bar{{\mathcal{L}}}}$-Lipschitz with respect to $\Vert \cdot \Vert_1$. It holds that $$\begin{aligned}
- G_{\bar{{\mathcal{L}}}} T \Vert x^*_p - x^*\Vert_1
\leq {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y) - {\min_{x\in \Delta_{X,\theta}}\max_{y\in \Delta_{Y,\theta}}}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y)
\leq G_{\bar{{\mathcal{L}}}}T \Vert y^*_p - y^*\Vert_1.\end{aligned}$$
Let $(x^*,y^*)$ be any saddle point pair for $\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y)$ with $x^*\in \Delta_X, y^*\in \Delta_Y$. Let $(x^*_\theta,y^*_\theta)$ be any saddle point pair for $\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x,y)$ with $x^*_\theta \in \Delta_{X,\theta}, y^*_\theta \in \Delta_{Y,\theta}$. Let $x^*_p, y^*_p$ be the projection of $x^*, y^*$ onto $\Delta_{X,\theta}, \Delta_{Y,\theta}$ respectively, using the $\Vert \cdot \Vert_\infty$ norm. We first show the second inequality.
Since $(x^*,y^*)$ is a saddle point for $\sum_{t=1}^T \bar{{\mathcal{L}}}_t (x,y)$ over $\Delta_X$ and $\Delta_Y$, and Player 1 deviated to $x^*_\theta$, we have $$\begin{aligned}
\sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*, y^*) & \leq \sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*_\theta, y^*)\\
& \leq \sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*_\theta, y^*_p) + G_{\bar{{\mathcal{L}}}}T \Vert y^*_p - y^*\Vert_1 \quad \text{since $\bar{{\mathcal{L}}}_t$ is $G_{\bar{{\mathcal{L}}}}$-Lipschitz}\\
& \leq \sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*_\theta, y^*_\theta) + G_{\bar{{\mathcal{L}}}}T \Vert y^*_p - y^*\Vert_1,\end{aligned}$$ where the last inequality holds since $(x^*_\theta, y^*_\theta)$ is a saddle point for $\sum_{t=1}^T \bar{{\mathcal{L}}}_t (x,y)$ over $\Delta_{X,\theta}$ and $\Delta_{Y,\theta}$.
To show the first inequality in the statement of the lemma, by using similar argument, we have $$\begin{aligned}
\sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*, y^*) & \geq \sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*, y^*_\theta) \\
& \geq \sum_{t=1}^T \bar{{\mathcal{L}}}_t (x^*_p, y^*_\theta) - G_{\bar{{\mathcal{L}}}} T \Vert x^*_p - x^*\Vert_1\\
& \geq \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x^*_\theta, y^*_\theta) - G_{\bar{{\mathcal{L}}}} T \Vert x^*_p - x^*\Vert_1.\end{aligned}$$ This concludes the proof.
Combining the previous two lemmas and Theorem \[theorem:sp\_regret\_convex\_concave\], we can show the SP-Regret bound for holds. We are ready to prove Theorem \[thm:omg\_rftl\_regret\].
For convenience, we define $\bar{{\mathcal{L}}}_t(x,y) = x^\top A_t y$. Let $(x^*, y^*)$ be any saddle point of ${\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^\top A_t y$, and let $(x^*_p, y^*_p)$ be the respective projections onto $\Delta_{X,\theta}, \Delta_{Y,\theta}$ using $\Vert \cdot \Vert_\infty$ norm. Using Lemma \[lemma:dist\_proj\_sp\], Lemma \[lemma:sp\_val\_error\_theta\] and Theorem \[theorem:sp\_regret\_convex\_concave\], we have $$\begin{aligned}
& \quad \sum_{t=1}^T x_t^\top A_t y_t - {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^\top A_t y \\
& \leq \sum_{t=1}^T x_t^\top A_t y_t - {\min_{x\in \Delta_{X,\theta}}\max_{y\in \Delta_{Y,\theta}}}\sum_{t=1}^T x^\top A_t y + G_{\bar{{\mathcal{L}}}} T \Vert x^* - x^*_p\Vert_1 \quad \text{by Lemma \ref{lemma:sp_val_error_theta} }\\
& \leq \sum_{t=1}^T x_t^\top A_t y_t - {\min_{x\in \Delta_{X,\theta}}\max_{y\in \Delta_{Y,\theta}}}\sum_{t=1}^T x^\top A_t y + 2G_{\bar{{\mathcal{L}}}} T \theta (d_1-1) \quad \text{by Lemma \ref{lemma:dist_proj_sp}}\\
&\leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X\theta}} R_X(x)\\
&\qquad + 2G_{\bar{{\mathcal{L}}}} T \theta (d_1-1) \quad \text{by Theorem \ref{theorem:sp_regret_convex_concave}}.\end{aligned}$$
By Lemma \[lemma:negative\_entropy\_lipschitz\], we know that $G_{R_X},G_{R_Y} \leq \max\{\vert \ln(\theta),1\vert\}$. Our choice of $\theta$ will ensure that $1\leq \vert \ln(\theta)\vert$, so $\max(G_{R_X},G_{R_Y})\leq \vert \ln(\theta)\vert$. Therefore, we have $$\begin{aligned}
& 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T \theta (d_1-1)\\
&\leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{|\ln(\theta)|}{\eta}]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T \theta (d_1-1)\\
&\leq 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_1-1),
\end{aligned}$$ where the last inequality holds by the choice of $\theta$.
Notice that $\max_{z\in \Delta_\theta} R(z) \triangleq \max_{z\in \Delta_\theta} \sum_{i=1}^d z_i \ln(z_i) + \ln(d) \leq 0 +\ln(d)$. Therefore, we have $$\begin{aligned}
&\quad 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_1-1)\\
& \leq 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta}\ln(d_2) + \frac{T}{\eta} \ln(d_1)+ 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_1-1)\\
& \leq 32 G_{\bar{{\mathcal{L}}}} \sqrt{T} (1 + \ln(T)) + \sqrt{T} (\ln d_1+\ln d_2) + 2 d_1 G_{\bar{{\mathcal{L}}}} T e^{-\sqrt{T}} \\
& = O\left(\ln(T)\sqrt{T} + \sqrt{T} \max\{\ln d_1 ,\ln d_2\}\right) + o(1)\max\{d_1,d_2\}.\end{aligned}$$ The last line follows because $G_{\bar{{\mathcal{L}}}}\leq 1$, since each entry of $A_t$ is bounded between $[-1, 1]$.
We now prove the other side of the inequality: $$\begin{aligned}
& \quad {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^\top A_t y - \sum_{t=1}^T x_t^\top A_t y_t \\
& \leq {\min_{x\in \Delta_{X,\theta}}\max_{y\in \Delta_{Y,\theta}}}\sum_{t=1}^T x^\top A_t y - \sum_{t=1}^T x_t^\top A_t y_t + G_{\bar{{\mathcal{L}}}} T \Vert y^* - y^*_p\Vert_1 \quad \text{by Lemma \ref{lemma:sp_val_error_theta} }\\
& \leq {\min_{x\in \Delta_{X,\theta}}\max_{y\in \Delta_{Y,\theta}}}\sum_{t=1}^T x^\top A_t y - \sum_{t=1}^T x_t^\top A_t y_t + 2G_{\bar{{\mathcal{L}}}} T \theta (d_2-1) \quad \text{by Lemma \ref{lemma:dist_proj_sp}}\\
&\leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{1}{\eta}\max(G_{R_X}, G_{R_Y})]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x)\\
&\qquad + 2G_{\bar{{\mathcal{L}}}} T \theta (d_2-1) \quad \text{by Theorem \ref{theorem:sp_regret_convex_concave}}\\
&\leq 8 \eta [G_{\bar{{\mathcal{L}}}}+ \frac{|\ln(\theta)|}{\eta}]^2 ( 1 + \ln(T) )+ \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T \theta (d_2-1)\\
&\leq 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_2-1),
\end{aligned}$$\
where the last inequality holds by the choice of $\theta$. Again, notice that $\max_{z\in \Delta_\theta} R(z) \triangleq \max_{z\in \Delta_\theta} \sum_{i=1}^d z_i \ln(z_i) + \ln(d) \leq 0 +\ln(d)$. We have $$\begin{aligned}
& 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta} \max_{y\in \Delta_{Y,\theta}} R_Y(y) + \frac{T}{\eta} \max_{x\in \Delta_{X,\theta}} R_X(x) + 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_2-1)\\
& \leq 32 \eta G_{\bar{{\mathcal{L}}}}^2 (1+\ln(T)) + \frac{T}{\eta}\ln(d_2) + \frac{T}{\eta} \ln(d_1)+ 2G_{\bar{{\mathcal{L}}}} T e^{-\eta G_{\bar{{\mathcal{L}}}}} (d_2-1)\\
& \leq 32 G_{\bar{{\mathcal{L}}}} \sqrt{T} (1 + \ln(T)) + \sqrt{T} (\ln d_1+\ln d_2) + 2 d_2 G_{\bar{{\mathcal{L}}}} T e^{-\sqrt{T}} \\
& = O\left(\ln(T)\sqrt{T} + \sqrt{T} \max\{\ln d_1 ,\ln d_2\}\right) + o(1)\max\{d_1,d_2\}.\end{aligned}$$ The last line follows because $G_{\bar{{\mathcal{L}}}}\leq 1$, since each entry of $A_t$ is bounded between $[-1, 1]$. This concludes the proof.
Online Matrix Games with Bandit Feedback {#section:omg_bandit}
----------------------------------------
The results we proved for the OMG problem can be extended to a setting with bandit feedback. In the bandit setting, the players observe in every round only the payoff corresponding to the chosen actions. In other words, if Player 1 chooses action $i$, Player 2 chooses action $j$, and the payoff matrix at that time step is $A_t$, then the players observe only $(A_t)_{ij}$ instead of the full matrix $A_t$. The limited feedback makes the problem significantly more challenging than the full information one, as the players must balance the exploration-exploitation tradeoff. This problem resembles that of Online Bandit Optimization [@flaxman2005online; @auer1995gambling; @bubeck2016kernel; @hazan2016optimal], albeit with two players. For convenience, we define some useful notation. For $i=1,...,d$, let $e_i \in \mathbb{R}^d$ be the collection of standard unit vectors i.e. $e_i$ is the vector that has a $1$ in the $i$-th entry and $0$ in the rest. Let $e_{x,t}$ be the standard unit vector corresponding to the decision made by Player 1 for round $t$, define $e_{y,t}$ similarly. Notice that under bandit feedback, in round $t$ both players only observe the quantity $e_{x,t}^{\top} A_t e_{y,t}$.
### One-Point Estimate for Payoff Function
As explained previously, in each round $t$ the players must estimate $A_t$ by observing only one of its entries. To this end, we allow the players to share with each other their decisions and to randomize *jointly* (a similar assumption is used to define correlated equilibria in zero-sum games, see [@aumann1987correlated]). The following result shows how to build a random estimate of $A$ by observing only one of its entries.
\[thm:hess\_estimate\] Let $x\in \Delta_{X,\delta}, y\in \Delta_{Y,\delta}$ with $d_1,d_2\geq2$ and $\delta >0$. Sample $i' \sim x, j'\sim y$. Let $\hat{A}$ be the $d_1\times d_2$ matrix with $\hat{A}_{i,j}=0$ for all $i,j$ such that $i\neq i'$ and $j\neq j'$ and $\hat{A}_{i',j'} = \frac{A_{i',j'}}{x_{i'}y_{j'}}$. It holds that $$\mathbb{E}_{i' \sim x, j' \sim y} [ \hat{A}] = A.$$
Let $B_{i,j}$ be the matrix of zeros everywhere except in the $i,j$ entry where it is equal to $ \frac{A_{i,j}}{x(i)y(j)}$. We have $$\begin{aligned}
\mathbb{E}_{i' \sim x, j' \sim y} [ \hat{A}] = \sum_{i=1}^{d_1} \sum_{j=1}^{d_2} x_i y_j B_{i,j} = A. \end{aligned}$$
### Algorithm under Bandit Feedback
We now present an algorithm that ensures sublinear (i.e. $o(T)$) SP-Regret under bandit feedback for the OMG problem that holds against an adaptive adversary. By adaptive adversary, we mean that the payoff matrices $A_t$ can depend on the players’ actions up to time $t-1$; in particular, we assume the adversary does not observe the actions chosen by the players for time period $t$ when choosing $A_t$. We consider an algorithm that runs on a sequence of functions $\hat{{\mathcal{L}}}_t \triangleq x^\top \hat{A}_t y$, where $\hat{A}_t$ is the unbiased one-point estimate of $A_t$ derived in Theorem \[thm:hess\_estimate\]. Recall that the iterates of algorithm are distributions over the possible actions of both players. In order to generate the estimate $\hat{A}_t$, both players will sample an action from their distributions and weigh their observation with the inverse probability of obtaining that observation.
$x_1 \in \Delta_{X,\delta}\subset \mathbb{R}^{d_1}$, $y_1 \in \Delta_{Y,\delta}\subset \mathbb{R}^{d_2}$, parameters: $\eta>0$, $0<\delta<\min \{\frac{1}{d_1}, \frac{1}{d_2}\}$. Sample independently $e_{x,t} \sim x_t$ and $e_{y,t} \sim y_t$ Observe $e_{x,t}^{\top} A_t e_{y,t}$ Build $\hat{A_t}$ as in Theorem \[thm:hess\_estimate\] using $e_{x,t}^{\top} A_t e_{y,t}, x_t, y_t$ $\hat{{\mathcal{L}}}_t \gets x^\top \hat{A}_t y$ ${\mathcal{L}}_t(x,y) \gets \hat{{\mathcal{L}}}_t + \frac{1}{\eta}R_X(x)-\frac{1}{\eta}R_Y(y)$ $x_{t+1}\leftarrow \arg \min_{x\in \Delta_{X,\theta}} \max_{y \in \Delta_{Y, \theta}} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y)$ $y_{t+1}\leftarrow \arg \max_{y \in \Delta_{Y, \theta}} \min_{x\in \Delta_{X,\theta}} \sum_{\tau=1}^t {\mathcal{L}}_t(x,y) $
We have the following guarantee for .
\[no\_bandit\_regret\] Let $\{A_t\}_{t=1}^T$ be any sequence of payoff matrices chosen by an adaptive adversary, where $A_t\in [-1,1]^{d_1 \times d_2}$ for all $t=1,...,T$. Let $\{e_{x,t},e_{y,t}\}_{t=1}^T$ be the iterates generated by . Setting $\delta = \frac{1}{T^{1/6}}$, $\eta = T^{1/6}$ ensures $$\begin{aligned}
\left| \mathbb{E} \left[\sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t} - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T x^{\top}A_t y \right] \right| \leq O((d_1 + d_2) \ln(T) T^{5/6})
\end{aligned}$$ where the expectation is taken with respect to all the randomization used in the algorithm.
The full proof of this Theorem will be given shortly. We now present a few lemmas. The total payoff given to each of the players is given by $\sum_{t=1}^T e_{x,t}^\top A_t e_{y,t}$ so we must relate this quantity to the iterates $\{x_t,y_t\}_{t=1}^T$ of when run on sequence of matrices $\{\hat{A}_t \}_{t=1}^T$. The following two lemmas will allow us to do so.
\[e\_to\_x\] Let $\{e_{x,t}, e_{y,t}\}_{t=1}^T$ be the sequence of iterates generated by . It holds that $$\begin{aligned}
\textstyle
\mathbb{E}\left[ \sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t}\right] = \mathbb{E}\left[ \sum_{t=1}^T x^{\top}_t A_t y_t\right],
\end{aligned}$$ where the expectation is taken with respect to the internal randomness of the algorithm.
Let $\mathbb{E}[X | \tau=1,...,T-1]$ be the expectation of random variable $X$ conditioned on all the randomness from time steps $\tau=1,...,T-1$. $$\begin{aligned}
&\quad \mathbb{E}[\sum_{t=1}^T e_{x,t}^{\top} A_t e_{y,t}]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} e_{x,t}^{\top} A_t e_{y,t}] + \mathbb{E}[e_{x,T}^{\top} A_T e_{y,T}] \\
& = \mathbb{E}[\sum_{t=1}^{T-1} e_{x,t}^{\top} A_t e_{y,t}] + \mathbb{E}[ \mathbb{E}_{e_{x,T} \sim x_t, e_{y,T} \sim y_t}[e_{x,T}^{\top} A_T e_{y,T}|\tau=1,...,T-1]].\end{aligned}$$
Since the adversary can can not observe $e_{x,T}, e_{y,T}$ when selecting $A_T$, $e_{x,T}, e_{y,T}$ and $A_T$ are all independent from each other, thus it holds that $\mathbb{E}[ \mathbb{E}_{e_{x,T} \sim x_t, e_{y,T} \sim y_t}[e_{x,T}^{\top} A_T e_{y,T}|\tau=1,...,T-1]] = \mathbb{E}[ x_T^{\top} \mathbb{E}_{e_{x,T} \sim x_t, e_{y,T} \sim y_t}[ A_T |\tau=1,...,T-1]y_T ]$. Therefore
$$\begin{aligned}
&\quad \mathbb{E}[\sum_{t=1}^{T-1} e_{x,t}^{\top} A_t e_{y,t}] + \mathbb{E}[ \mathbb{E}_{e_{x,T} \sim x_t, e_{y,T} \sim y_t}[e_{x,T}^{\top} A_T e_{y,T}|\tau=1,...,T-1]] \\
& = \mathbb{E}[\sum_{t=1}^{T-1} e_{x,t}^{\top} A_t e_{y,t}] + \mathbb{E}[ x_T^{\top} \mathbb{E}_{e_{x,T} \sim x_t, e_{y,T} \sim y_t}[ A_T |\tau=1,...,T-1]y_T ]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} A_t y_t] + \mathbb{E}[ x_T^{\top} A_T y_T ] .\end{aligned}$$
Repeating the argument $T-1$ more times yields the result.
\[A\_hat\_no\_hat\] Let $\{A_t\}_{t=1}^T$ be any sequence of payoff matrices chosen by an adaptive adversary, where $A_t\in \mathbb{R}^{d_1 \times d_2}$ for all $t=1,...,T$. Let $\{x_t,y_t,\hat{A}_t\}_{t=1}^T$ be generated by . It holds that $$\begin{aligned}
\textstyle
\mathbb{E}\left[\sum_{t=1}^T x_t^{\top} \hat{A}_t y_t\right] = \mathbb{E}\left[\sum_{t=1}^T x_t^{\top} A_t y_t\right],\end{aligned}$$ where the expectation is with respect to all the internal randomness of the algorithm.
We have $$\begin{aligned}
& \quad \mathbb{E}[\sum_{t=1}^T x_t^{\top} \hat{A}_t y_t]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} \hat{A}_t y_t] + \mathbb{E}[x_T^{\top} \hat{A}_T y_T]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} \hat{A}_t y_t] + \mathbb{E}[ \mathbb{E}[x_T^{\top} \hat{A}_T y_T|\tau=1,...,T-1]].\end{aligned}$$ Since $(x_T,y_T)$ is deterministic conditioned on everything that has happened up to time $T-1$ it holds that $\mathbb{E}[ \mathbb{E}[x_T^{\top} \hat{A}_T y_T|\tau=1,...,T-1]] = \mathbb{E}[ x_T^{\top} \mathbb{E}[ \hat{A}_T |\tau=1,...,T-1]y_T ]$. It then follows that $$\begin{aligned}
& \quad \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} \hat{A}_t y_t] + \mathbb{E}[ \mathbb{E}[x_T^{\top} \hat{A}_T y_T|\tau=1,...,T-1]]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} \hat{A}_t y_t] + \mathbb{E}[ x_T^{\top} \mathbb{E}[ \hat{A}_T |\tau=1,...,T-1]y_T ]\\
& = \mathbb{E}[\sum_{t=1}^{T-1} x_t^{\top} \hat{A}_t y_t] + \mathbb{E}[ x_T^{\top} A_T y_T ],\end{aligned}$$ where the last equality holds by Theorem \[thm:hess\_estimate\]. Repeating the argument $T-1$ more times yields the result.
We will then bound the difference between the comparator term ${\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top} A_t y$ and the comparator term Theorem \[thm:omg\_rftl\_regret\] gives us by running on functions $\{\hat{{\mathcal{L}}}\}_{t=1}^T$ and sets $\Delta_{X,\delta}, \Delta_{Y,\delta}$, ${\min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}}}\sum_{t=1}^T x^{\top} \hat{A}_t y$. Special care must be taken to ensure this difference holds even against an adaptive adversary. To this end, we use the next two lemmas; as we will see, the proof of Lemma \[lemma\_with\_alphas\] relies heavily on Theorem \[thm:hess\_estimate\].
\[close\_sp\_vals\] With probability 1, for any $y \in \Delta_{Y,\delta}$, it holds that $$\begin{aligned}
\left| \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y - \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}\hat{A}_t y \right|
\leq \max_{y\in \Delta_{Y,\delta}} \left \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \right \Vert_2.\end{aligned}$$
Let us fist bound $|\sum_{t=1}^T x^\top A_t y - \sum_{t=1}^T x^\top \hat{A}_t y|$ for any $x \in \Delta_X$ and $y \in \Delta_Y$ with probability 1. For any $x \in \Delta_X$ and $y \in \Delta_Y$ we have $$\begin{aligned}
|\sum_{t=1}^T x^\top A_t y - \sum_{t=1}^T x^\top \hat{A}_t y|
& = | x^{\top} ( \sum_{t=1}^T A_t y - \sum_{t=1}^T \hat{A}_t y ) |\\
& \leq \Vert x\Vert_2 \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \text{by Cauchy-Schwarz}\\
& \leq \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \text{since $x\in \Delta_X$}.\end{aligned}$$
This implies that $$\begin{aligned}
& \sum_{t=1}^T x^{\top} \hat{A}_t y \leq \sum_{t=1}^T x^{\top}A_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \text{$ \forall x \in \Delta_X$, $ \forall y \in \Delta_Y$},\end{aligned}$$ which implies that $$\begin{aligned}
& \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y \leq \sum_{t=1}^T x^{\top}A_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \forall x \in \Delta_{X,\delta} ,y \in \Delta_{Y,\delta}.\end{aligned}$$ Therefore, it holds that $$\begin{aligned}
&\min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y \leq \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \forall x \in \Delta_{X,\delta},y \in \Delta_{Y,\delta},\end{aligned}$$ Thus $$\begin{aligned}
& \max_{y\in \Delta_{Y,\delta}} \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y \leq \min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y + \max_{y\in \Delta_{Y,\delta}}\Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 . \end{aligned}$$
Since $\max_{y\in \Delta_{Y,\delta}} \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y = \min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y$ (the function is convex-concave and the sets $ \Delta_Y^\delta$ and $\Delta_X^\delta$ are convex and compact), we have shown that $$\begin{aligned}
\min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y \leq \min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y + \max_{y\in \Delta_{Y,\delta}} \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \forall y \in \Delta_{Y,\delta}.\end{aligned}$$ The other side of the inequality follows from a similar argument. Indeed we know that $$\begin{aligned}
& \sum_{t=1}^T x^{\top} A_t y \leq \sum_{t=1}^T x^{\top} \hat{A}_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \text{$ \forall x \in \Delta_X$, $ \forall y \in \Delta_Y$}.\end{aligned}$$ The previous inequality implies that $$\begin{aligned}
& \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} A_t y \leq \sum_{t=1}^T x^{\top}\hat{A}_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \forall x \in \Delta_{X,\delta} ,y \in \Delta_{Y,\delta},
\end{aligned}$$ or $$\begin{aligned}
& \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} A_t y \leq \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y + \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 \quad \forall x \in \Delta_{X,\delta},y \in \Delta_{Y,\delta}.\end{aligned}$$ Therefore, we have $$\begin{aligned}
& \max_{y\in \Delta_{Y,\delta}} \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} A_t y \leq \min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top} \hat{A}_t y + \max_{y\in \Delta_{Y,\delta}} \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 . \end{aligned}$$
Since $\max_{y\in \Delta_{Y,\delta}} \min_{x\in \Delta_{X,\delta}} \sum_{t=1}^T x^{\top} A_t y = \min_{x\in \Delta_{X,\delta}} \max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top} A_t y$ we get the result.
\[lemma\_with\_alphas\] Let $\{A_t\}$ be any sequence of payoff matrices chosen by an adaptive adversary, where with $A_t\in \mathbb{R}^{d_1 \times d_2}$ for all $t=1,...,T$. Let $\{\hat{A}_t\}$ be the sequence of matrices generated by <span style="font-variant:small-caps;">Bandit-OMG-RFTL</span>. For any $y\in \Delta_{Y}$, it holds that $$\begin{aligned}
\mathbb{E}\left[ \left \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \right \Vert_2 \right] \leq \frac{2 \sqrt{T} \min (d_1,d_2)}{\delta^2},\end{aligned}$$ where the expectation is taken with respect to the internal randomness of the algorithm.
For any $y$ define $\alpha_t \triangleq A_t y - \hat{A}_t y$. We first show that for all $t, t'$ such that $t<t'$ it holds that $\mathbb{E}[\alpha_t^{\top}\alpha_{t'}]=0$. Indeed $$\begin{aligned}
\mathbb{E}[\alpha_t^{\top} \alpha_{t'}] &= \mathbb{E}[(A_t y -\hat{A}_t y )^{\top}(A_{t'} y -\hat{A}_{t'} y )]\\
&= \mathbb{E}[(A_t y)^{\top}A_{t'} y - (A_t y)^{\top}\hat{A}_{t'} y - (\hat{A}_t y )^{\top}A_{t'} y + (\hat{A}_t y )^{\top} \hat{A}_{t'} y ]\\
& = (A_t y)^{\top}A_{t'} y - (A_t y)^{\top}A_{t'} y - (A_t y )^{\top}A_{t'} y + \mathbb{E}[(\hat{A}_t y )^{\top} \hat{A}_{t'} y ]\\
& = (A_t y)^{\top}A_{t'} y - (A_t y)^{\top}A_{t'} y - (A_t y )^{\top}A_{t'} y + (A_t y )^{\top} A_{t'} y\\
&= 0,\end{aligned}$$ where the second to last line follows since $$\begin{aligned}
\mathbb{E}[(\hat{A}_t y )^{\top} \hat{A}_{t'} y ] &= \mathbb{E}_{1,...,t'-1}[ \mathbb{E}[(\hat{A}_t y )^{\top} \hat{A}_{t'} y |\tau = 1,..., t'-1]]\\
&= \mathbb{E}_{1,...,t'-1}[ (\hat{A}_t y )^{\top} \mathbb{E}[ \hat{A}_{t'} y |\tau = 1,..., t'-1]]\\
&= \mathbb{E}_{1,...,t'-1}[ (\hat{A}_t y )^{\top} A_{t'}y ]\\
&= (A_t y )^{\top} A_{t'}y.\end{aligned}$$
Now, we have $$\begin{aligned}
\mathbb{E}[ \Vert \sum_{t=1}^T A_t y - \hat{A}_t y \Vert_2 ] & = \sqrt{\mathbb{E}[\Vert \sum_{t=1}^T \alpha_t \Vert_2 ]^2}\\
& \leq \sqrt{ \mathbb{E}[\Vert \sum_{t=1}^T \alpha_t \Vert_2^2 ]} \qquad \text{by Jensen's Inequality}\\
& = \sqrt{ \sum_{t=1}^T \mathbb{E}[\Vert \alpha_t\Vert_2^2] + 2 \sum_{t < t'} \mathbb{E}[\alpha_t^{\top}\alpha_{t'}]}\\
& = \sqrt{\sum_{t=1}^T \mathbb{E}[\Vert A_t y - \hat{A}_t y\Vert_2^2]}\\
& \leq \sqrt{ \sum_{t=1}^T \mathbb{E}[2 \Vert A_t y\Vert ^2 + 2\Vert \hat{A}_t y\Vert_2^2 ] }.\end{aligned}$$ We proceed to bound $\Vert \hat{A}_t y\Vert_2$, the upper bound we obtain will also bound $\Vert A_t y\Vert $ because of the following fact. If the random vector $\tilde{a}$ satisfies $\Vert \tilde{a}\Vert \leq c$ for some constant c with probability 1 then $\Vert \mathbb{E}\tilde{a}\Vert \leq c$. Indeed by Jensen’s inequality, we have $\Vert \mathbb{E}\tilde{a}\Vert \leq \mathbb{E} \Vert \tilde{a}\Vert \leq c$. Let us omit the subscript $t$ for the rest of the proof. Let $\hat{A}_{[i,:]}$ be the $i$-th row of matrix $\hat{A}$.
$$\begin{aligned}
\Vert \hat{A}y\Vert_2 &= \sqrt{\sum_{i=1}^{d_1} \big[\sum_{j=1}^{d_2} \hat{a}_{i,j} y_j\big]^2}\\
&\leq \sum_{i=1}^{d_1} \sqrt{\big[\sum_{j=1}^{d_2} \hat{a}_{i,j} y_j\big]^2}\\
&= \sum_{i=1}^{d_1} \big| \sum_{j=1}^{d_2} \hat{a}_{i,j} y_j \big| \\
& \leq \sum_{i=1}^{d_1} \Vert \hat{A}_{[i,:]} \Vert_{\infty} \Vert y\Vert_1 \quad \text{by generalized Cauchy Schwarz}\\
& \leq d_1 \max_{i,j} |\frac{A_{i,j}}{\delta^2}| \quad \text{by definition of $\hat{A}$ and using the fact that $ x_t \in \Delta_{X,\delta}$ and $y_t \in \Delta_{Y,\delta}$}\\
& \leq \frac{d_1}{\delta^2}.\end{aligned}$$
Notice the upper bound $\frac{d_2}{\delta^2}$ can also be obtained by interchanging the summations and repeating the argument. This yields the desired result.
Before we prove Theorem \[no\_bandit\_regret\], we need the following lemma.
\[bilinear\_lipschitz\] Consider a matrix $A\in \mathbb{R}^{d_1\times d_2}$. If the absolute value of each entry of $A$ is bounded by $c>0$, then the function ${\mathcal{L}}(x,y) = x^{\top}A y$ is $ G_{{\mathcal{L}}}^{\Vert \cdot \Vert_2}$-Lipschitz continuous with respect to $\Vert \cdot \Vert_2$ over the sets $\Delta_X\subset \mathbb{R}^{d_1}$ and $\Delta_Y\subset \mathbb{R}^{d_2}$ , where $ G_{{\mathcal{L}}}^{\Vert \cdot \Vert_2} = \sqrt{c}\left(\sqrt{d_1}+\sqrt{d_2}\right)$. The function ${\mathcal{L}}$ is also $G_{{\mathcal{L}}}^{\Vert \cdot \Vert_1}$-Lipschitz continuous (over the same sets) with respect to norm $\Vert \cdot \Vert_1$, where $G_{{\mathcal{L}}}^{\Vert \cdot \Vert_1} = c$.
$$\begin{aligned}
\Vert \nabla x^\top Ay \Vert_2 &=
\left\Vert
\begin{bmatrix}
\nabla_x x^\top Ay;
\nabla_y x^\top Ay
\end{bmatrix}^\top \right\Vert_2 \\
&=
\left\Vert \begin{bmatrix}
A_{[1,:]}^\top y;
... ;
A_{[d_1,:]}^\top y;
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_2 \\
& \leq
\left\Vert \begin{bmatrix}
A_{[1,:]}^\top y ;
... ;
A_{[d_1,:]}^\top y\\
\end{bmatrix}^\top \right\Vert_2
+
\left\Vert \begin{bmatrix}
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_2 \\
&\leq
\sqrt{\sum_{i=1}^{d_1}(A_{[i,:]}^\top y)^2}
+
\left\Vert \begin{bmatrix}
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_2 \\
&\leq
\sqrt{d_1(\Vert A_{[i,:]}\Vert_{\infty} \Vert y\Vert_1)^2}
+
\left\Vert \begin{bmatrix}
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_2 \quad \text{by Generalized Cauchy Schwarz}\\
&\leq
\sqrt{c d_1}
+
\left\Vert \begin{bmatrix}
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_2 \\
&\leq
\sqrt{c d_1}
+
\sqrt{c d_2} \quad \text{by using the same reasoning}.\end{aligned}$$
We now prove the second part of the claim by bounding $\Vert \nabla x^\top Ay \Vert_\infty$. $$\begin{aligned}
\Vert \nabla x^\top Ay \Vert_\infty &=
\left\Vert
\begin{bmatrix}
\nabla_x x^\top Ay ;
\nabla_y x^\top Ay
\end{bmatrix}^\top \right\Vert_\infty \\
&=
\left\Vert \begin{bmatrix}
A_{[1,:]}^\top y ;
... ;
A_{[d_1,:]}^\top y;
A_{[:,1]}^\top x ;
... ;
A_{[:,d_2]}^\top x
\end{bmatrix}^\top \right\Vert_\infty\end{aligned}$$ By Cauchy-Schwarz inequality, for any $i=1,...,d_1$, we have $A_{[i,:]}^\top y \leq \Vert A_{[i,:]} \Vert_\infty \Vert y\Vert_1 \leq c \Vert y\Vert_1 \leq c$, since $y \in \Delta_Y$. Similarly, for any $j=1,...,d_2$ $A_{[:,j]}^\top x \leq c$. This shows that $\Vert \nabla x^\top Ay \Vert_\infty\leq c$.
The proof of Theorem \[no\_bandit\_regret\] follows by combining Lemmas \[e\_to\_x\] through \[lemma\_with\_alphas\], with careful choice of tuning parameters.
We first focus on one side of the inequality, $$\begin{aligned}
& \quad \mathbb{E}[ \sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t} - {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y ] \\
& = \mathbb{E}[ \sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t}] - \mathbb{E}[{\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y ]\\
& = \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t] - \mathbb{E}[{\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y ] \quad \text{by Lemma \ref{e_to_x} }\\
& = \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t] - \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y ] + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_1-1) T \quad {\text{by Lemmas \ref{lemma:dist_proj_sp} and \ref{lemma:sp_val_error_theta} }}\\
& \leq \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t] - \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}\hat{A}_t y ] + \frac{2 \sqrt{T}\min(d_1,d_2)}{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_1-1) T\\
& \qquad \text{by Lemmas \ref{close_sp_vals} and \ref{lemma_with_alphas}}\\
& \leq \mathbb{E}[ \sum_{t=1}^T x_t^{\top}\hat{A}_t y_t] - \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}\hat{A}_t y ] + \frac{2 \sqrt{T}\min(d_1,d_2)}{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_1-1) T\\
& \qquad \text{by Lemma \ref{A_hat_no_hat}}\\
& \leq 8 \eta [G_{\hat{{\mathcal{L}}}}^{\Vert \cdot \Vert_1} + \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T}\min(d_1,d_2) }{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_1-1) T.\end{aligned}$$
The last inequality follows by the same reasoning we used in the proof of Theorem \[thm:omg\_rftl\_regret\]. Indeed $\{(x_t,y_t)\}_{t=1}^T$ are the iterates of run on the sequence of payoff functions $\{x^\top A_t y\}_{t=1}^T$ so the same proof holds. By Lemma \[bilinear\_lipschitz\], since the absolute value of all the entries in $A_t$ is bounded above by 1, it holds that $$\begin{aligned}
& 8 \eta [G_{\hat{{\mathcal{L}}}}^{\Vert \cdot \Vert_1} + \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T}\min(d_1,d_2) }{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_1-1) T\\
& = 8 \eta [\frac{1}{\delta^2}+ \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T} \min(d_1,d_2) }{\delta^2} + 2\delta (d_1-1) T\\
& = O((d_1 + d_2) \ln(T) T^{5/6}),\end{aligned}$$ where the last equality holds since we use $\delta = \frac{1}{T^{1/6}}$, $\eta = T^{1/6}$.
We now show the other side of the inequality. $$\begin{aligned}
& \quad \mathbb{E}[ {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y - \sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t} ] \\
& = \mathbb{E}[{\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y ] - \mathbb{E}[ \sum_{t=1}^T e_{x,t}^{\top}A_t e_{y,t}] \\
& = \mathbb{E}[{\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^T x^{\top}A_t y ] - \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t]\quad \text{by Lemma \ref{e_to_x} }\\
& = \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}A_t y ] - \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t] + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_2-1) T \quad {\text{by Lemmas \ref{lemma:dist_proj_sp} and \ref{lemma:sp_val_error_theta} }}\\
& \leq \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}\hat{A}_t y ] - \mathbb{E}[ \sum_{t=1}^T x_t^{\top}A_t y_t] + \frac{2 \sqrt{T}\min(d_1,d_2)}{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_2-1) T\\
& \qquad \text{by Lemmas \ref{close_sp_vals} and \ref{lemma_with_alphas}}\\
& \leq \mathbb{E}[ \min_{x\in \Delta_{X,\delta}}\max_{y\in \Delta_{Y,\delta}} \sum_{t=1}^T x^{\top}\hat{A}_t y ] - \mathbb{E}[ \sum_{t=1}^T x_t^{\top}\hat{A}_t y_t] + \frac{2 \sqrt{T}\min(d_1,d_2)}{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_2-1) T\\
& \qquad \text{by Lemma \ref{A_hat_no_hat}}\\
& \leq 8 \eta [G_{\hat{{\mathcal{L}}}}^{\Vert \cdot \Vert_1} + \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T}\min(d_1,d_2) }{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_2-1) T.\end{aligned}$$
The last inequality follows by the same reasoning we used in the proof of Theorem \[thm:omg\_rftl\_regret\]. Indeed $\{(x_t,y_t)\}_{t=1}^T$ are the iterates of run on the sequence of payoff functions $\{x^\top A_t y\}_{t=1}^T$ so the same proof holds. By Lemma \[bilinear\_lipschitz\], since the absolute value of all the entries in $A_t$ is bounded above by 1, it holds that $$\begin{aligned}
& 8 \eta [G_{\hat{{\mathcal{L}}}}^{\Vert \cdot \Vert_1} + \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T}\min(d_1,d_2) }{\delta^2} + 2\delta G_{\bar{{\mathcal{L}}}}^{\Vert \cdot \Vert_1}(d_2-1) T\\
& = 8 \eta [\frac{1}{\delta^2}+ \frac{\vert \ln(\delta)\vert}{\eta}]^2 (1+\ln(T)) +\frac{T}{\eta} (\ln(d_1)+ \ln(d_2)) + \frac{2 \sqrt{T} \min(d_1,d_2) }{\delta^2} + 2\delta (d_2-1) T \\
& = O((d_1 + d_2) \ln(T) T^{5/6}).\end{aligned}$$ The last equality holds since we use $\delta = \frac{1}{T^{1/6}}$, $\eta = T^{1/6}$. This completes the proof.
Relationship between SP-regret and Individual-regret {#choose_one}
====================================================
We have defined two regret metrics for the OSP problem, namely the *SP-regret* and the *individual-regret* . In the previous subsection, we proposed an algorithm () with sublinear SP-regret. We have also mentioned that any OCO algorithm (e.g., online gradient descent, online mirror descent, Follow-the-Leader) can achieve sublinear individual-regret. A natural question is whether there exists a *single* algorithm that has both sublinear SP-regret and individual-regret. Surprisingly, the answer is negative.
\[thm:impossible\] Consider any algorithm that selects a sequence of $x_t, y_t$ pairs given the past payoff matrices $A_1, \ldots, A_{t-1}$. Consider the following three objectives: $$\begin{aligned}
\label{eq:xyregret} \left\vert \sum_{t=1}^Tx_t^\top A_t y_t - {\min_{x\in \Delta_X}\max_{y\in \Delta_Y}}\sum_{t=1}^Tx^\top A_t y \right \vert & = & o(T), \\
\label{eq:xregret} \sum_{t=1}^Tx_t^\top A_t y_t - \min_{x\in \Delta_X} \sum_{t=1}^T x^\top A_t y_t & = & o(T), \\
\label{eq:yregret} \max_{y\in \Delta_Y} \sum_{t=1}^T x_t^\top A_t y - \sum_{t=1}^Tx_t^\top A_t y_t & = & o(T).\end{aligned}$$ Then there exists an (adversarially-chosen) sequence $A_1, A_2, \ldots$ such that not all of , , and , are true.
A formal proof of the result is shown shortly, but here we give a sketch. The main idea is to construct two parallel scenarios, each with their own sequences of payoff matrices. The two scenarios will be identical for the first $T/2$ periods but are different for the rest of the horizon. In our particular construction, in both scenarios the players play the well known “matching-pennies” game for the first $T/2$ periods, then in first scenario they play a game with equal payoffs for all of their actions and in the second scenario they play a game where Player 1 is indifferent between its actions. One can show that if all three quantities in the statement of the theorem are $o(T)$ in the first scenario, then we prove that at least one of them is $\Omega(T)$ in the second one which yields the result. This suggests that the machinery for OCO, which minimizes individual regret, cannot be directly applied to the OMG problem.
We note that despite the negative result in Theorem \[thm:impossible\], it is possible to achieve both sublinear SP-Regret and individual-regret with further assumptions on the payoff functions $\{{\mathcal{L}}_t\}_{t=1}^T$.
One such example is where ${\mathcal{L}}_t(x,y)$ is sampled i.i.d. ; this case is discussed in §\[OCOwK\_setup\]. However, in light of Theorem \[thm:impossible\], in the general case where $\{{\mathcal{L}}_t\}_{t=1}^T$ is an arbitrary sequence, the best one can hope for is achieve either SP-Regret or individual-regret, but not both. In §\[subsec:regret-simulation\], we include a numerical example to further illustrate the relationship between SP-Regret and individual-regret.
Proof of the Impossibility Result
---------------------------------
We now present a formal proof of the impossibility result.
We assume there exists an algorithm such that $$\begin{aligned}
\vert \sum_{t=1}^Tx_t^\top A_t y_t - \min_{x\in \Delta_X}\max_{y \in \Delta_Y} \sum_{t=1}^Tx^\top A_t y\vert &\leq o(T),\\
\sum_{t=1}^Tx_t^\top A_t y_t - \min_{x\in \Delta_X} \sum_{t=1}^T x^\top A_t y_t &\leq o(T), \\
\max_{y\in \Delta_Y} \sum_{t=1}^T x_t^\top A_T y - \sum_{t=1}^Tx_t^\top A_t y_t &\leq o(T),\end{aligned}$$ for all possible sequences of matrices $\{A_t\}_{t=1}^T$ with bounded entries between $[-1,1]$. We now construct two sequences of functions for which all the three guarantees hold and lead that to a contradiction. Let $T$ be divisible by $2$.
In scenario 1: $A_t =
\begin{bmatrix}
1 &-1 \\
-1 &1\\
\end{bmatrix}$ for $1\leq t \leq \frac{T}{2}$ and $A_t=
\begin{bmatrix}
0 &0 \\
0 &0\\
\end{bmatrix}$ for $\frac{T}{2}<t\leq T$.
In scenario 2: $A_t =
\begin{bmatrix}
1 &-1 \\
-1 &1\\
\end{bmatrix}$ for $1\leq t \leq \frac{T}{2}$ and $A_t=
\begin{bmatrix}
1 &-1 \\
1 &-1\\
\end{bmatrix}$ for $\frac{T}{2}<t\leq T$.
It is easy to see that for both scenarios it holds that ${\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T x^\top A_t y = 0$. Since $d_1=d_2=2$ and we can parametrize any $x\in \Delta_X$ as $x = [\alpha; 1-\alpha]$ and any $y\in \Delta_Y$ as $y=[\beta;1-\beta]$ for some $0\leq \alpha, \beta \leq 1$. By assumption, we have $$\max_{y\in \Delta_Y} \sum_{t=1}^T x_t^\top A_t y - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T x^\top A_t y \leq o(T)$$ for all sequences of matrices $\{A_t\}_{t=1}^T$. This implies for scenario 1 that $$\max_{0\leq \beta \leq 1} \sum_{t=1}^{\frac{T}{2}}4\alpha_t\beta - 2\beta + 1 -2\alpha_t \leq o(T),$$ which also implies that $\sum_{t=1}^{\frac{T}{2}}2\alpha_t -1 \leq o(T)$ and $\sum_{t=1}^{\frac{T}{2}} 1-2\alpha_t \leq o(T)$ since $\sum_{t=1}^{\frac{T}{2}}4\alpha_t\beta - 2\beta + 1 -2\alpha_t$ is a linear function of $\beta$ and thus its maximum occurs at $\beta=0$ or $\beta=1$.
For scenario 2 $\max_{y\in \Delta_Y} \sum_{t=1}^T x_t^\top A_t y - {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T x^\top A_t y \leq o(T)$ reduces to $$\max_{0\leq \beta \leq 1} \sum_{t=1}^{\frac{T}{2}} 4\alpha_t \beta - 2\beta + 1 - 2\alpha_t + \frac{T}{2} (2\beta-1)\leq o(T),$$ which implies $\sum_{t=1}^{\frac{T}{2}} 2\alpha_t -1 + \frac{T}{2} \leq o(T)$ and $\sum_{t=1}^{\frac{T}{2}} 1 -2\alpha_t + \frac{T}{2} \leq o(T)$. Finally, notice that $\sum_{t=1}^{\frac{T}{2}} 2\alpha_t -1 + \frac{T}{2} \leq o(T)$ implies $\frac{T}{2} \leq o(T) +\sum_{t=1}^{\frac{T}{2}} 1 - 2\alpha_t$. But from scenario 1, we have $\sum_{t=1}^{\frac{T}{2}} 1 - 2\alpha_t \leq o(T)$ since $\frac{T}{2}\leq o(T)$ is a contradiction we get the result.
Online Convex Optimization with Knapsacks {#OCOwK_setup}
=========================================
In this section, we consider the online convex optimization with knapsacks (OCOwK) problem. This problem is motivated by various applications in dynamic pricing, online ad auctions, and crowdsourcing (see [@badanidiyuru2018bandits] and discussion in §\[Intro\]). The OCOwK model generalizes the standard OCO framework by having an additional set of resource constraints. We will show that OCOwK is closely related to the OSP problem studied in §\[setup\_problem\].
In the OCOwk problem, the decision maker has a set of resources $i=1,\ldots,m$ with given budgets $b=[b_1;b_2;...;b_m]$. There are $T$ time periods. At each time period, the decision maker chooses $x_{t}\in X\subset\mathbb{R}^{n}$, where $X$ is a convex compact set. [After the decision is chosen, Nature reveals two functions: a concave reward function $r_{t}: X\to\mathbb{R}_+$, which is assumed to be $G$-Lipschitz (with respect to $\Vert \cdot \Vert_2$), and a vector-valued resource consumption function $c_{t}: X\to\mathbb{R}^{m}_+$, where each entry of $c_t$ is a convex, $G$-Lipschitz function (with respect to $\Vert \cdot \Vert_2$). ]{}
The objective is to maximize cumulative reward while satisfying the budget constraints. In particular, we assume that if a decision $x_t$ violates any of the budget constraints, no reward is collected at period $t$. Therefore, the decision maker’s cumulative reward is given by $$\label{eq:ocowk-reward}
R(x_{1},x_{2},\cdots,x_{T})=\sum_{t=1}^{T} \left( r_{t}(x_{t})\mathbb{I}\bigl[\sum_{\tau =1}^{t}c_{\tau}(x_{\tau})\leq b \bigr] \right),$$ where $\mathbb{I}[\cdot]$ denotes the indicator function. In , if $b = +\infty$, the problem reduces to the standard OCO setting. In the special case where $r_t$ and $c_t$ are linear functions, our problem is related to the Bandits with Knapsack (BwK) model studied in [@badanidiyuru2018bandits]. A similar problem with general concave reward and convex constraints is studied by [@agrawal2014bandits]. However, unlike our model, both of these papers assume bandit feedback.
In order to guarantee that the budget constraint can always be satisfied, we assume there exists a “null action” that doesn’t consume any resource or generate any reward.
\[assumption:null\] There exists an action $x_0 \in X$ such that $r_t(x_0)\equiv 0$ and $c_t(x_0) \equiv 0$ for all $t=1,\ldots,T$.
The “null action” assumption is often satisfied in real-world applications of OCOwK. For example, in dynamic pricing, the “null action” is equivalent to charging an extremely high price so there is no customer demand; in online auctions (see [@balseiro2017learning]), the “null action” corresponds to bidding at \$0.
If the reward and consumption functions are chosen arbitrarily, it can be shown that no algorithm can achieve sublinear regret for OCOwK. Intuitively, if the reward and consumption functions shift at $\lfloor T/2 \rfloor$, no algorithm can recover the mistake made before $T/2$ in the remaining periods (which is similar to the case in §\[choose\_one\]). Therefore, we consider the setting where the reward and consumption functions are stochastic.
\[assumption:iid\] For $t=1,\ldots,T$, the reward function $r_t$ and consumption function $c_t$ are sampled i.i.d. from a fixed and unknown joint distribution.
Notice that even when the reward and consumption distribution is known, the optimal policy for the OCOwK problem is not a static decision, as the optimal decision depends on the remaining time and remaining budget. Therefore, defining the offline benchmark for OCOwK is not as straightforward as in the stochastic OCO setting. However, it has been shown in the literature that the following offline convex problem provides an *upper bound* of the expected reward of the optimal offline policy under Assumption \[assumption:iid\] (see e.g. [@badanidiyuru2018bandits; @besbes2012blind]): $$\begin{aligned}
r^{*}\triangleq \max_{x\in X} &\left\{ \sum_{t=1}^{T}\Ex[r_t(x)],\; \text{subject to} \sum_{t=1}^{T}\Ex[c_{t}(x)]\leq b \right\}.
\label{eq:hindsight}\end{aligned}$$ Therefore, we define the expected regret for the OCOwK problem as $$\begin{aligned}
{\mathsf{Regret}}(T) \triangleq r^* - \Ex [ R(x_{1},x_{2},\cdots,x_{T})],\end{aligned}$$ where the expectation is taken with respect to the random realizations of functions $r_t$ and $c_t$.
Reduction to a Saddle Point Problem
-----------------------------------
We relate the OCOwK problem to the OSP problem studied in §\[setup\_problem\] by defining the function $$\label{eq:def_L_t}
L_{t}(x,y)\triangleq -r_{t}(x)-y^{\top}(b/T-c_{t}(x)),$$ with $y \in \mathbb{R}^m_+$. Note that $L_t(x,y)$ is convex in $x$ and concave in $y$, so we can treat $L_{t}(x,y)$ as a payoff function in the OSP problem. Here, $y$ can be viewed as the dual prices associated with the budget constraints in , and the function $L_t(x,y)$ penalizes the payoff if consumption at iteration $t$ exceeds the average budget per period.
We let constant $y_{max,i}$ be the maximum reward that can be gained by adding one unit of resource $i$ ($\forall i \in [m]$), and define set $Y = \prod_{i=1}^m [0,y_{max,i}]$. Namely, $y_{max,i}$ is an upper bound on the dual variables for problem . [We also define vector $y_{max} = [y_{max,1}; \cdots; y_{max,m} ]$.]{} For any sequence of decisions $x_1,\cdots, x_T$, we claim that the decision maker’s total reward is bounded by $$\label{eq:reward}
R(x_{1},x_{2},\cdots,x_{T})\geq \sum_{t=1}^{T}r_{t}(x_{t}) + \min_{y \in Y}\left\{ y^{\top}\sum_{t=1}^{T}(b/T-c_{t}(x_{t}))\right\} =- \max_{y \in Y}\sum_{t=1}^{T}L_{t}(x_{t},y).$$ To see this, consider a modified OCOwK problem where resource consumption is allowed to go over the budget, but the decision maker must pay $y_{max,i}$ for each additional unit of resource $i$ used over $b_i$. By the definition of $y_{max,i}$, the decision maker’s profit under the modified problem is given by the right-hand side of , which can be no more than the reward in the original problem.
We now consider the benchmark . By Assumption \[assumption:null\], the Slater condition holds for the convex optimization problem ; so by using strong duality, we have $$r^{*}= - \min_{x\in X}\max_{y \in Y}\left\{ -\sum_{t=1}^{T}\Ex[r_{t}(x)]-y^{\top}\sum_{t=1}^{T}(b/T-\Ex[c_{t}(x)])\right\} = - \min_{x\in X}\max_{y \in Y} \Ex[\sum_{t=1}^{T}L_{t}(x,y)].$$ Therefore, the expected regret for OCOwK is bounded by $$\begin{aligned}
& {\mathsf{Regret}}(T) = r^* - \Ex [ R(x_{1},x_{2},\cdots,x_{T})] \nonumber \\
\leq & \Ex\Bigl[\max_{y \in Y}\sum_{t=1}^{T}L_{t}(x_{t},y)\Bigr] - \min_{x\in X}\max_{y \in Y} \Ex\Bigl[\sum_{t=1}^{T}L_{t}(x,y)\Bigr] \nonumber\\
= & \underbrace{\Ex\Bigl[\max_{y\in Y}\sum_{t=1}^{T}L_{t}(x_{t},y)-\sum_{t=1}^{T}L_{t}(x_{t},y_{t})\Bigr]}_{(\dagger)}+\underbrace{\Ex\Bigl[\sum_{t=1}^{T}L_{t}(x_{t},y_{t})\Bigr]-\min_{x\in X}\max_{y\in Y}\Ex\Bigl[\sum_{t=1}^{T}L_{t}(x,y)\Bigr]}_{(\ddagger)}. \label{eq:regret-bound}\end{aligned}$$ We have bounded the regret of the OCOwK problem by two quantities in a related OSP problem. In particular, the term $(\dagger)$ is equal to the expectation of player 2’s individual-regret (see Eq ), and the second term is related to the SP-Regret.
Algorithms for OCOwK {#subsec:algorithms-for-ocowk}
--------------------
Motivated by its connection to the OSP problem, we propose two algorithms for the OCOwK problem. For clarity we defer all proofs to the next two subsections.
First, we consider defined in Algorithm \[alg: SP-FTL\]. In view of Eq , we can bound the regret for OCOwK by the sum of an individual-regret and the SP-Regret. Theorem \[theorem:sp\_regret\_convex\_concave\] has already provided a SP-Regret bound for , so we just need to prove a sublinear individual-regret bound. In general, this is impossible due to the negative result in Theorem \[thm:impossible\]. However, since we made the additional assumption that $r_t$ and $c_t$ are sampled i.i.d., we are able to get a sublinear individual-regret for in the OCOwK problem.
We start by establishing a high probability bound on the individual-regret for the general OSP problem when the payoff function ${\mathcal{L}}_t(x,y)$ is strongly convex-concave.
\[no\_sp\_indiv\_reg\_str\] Let $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$ be an i.i.d. sequence of functions that is $H$-strongly convex-concave with respect norm $\Vert \cdot \Vert_2$, and $G$-Lipschitz continuous with respect norm $\Vert \cdot \Vert_2$. Here ${\mathcal{L}}_t:X\times Y \rightarrow \mathbb{R}$. Let $d$ be the dimension of $X \times Y$, and $D_{XY}>0$ be some constant such that $\max_{z_1,z_2\in X\times Y}\Vert z_1 - z_2\Vert_2\leq D_{XY}$. run on the sequence of functions $\{{\mathcal{L}}_t\}_{t=1}^T$ guarantees that with probability at least $1-1/T$ $$\begin{aligned}
& \max_{y\in Y}\sum_{t=1}^T {\mathcal{L}}_t(x_t,y) -\sum_{t=1}^T{\mathcal{L}}_t(x_t,y_t) \leq \frac{8G^2}{H}(1+\ln(T)) + O\big( \frac{G^{3/2} D_{XY}^{1/2} (d \ln(T) \ln(dT) )^{1/4} T^{3/4}\ln^{1/4}(T)}{H^{1/2}}\big).\end{aligned}$$
The proof of Lemma \[no\_sp\_indiv\_reg\_str\] uses a concentration inequality for Lipschitz functions by Shalev-Shwartz et al. [@shalevstochastic]. The key step in the proof is to show that the solution of the sample average approximation at step $t$ i.e. $x_t$ is close to $x^*$, the saddle point of the expected game.
However, we cannot directly use Lemma \[no\_sp\_indiv\_reg\_str\] to bound the individual-regret term $(\dagger)$ in because the function $L_t(x,y)$ is linear in $y$ and thus not strongly convex-concave. We add a regularization term to $L_t(x,y)$ to make it $H$-strongly convex-concave. Notice our choice of the regularization term here is not the same as in Theorem \[theorem:sp\_regret\_convex\_concave\], which leads to a $\tilde{O}(T^{5/6})$ bound in the following theorem.
\[no\_sp\_indiv\_reg\] Let $X\subset \mathbb{R}^n$ be a convex compact set. Let $\{r_t(x)\}_{t=1}^T$ be an i.i.d. sequence of concave reward functions with $r_t:X\rightarrow \mathbb{R}_+$ which is $G$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. Let $\{c_t(x)\}_{t=1}^T$ be an i.i.d. sequence of vector-valued functions with $c_t:X \rightarrow \mathbb{R}^m$ where each entry of $c_t$ is convex in $x$ and $G$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. Let $L_t$ be defined as in equation (\[eq:def\_L\_t\]). Define $\bar{{\mathcal{L}}}_t(x,y) \triangleq L_t(x,y) + H\Vert x \Vert^2- H\Vert y \Vert^2$, where $H \triangleq T^{-1/6}$. Applying the algorithm on functions $\{\bar{{\mathcal{L}}}_t\}_{t=1}^T$ guarantees that with probability at least $1-1/T$ it holds that $$\begin{aligned}
\label{eq:high-probability-bound}
& \quad\max_{y\in Y} \sum_{t=1}^T L_t(x_t,y) - \sum_{t=1}^T L_t(x_t,y_t)\\
& \leq \frac{8(G+2H(D_X+D_Y))^2}{H}(1+\ln(T)) \\
&\quad + O\big( \frac{\beta(G,H,D_X,D_Y) D_{XY}^{1/2} (d \ln(T) \ln(dT) )^{1/4} T^{3/4}\ln^{1/4}(T)}{H^{1/2}}\big) + TH(D_X^2 + D_Y^2)\\
& = O\left(poly(G,D_X, D_Y) \ln^{1/2}(T) T^{5/6} \right),\end{aligned}$$ where $\beta(G,H,D_X,D_Y)$ is defined in the proof of the theorem. Additionally, for the OCOwK problem, with probability at least $1-1/T$ it holds that
$$\begin{aligned}
{\mathsf{Regret}}(T) = O\left(poly(G,D_X, D_Y) \ln^{1/2}(T) T^{5/6} \right).
\end{aligned}$$
Next, we present an algorithm for OCOwK that improves the regret bound in Theorem \[no\_sp\_indiv\_reg\]. The key idea of this algorithm is to update primal variable $x$ and dual variable $y$ of $L_t(x,y)$ in parallel. Each variable can be updated using any algorithm for Online Convex Optimization such as Online Gradient Descent [@zinkevich2003online] or Regularized Follow the Leader [@hazan2007logarithmic; @hazan2016introduction]. We call this algorithm Primal-Dual Regularized-Follow-the-Leader () (see Algorithm \[alg: PD-RFTL\]).
[ Convex compact decision sets $X$,$Y=\Pi_{i=1}^m[0,y_{\max,i}]$. Parameters $\eta_1,\eta_2$. $x_1 \leftarrow \arg \min_{x\in X} \frac{1}{2}\Vert x\Vert_2^2$ $y_1 \leftarrow \arg \max_{y\in Y} -\frac{1}{2}\Vert y\Vert_2^2 $ Play $(x_t,y_t)$ Observe $L_t$; define $f_{t}(x)\triangleq L_{t}(x,y_{t})$ and $g_{t}(y)\triangleq L_{t}(x_{t},y)$ $x_{t+1} \leftarrow \arg\min_{x\in X} \left\{\sum_{\tau=1}^{t} \nabla f_\tau (x_\tau)^\top x +\frac{1}{2\eta_1}\Vert x \Vert_2^2\right\}$ $y_{t+1} \leftarrow \arg\max_{y \in Y} \left\{\sum_{\tau=1}^{t} \nabla g_\tau (y_\tau)^\top y-\frac{1}{2\eta_2} \Vert y \Vert_2^2\right\}$ ]{}
We will bound the regret of using Eq . Before we proceed we state the well known individual-regret guarantee of .
\[thm:rftl\_bound\] Let $\{f_t(x)\}_{t=1}^T$ be any sequence of convex and $G_f$-Lipschitz functions where $f_t:X\rightarrow \mathbb{R}$ and $X\subset\mathbb{R}^d$ is a convex compact set such that $\max_{x_1,x_2}\Vert x_1-x_2\Vert_2\leq D_X$. The algorithm: $x_1 \leftarrow \arg\min_{x\in X}\frac{1}{2}\Vert x \Vert_2^2$, $x_t\leftarrow \arg \min_{x\in X} \sum_{\tau=1}^t \nabla f_t(x_t)^\top x + \frac{1}{2\eta}\Vert x \Vert_2^2$ guarantees that $$\begin{aligned}
\sum_{t=1}^T f_t(x_t) - \min_{x\in X} \sum_{t=1}^T f_t(x)\ \leq 2\eta G_f^2 T + \frac{D_X^2}{\eta}.\end{aligned}$$
Recall that for , it was more challenging to bound the first term $(\dagger)$ and relatively easy to bound the term $(\ddagger)$. For it is quite the opposite. By defining $g_{t}(y)\triangleq L_{t}(x_{t},y)$, the first term $(\dagger)$ can be written as $\Ex[\max_{y\in Y} \sum_{t=1}^T g_t(y) - \sum_{t=1}^T g_t(y_t)]$, so we immediately have $(\dagger) = O(\sqrt{T})$ using the regret bound for Regularized Follow-the-Leader in the OCO setting. To bound the second term $(\ddagger)$, we have the following result.
\[no\_ocowk\_regret\_pdftl\] Let $X\subset \mathbb{R}^n$ be a convex compact set. Let $\{r_t(x)\}_{t=1}^T$ be an i.i.d. sequence of concave reward functions with $r_t:X\rightarrow \mathbb{R}_+$ which is $G$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. Let $\{c_t(x)\}_{t=1}^T$ be an i.i.d. sequence of vector-valued functions with $c_t:X \rightarrow \mathbb{R}^m$ where each entry of $c_t$ is convex in $x$ and $G$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. run with $\eta_1 = \frac{D_X}{G(1+\Vert y_{max} \Vert_2)\sqrt{T}}, \eta_2 =\frac{\Vert y_{max}\Vert_2}{(\frac{1}{T}\Vert b\Vert_2 + \sqrt{m G D_X})\sqrt{T}} $ guarantees that $$\begin{aligned}
{\mathsf{Regret}}(T) \leq 5 G (1+\Vert y_{max}\Vert_1) D_X \sqrt{T} + 5 (\frac{1}{T} \Vert b \Vert_2 + \sqrt{m G D_X}) \Vert y_{max}\Vert_2 \sqrt{T}.\end{aligned}$$
Compared to other algorithms for OCOwK, including the UCB-based algorithm in [@badanidiyuru2018bandits; @agrawal2014bandits] and Thompson sampling-based algorithm in [@ferreira2017online], the proof for Theorem \[no\_ocowk\_regret\_pdftl\] is surprisingly simple, as we are able to exploit the connection between OCOwK and the OSP problem.
The $O(\sqrt{T})$ regret bound in Theorem \[no\_ocowk\_regret\_pdftl\] also gives the best possible rate in $T$, since OCO is a special case of OCOwK, and it is well-known that any algorithm must have $\Omega(\sqrt{T})$ regret for the general OCO problem. In Section \[subsec:numerical-ocowk\], we compare the performance of and in a numerical experiment.
\[rmk:bandit-OCOwK\] Our proof for Theorem \[no\_ocowk\_regret\_pdftl\] allows the RFTL subroutine in Algorithm \[alg: PD-RFTL\] to be replaced with other OCO algorithms with $O(\sqrt{T})$ regret. In addition, we can extend Algorithm \[alg: PD-RFTL\] to the bandit setting of OCOwK, where we only observe the values $r_t(x_t)$ and $c_t(x_t)$ after $x_t$ is chosen. By replacing the RFTL subroutine with any Bandit Convex Optimization (BCO) algorithm [@bubeck2017kernel], we can also establish sublinear regret bounds for OCOwK in the bandit setting.
Proof of SP-FTL for OCOwK
-------------------------
In this section we present the analysis of , applied to the OCOwK problem. The following result from Shalev-Shwartz et al. [@shalevstochastic] (Theorem 5) will be useful.
\[shai\_stoch\] Let $f(w,\xi):W \times \Xi \rightarrow \mathbb{R}$ be $G$-Lipschitz in $w$ with respect to norm $\Vert \cdot \Vert_2$, where $W\subset \mathbb{R}^d$ is bounded set such that there exists a constant $D_W>0$ such that $\max_{w_1,w_2 \in W}\Vert w_1 - w_2 \Vert_2 \leq D_W$. Then with probability at least $1-\delta$, for all $w\in W$, it holds that $$\begin{aligned}
\Bigl| \sum_{t=1}^T f(w,\xi_t) - T \mathbb{E}_{\xi}[f(x,\xi)]\Bigr| \leq O\big( G D_W \sqrt{d \ln(T) \ln(\frac{d}{\delta}) T}\big).\end{aligned}$$
First, we prove the following lemma.
\[close\_to\_true\_sp\] Let $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$ be a sequence of i.i.d. functions which are $H$-strongly convex concave with respect $\Vert \cdot \Vert_2$ and $G$-Lipschitz with respect to norm $\Vert \cdot \Vert_2$. Here ${\mathcal{L}}_t: X \times Y \rightarrow \mathbb{R}$, where $X\subset \mathbb{R}^{d_1}$, $Y \subset \mathbb{R}^{d_2}$ are convex compact sets. Let $\{(x_t,y_t)\}_{t=1}^T$ be the the iterates of SP-FTL when run on functions $\{{\mathcal{L}}_t(x,y)\}_{t=1}^T$. With probability at least $1-\delta$, for any $t=1,...T$ it holds that $$\begin{aligned}
\Vert x_{t+1}-x^*\Vert_2 \leq O\big( \frac{G^{1/2} D_{XY}^{1/2} (d \ln(t) \ln(\frac{d}{\delta}) )^{1/4}}{H^{1/2} t^{1/4}}\big),\end{aligned}$$ where $D_{XY}>0$ is a constant such that $\max_{z_1,z_2 \in X \times Y}\Vert z_1 - z_2\Vert_2 \leq D_{XY}$ and $d$ is the dimenssion of $X\times Y$.
Define the concentration error at time $t$ as $$\begin{aligned}
CE_t \triangleq O\big( GD_{XY}\sqrt{d \ln(t) \ln(\frac{d}{\delta})t}\big). \end{aligned}$$ Notice that ${\mathcal{L}}_\tau$ satisfies all the assumptions of Theorem \[shai\_stoch\], so with probability at least $1-\delta$, for all $x\in X, y\in Y$ we have $$\begin{aligned}
\label{eq_el_concentration}
\big| \sum_{\tau=1}^t {\mathcal{L}}_\tau (x,y) - t \bar{{\mathcal{L}}}(x,y) \big| \leq CE_t.\end{aligned}$$ We now derive some consequences of this fact. Recall that $\bar{{\mathcal{L}}}(x,y) \triangleq \mathbb{E}[{\mathcal{L}}_1(x,y)]$ and $(x^*,y^*)$ is the saddle point of $\bar{{\mathcal{L}}}$. With probability at least $1-\delta$. $$\begin{aligned}
t \bar{{\mathcal{L}}}(x^*,y^*) &\leq t \bar{{\mathcal{L}}}(x_{t+1},y^*) & \text{by definition of $x^*$} \\
&\leq \sum_{\tau=1}^t {\mathcal{L}}_\tau(x_{t+1},y^*) + CE_t & \text{by Equation \eqref{eq_el_concentration}}\\
&\leq \sum_{\tau=1}^t {\mathcal{L}}_\tau(x_{t+1},y_{t+1}) + CE_t. & \text{by definition of $y_{t+1}$}\\\end{aligned}$$ This implies that $$\begin{aligned}
\label{conseq_one}
t \bar{{\mathcal{L}}}(x^*,y^*) - \sum_{\tau=1}^t {\mathcal{L}}_\tau(x_{t+1},y_{t+1}) \leq CE_t.\end{aligned}$$ We now show that $$\begin{aligned}
\label{conseq_two}
\sum_{\tau=1}^t {\mathcal{L}}_\tau(x^*,y_{t+1}) - t\bar{{\mathcal{L}}}(x^*,y^*) \leq CE_t.\end{aligned}$$ Indeed, it holds that $$\begin{aligned}
\sum_{\tau=1}^t {\mathcal{L}}_\tau(x^*,y_{t+1}) &\leq t \bar{{\mathcal{L}}}(x^*,y_{t+1}) + CE_t & \text{by Equation \eqref{eq_el_concentration}}\\
&\leq t \bar{{\mathcal{L}}}(x^*,y^*) + CE_t & \text{by definition of $y^*$}.\end{aligned}$$ Now, using the fact that $x_{t+1}$ is the saddle point of $\sum_{\tau=1}^t {\mathcal{L}}_\tau(x,y)$, which is $(Ht)$-strongly convex, we have $$\begin{aligned}
\frac{Ht}{2} \Vert x_{t+1} - x^*\Vert_2^2 &\leq \sum_{\tau=1}^t {\mathcal{L}}_\tau(x^*,y_{t+1}) - \sum_{\tau=1}^t {\mathcal{L}}_\tau(x_{t+1},y_{t+1}) \\
&= \sum_{\tau=1}^t {\mathcal{L}}_\tau(x^*,y_{t+1})-t\bar{{\mathcal{L}}}(x^*,y^*) + t\bar{{\mathcal{L}}}(x^*,y^*) - \sum_{\tau=1}^t {\mathcal{L}}_\tau(x_{t+1},y_{t+1})\\
&\leq 2 CE_t \quad \text{by Equations \eqref{conseq_one} and \eqref{conseq_two}}.\end{aligned}$$ It follows that $$\begin{aligned}
\Vert x_{t+1}-x^*\Vert_2 \leq O\big( \frac{G^{1/2} D_{XY}^{1/2} (d \ln(t) \ln(\frac{d}{\delta}) )^{1/4}}{H^{1/2} t^{1/4}}\big).\end{aligned}$$
We now prove Lemma \[no\_sp\_indiv\_reg\_str\] in Section \[subsec:algorithms-for-ocowk\].
For all $y\in Y$, it holds that $$\begin{aligned}
\sum_{t=1}^T {\mathcal{L}}_t(x_t,y) &\leq \sum_{t=1}^{T-1} {\mathcal{L}}_t(x_t,y) + {\mathcal{L}}_T(x^*,y) + G \Vert x_T-x^*\Vert_2 \qquad \text{since ${\mathcal{L}}_T$ is $G$-Lipschitz}\\
&\leq \sum_{t=1}^T{\mathcal{L}}_t(x^*,y) + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2 \quad \qquad \text{since each ${\mathcal{L}}_t$ is $G$-Lipschitz}\\
&\leq \sum_{t=1}^T{\mathcal{L}}_t(x_{T+1},y)+ GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2.\end{aligned}$$
It follows that $$\begin{aligned}
\max_{y\in Y}\sum_{t=1}^T {\mathcal{L}}_t(x_t,y) &\leq \max_{y\in Y} \sum_{t=1}^T{\mathcal{L}}_t(x_{T+1},y)+ GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2\\
&=\sum_{t=1}^T{\mathcal{L}}_t(x_{T+1},y_{T+1})+ GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2 \\
& \qquad \text{(by definition of $y_{T+1}$)}\\
& = {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) + GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2.\end{aligned}$$
Subtracting in both sides $\sum_{t=1}^T{\mathcal{L}}_t(x_t,y_t)$, we get $$\begin{aligned}
&\quad \max_{y\in Y}\sum_{t=1}^T {\mathcal{L}}_t(x_t,y) -\sum_{t=1}^T{\mathcal{L}}_t(x_t,y_t)\\
&\leq {\min_{x\in X}\max_{y\in Y}}\sum_{t=1}^T {\mathcal{L}}_t(x,y) -\sum_{t=1}^T{\mathcal{L}}_t(x_t,y_t) + GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2\\
&\leq \frac{8G^2}{H}(1+\ln(T)) + GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2 \quad \text{By Theorem~\ref{thm:sp_regret_str}}.\end{aligned}$$
By Lemma \[close\_to\_true\_sp\], and a simple union bound, we have that with probability at least $1-\delta T$ $$\begin{aligned}
& \quad \frac{8G^2}{H}(1+\ln(T)) + GT\Vert x^*-x_{T+1}\Vert_2 + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2\\
&\leq \frac{8G^2}{H}(1+\ln(T)) + \frac{G^{3/2}D_{XY}^{1/2}(d\ln(T)\ln(\frac{d}{\delta}))^{1/4}T^{3/4}}{H^{1/2}} + G\sum_{t=1}^T \Vert x_t-x^*\Vert_2 \\
&\leq \frac{8G^2}{H}(1+\ln(T)) + \frac{G^{3/2}D_{XY}^{1/2}(d\ln(T)\ln(\frac{d}{\delta}))^{1/4}T^{3/4}}{H^{1/2}} + \frac{G^{3/2} D_{XY}^{1/2}(d\ln(\frac{d}{\delta}))^{1/4}}{H^{1/2}} \sum_{t=1}^T \frac{\ln^{1/4}(t)}{t^{1/4}}.\end{aligned}$$
Therefore, with probability at least $1-\delta T$ it holds that $$\begin{aligned}
&\quad \max_{y\in Y}\sum_{t=1}^T {\mathcal{L}}_t(x_t,y) -\sum_{t=1}^T{\mathcal{L}}_t(x_t,y_t)\\
&\leq \frac{8G^2}{H}(1+\ln(T)) + \frac{G^{3/2}D_{XY}^{1/2}(d\ln(T)\ln(\frac{d}{\delta}))^{1/4}T^{3/4}}{H^{1/2}} + \frac{G^{3/2} D_{XY}^{1/2}(d\ln(\frac{d}{\delta}))^{1/4}}{H^{1/2}} \int_{1}^T \frac{\ln^{1/4}(t)}{t^{1/4}} dt\\
&= \frac{8G^2}{H}(1+\ln(T)) + O\big( \frac{G^{3/2} D_{XY}^{1/2} (d \ln(T) \ln(\frac{d}{\delta}) )^{1/4} T^{3/4}\ln^{1/4}(T)}{H^{1/2}}\big).\end{aligned}$$ Setting $\delta=1/T^2$ yields the result.
We are ready to prove Theorem \[no\_sp\_indiv\_reg\].
Recall that $\bar{{\mathcal{L}}}_t(x,y) \triangleq L_t(x,y) + H \Vert x \Vert_2^2 - H \Vert y \Vert_2^2$. If $D_X$, $D_Y$ are constants such that $\max_{x\in X} \Vert x\Vert_2 \leq D_X$ and $\max_{y\in Y} \Vert y\Vert_2 \leq D_Y$, it holds that for all $x\in X$, $y\in Y$ $$\label{dif_el_bar_el}
- HD_{Y}^2 \leq \bar{{\mathcal{L}}}_t(x,y) - L_t(x,y) \leq HD_X^2, \quad \forall t=1,\ldots,T.$$
By adding up equation , we have that for any $y\in Y$ $$\begin{aligned}
\sum_{t=1}^T L_t(x_t,y) \leq \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y) + HD_Y^2 T.\end{aligned}$$ This implies that $$\begin{aligned}
\max_{y\in Y}\sum_{t=1}^T L_t(x_t,y) \leq \max_{y \in Y}\sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y) + HD_Y^2 T.\end{aligned}$$
Therefore, we have $$\begin{aligned}
& \quad\max_{y\in Y} \sum_{t=1}^T L_t(x_t,y) - \sum_{t=1}^T L_t(x_t,y_t)\\
&\leq \max_{y\in Y} \sum_{t=1}^T L_t(x_t,y) - \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y_t) + T H D_X^2 \qquad \text{by Equation \eqref{dif_el_bar_el}}\\
& \leq \max_{y\in Y} \sum_{t=1}^T \bar{{\mathcal{L}}}_t(x_t,y) - \sum_{t=1}^T\bar{{\mathcal{L}}}_t(x_t,y_t) + T H D_X^2 + THD_Y^2.\end{aligned}$$
Thus, by Lemma \[no\_sp\_indiv\_reg\_str\] it holds that with probability at least $1-\frac{1}{T}$
$$\begin{aligned}
& \quad\max_{y\in Y} \sum_{t=1}^T L_t(x_t,y) - \sum_{t=1}^T L_t(x_t,y_t)\\
&\leq \frac{8(G+2H(D_X+D_Y))^2}{H}(1+\ln(T)) \\
&\quad \quad + O\big( \frac{(G+2H(D_X+D_Y))^{3/2} D_{XY}^{1/2} (d \ln(T) \ln(dT) )^{1/4} T^{3/4}\ln^{1/4}(T)}{H^{1/2}}\big) + TH(D_X^2 + D_Y^2).\end{aligned}$$
Since $$\begin{aligned}
&(G+2H(D_X+D_Y))^{3/2} \\
=\ & (G^3 + 6G^2H(D_X+D_Y) + 12GH^2(D_X+D_Y)^2 + 8H^3(D_X+D_Y)^3 )^{1/2}\\
\leq\ & G^{3/2} + (6G^2H(D_X+D_Y))^{1/2} + (12GH^2(D_X+D_Y)^2)^{1/2} + (8H^3(D_X+D_Y)^3)^{1/2} \\
\triangleq\ & \beta(G,H,D_X,D_Y),\end{aligned}$$ we have that with probability at least $1-\frac{1}{T}$, $$\begin{aligned}
& \quad\max_{y\in Y} \sum_{t=1}^T L_t(x_t,y) - \sum_{t=1}^T L_t(x_t,y_t)\\
&\leq \frac{8(G+2H(D_X+D_Y))^2}{H}(1+\ln(T))\\
&\qquad + O\big( \frac{\beta(G,H,D_X,D_Y) D_{XY}^{1/2} (d \ln(T) \ln(dT) )^{1/4} T^{3/4}\ln^{1/4}(T)}{H^{1/2}}\big) + TH(D_X^2 + D_Y^2)\\
& = O\left(poly(G,D_X, D_Y) \ln^{1/2}(T) T^{5/6} \right),\end{aligned}$$
where we plugged in $H=T^{-1/6}$. This concludes the first part of the proof.
We now prove the second claim. Recall from Eq , that $$\begin{aligned}
& \quad {\mathsf{Regret}}(T) \\
& \leq \Ex\Bigl[\max_{y\in Y}\sum_{t=1}^{T}L_{t}(x_{t},y)-\sum_{t=1}^{T}L_{t}(x_{t},y_{t})\Bigr] + \Ex\Bigl[\sum_{t=1}^{T}L_{t}(x_{t},y_{t})\Bigr]-\min_{x\in X}\max_{y\in Y}\Ex\Bigl[\sum_{t=1}^{T}L_{t}(x,y)\Bigr] \\
& = \Ex\Bigl[\max_{y\in Y}\sum_{t=1}^{T}L_{t}(x_{t},y)-\sum_{t=1}^{T}L_{t}(x_{t},y_{t})\Bigr] + \Ex\Bigl[\sum_{t=1}^{T}L_{t}(x_{t},y_{t}) - \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T L_t(x,y)\Bigr]\\
&\quad + \Ex\Bigl[ \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T L_t(x,y)\Bigr] -\min_{x\in X}\max_{y\in Y}\Ex\Bigl[\sum_{t=1}^{T}L_{t}(x,y)\Bigr].\end{aligned}$$
Let $$\begin{aligned}
A\triangleq & \max_{y\in Y}\sum_{t=1}^{T}L_{t}(x_{t},y)-\sum_{t=1}^{T}L_{t}(x_{t},y_{t}), \\
B \triangleq & \sum_{t=1}^{T}L_{t}(x_{t},y_{t}) - \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T L_t(x,y), \\
C \triangleq & \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T L_t(x,y) -\min_{x\in X}\max_{y\in Y}\Ex\Bigl[\sum_{t=1}^{T}L_{t}(x,y)\Bigr].\end{aligned}$$
Notice that we already have an upper bound for $A$ that holds with high probability, the term $B$ can be upper bounded using Theorem \[theorem:sp\_regret\_convex\_concave\]. Let us upper bound $C$ with high probability. As in the proof of Lemma \[close\_to\_true\_sp\] we know that with probability at least $1-\delta$ it holds that for all $x\in X, y \in Y$ $\left\vert \sum_{t=1}^T L_t(x,y) - \sum_{t=1}^T \mathbb{E}[L(x,y)] \right \vert \leq CE_T $. Therefore, with probability at least $1-\delta$, we have $$\begin{aligned}
\sum_{t=1}^T L_t(x,y) \leq \sum_{t=1}^T \mathbb{E}[L(x,y)] + CE_T \quad \forall x\in X, y\in Y.\end{aligned}$$ This implies that with probability at least $1-\delta$ $$\begin{aligned}
\min_{x\in X}\sum_{t=1}^T L_t(x,y) \leq \sum_{t=1}^T \mathbb{E}[L(x,y)] + CE_T \quad \forall x\in X, y\in Y,\end{aligned}$$ which implies that with probability at least $1-\delta$ $$\begin{aligned}
\max_{y\in Y} \min_{x\in X}\sum_{t=1}^T L_t(x,y) \leq \max_{y\in Y} \sum_{t=1}^T \mathbb{E}[L(x,y)] + CE_T \quad \forall x\in X.\end{aligned}$$ Therefore it holds that with probability at least $1-\delta$ $$\begin{aligned}
\min_{x\in X}\max_{y\in Y} \sum_{t=1}^T L_t(x,y) \leq \min_{x\in X} \max_{y\in Y} \sum_{t=1}^T \mathbb{E}[L(x,y)] + CE_T,\end{aligned}$$ thus we have a high probability bound for $C$. Our high probability bound for $A$ scales as $\asymp\ln^{1/2}(T)T^{5/6}$, the deterministic bound for $B$ scales as $\asymp\ln^{1/2}(T)T^{5/6}$, and the high probability bound for $C$ scales as $\asymp\sqrt{\ln(T)T}$. The high probability bounds imply bounds in expectation (please see Lemmas 8 and 9 in [@pmlr-v89-cardoso19a] that show how to convert high probability bounds into bounds that hold in expectation). It follows that [ $$\begin{aligned}
{\mathsf{Regret}}(T) &\leq \mathbb{E}[A]+ \mathbb{E}[B] + \mathbb{E}[C]\\
& = O\left(poly(G,D_X, D_Y) (\ln^{1/2}(T)T^{5/6} + \ln^{1/2}(T)T^{5/6} + \sqrt{\ln(T)T}) \right)\\
& = O\left(poly(G,D_X, D_Y)\ln^{1/2}(T)T^{5/6}\right).\end{aligned}$$ ]{} This concludes the proof.
Proof of PD-RFTL for OCOwK
--------------------------
In this section we present the analysis of , applied to the OCOwK problem. Notice that is using two instances of . One is run on convex functions $\{f_t(x)\}_{t=1}^T$ and another one on concave functions $\{g_t(y)\}_{t=1}^T$ (thus the maximization and the negative sign in the regularizer). Let $R_1 \triangleq 2\eta_1 G_f^2 T + \frac{D_X^2}{\eta_2}$ be the individual-regret guarantee of RFTL on functions $\{f_t(x)\}_{t=1}^T$ from Lemma \[thm:rftl\_bound\]. Let $R_2 \triangleq 2\eta_2 G_g^2 T + \frac{D_Y^2}{\eta_2}$ be the individual-regret guarantee of RFTL on functions $\{g_t(y)\}_{t=1}^T$ from Lemma \[thm:rftl\_bound\]. We have $$\label{eq:oco-ftl}
\sum_{t=1}^{T}L_{t}(x_{t},y_{t})=\sum_{t=1}^{T}f_{t}(x_{t})\leq\min_{x}\sum_{t=1}^{T}f_{t}(x)+R_1=\min_{x}\sum_{t=1}^{T}L_{t}(x,y_{t})+R_1,$$ and $$\label{eq:oco-ftl-2}
\sum_{t=1}^{T}L_{t}(x_{t},y_{t})=\sum_{t=1}^{T}g_{t}(y_{t})\geq\max_{y}\sum_{t=1}^{T}g_{t}(y)-R_2=\max_{y}\sum_{t=1}^{T}L_{t}(x_{t},y)-R_2.$$
Let $\bar{L}(x,y) = \Ex[L_t(x,y)]$ for any $x\in X,y\in Y$. Let $(x^*, y^*)$ be the saddle point of $\bar{L}$, satisfying $$\label{eq:def-y*}
\bar{L}(x^*, y^*) = \max_{y \in Y} \bar{L}(x^*,y)= \min_{x\in X} \max_{y \in Y} \bar{L}(x,y).$$
[ Notice that $$\label{eq:interchange_expect}
\mathbb{E}\left[\sum_{t=1}^{T}L_{t}(x^*,y_{t})\right] = \mathbb{E}\left[\sum_{t=1}^{T}( \overline{L}(x^*,y_{t}) ) \right],$$ where the expectation is taken with respect to the random draw of functions $\{L_t\}_{t=1}^T$. Indeed, we have $$\begin{aligned}
& \quad \mathbb{E}\left[\sum_{t=1}^{T}L_{t}(x^*,y_{t})\right] \\
&= \mathbb{E}\left[\sum_{t=1}^{T-1}L_{t}(x^*,y_{t})\right] + \mathbb{E}\left[L_{T}(x^*,y_{T})\right]\\
&= \mathbb{E}\left[\sum_{t=1}^{T-1}L_{t}(x^*,y_{t})\right] + \mathbb{E}_{\{L_t\}_{t=1}^{T-1}}\left[ \mathbb{E}_{L_T} \left[L_{T}(x^*,y_{T})\vert\{L_t\}_{t=1}^{T-1}\right] \right]\\
&= \mathbb{E}\left[\sum_{t=1}^{T-1}L_{t}(x^*,y_{t})\right] + \mathbb{E}_{\{L_t\}_{t=1}^{T-1}}\left[ \bar{L}_{T}(x^*,y_{T}) \right],\end{aligned}$$ where the last equality holds since for any $x\in X, y \in Y$ it holds that $\bar{L}(x,y) = \Ex[L_t(x,y)]$ and $y_T$ is deterministic given $\{L_t\}_{t=1}^{T-1}$. We have $\mathbb{E}\left[\sum_{t=1}^{T}L_{t}(x^*,y_{t})\right] = \mathbb{E}\left[\sum_{t=1}^{T-1}L_{t}(x^*,y_{t})\right] + \mathbb{E}\left[ \bar{L}_{T}(x^*,y_{T}) \right]$, repeating the argument $T-1$ more times shows that $\mathbb{E}\left[\sum_{t=1}^{T}L_{t}(x^*,y_{t})\right] = \mathbb{E}\left[\sum_{t=1}^{T}( \overline{L}(x^*,y_{t}) ) \right]$. ]{}
We are ready to prove the statement of the theorem. $$\begin{aligned}
& \quad \Ex\left[\max_{y}\sum_{t=1}^{T}L_{t}(x_{t},y)\right]-\min_{x\in X}\max_{y\geq0}\Ex\left[\sum_{t=1}^{T}L_{t}(x,y)\right]\\
& \leq\Ex\left[\min_{x}\sum_{t=1}^{T}L_{t}(x,y_{t})\right]-\min_{x\in X}\max_{y\geq0}\Ex\left[\sum_{t=1}^{T}L_{t}(x,y)\right]+R_1 +R_2 \qquad \text{by Equations \eqref{eq:oco-ftl}, \eqref{eq:oco-ftl-2}}\\
& =\Ex\left[\min_{x}\sum_{t=1}^{T}L_{t}(x,y_{t})\right] -\sum_{t=1}^T \overline{L}(x^{*},y^{*}) +R_1 +R_2 \qquad \text{by Equation \eqref{eq:def-y*}}\\
& \leq\Ex\left[\sum_{t=1}^{T}L_{t}(x^*,y_{t})\right] -\sum_{t=1}^T \overline{L}(x^{*},y^{*}) +R_1 +R_2 \qquad \text{because } \min_{x}\sum_{t=1}^{T}L_{t}(x,y_{t}) \leq \sum_{t=1}^{T} L_{t}(x^*,y_{t}) \\
& = \Ex\left[\sum_{t=1}^{T}( \overline{L}(x^*,y_{t}) - \overline{L}(x^{*},y^{*}) ) \right] +R_1 +R_2 \qquad \text{by Equation \eqref{eq:interchange_expect}}\\
& \leq0+R_1 +R_2. \qquad \text{by Equation~\eqref{eq:def-y*}}\end{aligned}$$
[ By Equation , the above inequality implies that ${\mathsf{Regret}}(T) \leq 2\eta_1 G_f^2 T + \frac{D_X^2}{\eta_1} + 2\eta_2 G_g^2 T + \frac{D_Y^2}{\eta_2}$. Let us now bound $G_f, G_g$ and $D_Y$ from above. We start with $G_f$. To bound $G_f$, it suffices to bound $\Vert \nabla_x f_t(x) \Vert_2$. Let $y_{t,i}$ be the $i$-th entry of vector $y_t$ and $c_t(x)_i$ be the $i$-th entry of vector valued function $c_t$, which by assumption is $G$-Lipschitz continuous. We have $$\begin{aligned}
\Vert \nabla_x f_t(x) \Vert_2 &= \Vert -\nabla_x r_t(x) + \sum_{i=1}^m y_{t,i}\nabla_x c_t(x)_i\Vert_2\\
& \leq G + \sum_{i=1}^m y_{t,i} \Vert \nabla_x c_t(x)_i \Vert_2\\
& \leq G + G \Vert y_{max} \Vert_1.\end{aligned}$$ The previous line implies that $G_f\leq G + G \Vert y_{max} \Vert_1$. Let us now upper bound $\Vert \nabla_y g_t(y) \Vert_2$ to obtain a bound on $G_g$. Recall $g_t(y) = -r_t(x) - y^\top(\frac{b}{T}-c_t(x_t))$, and therefore we have $$\begin{aligned}
\Vert \nabla_y g_t(y) \Vert_2 &= \Vert \frac{b}{T} + c_t(x_t)\Vert_2\\
&\leq \frac{1}{T} \Vert b \Vert_2 + \Vert 0 - c_t(x_t)\Vert_2\\
&\leq \frac{1}{T} \Vert b \Vert_2 + \Vert [G \Vert x_0 - x_t\Vert_2;G \Vert x_0 - x_t\Vert_2;...,G \Vert x_0 - x_t\Vert_2] \Vert_2\\
&\leq \frac{1}{T} \Vert b \Vert_2 + \sqrt{\sum_{i=1}^m G D_X}\\
& = \frac{1}{T} \Vert b \Vert_2 + \sqrt{m G D_X}.\end{aligned}$$ Plugging in $\eta_1 = \frac{D_X \sqrt{T}}{G_f}, \eta_2 = \frac{D_Y \sqrt{T}}{G_g}$ we get $$\begin{aligned}
{\mathsf{Regret}}(T) &\leq 5 G_f D_X \sqrt{T} + 5 G_g D_Y \sqrt{T}\\
&= 5 G (1+\Vert y_{max}\Vert_1) D_X \sqrt{T} + 5 (\frac{1}{T} \Vert b \Vert_2 + \sqrt{m G D_X}) \Vert y_{max}\Vert_2 \sqrt{T}.\end{aligned}$$ This concludes the proof. ]{}
Numerical Experiments
=====================
Individual-Regret and SP-Regret {#subsec:regret-simulation}
-------------------------------
To further illustrate the relationship between SP-regret and individual-regret and the impossibility result of Theorem \[thm:impossible\], we compare the performance of two online algorithms numerically. The first algorithm is defined in Algorithm \[alg: SP-FTL\]. In the second algorithm, which we call , player 1 applies online gradient descent to function ${\mathcal{L}}_t(\cdot, y_t)$ and player 2 applies online gradient ascent to function ${\mathcal{L}}_t(x_t, \cdot)$.
We generated two different instances. In both instances, we assume $X=Y=[-10,10]$. The payoff functions in both instances are the same for $t=1,...,\lfloor T/3\rfloor$, given by $
{\mathcal{L}}_t(x,y) = xy + \frac{1}{2}\Vert x-2 \Vert^2 - \frac{1}{2}\Vert x+1 \Vert^2.
$ In Instance 1, for $t=\lfloor T/3\rfloor +1,...T $, we define $
{\mathcal{L}}_t(x,y) = xy + \frac{1}{2}\Vert x+1 \Vert^2 - \frac{1}{2}\Vert x+2 \Vert^2.
$ In Instance 2 for $t=\lfloor T/3\rfloor +1,...T $, we define $
{\mathcal{L}}_t(x,y) = xy + \frac{1}{2}\Vert x+1\Vert^2 - \frac{1}{2}\Vert x-3\Vert^2.
$ Since these functions are strongly convex-concave, when players use OGDA with step size $O(\frac{1}{t})$, they are both guaranteed logarithmic individual-regret.
In Figure \[fig:fig\_one\], we plot the SP-regret of the two instances. On the left, it can be seen that the SP-regret of increases significantly after the payoff function switches at $\lfloor T/3\rfloor$, while the SP-regret of remains small throughout the entire horizon.
From Figure \[fig:fig\_two\], we can observe when both players use the algorithm, their individual-regrets are small. However, when when they use , at least one player suffers from high individual-regret. Figures \[fig:fig\_one\] and \[fig:fig\_two\] verify Theorem \[thm:impossible\], which states that no algorithm can achieve both sublinear SP-regret and sublinear individual-regret.
SP-FTL and OGDA for the OCOwK problem {#subsec:numerical-ocowk}
-------------------------------------
In this section, we compare the numerical performance of and (Online Gradient Descent/Ascent) for solving a OCOwK problem. In , player 1 applies online gradient descent to function ${\mathcal{L}}_t(\cdot, y_t)$ and player 2 applies online gradient ascent to function ${\mathcal{L}}_t(x_t, \cdot)$. The proof for Theorem \[no\_ocowk\_regret\_pdftl\] can also show that has a regret of $O(\sqrt{T})$ (see Remark \[rmk:bandit-OCOwK\]).
We construct a numerical example where for each iteration $t=1,...,T$, the decision maker chooses an action $x_t \in X = [0,20]$. The reward function is $r_t = -x^2 + b_t x$ where $b_t \sim U[0,20]$. There are two types of resources with budgets $B_1$ and $B_2$. The consumption function for the first resource is given by $c_{t,1} = (a_t x)^2 + 50x$ where $a_t \sim U[0,3]$, and the consumption function for the is $c_{t,2}=x$. We assume the budgets are some linear functions of $T$, $B_1(T)$ and $B_2(T)$ respectively. In our simulations $B_1$ and $B_2$ are chosen so that playing the optimal solution to the problem without budgets is no longer optimal.
Figure \[fig:fig\_three\] compares the performance of vs on the OCOwK instance defined above. Performance is measured as the ratio of total reward incurred by the algorithm and the solution to Equation across 25 simulation runs. It can be observed that both algorithms indeed improve their performance as $T$ increases. Moreover, it can be observed that while has worse performance for small values of $T$, the rate at which performance improves is greater than that for , which is consistent with our theoretical results that has $\tilde{O}(T^{5/6})$ regret and (or ) has $O(\sqrt{T})$ regret.
Conclusion
==========
In this paper we introduced the Online Saddle Point problem. In this problem, we consider two players that jointly play an arbitrary sequence of convex-concave games against Nature. This problem is a generalization of the classical Online Convex Optimization problem, which focuses on a single player. The objective is to minimize the saddle-point regret (SP-Regret), defined as the absolute difference between the cumulative payoffs and the saddle point value of the game in hindsight.
We proposed an algorithm for the Online Saddle Point problem and showed that it achieves $\tilde{O}(\sqrt{T})$ SP-Regret for a game with $T$ periods. In the special case where the payoff functions are strongly convex-concave, we showed that the algorithm attains $O(\log T)$ SP-Regret. Furthermore, we proved that if the sequence of payoff functions are chosen arbitrarily, any algorithm with $o(T)$ regret for the Online Convex Optimization problem may incur $\Omega(T)$ SP-Regret in the worst case. [ We also consider the special case where the payoff functions are bilinear and the decision sets are the probability simplex. In this setting we are able to design algorithms that reduce the bounds on SP-Regret from a linear dependence in the dimension of the problem to a *logarithmic* one. We also study the problem under bandit feedback and provide an algorithm that achieves sublinear SP-Regret.]{} This implies that all existing algorithms for the Online Convex Optimization problem cannot be applied to the Online Saddle Point problem. Moreover, we showed how our algorithm can be applied to solve the problem of Stochastic Online Convex Optimization with Knapsacks. Finally, we performed some numerical simulations to validate our results.
[^1]: We will also use the notation $\Delta_{X,\theta}$ and $\Delta_{Y,\theta}$ to mean the restricted simplex of Player 1 and 2, respectively
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Magneto-transport and drag measurements on a quasi-Corbino 2D electron bilayer at the systems total filling factor 1 ($\nu_T$=1) reveal a drag voltage that is equal in magnitude to the drive voltage as soon as the two layers begin to form the expected $\nu_T$=1 exciton condensate. The identity of both voltages remains present even at elevated temperatures of 0.25 K. The conductance of the drive layer vanishes only in the limit of strong coupling between the two layers and at T$\rightarrow$0 K which suggests the presence of an excitonic circular current.'
author:
- 'L. Tiemann'
- 'J. G. S. Lok'
- 'W. Dietsche'
- 'K. von Klitzing'
- 'K. Muraki'
- 'D. Schuh'
- 'W. Wegscheider'
title: Exciton condensate at a total filling factor of 1 in Corbino 2D electron bilayers
---
When two closely spaced two-dimensional electron systems (electron bilayer) are exposed to a perpendicular magnetic field $B$ so that each layer has a filling factor close to 1/2 and the relative distance between interlayer electrons, parameterized by the ratio $d/l_B$ ($d$: layer separation, $l_B=\sqrt{\hbar /eB}=1/\sqrt{2\pi
n_T}$: magnetic length with $n_T$ as the total density), is sufficiently small, a new quantum Hall (QH) state characterized by the total filling factor 1 ($\nu_T$=1) can be observed. Both theoretically and experimentally it is found that this novel state in bilayers occurs at a $d/l_B$ ratio of less than $\approx$ 2 [@Spielmann00; @Kellogg02]. In the limit of comparably small tunneling, its origin is dominated by Coulomb interactions, where the electrons in the two layers form a strongly correlated many-body state to minimize their exchange energy. In this state and in the low temperature limit, spontaneous interlayer phase coherence develops, driving all electrons in a quantum mechanical superposition of the layer eigenstates sharing the same macroscopic phase $\phi $ [@MacDonald01; @Wen92]. However, the predicted Kosterlitz-Thouless type of phase-transition [@Moon1995] has not yet been unequivocally demonstrated in experiment. After a particle-hole transformation that changes the sign of the interactions from repulsive to attractive, this state can be regarded as an excitonic condensate, where each electron is bound to a vacant state in the opposite layer. Interlayer drag experiments, where a constant current is passed through one of the layers (drive layer) and the induced longitudinal and transverse voltage drop in the other layer (drag layer) is measured [@Kellogg02], have revealed a Hall drive and drag which approaches a quantized value of $h/e^{2}$ at $\nu_T=1$ in a temperature-activated fashion. The quantization of the Hall drag is an indirect indication of a superfluid mode of excitons [@Eisenstein03], which can be viewed as either an uniform flow of interlayer excitons or as a counter flow of electrons in the opposite layers.
In standard Hall bars the occurrence of the ordinary integer QH effect with the vanishing of the longitudinal resistance and the quantization of the Hall resistance can be explained in terms of one-dimensional (dissipationless) edge channels [@LB85; @LB88]. However, in the case of the $\nu_T=1$ and its associated superfluid transport mode, it cannot be ruled out from the Hall bar data that a dissipationless quasi-particle current at the sample edges is responsible for the observed effects [@Rossi2005].
In this paper, we report on interlayer drag measurements on a *quasi*-Corbino electron bilayer with independent contacts to both layers. An *ideal* Corbino structure allows direct measurement of the conductivity $\sigma_{xx}$ in contrast to the common Hall bar geometry where the resistivities are measured. We observe that at $\nu_T=1$ a voltage develops in the drag layer that equals in sign and magnitude the voltage across the drive layer. We find that the identity of the drag and drive voltages is maintained up to high temperatures or large $d/l_B$ where the single layer current flow shows nearly no trace of the $\nu_T=1$ QH effect. Due to the absence of sample edges connecting source and drain contacts in a ring, the current is driven selectively through the bulk of the $\nu_T=1$ system. At low temperatures, drive and drag voltage remain identical and the drive current nearly vanishes. The Corbino experiments thus open a new venue to explore the bulk property of the $\nu_T=1$ system.
![Schematic view of the Corbino geometry used in this experiment. Application of appropriate voltages to the back gates (marked as ”BG”) and front gates (”FG”) will lead to contact separation, i.e. contacts 1 through 5 will connect to the upper quantum well and 1\* through 5\* to the lower one.[]{data-label="fig:Corbino"}](Fig1.ps){width="48.00000%"}
Our two-dimensional (2D) electron bilayer is confined in two 19-nm GaAs quantum wells, separated by a 9.9 nm superlattice barrier composed of alternating layers of AlAs (1.70 nm) and GaAs (0.28 nm). Each quantum well has an intrinsic electron density of about $4.3\times 10^{14}$ m$^{-2}$ and a low-temperature mobility of 67 (45) m$^{2}$/Vs for the upper (lower) quantum well (measured on a Hall bar fabricated from the same wafer). Since the ideal Corbino geometry is not compatible with the selective-depletion technique [@Eisenstein90; @Rubel97] for independently contacting each layer, we instead employ a quasi-Corbino geometry with four contact arms attached to each ring as depicted in Fig. \[fig:Corbino\]. The back gates were patterned *ex situ* from a Si-doped GaAs epilayer before growing an insulating GaAs/AlGaAs superlattice and the bilayer on top. Electrical isolation between the two layers is achieved by applying appropriate negative voltages to the buried back gates and metallic front gates crossing the contact arms. One set of contacts can then be used to pass a current and another one to measure the voltage across the ring. The densities in each layer can be adjusted independently by using another set of front and back gates (not shown) covering the active region of the Corbino ring including the ring edges.
Below we present data from two samples from the same wafer which show essentially the same behavior. Sample A consists of a quasi-Corbino ring with an outer diameter of $d_O=600$ $\mu $m and a ring width of $w=140$ $\mu $m, while sample B is characterized by $d_{O}=780$ $\mu $m and $w=230$ $\mu $m. For all samples interlayer tunneling is small; the interlayer resistance (at zero magnetic field and 0.25 K) is of the order $10^{7}\ $-$10^{8}$$\Omega$. Transport measurements were performed by using a standard lock-in technique with the sample mounted at the cold finger of a dilution refrigerator or a $^3$He system. For all measurements the electron densities in the two layers were adjusted to be equal. A small excitation voltage $V_{exc}$ (60-65$\mu $V, 3-5Hz [^1]) was applied radially across one layer (the drive layer) through an isolation transformer and the induced current through this layer was measured with a small resistance connected in series. We would like to stress that the total current has a radial and an azimuthal part. These two parts oscillate anti-cyclically as a function of the magnetic field, i.e. in a QH state the radial fraction is zero while the azimuthal (circular) part is maximal. Hence, the (radial) voltage dropping over the drive layer changes in response to the radial current as well. For that reason, the voltage across the drive layer was monitored using a separate pair of contacts in a quasi four-terminal geometry together with the induced voltage in the drag layer. This excludes also the effects of the finite resistances of the ohmic contacts and the contact arms. The measurements were reproducible upon interchanging contacts and upon interchanging drive and drag layer.
![Top panel: Measured drive (solid line) and drag (dash-dot line) voltages at $T_{bath}$=15 mK on sample B. The (integer) filling factors $\nu\leq2$ and $\nu_T=1$ ($d/l_B$=1.62) are labeled. Bottom panel: Measured current in the drive layer. The inset plots the temperature dependence of the radial conductance G; the line is a fit using G $\propto$ exp(-$E_{gap}$/T).[]{data-label="fig:fullsweep1"}](Fig2.eps){width="50.00000%"}
We start by showing data at lowest temperatures. Fig \[fig:fullsweep1\] presents data at $T_{\mathrm{bath}}$ = 15 mK on sample B. The bottom and top panel plot the measured (radial) current in the drive layer and the corresponding drive and drag voltages as a function of the magnetic field. The electron densities in the two layers were adjusted to be equal, producing a total density of $4.8\times 10^{14}$m$^{-2}$. Below 1.5 T, the current oscillates reflecting the varying filling factors and integer QH states. At total filling factor 1 which occurs here at about 2.0 T, we observe a strong minimum in the current like at the ordinary QH states at lower magnetic fields. As a result, the voltage drop over the drive layer almost equals the source voltage (top panel). Meanwhile, a large drag voltage develops over the region of the correlated $\nu_T=1$ phase, with the sign and magnitude identical to that of the drive layer. Since the radial component of the current in the drive layer is nearly zero, one possible explanation for the observed drag voltage is the existence of an azimuthal (i.e. circling) current in the drive layer, in analogy to the ordinary QH states. Owing to the excitonic coupling it would trigger an azimuthal current of the same magnitude in the drag layer, leading to identical voltages across both layers. However, we neither know the nature of this excitonic current nor where it flows. It could be homogeneously distributed throughout the bulk or rather concentrated at the sample edges. Nevertheless, the well-established model of electron-hole pairing around $\nu_T=1$ implies that such a transport mode in Corbino bilayers might be possible. Supported is that notion by the fact that the ohmic contacts of the drag layer in our geometry are located at the opposite side of the ring, i.e. approximately 1 mm away from the ohmic contacts of the drive layer. In previous drag experiments using a standard Hall bar geometry, identical Hall voltages in the drag and drive layers were also considered to be signaling the underlying excitonic superfluidity.
The origin of identical voltages could equivalently be attributed to the special nature of the excitonic state. Since an excitonic (quasi-particle) wave function would have to exist across the barrier, quasi-particle transfer between the layers would become possible as soon as the system reaches a total filling factor of 1. While standard tunneling spectroscopy experiments [@Spielmann00] performed on very similar electron bilayer samples indeed indicate that tunneling becomes resonantly enhanced in the vicinity of $\nu_T=1$, identical voltages could only be explained if the interlayer resistance became insignificantly small compared to the bulk resistance. This, however, is inconsistent with tunneling experiments [@Spielmann00; @Spielmann04; @Wiersma06] on common electron bilayer samples, showing resistances within the M$\Omega$ range instead.
![Top panel: Drive (solid line) and drag (dash-dot line) voltage versus the magnetic field at T = 0.25 K measured on sample A. Bottom panel: The current in the drive layer measured simultaneously. The inset illustrates the conductance G of the drive layer as a function of the magnetic field and the total density. Clearly visible is how the conductance at $\nu_T=1$ decreases as the total density $n_T$ is reduced.[]{data-label="fig:fullsweep2"}](Fig3.ps){width="50.00000%"}
Fig. \[fig:fullsweep2\] plots data taken at a temperature of $T$=0.25 K on sample A. The densities in both layers are still equal but reduced to a total electron density of approximately $4.2\times
10^{14}$ m$^{-2}$. Now $\nu_T=1$ occurs at $B$=1.76 T which corresponds to $d/l_B=1.49$. At 0.25 K the minimum in the current has almost entirely disappeared (bottom panel). Nonetheless, there is still a sizeable peak in the drive voltage at $\nu_T=1$ (solid line in the top panel). Surprisingly, the voltage over the drag layer (dash-dot line) also displays a peak [*with the same amplitude*]{}. This striking observation of a nearly doubled dissipation in the drive layer accompanied by an identical drag voltage can be interpreted as evidence that both layers are in a state of commencing interlayer correlation. A previous report on drag experiments on Hall bar bilayers [@Kellogg02] has shown that identical voltages, i.e. the quantization of the Hall resistance to $h/e^{2}$, are only observable at lowest temperatures and low $d/l_B$ ratios when the $\nu_T=1$ QH state is fully developed. While this is in direct contrast to our data and might indicate a geometry-dependence, the resilience of the $\nu_T=1$ QH state to increasing temperatures yet is a behavior reminiscent of results obtained on bilayer 2-dimensional *hole* gas samples [@Tutuc06] in counter-flow configuration. We cannot offer any explanation for these similarities, however, it might simply be owing to the reported interlayer leakage or the larger effective mass of the holes.
We find that the ratio of both voltages remains 1 until $d/l_B$ approaches a critical limit. We have traced the drag and drive voltages for a number of different (but matched) total densities at 0.25 K. The results are summarized in Fig. \[fig:Vvsdlb\] which plots drag and drive voltage at $\nu_T$=1 versus $d/l_B$. At 0.25 K, the identity of both drag and drive voltages can be tracked up to a $d/l_B$ ratio of about 1.65 where the $\nu_T=1$ QH state is collapsing owing to thermal fluctuations. For $d/l_B>1.65$ small peaks of different amplitude can be observed as illustrated in the inset.
![Drive (solid dots) and drag (open dots) voltages at $\nu_T=1$ versus the parameter $d/l_B$ at $T_{bath}$ = 0.25 K. Above $d/l_B$=1.65, the amplitudes of drive and drag voltage at total filling factor 1 diverge. The inset shows the corresponding field sweep for the last pair of points at $d/l_B$=1.73.[]{data-label="fig:Vvsdlb"}](Fig4.eps){width="50.00000%"}
At some finite temperature, the collapse of the excitonic condensate at $\nu_T=1$ in the bilayer can be observed. For sample B and temperatures below 0.25 K, the conductance G=I/V is well described by thermal activation, i.e. G $\propto$ exp(-$E_{gap}$/T), with an activation energy gap of approximately 0.5 K as shown in the inset of Fig \[fig:fullsweep1\]. The magnitude of the extracted energy gap is in good agreement with earlier reports on comparable double quantum well structures [@Kellogg02; @Kellogg04; @Tutuc04; @Wiersma04], where the activation energy was extracted from measurements of the temperature dependence of the longitudinal resistance in Hall bars.
In a theoretical Letter, Stern and Halperin [@Stern02] suggested that the electron bilayer system at high $d/l_B$ ratios is composed of puddles of strong interlayer correlation incorporated in the compressible fluids of the individual layers. Their model, albeit addressing specifically Hall bar geometries, appears to be connected with our observations as well. As long as these puddles are small in number and/or unrelated, a sizeable current could flow through the bulk between source and drain contacts. As $d/l_B$ is decreased their number and/or size will increase until they eventually percolate, while the current through the bulk slowly diminishes. The smooth transition we observe in Corbino samples from a compressible to a nearly fully incompressible state upon decreasing the temperature and/or the parameter $d/l_B$ appears to signify such a percolation.
In conclusion, we have conducted interlayer drag experiments on quasi-Corbino electron bilayers. At the lowest temperature and strong coupling, the ratio of drag and drive voltages is 1 while the conductance in the drive layer vanishes. These data imply a circular potential distribution along the sample edges owing to a circling (azimuthal) excitonic current in both layers. At elevated temperatures, the identity of both voltages is still present.
We thank Allan H. MacDonald for inspiring conversations. Also we would like to acknowledge Maik Hauser for providing some of our MBE wafers, J. H. Smet for giving us access to his dilution system, S. Schmult for his help with the manuscript and the German Ministry of Education and Research (BMBF) for its financial support (BMBF 01BM456).
[15]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , , , , , ****, ().
, , , , , ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , , , , , , ****, ().
, ****, ().
[^1]: The excitation voltage is generated with the internal oscillator of one lock-in amplifier and sized down with a voltage divider.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a calculation for the diffractive structure functions $x_{\rm p} F_{\rm T}^{D(3)}$ and $x_{\rm p} F_{\rm L}^{D(3)}$ at low values of $Q^2$ using an eikonal approximation. The non-perturbative model of the stochastic QCD vacuum is used to evaluate the interactions between the quarks (antiquarks) and the color field. The diffractive final state contains only a quark-antiquark pair. We show the behavior of the diffractive structure functions against $\beta$ and $Q^2$.'
address: |
Institut für Theoretische Physik der Universität Heidelberg,\
Philosophenweg 16, D-69120 Heidelberg
author:
- 'Oscar E. Ramírez del Prado [^1]$^{,}$ [^2]$^{,}$'
title: '[**Diffractive Structure Functions in the Model of the Stochastic Vacuum**]{} [^3]'
---
Introduction
============
A calculation for the diffractive structure functions in a non-perturbative region using a leading eikonal approximation schema [@Nachtmann] and the model of the stochastic vacuum (MSV) [@Dosch1; @DoschSimonov] is presented. The process under consideration is the diffractive dissociation of a photon by a proton. The proton is taken in a quark-diquark representation and at high energies the photon can be represented as a quark-antiquark pair. The quarks and the diquark pick up an eikonal phase factor in the vacuum background field and the interaction is obtained by average over the vacuum field by means of the MSV. If we draw the trajectories of the $q-\bar{q}$ pair in a space-time diagram and close the extremes to have a colorless object, we get the so called Wegner-Wilson Loops.
Diffractive Structure Functions
===============================
We want to calculate the diffractive dissociation process (figure \[fig1\]) where a virtual photon scatters with a proton and in the final state we get a proton $+$ anything ($X$):
$$\label{difdis}
\gamma^{*}+p\rightarrow p+X \; ,$$
we describe the “anything” by a $q-\bar{q}$ pair in a color singlet state. In the center of mass frame the initial transverse momentum of both photon and proton is $\vec{0}$, furthermore for the photon $q^{+}$ and for the proton $p^{-}$ become large. After the collision the proton remains almost intact with momentum $p' \approx p$ and the photon dissociates in a quark carrying a fraction $z$ of the photon’s original momentum and an antiquark carrying the remaining $(1-z)$ fraction of the momentum. This configuration is very sensitive to the extreme kinematical regions where the quark or the antiquark carries most of the photon momentum i.e, $z \rightarrow 1,0$ [@BuchmullerMcDemontHebecker].
{width="6.5cm"}
The momentum transfer is purely transversal, and it will be denoted by $\vec{\Delta}_{\perp}$.
We are working at ISR energies in $s$ and in the regime where $s\gg t$ with a momentum transfer $|t|\leq 1 \; GeV^2$.
We define the following invariants:
$$\begin{aligned}
\label{invariants}
s&=&W^2=(p+q)^2 \; ,\\
t&=&(p'-p)^2=\Delta_{\perp}^2 \; ,\end{aligned}$$
and the standard variables for diffraction:
$$\label{xpomeron}
x_p=\frac{Q^2+M^2}{Q^2+W^2}=\frac{x}{\beta} \hspace*{1em} , \hspace*{1em} x=\frac{Q^2}{s} \; ,$$
$$\label{beta}
\beta=\frac{Q^2}{Q^2+M^2} \; ,$$
where $M^2$ is the invariant mass of the $q-\bar{q}$ pair.
In a pomeron picture $x_p$ can be interpretated as the momentum fraction of the pomeron within the proton and $\beta$ as the momentum fraction of the struck quark within the pomeron.
The phase space is given by:
$$\label{phasespacedef}
d\Phi_3=\frac{1}{2^3(2\pi)^5}\frac{Q^2}{2s}\frac{dx_p}{x_p}\frac{dz}{\beta}\;d^2\Delta_{\perp}\;d\theta_{\vec{l}} \; ,$$
The cross section is
$$\label{crosssectiondef}
d\sigma^{\lambda}=\frac{1}{2s}\left|T^{\lambda}\right|^2d\Phi_3 \; ,$$
and the scattering amplitude $T^{\lambda}$, following [@Nachtmann; @DoschFerreraKramer], is given by
$$\begin{aligned}
\label{tamplitude}
\lefteqn{T^{\lambda} = i2s(2\pi)^{\frac{1}{2}}\sqrt{2z(1-z)}\int{d^2b} \; e^{i{\vec{\Delta}}_{\perp}\cdot\vec{b}}} \nonumber \\
&&\times \int{d^2r_{\perp}}\Psi^{\lambda}_{\gamma^{*}}(z,r_{\perp})W(\vec{b},\vec{r_{\perp}},z)e^{-i\vec{l}\cdot\vec{r_{\perp}}} \; ,\end{aligned}$$
the index $\lambda$ in the photon wave function $\Psi^{\lambda}_{\gamma^{*}}(z,r_{\perp})$ refers to its polarization ($\lambda=0$ longitudinal, $\lambda=\pm 1$ transversal).
The function
$$\label{functionw}
W(\vec{b},\vec{r_{\perp}},z) = \hspace*{-.5em} \int \hspace*{-.5em} \frac{d^2 r_2}{4\pi} \left|
\Psi_P(r_2)\right|^2
\tilde{J}\left(\vec{r}_1,\vec{r}_2,\vec{b},z\right),$$
involves the proton-dipole interaction and contains the loop-loop scattering term $\tilde{J}\left(\vec{r}_1,\vec{r}_2,\vec{b},z\right)$ which is evaluated in the MSV.
We take a Gaussian function for the proton quark-diquark wave function
$$\label{protonwf}
\Psi_P(r_2)=\frac{\sqrt{2}}{S_P}e^{-\frac{r_2^2}{4S_P^2}} \; ,$$
and the photon wave function has the form given in [@DoschGousKulPir] according to light-cone perturbation theory.
Now, writing everything together and integrating over $\vec{\Delta}_{\perp}$ and $b$ we have
$$\begin{aligned}
\label{crossection}
\lefteqn{d\sigma^{\lambda}=\frac{Q^2}{2^2(2\pi)^2}\frac{dx_p}{x_p}\frac{dz}{\beta}d\theta_{\vec{p}} \; z(1-z)} \\
&&\hspace*{-2em} \times \int{d^2b}\left| \int d^2r_{\perp} \Psi^{\lambda}_{\gamma^{*}}(z,r_{\perp})W(\vec{b},\vec{r_{\perp}},z)e^{-i\vec{l}\cdot\vec{r_{\perp}}}\right|^2 \; . \nonumber \end{aligned}$$
With the usual definition of the structure functions and taking into account the symmetries in the angular integrations and the different polarizations we obtain:
a)
: Longitudinal case: $$\begin{aligned}
\label{longitudinal}
\lefteqn{x_P F_{\rm L}^{D(3)}=\frac{16}{(2\pi)^3}\frac{Q^6}{\beta}N_{c}\int_0^1
dz\, (z (1-z))^3 } \nonumber \\
&&\hspace*{-2em} \times \int_0^\infty db\, b \int_0^{2\pi} d\theta \nonumber \\
&&\hspace*{-2em} \times \left| \int \frac{d^2 r_1}{4\pi}\; e^{\vec{l}\cdot\vec{r}_1} {\rm K}_0(\epsilon r_1)W\left(\vec{r}_1,b,z\right)\right|^2 \; .\end{aligned}$$
b)
: Transversal case: $$\begin{aligned}
\label{transversal}
\lefteqn{x_P F_{\rm T}^{D(3)}=\frac{4}{(2\pi)^3}\frac{Q^6}{\beta}N_{c}\int_0^1 dz\, (z (1-z))^2} \nonumber \\
&&\hspace*{-2em} \times \; (z^2+ (1-z)^2) \int_0^\infty db\, b \int_0^{2\pi} d\theta \nonumber \\
&&\hspace*{-2em} \times \left| \int \frac{d^2 r_1}{4\pi}\; e^{i\theta_{r_1}}\; e^{\vec{l}\cdot\vec{r}_1}{\rm K}_1(\epsilon r_1) W\left(\vec{r}_1,b,z\right)\right|^2 \; .\end{aligned}$$
where ,\
is the transverse momentum of the $q-\bar{q}$ pair.
These results have the same structure as those of Buchmüller et al. [@BuchmullerMcDemontHebecker], but we go further and calculate the dipole-proton scattering in the MSV and give numerical results.
Results
=======
In this section we present the numerical evaluation of equations (\[longitudinal\]) and (\[transversal\]). The error bars in the plots come from the numerical computation. We have taken the parameters of the model, $$a=0.31fm,\,<g^2FF>=3GeV^4,\,S_P=0.85fm$$ as in [@RueterPhD]. The structure functions are calculated in the range from $1.0$ to $12.0$ $GeV^2$ for $Q^2$. For the $\beta$ dependence we take values between $0.5$ and $1.0$ where the non-perturbative contribution becomes more important because the transverse momentum of the quarks gets small making that the eikonal factors “feel” large transverse distances. Since the extreme regions on the $z$ integral and small dipole radius might give an important contribution we present also results using the modifications to the model proposed by Rüter [@Rueterthis].
For the behavior of the structure functions against the photon’s momentum $Q^2$ at fixed $\beta = 0.65$ we found out that the transversal structure function, as shown in figure \[tb65qcd\], gets constant at big values of $Q^2$ revealing a leading twist behavior. The same $Q^2$ independence can be seen for different values of $\beta$.
(200,125) (0,70) (10,0)[{width="6.5cm"}]{} (90,0)[$Q^2 \; \; GeV^2$]{}
The longitudinal structure function (figure \[lb65qcd\]) as a function of $Q^2$ for fixed $\beta=0.65$ has an important contribution at low values of $Q^2$ and decreases with increasing $Q^2$ showing a higher twist behavior. The same behavior is observed for different values of $\beta$.
The $\beta$ dependence of the structure functions at fixed values of $Q^2$ is plotted in figure \[figura4\] for $Q^2=4.5 \; GeV^2$ and in figure \[figura6\] for $Q^2=12.0 \; GeV^2$. The longitudinal structure function grows up with $\beta$ whereas the transversal decreases. As $\beta \rightarrow 1$ the longitudinal contribution to $x_{\rm p} F_{\rm 2}^{D(3)}$ becomes important. These results are not changed by the modifications introduced in the model.
(200,125) (0,70) (10,0)[{width="6.5cm"}]{} (90,0)[$Q^2 \; \; GeV^2$]{}
(200,125) (10,0)[{width="6.5cm"}]{} (100,0)[$\beta$]{}
We also compared our predictions for $x_{\rm p} F_{\rm 2}^{D(3)}$ and $x_{\rm p} F_{\rm T}^{D(3)}$ with the experiment [@HERAH1]. We see that at $Q^2=4.5 \; GeV^2$ (figure \[figura8\]) we have a good agreement with experiment, but for $Q^2=12.0 \; GeV^2$ (figure \[figura9\]) the result of the no modified model is by a factor 2 to large. Introducing modifications for small dipoles as proposed in ref [@Rueterthis] leads however to very good agreement.
(200,125) (10,0)[{width="6.5cm"}]{} (100,0)[$\beta$]{}
(200,125) (10,0)[{width="6.5cm"}]{} (100,0)[$\beta$]{}
Conclusions
===========
Our approach for calculating the diffractive structure functions is similar to the semi-classical approach of Buchmüller et al. [@BuchmullerMcDemontHebecker]. Our calculation reproduces the leading twist behavior for $x_{\rm p} F_{\rm T}^{D(3)}$, but gives a not negligible contribution at $\beta$ near $1$. It was possible to get good agreement with experiment introducing the modifications presented in [@Rueterthis]. Further work to calculate the structure functions for values of $Q^2 < 1.0 \; GeV^2$ is in progress.
(200,125) (10,0)[{width="6.5cm"}]{} (100,0)[$\beta$]{}
Acknowledgments
===============
I would like to thank H.G. Dosch, A. Hebecker, M. Rüter, G. Kulzinger and E. Berger for useful discutions and comments.
[99]{} O. Nachtmann, Ann. Phys. [**209**]{}, 436 (1990). H.G. Dosch, Phys. Lett. [**B190**]{}, 177 (1987). H.G. Dosch and Y.A. Simonov, Phys. Lett. [**B205**]{}, 339 (1988). W. Buchmüller, M.F. McDermott and A. Hebecker, Nucl. Phys. [**B487**]{} (1997) 283. H.G. Dosch, E. Ferreira, A. Krämer, Phys. Rev. [**D50**]{}, 1992 (1994). H.G. Dosch, T. Gousset, G. Kulzinger and H.J. Pirner, Phys. Rev. [**D55**]{}, 2602 (1997). M. Rüter, [*Quark-Confinement und diffraktive Hadron-Streuung im Modell des stochastischen Vakuums*]{}, Ph.D. Thesis (Univ. Heidelberg), 1997. M. Rüter, These proceedings. C. Adloff et al. Z. Phys. [**C76**]{} (1997) 613.
[^1]: E-mail: [email protected]
[^2]: supported by the Deutscher Akademischer Austauschdienst
[^3]: Talk presented at the QCD’98 Euroconference, Montpellier, July 1998.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Recent massive sky surveys in different bandwidths are providing new opportunities to modern astronomy. The Virtual Observatory (VO) represents the adequate framework to handle the huge amount of information available and filter out data according to specific requirements.
In this work, we applied a selection strategy to find new, uncatalogued hot subdwarfs making use of VO tools. We used large area catalogues (GALEX, SDSS, SuperCosmos, 2MASS) to retrieve photometric and astrometric information of stellar objects. To these objects, we applied colour and proper motion filters, together with an effective temperature cutoff, aimed at separating hot subdwarfs from other blue objects such as white dwarfs, cataclysmic variables or main sequence OB stars. As a result, we obtained 437 new, uncatalogued hot subdwarf candidates. Based on previous results, we expect our procedure to have an overall efficiency of at least 80 per cent. Visual inspection of the 68 candidates with SDSS spectrum showed that 65 can be classified as hot subdwarfs: 5 sdOs, 25 sdOBs and 35 sdBs. This success rate above 95 per cent proves the robustness and efficiency of our methodology.
The spectral energy distribution of 45 per cent of the subdwarf candidates showed infrared excesses, a signature of their probable binary nature. The stellar companions of the binary systems so detected are expected to be late-type main sequence stars. A detailed determination of temperatures and spectral classification of the cool companions will be presented in a forthcoming work.
author:
- |
E. Pérez-Fernández $^{1,2}$[^1], A. Ulla$^{2}$, E. Solano$^{3,4}$, R. Oreiro$^{5}$ and C. Rodrigo$^{3,4}$\
$^{1}$ IES de Beade, Consellería de Educación e O.U., Camiño de Outeiro 10, 36312 Vigo, Spain\
$^{2}$Departamento de Física Aplicada, Universidade de Vigo, Campus Lagoas-Marcosende, 36310 Vigo, Spain\
$^{3}$Departamento de Astrofísica, Centro de Astrobiología (INTA-CSIC), PO Box 78, E-28691 Villanueva de la Cañada (Madrid)\
$^{4}$ Spanish Virtual Observatory\
$^{5}$ Instituto de Astrofísica de Andalucía (IAA-CSIC), Glorieta de la Astronomía, s/n, 18008 Granada.
bibliography:
- 'references.bib'
date: 'Accepted 2016 January 21; Received 2016 January 20; in original form 2015 July 6'
title: A search for new hot subdwarf stars by means of Virtual Observatory tools II
---
\[firstpage\]
stars:early type – hot subdwarfs – Virtual Observatory tools – astronomical databases:miscellaneous
Introduction
============
Hot subdwarf (hot sd) stars are core-helium burning stars at the end of the horizontal branch or even beyond that stage. The origin of these faint, blue objects is still a matter of controversy. With effective temperatures exceeding 19000 K and logg $\geq$ 5, hot sds are objects that have lost most of their H envelope in previous evolutionary stages, leading to a $\sim 0.5M_{\sun}$ star. They are unable to follow canonical evolution through the Asymptotic Giant Branch (AGB) proceeding, instead, directly towards the white dwarf cooling track. Circumstances that lead to the removal of all but a tiny fraction of the hydrogen envelope, at about the same time as the core has achieved the mass required for the He flash ($\sim 0.5 M_{\sun}$ ), are still a matter of debate. Theoretical evolution scenarios proposed so far include enhancement of the mass loss efficiency near the red giant branch (RGB) tip [@Cruz96] or mass transfer through binary interaction [@Mengel96]. See [@Heber_rev] for a review on observational and theoretical aspects of hot subdwarfs, or [@Geier13b] for more recent discoveries.
Hot subdwarfs are found in the field, both in the disk and halo, but also populating the most Extreme part of the Horizontal Branch (EHB) of some Galactic clusters. Based on this observational evidence, they have been proposed as an explanation for the UV-upturn phenomenon shown in some elliptical galaxies [@Brown1997].
Hot sds are divided in two main classes, sdBs and sdOs, according to composition. SdB spectra are dominated by the Balmer series, while sdOs are hotter objects caracterized by the presence of He[ii]{} 4686Å and the Pickering series. Additionally, a variety of He[i]{} lines may appear in both classes, and some sdOs show metallic C or N lines. More complex classification schemes have been proposed in [-@PGcat] or more recently in [@Drilling13].
The subdwarf database [@Roydatabase] catalogues 1600 sdBs and 500 sdOs spectroscopically confirmed hot subdwarfs. A significant number of new hot sds have been discovered in more recent studies like [@Vennes2011], [@MUCHFUSS; @Geier15_RVcat], [@Nemeth2012] , [@Kleinman], [@Kupfer2015] or [@Kepler15]. Increasing the number of hot sds is important for a robust statistical confrontation with theoretical evolutionary scenarios. It may also lead to the discovery of interesting objects that are still scarce, such as pulsating sdBs or sdOs [@pulsating], eclipsing or reflecting hot sd binaries [@For2010; @Derekas15], and hot sds as central stars of planetary nebulae [@Alba15]. All of them are particularly interesting for studying the stellar interiors, the mass transfer mechanism at work, as well as the evolutionary formation channels, and would contribute to better understand these evolved objects.
Hot subdwarfs were first found analyzing faint blue stars, starting with the [@Humason47] survey or the Palomar-Green (PG) catalogue [@PGcat]. At present we have at our disposal deeper and more extensive surveys, covering large regions of the sky and wide spectral ranges. Besides, with online access tools like the Virtual Observatory (VO)[^2] we can access data from most of these surveys in a very efficient way, crossmatching information to select objects with particular characteristics.
In this regard, the aim of the work here presented is to obtain a number of new hot sd candidates as large as possible. We apply the selection process developed in [@Oreiro2011] (hereafter paper I), that combines photometric and proper motion information from different surveys, making use of VO tools, with the intention of discriminating hot sds from other types of objects of similar colours, mainly white dwarfs (WD), cataclysmic variables (CV) and main-sequence O and B stars, considered as contamination sources in this work.
Section 2 describes the methodology employed, Section 3 the results obtained together with their analysis, Section 4 a summary of our main achievements on spectral classification of the new hot sd candidates discovered, and Section 5 ends with a general summary and conclusions.
Methodology
===========
The methodology described in paper I is reproduced here. In that work, a hot sd selection procedure was defined and tested by means of a thorough retrieval, with the aid of VO tools, of multi-colour photometry and astrometric information from stellar catalogues. A filtering procedure to distinguish among different types of objects was designed to obtain a hot subdwarf sample with a low contamination factor. The method was tested on two sky regions: the Kepler FoV[^3] and a region of 300 $deg^2$ around ($\alpha$:225, $\delta$:5 deg) obtaining a high rate of success (above $80$ per cent) in finding new uncatalogued hot subdwarfs. Temperatures were provided by fitting their spectral energy distribution (SED), and considering two-atmosphere fits for those objects with a clear infrared excess, a signature of the possible presence of a cool companion.
Once tested the validity of the proposed strategy, in this work we apply it to a wider sky region. The extension of this region is about 11663$deg^2$, limited by the SDSS[^4] DR7 survey coverage [@DR7]. Most of the Galactic northern cap down to b=+30 is covered, as well as some strip-shaped areas crossing the Galactic plane and reaching the southern Galactic hemisphere. Fig. \[dr7\] shows the footprint in Galactic coordinates of the regions studied.
We remind the reader that the principal aim of this work is to apply a reliable selection procedure to find a large number of new hot subdwarf candidates, and not to perform a deep analysis of each star physical parameters. In this regard, we computed effective temperature estimates using an automated procedure provided by VOSA [@VOSA], a very useful online facility of the Spanish Virtual Observatory[^5].
For the ease of the reader, we outline here the selection procedure.
- [**Hot subdwarf selection filters:**]{} First step was to crossmatch the photometric and proper motion surveys and apply to the retained sources the cuts aimed at selecting hot subdwarfs.
The surveys employed were SDSS DR7, GALEX GR6/GR7[^6] [@Galex], 2MASS Point Source Catalogue[^7] [@2MASS] and SuperCosmos[^8] [@SuperCosmos]. We required the GALEX sources to have measured magnitudes in both filters ($FUV>0$, $NUV>0$) and to be brighter than 5$\sigma$ of the magnitude limit ($FUV<19.9$, $NUV<20.8$). Sources must also be classified as point objects by SDSS (cl=6). We retained sources with counterparts in all the surveys within a maximum distance of 5 arcsec. To the selected sources we applied the following cuts, as discussed in paper I: $$\begin{aligned}
&& -4<(FUV_0-Ks_0) <0.5 \label{filter1} \\
&&-2 <(FUV_0 - NUV_0)<0.2 \label{filter2} \\
&& 19 < H(NUV_0) < 27 \label{filter3}\end{aligned}$$ where the 0 subscript stands for Galactic extinction corrected magnitudes, and $H(NUV_0)$ for the reduced proper motion of the $NUV$ filter.
For bright stars, severe calibration problems in the GALEX photometry have been pointed out by [@Camarota2014] who, using a well studied sample of WDs with UV spectra, derived empirical corrections to the GALEX magnitudes in the non-linear range. The corrections are valid within the $9.321 < NUV < 17.5$ or $10.509 < FUV<17.5$ ranges. We have thus identified the stars in our sample lying within those limits, and applied to them the correction factors established in that paper.
- [**Discriminating new from already classified objects:**]{} We crossmatched our list with published and well-established catalogues of spectroscopically confirmed subdwarfs, white dwarfs, cataclysmic variables and OB stars. These include:
- [*The subdwarf database for hot sds*]{} [@Roydatabase].
- [*A selection of hot subluminous stars in the Galex survey*]{} [@Vennes2011; @Nemeth2012]
- [*The photometric and spectroscopic catalogue for luminous stars*]{} [@Reed].
- [*The catalogue of Cataclysmic Variables*]{}, version 2006 [@Downes].
- [*The SDSS DR7 white dwarf catalogue*]{} [@Kleinman].
- [*A Catalogue of Spectroscopically Identified White Dwarfs*]{}, version 2008 [@Cook]
Sources already available in these catalogues were discarded. The remaining objects were searched in SIMBAD[^9], VIZIER.[^10] and any catalogue available through online VO tools. Very recent catalogues like [@Geier15_RVcat], [@Kupfer2015] and [@Gentile2015] were considered in the Vizier search. Other catalogues containing spectroscopically confirmed hot subdwarfs but not included in Vizier [@Vennes2011; @Nemeth2012; @Kawka2015; @Kepler15] were also inspected. Any source already spectroscopically classified in these catalogues was discarded.
- [**Spectral distribution fit:**]{} For each object in the pre-candidate list, we used VOSA to accomplish the following steps:
- Gathering of additional photometry: GALEX-SDSS-2MASS photometry was complemented with additional photometry from UKIDSS[^11] LAS DR9 [@UKIDSS], Tycho-2 [@Tycho] and WISE[^12] [@WISE]. Some candidates had saturated or bad SDSS photometry. In these cases we replaced SDSS by UCAC4 [@UCAC4] photometry, if available.
- Magnitude-to-flux transformation: VOSA used the gathered photometric information to calculate the absolute fluxes and their associated errors taking advantage of the Filter Profile Service[^13] (FPS), a service developed by the Spanish Virtual Observatory to provide VO access and representation of many of the most common photometric systems in astrophysics. Fluxes were then dereddened using the extinction law by [@Fitz99] and the $E(B-V)$ values available in the GALEX catalogue, which in turn have been taken from the [-@Schlegel] extinction maps.
- Model comparison: The flux-dereddened observational SEDs were then compared to the TLUSTY OSTAR2002+BSTAR2006 NTLE models for O and B stars [@Hubeny1995; @Lanz2003; @Lanz2007] implemented at VOSA to derive effective temperatures. We considered the whole model grid, with $T_{\rm eff}$ ranging from 15000 to 55000K. In the SEDs’ fitting procedure both surface gravity and metallicity were simply left as free parameters, as their impact on the effective temperature determination can be considered as negligible. Therefore, we warn the reader that the gravity and metallicity values obtained from the SED fitting cannot be considered as the real physical parameters of the objects listed in Tables \[table:1\] - \[table:4\] below.
[@Heber2000] have shown that the use of LTE vs NLTE model atmospheres yields almost identical $T_{\rm eff}$’s and only systematic $\log g$ differences, at least when fitting hot sd spectral lines. We do not expect other result in our procedure of fitting SEDs.
We have performed, anyway, a comparison between the effective temperatures calculated using the TLUSTY and Kurucz [@Castelli97] grids of atmospheric models. Only objects with a good SED fit flag (5XX) and a TLUSTY temperature value lower than 35000K (to avoid boundary problems with the maximum temperature of the grids) were considered. We obtained a difference in effective temperatures below $10$ per cent for $90$ per cent of the objects (or $82$ per cent of objects for an up to $5$ per cent difference), indicating that, as expected, the NLTE effects on $T_{\rm eff}$ determination can be neglected.
We also attempted to leave $A_v$ as a free parameter in the SED fitting process. Nevertheless, due to the $A_v-T_{\rm eff}$ degeneracy, this exercise rendered multiple solutions and we finally decided to include extinction as a fixed parameter.
- [**Source image checking:**]{} Finally, we visually inspected using Aladin[^14] the SDSS images and catalogue data of our pre-candidate list of targets to discard instrumental features, bad crossmatches or contamination from nearby, bright sources.
In fact, we found some cases with a clear mismatch between GALEX, SDSS and 2MASS sources. These pathological cases are mostly due to the different spectral coverage and limiting magnitude of the surveys. We kept these objects without infrared photometry in a separate list, as they appear to be very hot and blue objects, and thus interesting from our point of view (Table \[table:4\]).
Results {#sec:results}
=======
After crossmatching the photometric surveys and applying the selection filters in equations (\[filter1\])-(\[filter3\]), we ended up with a list of 1242 pre-candidates. 638 of them were already classified in the literature, with the following percentages: $83$ per cent hot subdwarfs, $12$ per cent WD, $2$ per cent CV, $2$ per cent B stars, and less than $0.5$ per cent other main sequence stars. These numbers agree with those obtained in paper I, demonstrating the robustness of our selection procedure. The remaining 604 pre-candidates where not found in the literature.
In Fig. \[galactic\] the classified and unclassified objects selected by the photometric and proper motion cuts are pictured in Galactic coordinates. Notice that a large fraction of the unclassified objects lay in the bands near the Galactic plane, as these tend to be less studied regions.
Effective temperatures {#subsec:fit}
----------------------
Effective temperatures were obtained from the comparison between the observational SEDs and the TLUSTY models. After the fitting, we kept candidates with $T_{\rm eff} > 19000K$, provided that the fit was good. Sources with $T_{\rm eff} < 19000K$ and a bad SED fitting were also kept, as this could be a signal of a binary candidate. 167 out of the 604 unclassified sources did not pass the cut, leaving us with a list of 437 final subdwarf candidates.
The bad fits are of mainly three different sorts: excess in the red part of the spectrum (IR), ultraviolet (UV) excess, and both IR and UV excesses in the same source. IR excesses are probably a signature of a binary system. UV excess could also indicate the presence of a very hot companion, but uncertainties associated to the ultraviolet extinction correction cannot be discarded. $E(B-V)$ values have been taken from [@Schlegel], who seem to overestimate the reddening to lines of sight where $A_V \geq 0.5$ mag [@Arce99].
As a further check, we also performed in VOSA a Bayesian analysis of the model fits. We found that, for 356 sources, the probability associated to the $T_{\rm eff}$ value obtained from the chi-square fitting was over 80 per cent. For the rest of sources (81), we provide an effective temperature interval covering an accumulated probability of, at least, 80 per cent. With this procedure we obtain fairly temperature estimations for the whole sample. The only exception to this were the targets whose temperature estimate reached the upper limit of the TLUSTY models, 55000K. This is not suprising as we know that some sdOs can achieve very high temperatures [@Stro2007]. For these objects, just a lower limit in effective temperatures is provided.
To classify the quality of the fits we tagged each target with a three digits quality flag (see Tables \[table:1\]-\[table:4\]). The first digit ranges from one to five: ‘5’ represents good SED fitting, ‘4’ stands for excess in the red part of the SED (IR excess), ‘3’ for both IR and UV excesses, ‘2’ for UV excess only, and ‘1’ for a bad fitting of any other sort. Excesses in the infrared or ultraviolet part of the SED were defined whenever the relative difference between the model and observed (dereddened) values was above $20$ per cent, and the difference increased with decreasing/increasing wavelength, for UV and IR, respectively. We found this criterion matched quite well with a visual inspection of the SED fits. Differences without a clear pattern, in the middle or any part of the SED, were considered bad fits of type ‘1’.
Second and third digits refer to the quality of the GALEX data: a ‘1’ in the second position represents a problematic GALEX artifact, and a ‘1’ in the third position stands for a bad flag in the photometry extraction[^15]. In both cases, ‘0’ stands for a good GALEX flag. 2MASS quality flags were also considered: Photometric values with an $U$ flag ($U$ standing for [*upper limit*]{} in magnitude) were not taken into account to perform the SED fitting.
Examples of the different quality fits can be seen in Fig. \[vosas\]. Red points (grey in the grey-scale version of the figure) represent the derredened magnitudes of the given object and the conected blue points (dark grey) the synthetic magnitudes that best fit. Under each graph we represent the residuals of the observed data and models.
In Tables \[table:1\]-\[table:4\] we present a sample of the hot subdwarf candidates found by our selection method. Table \[table:1\] includes good fitted objects with photometric data ranging from the ultraviolet to the infrared, and thus represents clear single candidates (192 objects). Table \[table:2\] shows sources with excess in the red part of the SED, the most clear binary candidates (110 objects). In Table \[table:3\] we included the rest of the bad fitted objects (115). Finally, Table \[table:4\] includes hot objects with no infrared photometry available (20).
$T_{\rm eff}$ estimates for candidates with flags 4XX, 3XX, 2XX and 1XX must be treated with caution, which is warned by means of one of these bad fit flags. We remind the reader that, although the estimated $T_{\rm eff}$ are below 19000K, these objects are kept in the candidate list because the combination of bad fit and low temperature is used as indicator for the presence of binary systems.
In all tables, $FUV$ and $NUV$ were taken from the GALEX archive, and corrected as explained above, if necessary; $u$, $g$, $r$ are from SDSS Data Release 7 and $J$, $H$, $K$ from the 2MASS Point Source Catalogue. We included a column with the 2MASS quality flags of the source. $T_{\rm eff}$ is obtained from the best SED fit performed by VOSA. As explained above, an interval in the temperature column is given whenever the Bayes analysis gave the most likely $T_{\rm eff}$ value with a probability below $80$ per cent. The [*fit flag*]{} column shows our notation for the different qualities in the VOSA $T_{\rm eff}$ fit. In Tables \[table:2\] and \[table:3\] we also included the filter where the excess begins and the expected spectral type of the stellar companion, according to the criterium explained in the next section.
Full tables, with all the photometric filters and other data, including links to the SED fitting diagrams and the SDSS spectrum, when available, can be accessed using the SVO hot subdwarf archive (see Appendix \[sec:apen\]).
Fig. \[histogram\] shows a histogram of the effective temperatures obtained for both the good fitted single candidates (fit flag 5XX) and the most clear binary candidates (fit flag 4XX). The majority of stars within the single sample lay in the temperature range 20000-30000K while the effective temperatures of the binary sample are shifted towards lower values. This is not surprising, as we are including in the 4XX category objects with effective temperatures below 19000 K (see above). On the other hand, the peaks at 15000K and 55000K are signaling the limits in $T_{\rm eff}$ of the TLUSTY grid of models.
---------- ----------- -------- -------- -------- -------- -------- -------- -------- -------- ------- --------------- ------
RA DEC NUV FUV u g r J H K 2MASS $T_{\rm eff}$ Fit
(J2000) (J2000) flag (VOSA) flag
00:03:07 +24:12:12 16.028 15.916 16.105 16.148 16.533 16.628 16.085 17.106 BUU 25000 500
00:11:43 -10:40:34 14.044 13.947 14.686 14.987 15.471 15.95 15.601 16.524 ABU 32-35000 501
00:50:40 -09:25:46 12.845 12.553 13.818 14.174 14.716 15.216 15.307 15.402 AAC 55000 500
01:11:56 +15:17:53 14.866 14.499 15.154 15.206 15.627 15.74 15.948 15.615 ACD 25000 501
01:30:32 +52:33:50 16.679 16.402 16.072 16.067 16.18 15.931 15.965 15.998 ACD 37500 500
01:32:33 +51:57:57 15.281 15.309 14.935 14.996 15.167 15.202 15.186 15.291 AAB 37500 500
02:20:35 +17:04:07 14.912 14.615 15.011 14.957 15.274 15.241 15.103 14.992 AAC 24000 500
02:31:45 +22:08:30 16.749 16.597 16.545 16.392 16.646 16.319 16.277 15.914 ADD 23-24000 510
02:34:56 -06:09:13 14.884 14.167 15.604 15.963 16.475 16.888 16.566 16.862 CDU 42-55000 500
02:35:48 +05:35:24 14.725 14.269 15.407 15.518 15.999 16.806 16.327 15.609 CCU 30000 511
---------- ----------- -------- -------- -------- -------- -------- -------- -------- -------- ------- --------------- ------
---------- ----------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ---------- ------ -------- --------
RA DEC NUV FUV u g r J H K 2MASS Teff Fit Excess Binary
(J2000) (J2000) flag (VOSA) flag from class
00:14:40 +08:03:52 17.11 16.87 16.83 16.57 16.48 15.69 15.36 15.21 AAB 15000 410 z GK
01:23:41 +30:02:32 16.02 15.62 16.33 16.35 16.64 16.42 15.84 15.29 BCU 18-21000 410 z GK
01:24:59 +47:56:41 15.98 15.48 16.53 16.88 17.06 15.82 15.23 15.00 AAA 55000 400 i FGK
01:33:14 +48:57:28 13.33 13.15 14.00 14.57 15.04 12.10 12.14 12.16 AAA 18000 410 B F
02:28:23 +25:35:19 14.37 13.35 13.50 14.16 13.12 13.04 13.10 13.09 AAA 15000 400 B F
02:41:13 +21:57:43 14.14 14.28 14.38 12.92 13.06 12.67 12.68 12.72 AAA 15000 400 B F
02:44:14 +30:07:23 15.40 14.79 15.06 14.67 14.60 13.89 13.63 13.63 AAA 16000 400 i FGK
02:57:48 +37:15:35 15.42 14.93 16.07 16.41 16.87 16.74 16.28 15.71 CDU 55000 400 J GK
03:18:23 +41:55:22 14.36 13.99 14.49 14.63 14.92 14.69 14.52 14.34 AAA 40000 410 J GK
03:48:30 +16:39:46 17.89 17.58 17.41 17.22 17.07 16.15 15.57 15.79 ABD 21000 410 i FGK
---------- ----------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ---------- ------ -------- --------
---------- ----------- -------- -------- -------- -------- -------- -------- -------- -------- ------- --------------- ------
RA DEC NUV FUV u g r J H K 2MASS $T_{\rm eff}$ Fit
(J2000) (J2000) flag (VOSA) flag
02:12:44 +68:07:08 19.267 18.996 19.312 19.065 18.758 15.598 14.956 15.017 ABC 15000 100
02:25:44 +72:49:44 18.642 17.982 16.902 16.125 15.575 14.055 13.661 13.361 AAA 29-37500 200
02:27:18 +73:36:11 17.864 17.382 16.32 15.454 14.907 13.483 12.976 12.969 AAA 19000 200
02:51:46 +75:09:04 19.360 19.049 18.504 18.023 17.711 16.597 15.767 15.942 BUU 15000 200
03:33:56 +17:56:36 17.699 17.156 17.975 18.011 18.26 16.703 15.656 16.914 BUU 18000 300
03:53:07 +16:48:49 15.646 15.113 15.036 14.675 14.598 13.577 13.11 13.046 AAA 15000 300
04:07:24 +14:44:06 18.995 18.562 17.535 16.905 16.51 15.156 14.639 14.566 AAA 15000 300
04:31:18 +55:53:08 18.696 18.598 17.365 17.054 16.889 15.899 15.531 15.277 ABB 37500 200
04:35:42 +56:24:15 19.770 19.514 18.022 17.523 17.105 15.931 15.644 15.444 ABC 18000 200
04:38:22 +19:03:06 17.079 16.814 16.619 16.34 16.283 15.546 15.14 15.003 AAB 18000 310
04:41:41 -06:11:29 15.389 15.132 15.735 15.54 15.614 15.061 14.833 14.867 AAB 15000 300
---------- ----------- -------- -------- -------- -------- -------- -------- -------- -------- ------- --------------- ------
------------- ------------- --------- --------- -------- -------- ---------- -------- ------
RA DEC NUV FUV u g r Teff Fit
(J2000) (J2000) (VOSA) flag
02:51:03.80 +75:15:03.4 16.7211 16.3976 17.243 17.194 17.564 42500 200
04:55:29.84 +24:45:07.6 18.8049 18.0702 18.501 18.631 18.853 42500 201
04:59:12.49 +60:51:56.5 19.0112 18.7774 19.017 18.631 18.929 37500 200
04:59:39.94 +59:48:53.5 17.5742 17.1230 17.219 17.25 17.56 42500 100
05:02:51.04 +13:49:26.9 20.4623 19.3643 19.512 19.522 19.552 42500 200
05:11:25.05 +15:03:01.0 18.9370 18.3932 17.986 18.012 18.026 40000 500
06:11:51.35 +34:04:01.5 19.3621 19.2375 19.924 20.152 20.544 42500 200
06:25:53.27 +34:54:28.2 16.8423 16.4591 17.304 17.298 17.666 42500 210
06:40:51.07 +26:44:28.0 12.9711 13.1221 13.524 11.024 11.064 15000 100
08:00:03.16 +07:40:43.3 14.6031 13.9893 15.507 15.907 16.432 55000 501
08:06:08.31 +10:24:20.1 16.4314 16.1312 17.044 17.195 17.682 28000 500
08:13:32.86 +05:54:30.1 14.8425 14.1695 15.608 16.093 16.639 55000 500
08:23:15.22 +00:18:46.0 15.6761 15.0236 16.502 17.019 17.385 8 55000 500
08:28:16.33 +22:32:26.5 14.3126 13.6464 13.754 12.389 12.777 15000 400
16:07:41.25 +25:42:20.6 16.4320 16.1035 16.825 16.971 17.423 28000 500
17:37:03.25 +50:40:41.1 15.4195 15.2508 15.982 16.041 16.539 27000 500
20:14:55.06 +08:42:13.9 18.8628 18.5533 19.143 19.03 19.424 30000 500
20:46:23.13 -06:59:26.8 16.4524 15.9148 17.081 17.455 17.89 55000 500
21:08:04.46 +05:15:28.5 16.2921 15.8421 16.777 17.042 17.462 42500 500
23:28:59.74 +52:16:24.1 18.4366 18.0371 18.742 18.888 19.255 42500 200
------------- ------------- --------- --------- -------- -------- ---------- -------- ------
Binary sample {#subsec:bin}
-------------
An important issue regarding hot subdwarfs is to know the binary fraction of these objects, as some of the proposed formation channels involve evolution in binary systems [@Han2002; @Han2003; @Clausen12].
The three main binary evolution channels, as proposed in these papers, are the common envelope (CE) ejection channel, the stable Roche lobe overflow (RLOF) channel and the double helium white dwarfs (WDs) merger channel. The CE ejection channel leads to the formation of subdwarfs in short-period binaries with typical orbital periods between 0.1 and 10 days and very thin hydrogen-rich envelopes. On the other hand, the stable RLOF channel produces stars with long orbital periods (400 to 1500 days) and with rather thick hydrogen-rich envelopes. The merger channel gives rise to single subdwarfs whose hydrogen-rich envelopes are extremely thin. This channel is believed to explain the formation of helium-rich hot subdwarfs [@Zhang2012].
Nevertheless, the contribution of the binary channels to the formation of the different subtypes of hot subdwarf stars is still unclear, with new discoveries challenging the standard binary evolution scenarios (see @Geier13b for a recent review).
Regarding sdBs, the binary fraction is estimated around $40$ per cent or higher, depending on the nature of the samples and the method used to detect the stellar companion (see @Nap2004, or @Heber_rev for a review). In [@Lisker2005] the estimated binary fraction is $32$ per cent, although this value should be taken as a lower limit as target selection was biased against composite spectrum objects. A recent study on IR excess in known radial velocity (RV) variable sdB binaries can be found in [@Kupfer2015].
The sdO binary fraction is more controversial, with different studies reaching opposite conclusions: [@Nap2004] found only one out of 23 RV variable sdO, while [@Green2008] and [@MUCHFUSS] found a similar distribution of RV variations between sdBs and sdOs. On the other hand, while [@Ulla98] found that 6 out of 14 (43 per cent) new hot sdOs with IR excesses, [@Stro2007] found 8 out of 52 (18 per cent) sdOs with photometric infrared excesses, although again this value should be taken as a lower limit.
A combination of optical and infrared photometry is commonly used to find late-type companions such as F, G or K types, because the hot subdwarf will shine in the blue, while the companion will have brighter red colours. [@Stark2003] found a $40$ per cent of binary systems in a magnitude-limited sample of hot subdwarfs from the Kilkenny catalogue [@Kil_cat], using 2MASS infrared filters and Johnson or Strömgren optical photometry. In [@Girven2012] they also combine GALEX ultraviolet photometry with optical and infrared filters to select subdwarf candidates in double systems. They complete the photometric study with a spectroscopic classification, finding a large fraction of composite systems with F, G or K companions. On the other hand, detection of radial velocity variations in sdB stars is used to find close binaries with invisible companions such as white dwarfs (see for instance @Morales and @Copperwheat).
The photometric data of our sample ranges from the ultraviolet to the far infrared (WISE colours), which enabled us to detect flux excesses from the $B$ magnitude. The companions found with our photometric methodology are expected to be late-type main sequence stars, such as F, G or K.
In our list of 437 hot subdwarf candidates, 20 of them have no infrared photometric data available (see Table \[table:4\]). For the other 417, we consider as [*binary candidates*]{} those with quality SED fit 4 or 3 (IR or both IR and UV excess, Tables \[table:2\] and \[table:3\]). There are 189 of them, making a total fraction of $45$ per cent. This fraction could be overestimated, as some of the flux excesses may be apparent, due to inaccurate photometric measures or bad SED fittings.
The possible spectral type of the companion star was estimated analyzing the excess of hot sds with main sequence companions catalogued in [@Roydatabase]. We have selected all subdwarfs classified as sds+F, sds+G or sds+K of this catalogue and identified from which band the excess is detected in VOSA. Our criterion for the companion spectral type was the following:
- Excess from $B$,$V$ or $g$ band: type F (17 objects)
- Excess from $r$ band: types F, G (6 objects)
- Excess from $i$ band: types F, G, K (86 objects)
- Excess from $z$ or $J$ band: types G, K (69 objects)
- Excess from $H$, $Ks$ or $W1$ band: type K (11 objects)
Hot sds can have companions of other spectral types. Close binaries formed by cool main sequence M-type or substellar objects may alter the measured photometric values. The variation is due to a reflection effect in the light curve caused by the irradiated surface of the much cooler companion. It is estimated that only 1/5 of short period sdBs contain a dM [@Roy2013]. In fact, few reflecting sdBs+dM/BD systems are known, and they show typical peak-to-peak photometric amplitudes of $\sim 0.2$ mag or less, diminishing towards blue wavelengths. This effect is not expected to significantly alter our procedure.
Less common are hot subdwarfs in eclipsing binaries, either with dM/BD or WD companions (see [@For2010]). In these cases, deep eclipses with $\sim < 0.8$ mag variations occur (although see the extraordinary sdO+dM system in [@Derekas15]). A raw estimate of the number of eclipsing sdB+dM systems in our sample leads to a negligible number of photometric disturbances caused by photometry being acquired on eclipse-phase. Of course, we can not rule out that chances of this effect occur, which could explain any of the bad fits encountered in this work.
A more rigorous check of all possible binary candidates, using a two-component fit, is presently being addressed, and will be presented in a future work.
Colour-colour diagrams
----------------------
Colour-colour or colour-magnitude diagrams allow us to separate sources of different nature using photometric colours. This is commonly used to detect ultraviolet or infrared excesses signaling a probable binary nature of the objects under study.
We have first plotted our candidates in the colour-colour plane $u_0-FUV_0$ versus $V_0-u_0$, as seen in Fig. \[our\_colors\]. The $0$ subscript stands for derredened magnitudes, with the $E(B-V)$ values taken from [@Schlegel], and the corresponding correction factors computed from the formulae in [-@Cardelli]. $FUV$ and $u$ are GALEX and SDSS filters, respectively. As we do not have data of the $V$ magnitude for all our candidates, we use a $V$ value computed with the transformation formula given in [@Jester2005] between the $ugriz$ and $UBV$ systems: $$V=g-0.58 (g-r)-0.01
\label{vsinc}$$
The figure shows the relative difference in the blue colours of our candidates. Most of the good fitted objects lay on the right side of the diagram, as would be expected, because they stand for relatively hot single candidates. The bluer stars lay on the bottom right corner of the diagram, where we can see objects labeled with ultraviolet excess (quality fit 200). To discard the UV excess was only an effect of a possible overestimation in the extinction correction, we checked the $A_V$ values of these objects. They turned up to be relatively high ($A_V>1.14$ mag), but not higher than many other objects without UV excess (single or binary candidates).
There are various references in the recent literature studying different samples of hot subdwarfs and proposing alternative ways to separate single from binary stars using colour-colour diagrams. In [@Stark2003] they plot $J-K_s$ versus $V-K_s$ and found single stars remain inside a box limited by $V-K_s \leq +0.2$ and $J-K_s \leq +0.05$, while the composite stars lay outside this rectangle (see fig. 1 in that paper). In [@Green2008] they plot $V-J$ versus $J-H$ of a sample of confirmed subdwarfs, finding a gap in the colour diagram, separating single from composite stars (fig.5 of that paper). Finally, in [@Girven2012] $FUV-r$ is plotted versus $r-K_s$, showing also separate regions for the single and composite subdwarfs.
Similarly to [@Green2008], to show the single or binary nature of our candidates we plotted $V_0-J_0$ versus $J_0-H_0$, the $V$ magnitude computed as in equation (\[vsinc\]) and $J$ and $H$ from 2MASS. The result is shown in Fig. \[green\].
Not all the candidates are represented there. Many of our sources are quite faint in the infrared, and thus do not have very good 2MASS flags: 173 sources ($41$ per cent) have at least an $U$ ([*upper limit*]{}) in one of the 2MASS filters; while other 46 sources ($11$ per cent) have at least one filter with a $D$ or $E$ flag[^16]. To avoid big errors, we restricted our plot to sources with $err(J)+err(H)<0.3$. For the rest of the objects, we substituted 2MASS data with UKIDSS, if available.
In Fig. \[green\], we can see that single candidates tend to concentrate around $V_0-J_0 \simeq -0.6$, and binary candidates, although more spread, around $V_0-J_0 \simeq 0.25$. In spite of the source spreading, the gap between composite and single systems is still visible. The source spreading may be due to a variety of factors. We first noted that the majority of binary candidates spread around the diagram are those where the excess begins at $B$ or $r$ magnitudes, causing the optical region of the SED to be untypical. Other uncertainties in the SDSS photometry, due to either faint magnitudes or bright values near the saturation limits, could cause the star to be misplaced. Finally, we can not discard other phenomena, like the presence of planetary nebulae.
Spectral classification
=======================
A first classification scheme for hot subdwarfs was proposed in [@PGcat], where an eight class system was defined. In [@Drilling13] an evolved MK-like system is developed. In this system hot subdwarfs are divided in four sequences: [*He-weak*]{}, [*He-normal*]{}, [*He-strong*]{} and [*He-strong C*]{}, with carbon or other metallic lines. The differences in helium content of each object are also quantified measuring ratios of hydrogen to helium line depths. In each of the four helium sequences, the spectral subclasses would range from sdO1 to sdB9, as in the MK system, line depths and ratios varying smoothly within each sequence (see figs. 1 to 4 in that paper). Stars former classified within the somehow [*ad-hoc*]{} sdOB subclass [@Moehler1990] are placed naturally in this scheme, in the transition between late sdO and early sdB subclasses. In the present work we will appply the [@Drilling13] system to classify the candidates with spectrum.
Only 68 stars of our list of subdwarf candidates ($16$ per cent) had SDSS spectrum. We begun with a visual inspection of each object’s whole spectrum. One spectrum was too noisy to allow identification. The rest were identified as one white dwarf, one probable cataclysmic variable and 65 subdwarfs: 5 sdOs ($8$ per cent), 25 sdOBs ($38$ per cent) and 35 proper sdBs ($54$ per cent). Note the effectiveness of our selection procedure improves to $95.6$ per cent within this subset.
We also inspected the presence of characteristic lines of cooler main sequence stars, to identify binary candidates. In particular, we looked for the Mg[i]{} triplet (5172, 5183, 5167Å), the G band (4300Å), and the Ca[ii]{} K line (3933Å). The Na[i]{} doublet (5889-5895Å) can also be an indicator, although it may be overlapped with a near He[ii]{} line. We found 23 probable binary systems: 22 binary sdBs (including sdOBs) and 1 binary sdOs. The binary fraction for sdBs obtained by visual inspection of the spectrum was of $37$ per cent, somehow lower than the photometric fraction. Besides the possible overestimation of the photometric fraction, as argued above, the relatively low signal-to-noise ratio of some SDSS spectra may be obscuring the binary nature of some candidates, hiding the cold companion lines. We point out that all the spectrum-detected binary systems are also binary candidates from the photometric excess point of view.
To classify the subdwarfs within the [@Drilling13] system we cut and normalized the spectra between 4000 and 5000 $\AA$ and then compared them with the standard stars defining the system. The complete classification of our candidates can be found in the electronic tables. A sample is shown in Figures \[normal\]-\[weak\]. Sumarizing our results, we have found:
- 1 star ($1.5$ per cent) belonging to the [*He-strong C*]{} sequence,
- 2 stars ($3.1$ per cent) in the [*He-strong* ]{} sequence,
- 48 stars ($73.9$ per cent) in the [*He-normal*]{} sequence and
- 14 stars ($21.5$ per cent) in the [*He-weak*]{} sequence.
Note that the signal-to-noise ratio of some spectra might be masking weak metallic carbon or nitrogen lines, affecting the [*He-strong C*]{} and [*He-strong*]{} relative abundances.
Calibrations made in [@Drilling13] demonstrate that [*He-weak*]{} stars have subsolar helium abundances, [*He-normal*]{} stars more nearly solar abundances and [*He-strong*]{} objects high helium abundances. The problem of helium abundances in hot subdwarfs has been addressed in [@Toole2008] and [@Geier13]. In [@Edelmann2003] a correlation between effective temperature and helium abundance in sdB stars was discovered, showing two sequences with approximately the same trend with increasing $T_{\rm eff}$ (fig. 5 in that paper). The ratio of objects in the lower sequence to those in the upper one is argued to be between 1:10 and 4:10 in [@Toole2008]. Our ratio 14:48 of weak to normal helium stars is consistent with these margins.
[@Drilling13] suggest their spectral sequences are temperature sequences, and find linear trends plotting the effective temperature against the spectral class (fig. 10 and 11 in that paper). For a check, we plotted our good SED fitted candidates in the $T_{\rm eff}$-spectral class plane, together with [@Drilling13] linear regressions. As a result, and although our temperatures appear in general subestimated, we found our candidates from classes sdO9 to sdB3 to roughly follow Drilling et al.’s trend lines. On the contrary, the hotter sdO3-sdO7 classes do not seem to follow this behavior. Such a result is not surprising, due to the effective temperature upper limit of the TLUSTY models ($55000K$).
Conclusions
===========
In this work we have extended the selection procedure developed in [@Oreiro2011] to identify hot subdwarfs, taking advantage of Virtual Observatory tools. The selection procedure includes photometric and proper motions filters and an effective temperature cutoff. We have identified 437 new subdwarf candidates from a 11663 $deg^2$ sky region, limited by the SDSS DR7 image coverage. We expect an effectiveness of at least $80$ per cent, although the subsample of objects with SDSS spectra reached a subdwarf identification above $95$ per cent, proving the accuracy of our selection filters.
From our spectral energy distribution analysis, we have estimated a photometric binary fraction of $45$ per cent, while identification of cool star metallic lines in the spectra yields a $37$ per cent of binaries among this subsample. Both numbers, although rough estimates, are in agreement with previous studies. Our method mainly selects binaries with late-type main sequence stellar companions, like F, G or K.
The colour-colour diagrams of Fig. \[our\_colors\] and \[green\] show the difference in, respectively, the blue and red colours of our subdwarf candidates. Objects in the lower right corner of Fig. \[our\_colors\] show a clear ultraviolet excess (labeled with SED fitting flag 200 – see text for details). In Fig. \[green\] we see a clear clustering of the single candidates around $V_0-J_0 \simeq -0.6$, and the colour gap between single and binary subdwarfs discovered in [@Green2008] is reproduced.
We have also performed a detailed spectral classification of the 68 candidates with SDSS spectra, following a recent MK-like system for hot subdwarfs developed by [@Drilling13]. We found our candidates perfectly suit in one of the four helium sequences proposed in that paper.
Much work remains to be done after the selection procedure developed here. Regarding the list of hot subdwarf candidates, a deeper analysis of the binary sample is presently under study. A two-body SED fitting will yield temperature values of the cold companion, which will aid to estimate its spectral class and its distance. As argued in [@Clausen12], composite systems of the type sdB + early F could be crucial in determining the binary formation channel of hot subdwarfs, depending on the periods measured for these systems. Recently, in [@Barlow13] and [@Vos13], long period sdB+F/G systems have been reported. It would thus be interesting to know if our sample contains this type of binaries, and to make a follow-up study of them, determining their orbital parameters.
Both the twenty candidates without available infrared photometry and the seventeen objects with apparent ultraviolet excess are very interesting from our point of view, as good candidates for very hot objects. The hotter sdOs are measured to have up to $100000 K$ effective temperatures [@Stro2007]. These objects are scarce between subdwarfs, and some of them have been proved to have planetary nebulae [@Alba13; @Alba15], a signal of a probable post AGB origin [@Heber91]. Photometric and spectroscopic accurate data would be needed to reach further conclusions about the origin of these stars.
Finally, a detailed line spectral analysis of the hot subdwarf candidates, to be performed using advanced/accurate NLTE atmospheric models, would yield more reliable values for the star effective temperatures, helium abundances and surface gravities. The position of our candidates in the $T_{\rm eff}-\log g$ plane would contribute to discriminate between the different origins and evolution paths proposed for hot subdwarfs.
Regarding the search of new hot subdwarf candidates, different approaches could be used. One of those would imply applying our selection procedure to new releases of some of the surveys already considered (e.g. SDSS, GALEX) or using new catalogues, both in the optical (e.g. Pan-Starrs, or J-PAS in the near future) and in the infrared (UKIDSS, VISTA).
Another possibility would be employing other catalogues containing astrometric information. At this moment, our routine discards any source without proper motion data in the SuperCosmos survey. This fact could be modified making the routine look for proper motion information in other catalogues, like UCAC4 or PPMXL [@PPMXL].
It would also be interesting to explore the possibilities of the data that GAIA[^17] would provide regarding this point. GAIA would measure distances with great accuracy, giving us information of star luminosities and their position in the HR diagram. This would yield estimations of both star masses and ages, an important information to track the evolutionary paths of hot subdwarfs.
Attention must be paid as well to other forthcoming missions such as CHEOPS [@Fortier14] or PLATO [@PLATO], as they are expected to have an impact on ultra-high precision photometry and stellar astroseismology for brigth targets, covering large fractions of the sky (up to 50 per cent in the case of PLATO), and widening, then, the possible detection of new pulsating hot sds.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank A. Aller for helping us improving the spectra classification. This publication makes use of VOSA, developed under the Spanish Virtual Observatory project supported from the Spanish MICINN through grant AyA2011-24052. It has also been partially supported by grant INCITE09312191PR (which includes FEDER funds), given by the Xunta de Galicia, and by grant 12VI20, given by the Universidade de Vigo. This research has made use of the SIMBAD database, the Vizier catalogue access tool (originally published in @Vizier) and the [*Aladin sky atlas*]{}, all operated at CDS, Strasbourg, France.
The SVO hot subdwarf archive {#sec:apen}
============================
In order to help the astronomical community on using the catalogue of subdwarfs identified in this paper we have developed an archive system that can be accessed from a Web page[^18] or through a Virtual Observatory ConeSearch[^19]
Web access
----------
The archive system implements a very simple search interface that permits queries by coordinates/radius as well as by other criteria of interest (object identifier, Teff, quality flag or excess). A selection of the astrometric, photometric and physical parameters to be displayed in the table of results can also be done assigning a type of verbosity: minimum, medium or maximum. The default search radius is set to 5 arcsec. The user can also select the maximum number of sources to return (with values ranging from 10 to unlimited) (Fig. \[figapen1\]).
The result of the query is a HTML table with all the sources found in the archive fulfilling the search criteria. Detailed information on the output fields can be obtained placing the mouse over the name of the column. The archive implements the SAMP (Simple Application Messaging Protocol). SAMP allows applications to communicate with each other in a seamless and transparent way for the user. This way, the results of a query can be easily transferred to other VO application, such as Topcat (Fig \[figapen2\]).
Virtual Observatory access
--------------------------
The Virtual Observatory (VO)[^20] is an international initiative whose primary goal is to provide an efficient access and analysis of the information hosted in astronomical archives and services. Having a VO-compliant archive is an important added value for an astronomical project to guarantee the optimum scientific exploitation of their datasets.
Our archive system has been designed following the IVOA standards and requirements. In particular, it implements the Cone Search protocol, a standard defined for retrieving records from a catalogue of astronomical sources. The query made through the Cone Search service describes a sky position and an angular distance, defining a cone on the sky. The response returns a list of astronomical sources from the catalogue whose positions lie within the cone, formatted as a VOTable.
\[figapen1\]
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: <http://www.ivoa.net/>
[^3]: <http://kepler.nasa.gov/science/about/targetFieldOfView/>
[^4]: <http://www.sdss.org/>
[^5]: <http://svo.cab.inta-csic.es>
[^6]: <http://galex.stsci.edu/>
[^7]: <http://www.ipac.caltech.edu/2mass/>
[^8]: <http://surveys.roe.ac.uk/ssa/>
[^9]: <http://simbad.u-strasbg.fr/simbad/>
[^10]: <http://vizier.u-strasbg.fr/viz-bin/VizieR>
[^11]: <http://surveys.roe.ac.uk/wsa/>
[^12]: <http://wise.ssl.berkeley.edu>
[^13]: <http://svo2.cab.inta-csic.es/theory/fps3/>
[^14]: <http://aladin.u-strasbg.fr/>
[^15]: see the GALEX documentation at <http://galex.stsci.edu/GR6/?page=ddfaq#6>
[^16]: <http://www.ipac.caltech.edu/2mass/releases/allsky/doc/sec1_6b.html> gives detailed explanations of their quality flags.
[^17]: <http://sci.esa.int/gaia/>
[^18]: <http://svo2.cab.inta-csic.es/vocats/hsa/>
[^19]: Try for instance <http://svo2.cab.inta-csic.es/vocats/hsa/cs.php?RA=0&DEC=0&SR=100&VERB=2>
[^20]: <http://www.ivoa.net>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Albert Ossó$^1$, Álvaro Corral$^{2}$ & Josep Enric Llebot$^1$'
title: ' Scaling of Tropical-Cyclone Dissipation\'
---
[ **The influence of climate variability and global warming on the occurrence of tropical cyclones (TC) is a controversial issue [@Goldenberg_Science; @Trenberth; @Emanuel_nature05; @Landsea_comment; @Webster_Science; @Chan_comment; @Klotzbach; @Shepherd_Knutson; @Kossin; @Elsner08]. Existing historical databases on the subject are not fully reliable [@Kossin; @Gray_comment; @Landsea_Science06; @Landsea_Eos07], but a more fundamental hindrance is the lack of basic understanding regarding the intrinsic nature of tropical cyclone genesis and evolution [@Emanuel_book]. It is known that tropical cyclones involve more than a passive response to changing external forcing [@Emanuel_bams08], but it is not clear which dynamic behaviour best describes them. Here we present a new approach based on the application of the power dissipation index ($PDI$), which constitutes an estimation of released energy [@Emanuel_nature05], to individual tropical cyclones. A robust law emerges for the statistics of $PDI$, valid in four different ocean basins and over long time periods. In addition to suggesting a novel description of the physics of tropical cyclones in terms of critical phenomena [@Bak_book; @Turcotte_book; @Christensen_Moloney], the law allows to quantify their response to changing climatic conditions, with an increase in the largest $PDI$ values with sea surface temperature or the presence of El Niño phenomenon, depending on the basin under consideration. In this way, we demonstrate that the recent upswing in North Atlantic hurricane activity does not involve TCs quantitatively different from those in other sustained high-activity periods prior to 1970.** ]{}
One important characterization of a complex phenomenon is given by the fluctuations in the “size” of the phenomenon over successive occurrences [@Bak_book; @Turcotte_book; @Malamud_hazards; @Christensen_Moloney]. We refer to neither spatial size (area, volume) nor something like the Saffir-Simpson category [@Kantha_Eos]; rather, we seek a physically relevant measure of released energy. For a tropical cyclone, a reasonable proxy for this energy has been proposed by Emanuel [@Emanuel_nature05], using the $PDI$, defined as $$PDI \equiv \sum_t v_t^3 \Delta t\ ,$$ where $t$ denotes time and runs over the entire lifetime of the storm and $v_t$ is the maximum sustained surface wind velocity at time $t$. In available best-track records, measurements are provided at intervals of $\Delta t=6$ hours. Note that in this paper the $PDI$ value is associated with an individual tropical cyclone, not with the total annual activity in some ocean basin [@Emanuel_nature05].
We analyze tropical cyclone best-track records for several ocean basins: the North Atlantic and Northeastern Pacific (from the National Hurricane Center [@NOAA]), and the Northwestern Pacific and the Southern Hemisphere (from the Joint Typhoon Warning Center [@ATCR_report]). We exclude the North Indian Ocean, due to the small number of storms in the most reliable portion of its records. \[For more details, see the Supplementary Information, including Fig. S2.\]
We display in Fig. 1(a) the $PDI$ probability density, $D(PDI)$, normalized in the usual way ($\int_0^\infty D(PDI) dPDI =1$), for all four basins. The distributions include all storms occurring during an extended period, either 1966-2007 or 1986-2007 as indicated in the legend. All four distributions (given vertical offsets for clarity) can be characterized by a power-law decay in their central regions, $$D(PDI) \propto 1/PDI ^\alpha,$$ where the exponent $\alpha$ is in between 0.95 and 1.25 (Supplementary Information, including Table S1). Deviations from the power law at small $PDI$ values can be attributed to the deliberate incompleteness of the records for “not significant” TCs, whereas the more rapid decay at large $PDI$ values is associated with the finite size of the basin. That is, the storms with the largest $PDI$ do not have enough room to last a longer time, as their tracks are limited by the size of the basin, which introduces a cutoff in the distribution (at its separation from the power-law fit, roughly) (Supplementary Information, Figs. S3-S7). Variations in the definition of the $PDI$, for example excluding times during which the storm attains subtropical or extratropical status, do not modify the shape of the $PDI$ distribution; the results are also unchanged when restricted to storms that do not make continental landfall (Supplementary Information, Figs. S8 and S9).
The degree of similarity between the basins is truly remarkable, given the variety of formative processes at work. TCs in the Western Pacific (typhoons) develop principally from the monsoon, for example, while North Atlantic hurricanes are mainly associated with easterly waves (and the degree of association depends on the intensity of the hurricane) [@Shepherd_Knutson; @Landsea93]. In addition, each regional centre or agency follows different protocols in obtaining their data, using techniques which have gradually improved [@Kossin; @Gray_comment; @Landsea_Science06; @Landsea_Eos07].
Nevertheless, the shape of the $PDI$ distribution is robust over long time periods. Figure 1(b) shows that this consistency holds over a period of at least 100 years in the North Atlantic (where the record is longest; corrections to the calibration of the maximum velocity do not alter this pattern, Supplementary Information, including Fig. S10). Thus, even though the database is certainly incomplete prior to the satellite era, and even more unreliable before aircraft reconnaissance began in 1944, the fraction of missed storms seems to be independent of $PDI$. This finding may appear counterintuitive until we consider the fact that a long-lasting tropical cyclone might be recorded as two shorter storms if its track is lost at some point. A power-law distribution is robust against such splitting of the data (Supplementary Information).
The existence of a simple statistical distribution that describes the whole spectrum of tropical cyclone sizes in different basins over a long period of time (apart from incompleteness and finite size effects) reflects a startling degree of unity in the phenomenon; the small tropical depressions are described in the same way as the full developed most severe storms. Moreover, a robust power-law distribution is the hallmark of scale invariance [@Christensen_Moloney]: there is no typical tropical cyclone $PDI$, up to the maximum allowed in a given ocean basin. The fact that all scales are equally important for energy dissipation poses a great challenge to the modelling of this complex phenomenon, and even to large-scale global climate simulations [@Emanuel_bams08].
Scale invariance can occur in processes where perturbations propagate through a critical (i.e., highly susceptible) medium [@Bak_book; @Turcotte_book; @Malamud_hazards]. Thus, our result could indicate that the atmosphere, or perhaps the ocean-atmosphere system, is close to a critical state. In fact, this idea is not new; already in the 70’s it was suggested that atmospheric convection takes place in a near-unstable environment [@Arakawa_schubert]. Much more recently, Peters and Neelin have demonstrated the existence of a non-equilibrium stability-instability transition to which the state of the atmosphere is attracted [@Peters_np]. Some properties of this transition can be obtained from static images of convecting cloud fields [@Peters_percolation] or local observations of precipitation [@Peters_prl]. These findings support our complex-system approach to tropical-cyclone evolution; in correspondence, we provide a complementary perspective to the puzzle that these atmospheric processes constitute.
In addition, an important property of critical systems is that perturbations can evolve while keeping a delicate balance between growth and attenuation, resulting in sudden intensifications and deintensifications. Although recent years have seen considerable improvement in the prediction of tropical cyclone trajectories, reliable forecasts of their intensities have not yet been achieved [@Willoughby_eye; @verification]. This failure may not be just due to technical limitations; it may be a fundamental feature of the criticality of tropical cyclone evolution.
Tropical cyclone activity shows large interannual variability. One important factor controlling such variability is sea surface temperature ($SST$). We average $SST$ from the Hadley Center [@SST] over the same spatial areas and months in the TC season than Webster [*et al.*]{} [@Webster_Science] in order to get an annual (so, seasonal) $SST$ value for each basin; then, we separate the $PDI$ density into two contributions, one for years with $SST$ above its long-term mean value $\langle SST \rangle$ (i.e., high $SST$) and another one for $SST$ below $\langle SST \rangle$ (low $SST$). Mathematically, $\langle SST \rangle \equiv \sum_y SST(y) /Y$, where $SST(y)$ refers to year $y$ and $Y$ is the total number of years. In this way we eliminate the effect of interannual variations in the number of TCs and concentrate on a comparison of the individual tropical cyclones characterizing each type of year.
Remarkably, for the North Atlantic and the Northeastern Pacific, the resulting distributions have essentially the same shape as the distribution for all years but with a difference in scale: high-$SST$ years are characterized by a larger value of the finite-size cutoff, and conversely for low-$SST$ years, as can be seen in Fig. 2(a). The other two basins show much minor $PDI$ variation with $SST$.
If we rescale each conditional distribution by a power of its mean value, $\langle PDI \rangle$, such that $PDI \rightarrow PDI/\langle PDI \rangle^\nu$ and $D(PDI) \rightarrow \langle PDI \rangle^\beta D(PDI)$, with $\nu=\beta=1$ for $\alpha \le 1$, and $\nu=1/(2-\alpha)$ and $\beta=\alpha/(2-\alpha)$ for $\alpha > 1$, it becomes apparent that, for each basin, both distributions share a similar shape, as shown in Fig. 2(b). (The reason of this rescaling is the fact that a distribution with $\alpha > 1$ does not scale linearly with its mean value, see Supplementary Information.) Then, the difference between high-$SST$ and low-$SST$ $PDI$ distributions rests mainly in the scale of the finite-size cutoff and not in the shape of the distribution. Years with high $SST$ are thus characterized by hurricanes with larger $PDI$ values. As the $PDI$ integrates the cube of the velocity over the storm lifetime, larger $PDI$ values can result from longer lifetimes, larger (6-hour) velocities, or both. An analysis of the distributions of these variables shows that the increase in $PDI$ comes mainly from an increase in the velocities, most apparent in the range above about 100 knots (i.e., corresponding to category 3 hurricanes and beyond [@Webster_Science]), in comparison with years of low $SST$ (Supplementary Information, Fig. S11).
An analogous study can be done as a function of the so-called $MEI$ index [@MEI2], which quantifies the strength of El Niño phenomenon. Taking annual values of $MEI$, years with $MEI > 0$ (corresponding to El Niño) lead to increased $PDI$ values in the Northeastern and Northwestern Pacific, but keeping again the same shape of the $PDI$ distribution, and the opposite for $MEI < 0$, see Figs. 2(a) and 2(b). In the case of the Northwestern Pacific, the contribution of TC durations to the increase in $PDI$ is larger than in the rest of the cases (Supplementary Information). This is somehow related to the findings of Refs. [@Lander; @Emanuel_07]. Very little variation with $MEI$ is observed in the other two basins. Other indices ($AMO$, $NAO$, etc.) do not seem to influence the $PDI$ distribution in any basin.
It is a well-known fact that individual years of high (or low) TC activity cluster into longer periods of predominantly high (or low) activity. For example, the North Atlantic has seen extraordinarily high activity between 1995 and 2005, which has been linked to global warming through an increase in the sea surface temperature [@Emanuel_nature05; @Webster_Science]. The issue is nonetheless controversial [@Landsea_comment; @Chan_comment; @Gray_comment]. According to our analysis, the $PDI$ distribution for the period 1995-2005 is indistinguishable from the distribution for years with high $SST$ between 1966 and 2007, as well as from the distribution for other periods of high activity, like 1926-1970 [@Goldenberg_Science], see Fig. \[highact\]. We conclude that the recent dramatic increase of activity does not lead to unprecedented energy releases by individual TCs (although even higher $SST$ could further increase the $PDI$ values).
[33]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix
Goldenberg, S. B., Landsea, C. W., Mestas-Nu[ñ]{}ez, A. M. & Gray, W. M. The Recent Increase in [Atlantic]{} Hurricane Activity: Causes and Implications. *Science* **293**, 474–479 (2001).
Trenberth, K. Uncertainty in Hurricanes and Global Warming. *Science* **308**, 1753–1754 (2005).
Emanuel, K. Increasing destructiveness of tropical cyclones over the past 30 years. *Nature* **436**, 686–688 (2005).
Landsea, C. W. Hurricanes and global warming. *Nature* **438**, E11–E12 (2005).
Webster, P. J., Holland, G. J., Curry, J. A. & Chang, H.-R. Changes in Tropical Cyclone Number, Duration, and Intensity in a Warming Environment. *Science* **309**, 1844–1846 (2005).
Chan, J. C. L. Comment on “[Changes]{} in Tropical Cyclone Number, Duration, and Intensity in a Warming Environment”. *Science* **311**, 1713b (2006).
Klotzbach, P. J. Trends in global tropical cyclone activity over the past twenty years (1986–-2005). *Geophys. Res. Lett.* **33**, L10805 (2006).
Shepherd, J. M. & Knutson, T. The Current Debate on the Linkage Between Global Warming and Hurricanes. *Geography Compass* **1**, 1–24 (2007).
Kossin, J. P., Knapp, K. R., Vimont, D. J., Murnane, R. J. & Harper, B. A. A globally consistent reanalysis of hurricane variability and trends. *Geophys. Res. Lett.* **34**, L04815 (2007).
Elsner, J. B., Kossin, J. P. & Jagger, T. H. The increasing intensity of the strongest tropical cyclones. *Nature* **455**, 92–95 (2008).
Gray, W. M. Comments on “[Increasing]{} destructiveness of tropical cyclones over the past 30 years”. *http://arxiv.org* **0601050** (2006).
Landsea, C. W., Harper, B. A., Hoarau, K. & Knaff, J. A. Can We Detect Trends in Extreme Tropical Cyclones? *Science* **313**, 452–454 (2006).
Landsea, C. W. Counting [Atlantic]{} Tropical Cyclones Back to 1900. *Eos* **88 (18)**, 197–202 (2007).
Emanuel, K. *Divine wind: the history and science of hurricanes* (Oxford University Press, New York, 2005).
Emanuel, K. The Hurricane-Climate Connection. *Bull. Am. Met. Soc.* **(5)**, ES10–ES20 (2008).
Bak, P. *How Nature Works: The Science of Self-Organized Criticality* (Copernicus, New York, 1996).
Turcotte, D. L. *Fractals and Chaos in Geology and Geophysics* (Cambridge University Press, Cambridge, 1997), 2nd edn.
Christensen, K. & Moloney, N. R. *Complexity and Criticality* (Imperial College Press, London, 2005).
Malamud, B. D. Tails of natural hazards. *Phys. World* **17 (8)**, 31–35 (2004).
Kantha, L. Time to Replace the [Saffir-Simpson]{} Hurricane Scale? *Eos* **87 (1)**, 3–6 (2006).
Jarvinen, B. R., Neumann, C. J. & David, M. A. S. A tropical cyclone data tape for the [North Atlantic]{} basin, 1886-1983: contents, limitations, and uses. *http://www.nhc.noaa.gov/pdf/NWS-NHC-1988-22.pdf* (1988).
Chu, J.-H., Sampson, C. R., Levine, A. S. & Fukada, E. The [Joint Typhoon Warning Center]{} Tropical Cyclone Best-Tracks, 1945-2000. *https://metocph.nmci.navy.mil/jtwc/best$\_$tracks/TC$\_$bt$\_$report.html* (2002).
Landsea, C. W. A climatology of intense (or major) hurricanes. *Mon. Weather Rev.* **121**, 1703–1713 (1993).
Arakawa, A. & Schubert, W. H. Interaction of a cumulus cloud ensemble with the large-scale environment, part [I]{}. *J. Atmos. Sci.* **31**, 674–701 (1974).
Peters, O. & Neelin, J. D. Critical phenomena in atmospheric precipitation. *Nature Phys.* **2**, 393–396 (2006).
Peters, O., Neelin, J. D. & Nesbitt, S. W. Mesoscale convective systems and critical clusters. *J. Atmos. Sci.* (2009).
Peters, O., Hertlein, C. & Christensen, K. A Complexity View of Rainfall. *Phys. Rev. Lett.* **88**, 018701 (2002).
Willoughby, H. E. Forecasting Hurricane Intensity and Impacts. *Science* **315**, 1232–1233 (2007).
National Hurricane Center Forecast Verification. *http://www.nhc.noaa.gov/verification* (2008).
Rayner, N. A. *et al.* Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. *J. Geophys. Res.* **108 (D14)**, 4407 (2003).
Wolter, K. & Timlin, M. S. Measuring the strength of [ENSO]{} events: how does 1997/98 rank? *Weather* **53(9)**, 315–324 (1998).
Lander, M. A. An Exploratory Analysis of the relationship between Tropical Storm Formation in the [Western North Pacific]{} and [ENSO]{}. *Mon. Weather Rev.* **122**, 636–651 (1994).
Emanuel, K. Environmental Factors Affecting Tropical Cyclone Power Dissipation. *J. Clim.* **20**, 5497–5509 (2007).
[**Supplementary Information**]{}\
Available.
[**Acknowledgements**]{}\
We have benefited from the expertise and kindness of A. Deluca, K. Emanuel, E. Fukada, A. González, J. Kossin, B. Mathiesen, M. Paczuski, O. Peters and A. Turiel. A.O. and A.C. were put in contact through G. Orriols. The initial part of our research has been financed by the EXPLORA - INGENIO 2010 program, and also partially by grants by other grants of the Spanish MEC and Generalitat de Catalunya.
[**Author Information**]{}\
Correspondence and requests for materials should be addressed to A.C. (ACorral at crm dot cat).
[**Competing financial interests**]{}\
The authors declare no competing financial interests.
{width="15cm"} 1(a)
{width="15cm"} 1(b)
{width="15cm"} 2(a)
{width="15cm"} 2(b)
{width="15cm"} 3
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Building on the rare pion and muon decay results of the PIBETA experiment, the PEN collaboration has undertaken a precise measurement of $B_{\pi\text{e}2}\equiv R^\pi_{\text{e}/\mu}$, the $\pi^+ \to
\text{e}^+\nu(\gamma)$ decay branching ratio, at the Paul Scherrer Institute, to reduce the present 40$\times$ experimental precision lag behind theory to $\sim 6-7\times$. Because of large helicity suppression, $R^\pi_{\text{e}/\mu}$ is uniquely sensitive to contributions from non-$(V-A)$ physics, making this decay a particularly suitable subject of study. Even at current precision, the experimental value of $B_{\pi\text{e}2}$ provides the most accurate test of lepton universality available. During runs in 2008–10, PEN has accumulated over $2\times 10^7$ [$\pi_{\text{e}2}$]{} events; a comprehensive maximum-likelihood analysis is currently under way. The new data will also lead to improved precision of the earlier PIBETA results on radiative $\pi$ and $\mu$ decays.
author:
- 'D. Počanić'
- 'A. Palladino'
- 'L. P. Alonzi'
- 'V. A. Baranov'
- 'W. Bertl'
- 'M. Bychkov'
- 'Yu.M. Bystritsky'
- 'E. Frlež'
- 'V.A. Kalinnikov'
- 'N.V. Khomutov'
- 'A.S. Korenchenko'
- 'S.M. Korenchenko'
- 'M. Korolija'
- 'T. Kozlowski'
- 'N.P. Kravchuk'
- 'N.A. Kuchinsky'
- 'M.C. Lehman'
- 'D. Mekterović'
- 'E. Munyangabe'
- 'D. Mzhavia'
- 'P. Robmann'
- 'A.M. Rozhdestvensky'
- 'S.N. Shkarovskiy'
- 'U. Straumann'
- 'I. Supek'
- 'P. Truöl'
- 'Z. Tsamalaidze'
- 'A. van der Schaaf'
- 'E.P. Velicheva'
- 'V.P. Volnykh'
title: New studies of allowed pion and muon decays
---
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{},altaddress=[Paul Scherrer Institut, Villigen PSI, CH-5232, Switzerland]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Paul Scherrer Institut, Villigen PSI, CH-5232, Switzerland]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Institut Rudjer Bošković, HR-10000 Zagreb, Croatia]{}]{}
[address=[NCBJ National Centre for Nuclear Research, Otwock, Poland]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Institut Rudjer Bošković, HR-10000 Zagreb, Croatia]{}]{}
[address=[Institute of Nuclear and Particle Physics, University of Virginia, Charlottesville, VA 22904, USA]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{},altaddress=[Institute for High Energy Physics, Tbilisi State University, GUS-380086 Tbilisi, Georgia]{}]{}
[address=[Physik-Institut, Universität Zürich, CH-8057 Zürich, Switzerland]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Physik-Institut, Universität Zürich, CH-8057 Zürich, Switzerland]{}]{}
[address=[Institut Rudjer Bošković, HR-10000 Zagreb, Croatia]{}]{}
[address=[Physik-Institut, Universität Zürich, CH-8057 Zürich, Switzerland]{}]{}
[address=[Institute for High Energy Physics, Tbilisi State University, GUS-380086 Tbilisi, Georgia]{}]{}
[address=[Physik-Institut, Universität Zürich, CH-8057 Zürich, Switzerland]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
[address=[Joint Institute for Nuclear Research, RU-141980 Dubna, Russia]{}]{}
Historically, the [$\pi$$\to$$\text{e}$$\nu$]{} (or [$\pi_{\text{e}2}$]{}) decay, provided an early confirmation of the $V-A$ nature of the electroweak interaction. Thanks to exceptionally well controlled theoretical uncertainties, its branching ratio is now understood at better than a part in $10^4$. The most recent independent theoretical calculations are in excellent agreement, and give: $$B_{\pi\text{e}2}^{\text{SM}} \equiv R_{\text{e}/\mu}^{\pi,\text{SM}} =
\frac{\Gamma(\pi \to \text{e}\bar{\nu}(\gamma))}
{\Gamma(\pi \to \mu\bar{\nu}(\gamma))}\bigg|_{\text{calc}} =
\begin{cases}
1.2352(5) \times 10^{-4} & \text{Ref.~\cite{Mar93},} \\
1.2354(2) \times 10^{-4} & \text{Ref.~\cite{Fin96},} \\
1.2352(1) \times 10^{-4} & \text{Ref.~\cite{Cir07},} \\
\end{cases}$$ where “$(\gamma)$” indicates that radiative decays are included. Marciano and Sirlin [@Mar93] and Finkemeier [@Fin96] took into account radiative corrections, higher order electroweak leading logarithms, short-distance QCD corrections, and structure-dependent effects, while Cirigliano and Rosell [@Cir07] used two-loop chiral perturbation theory. A number of exotic processes outside of the current standard model (SM) can produce deviations from the above predictions based on lepton universality, mainly through induced pseudoscalar (PS) currents. Prime examples are: charged Higgs in theories with multiple Higgs bosons, PS leptoquarks in theories with dynamical symmetry breaking, classes of vector leptoquarks, parameters of certain SUSY partner particles, as well as non-zero neutrino masses and their mixing (Refs. [@Rai08; @Bry11] give recent reviews of the subject). Thus, [$\pi_{\text{e}2}$]{} decay complements direct searches for new physics at modern colliders.
The two most recent measurements of the branching ratio [@Bri92; @Cza93] are mutually consistent and dominate the world average of $1.230(4) \times 10^{-4}$, which, however, trails the theoretical accuracy by a factor of 40. The PEN experiment [@pen06] is aiming to reach ($\Delta B/B)_{\pi\text{e}2} \simeq
5\times 10^{-4}$, and in doing so, to set new limits on the above non-SM processes. PEN also aims to improve the PIBETA results for radiative decays $\pi^+\to \text{e}^+\nu\gamma$ and $\mu^+\to
\text{e}^+\nu\bar{\nu}\gamma$. Meanwhile, PiENu [@PiENu], a complementary experiment currently under way at TRIUMF, has a similar goal for $(\Delta B/B)_{\pi\text{e}2}$.
The PEN experiment uses an upgraded version of the PIBETA detector system, described in detail in Ref. [@Frl04a], and used in a series of rare pion and muon decay measurements [@Poc04; @Frl04b; @Byc09]. The PEN apparatus, shown in Fig. \[fig:pen\_det\](a), consists of a large-acceptance ($\sim 3\pi$sr) electromagnetic shower calorimeter (pure CsI, 12 radiation lengths thick) with non-magnetic tracking in concentric cylindrical multi-wire proportional chambers (MWPC1,2) and plastic scintillator hodoscope (PH), surrounding a plastic scintillator active target (AT). Beam pions pass through an upstream detector (BC), lose energy in the active degrader (AD), are tracked in a mini time projection chamber (mTPC), and stop in the AT. Signals from the beam detectors are digitized in waveform digitizers, running at 2GS/s for BC, AD, and AT, and at 250MS/s for mTPC.
\
The [$\pi_{\text{e}2}$]{} branching ratio will be evaluated by normalizing the observed yield of [$\pi$$\to$$\text{e}$$\nu$]{} decays to the number of sequential decays [$\pi$$\to$$\mu$$\to$$\text{e}$]{}, within a 250ns gate starting some 40ns before the pion stop time [@pen06]. Assignment of detected events to either of the two processes, or to a background process, is made within a comprehensive, blind and unbinned maximum likelihood analysis (MLA) [@Pal12]. Key to achieving the goal uncertainty is in the control of the systematics.
One among many powerful tools available to develop reliable probability density functions for the MLA is provided by the beam counter waveform (wf) digitizer data, helping us to obtain clean samples of [$\pi_{\text{e}2}$]{} events (two pulses in the AT-wf) and sequential [$\pi$$\to$$\mu$$\to$$\text{e}$]{} decay events (three pulses in the AT-wf). The procedure and results are summarily illustrated in Fig. \[fig:pen\_det\](b,c). Another key component in the analysis is the comprehensive GEANT4 Monte Carlo simulation of the experiment that produces synthetic data fully equivalent to the measured data [@Alo12].
Furthermore, thanks to lower and better controlled backgrounds, the PEN data will enable us to improve on the precision of the previous PIBETA results on pion [@Byc09] and muon [@Van06; @Mun12] radiative decays, both of which are sensitive to non-$(V-A)$ interactions. Besides providing the most stringent limit on tensor interactions to date [@Bha12], the PIBETA $\pi\to\text{e}\nu\gamma$ results [@Byc09] also furnish fundamental inputs for chiral perturbation theory related to pion structure. On the other hand, muon radiative decay, being largely free from hadronic corrections, provides model-independent information on non-$(V-A)$ admixtures in the lagrangian.
Three PEN data runs have been completed, in 2008, 2009, and 2010, respectively, collecting over 20M [$\pi_{\text{e}2}$]{} events. Comprehensive data analysis, focusing on control of systematics, is currently under way.
This work has been supported by grants from the US National Science Foundation (most recently PHY-0970013), the Paul Scherrer Institute, and the Russian Foundation for Basic Research (Grant 08-02-00652a).
[00]{} W.J. Marciano and A. Sirlin, Phys. Rev. Lett. [**71**]{} (1993) 3629.
M. Finkemeier, Phys. Lett. B [**387**]{} (1996) 391.
V. Cirigliano and I. Rosell, Phys. Rev. Lett. (2007) 231801.
M. Raidal, A. van der Schaaf, I. Bigi, et al., Eur. Phys. J C [**57**]{} (2008) 13.
D. Bryman, W.J. Marciano, R. Tschirhart, T. Yamanaka, Ann. Rev. Nucl. Part. Sci. **61** (2011) 331.
D.I. Britton, S. Ahmad, D.A. Bryman, et al., Phys. Rev. Lett. [**68**]{} (1992) 3000.
G. Czapek, A. Federspiel, A. Flükiger, et. al., Phys. Rev. Lett. [**70**]{} (1993) 17.
“A precise measurement of the $\pi^+$$\to$$e^+\nu$ branching ratio,” PSI experiment proposal R-05-01, (2006).
PiENu: <http://pienu.triumf.ca>, and A. Sher, elsewhere in these Proceedings.
E. Frlež, D. Počanić, K. Assamagan, et al., Nucl. Inst. and Meth. A [**526**]{} (2004) 300.
D. Počanić, E. Frlež, V.A. Baranov, et al., Phys. Rev. Lett. [**93**]{} (2004) 181803.
E. Frlež, D. Počanić, V.A. Baranov, et al., Phys. Rev. Lett. [**93**]{} (2004) 181804.
M. Bychkov, D. Počanić, B.A. VanDevender, et al., Phys. Rev. Lett. **103** (2009) 051802.
A. Palladino, PhD thesis, University of Virginia (2012).
L.P. Alonzi, PhD thesis, University of Virginia (2012).
B.A. VanDevender, PhD thesis, University of Virginia (2006).
E. Munyangabe, PhD thesis, University of Virginia (2012), and article in preparation.
T. Bhattacharya, V. Cirigliano, S.D. Cohen, et al., Phys. Rev. D **85** (2012) 054512.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the nonparametric estimation of the jump density of a renewal reward process from one discretely observed sample path over $[0,T]$. We consider the regime when the sampling rate $\Delta=\Delta_T\rightarrow0$ as $T\rightarrow\infty$. The main difficulty is that a renewal reward process is not a Lévy process: the increments are non stationary and dependent. We propose an adaptive wavelet threshold density estimator and study its performance for the $L_p$ loss, $p\geq 1$, over Besov spaces. We achieve minimax rates of convergence for sampling rates $\Delta_T$ that vanish with $T$ at polynomial rate. In the same spirit as Buchmann and Grübel (2003) and Duval (2012), the estimation procedure is based on the inversion of the compounding operator. The inverse has no closed form expression and is approached with a fixed point technique.'
author:
- 'Céline Duval[^1]'
title: Nonparametric estimation of a renewal reward process from discrete data
---
**AMS 2000 subject classifications:** 62G99, 62M99, 60G50.\
**Keywords:** Renewal reward process, Continuous time random walk, Compound Poisson process, Discretely observed random process, Wavelet density estimation.
Introduction
============
Motivation and statistical setting
----------------------------------
Renewal reward processes are pure jump processes used in many application fields, for instance in seismology (see Alvarez [@Alvarez] or Helmstetter *et al.* [@Helmstetter]), to model rainfall (see Rodriguez-Iturbe *et al.* [@Rod]) or in mathematical insurance and finance (see for instance Scalas *et al.* [@scalas05; @Scalas] or Masolivier *et al.* [@Masoliver]). If many papers are devoted to the estimation of a discretely observed Lévy process (see for instance Bec and Lacour [@Lacour], Comte and Genon-Catalot [@Comte09; @Comte11], Figueroa-López [@Lopez] and Duval [@Du2] for the high frequency case and Neumann and Reiß[@Reiss] and Comte and Genon-Catalot [@Comte10] for the low frequency one), to the knowledge of the author, little exists on the estimation of a discretely observed renewal reward process. Vardi [@Vardi] estimates the density of a renewal process without rewards from the continuous observation of several independent trajectories. In this paper we estimate the compound law of a renewal reward process when one trajectory is observed at a sampling rate that goes to 0 arbitrarily slowly.
Let $J_1,\dots,J_i$ be nonnegative independent random variables where $J_2,\dots,J_i$ are identically distributed. Define $T_i$ the time of the $i$th jump as $T_i=J_1+...+J_i,$ $i\geq1.$ The associated counting process or renewal process $R$ is $$R_t=\sum_{i=1}^\infty\mathds{1}_{T_i\leq t},\ \ \ t\geq 0.$$ The Poisson process is a particular case of a renewal process, corresponding to exponentially distributed interarrivals $\big(J_i\big)$. That latter case excepted, $R$ does not have independent increments and is usually not stationary *i.e.* for all positive $t,h$ the law of $R_{t+h}-R_t$ depends on $t$. Assume that the common distribution $\tau$ of the $\big(J_i\big)$ has finite expectation $$\mu=\int_0^\infty t\tau(dt)<\infty,$$ define the distribution $$\begin{aligned}
\label{eq tau0}
\tau_0(x)&=\frac{1-\int_0^x\tau(dt)}{\mu}.\end{aligned}$$The process $R$ is stationary if and only if $J_1$ has distribution $\tau_0$ (see Lindvall [@Lindvall] p.70). Define the renewal reward process $X$ as $$\begin{aligned}
X_t&=\sum_{i=1}^{R_t}\xi_i,\ \ \ t\geq 0\end{aligned}$$ where the $\big(\xi_i\big)$ are independent and identically distributed random variables, independent of the interarrivals $\big(J_i\big)$. Renewal reward processes also correspond to decoupled continuous time random walks.
Assume that we have discrete observations of the process $X$ over $[0,T]$ at times $i\Delta$ for some $\Delta>0$ $$\begin{aligned}
\label{eq data}
\big(X_\Delta,\ldots,X_{\lfloor T\Delta^{-1}\rfloor\Delta}\big).\end{aligned}$$ We focus on the *microscopic regime* namely $$\Delta=\Delta_T \rightarrow 0 \ \ \ \ \ \mbox{ as }\ T\rightarrow\infty,$$ and work under the following assumption.
\[ass f\] The law of the $\xi_i$ has density $f$ which is absolutely continuous with respect to the Lebesgue measure.\
The law of the $J_i$, $i\geq 2$ has density $\tau$ which is absolutely continuous with respect to the Lebesgue measure and $J_1$ has density $\tau_0$.
The necessity of the last part of Assumption \[ass f\] is discussed in Section \[section discuss\].
We denote by $\mathcal{F}({\mathbb{R}})$ the space of densities with respect to the Lebesgue measure supported by ${\mathbb{R}}$. We investigate the nonparametric estimation of the density $f$ on a compact interval $\mathcal{D}$ of ${\mathbb{R}}$ from the observations . To that end we use wavelet threshold density estimators and study their rate of convergence, uniformly over Besov balls, for the following loss function $$\begin{aligned}
\label{eq loss} \big({\mathbb{E}}\big[\|\widehat {f}-f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p},\end{aligned}$$ where $\widehat{f}$ is an estimator of $f$, $p\geq1$ and $\|.\|_{L_p(\mathcal{D})}$ denotes $L_p$ loss over the compact set $\mathcal{D}$. We do not assume the interarrival distribution $\tau$ to be known: it is a nuisance parameter.
We estimate $f$ from the increments of $X$, which are dependent. By Assumption \[ass f\], on the event $\{X_{i\Delta}-X_{(i-1)\Delta}=0\}$ no jump occurred between $(i-1)\Delta$ and $i\Delta$ so that the increment $X_{i\Delta}-X_{(i-1)\Delta}$ gives no information on $f$. In the microscopic regime $\Delta=\Delta_T\rightarrow 0$ many increments are zero, therefore to estimate $f$ we focus on the nonzero increments. We denote by $N_T$ their number over $[0,T]$. In that statistical context different difficulties arise; the number of data $N_T$ used for the estimation is random, the increments are dependent, but more importantly on the event $\{X_{i\Delta}-X_{(i-1)\Delta}\ne0\}$, the density of $X_{i\Delta}-X_{(i-1)\Delta}$ is not $f$. Indeed even if $\Delta$ is small there is always a positive probability that more than one jump occurred between $(i-1)\Delta$ and $i\Delta$. Conditional on $\{X_{i\Delta}-X_{(i-1)\Delta}\ne0\}$, the law of $X_{i\Delta}-X_{(i-1)\Delta}$ has density given by (see Proposition \[PropDefOperator\] below) $$\begin{aligned}
\label{eq operator 1}\mathbf{P}_\Delta[f](x)=\sum_{m=1}^\infty{\mathbb{P}}\big(R_\Delta=m\big|R_\Delta\ne0\big)f^{\star m}(x), \ \ \ \ \mbox{ for }x\in{\mathbb{R}},\end{aligned}$$ where $\star $ is the convolution product and $f^{\star m}=f\star\ldots\star f$, $m$ times. Hereafter Lemma \[lem pmControl\] gives for $\Delta$ small enough $$\begin{aligned}
\label{eq lem1}1-2\tau(0)\Delta\leq {\mathbb{P}}\big(R_\Delta=m\big|R_\Delta\ne0\big)\leq 1.\end{aligned}$$ We deduce from the decomposition $$\mathbf{P}_\Delta[f]=f+r(\Delta),$$ where $r(\Delta)$ is a deterministic remainder of the order of $\Delta$. We will see in Theorem \[thm Renewal 1\] that if $\Delta=\Delta_T$ goes to 0 fast enough, namely $T\Delta_T^2=O(1)$ (up to logarithmic factor in $T$) $r(\Delta)$ is negligible and it is possible to estimate $f$ with optimal rates by ignoring the remainder $r(\Delta)$. Otherwise, when there exists $0<\delta< 1$ such that $T\Delta_T^2=O(T^\delta)$ (up to logarithmic factors in $T$) the remainder $r(\Delta)$ is no longer negligible. The condition $\delta<1$ ensures that $\Delta_T$ goes to 0 as $T$ tends to infinity. In the sequel we distinguish two different regimes that will be treated separately.
- *Fast microscopic rates* when –up to logarithmic factors in $T$– $$T\Delta_T^2=O(1).$$
- *Slow microscopic rates* when there exists $0<\delta<1$ such that –up to logarithmic factors in $T$– $$T\Delta_T^2=O(T^\delta).$$
Since all the results of the paper are given up to logarithmic factors in $T$, fast and slow microscopic rates cover all vanishing behaviours for $\Delta=\Delta_T$. We try to answer the following question: Is it possible to construct an adaptive wavelet estimator of $f$ in fast and slow microscopic regimes which is optimal? Papers which estimate nonparametrically the Lévy measure from a discretely observed Lévy process attain optimal rate estimators only for fast microscopic rates (see for instance Bec and Lacour [@Lacour], Comte and Genon-Catalot [@Comte09; @Comte10; @Comte11] and Figueroa-López [@Lopez]).
Our Results\[section our res\]
------------------------------
In Section \[section fast\] we estimate $f$ in the fast microscopic regime, the estimation procedure is based on the approximation $$f\approx\mathbf{P}_\Delta[f].$$ We construct an adaptive wavelet threshold density estimator from the observations . It achieves the minimax rate of convergence which is $T^{-\alpha(s,p,\pi)}$ if $f$ is of regularity $s$ measured with the $L_\pi$ norm, $\pi>0$, and where $\alpha(s,\pi,p)\leq 1/2$ (see hereafter). That procedure does not depend on the interarrival density $\tau$ apart from Assumption \[ass f\]. Moreover the estimator does not explicitly depend on the random quantity $N_T$, the number of nonzero increment.
In Section \[section slow\] we estimate $f$ in the slow microscopic regime, the estimation procedure is the analogue of the one used in Duval [@Du2]. The starting point is that $$f=\mathbf{P}_\Delta^{-1}\big[\mathbf{P}_\Delta[f]\big],$$ and we proceed in two steps to estimate $f$. The first step is the computation of the inverse of the operator $\mathbf{P}_\Delta$ defined in . That step can be referred as decoumpounding as introduced in Buchmann and Grübel [@Buchmann] or van Es *et al.* [@van; @es]. That inverse cannot be explicitly calculated, contrary to [@Du2], but can be approached using a fixed point method. Indeed $f$ is a fixed point of the operator $$\begin{aligned}
\mathbf{H}_{\Delta,f}:h\rightarrow \mathbf{P}_\Delta[f]+h-\mathbf{P}_\Delta[h]\end{aligned}$$ which is a contraction if $h$ and $f$ verifies suitable smoothness properties (see Proposition \[prop complet contract\] below). The Banach fixed point theorem guarantees that for $K$ in ${\mathbb{N}}$ and $p\geq 1$, $$\big\|\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]-f\big\|$$ is small in a sense that we precise later. Next we observe that the Taylor expansion of order $K$ in $\Delta$ of $\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]$ takes the form $$\begin{aligned}
\label{eq approx intro}\sum_{m=1}^{K+1} l_m(\Delta)\mathbf{P}_\Delta[f]^{\star m},\end{aligned}$$ where the $\big(l_m(\Delta)\big)$ depend on the unknown interarrival density $\tau$ (see Proposition \[PropDefOperator\] below). If $\tau$ is described by an unknown parameter $\vartheta\in {\mathbb{R}}$ then $l_m(\Delta)=l_m(\Delta,\vartheta)$ is estimated by plugging an estimator of $\vartheta$.
The second step consists in estimating the densities $\mathbf{P}_\Delta[f]^{\star m}$, for $m=1,\ldots,K+1$. For that we focus on the $N_T$ nonzero increments which have density $\mathbf{P}_\Delta[f]$. The difficulty here is that we have $N_T$ dependent observations where $N_T$ is a random sum of dependent variables. The dependency of the increments is treated using that at each renewal times the renewal process forgets its past. To cope with the randomness of $N_T$, we prove that $N_T/T$ concentrates for $T$ large enough around a deterministic limit using Bernstein type inequalities for dependent data (see Lemma \[lem Nconcentre Renewal\] in Section \[Section proof\] and Dedecker *et. al.* [@Doukhan]). In Theorem \[thm Renewal 2\] we show that wavelet threshold estimators of $\mathbf{P}_\Delta[f]^{\star m}$ attain a rate of convergence –up to logarithmic factors– in $T^{-\alpha(s,\pi,p)}$. We inject those estimators into and obtain an estimator of $f$ that we call *estimator corrected at order $K$.*
The study of the rate of convergence of the estimator corrected at order $K$ requires to control two distinct error terms. A deterministic one due the first step which is the error made when approximating $f$ by . And a statistical one due to the replacement of the ${\mathbf{P}_\Delta}[f]^{\star m}$ by estimators in the second step. The deterministic error decreases when $K$ increases. We choose $K$ sufficiently large for the deterministic error term to be negligible in front of the statistical one. We give in Theorem \[thm Renewal 2\] an upper bound for the rate of convergence of the estimator corrected at order $K$ which is in –up to logarithmic factors– $$\max\{T^{-\alpha(s,\pi,p)},\Delta_T^{K+1}\}.$$ Since $\alpha(s,\pi,p)\leq 1/2$ if there exists $K_0$ such that $$\begin{aligned}
T\Delta_T^{2K_0+2}\leq 1,\end{aligned}$$ the estimator corrected at order $K_0$ attains the optimal rate.
There is a slight difference of methodology between fast and slow microscopic rates to estimate $f$; for fast rates we estimate $f$ using all the increments but in slow rates we focus on nonzero ones. In that latter case, building an estimator using all the increments, even zero ones, achieving the rates of Theorem \[thm Renewal 2\] is possible but numerically unstable. And a technical constraint in the proof of the concentration of $N_T/T$ prevented us from having a unified procedure for fast and slow microscopic rates.
The paper is organised as follows. In Section \[section fast\] we give an adaptive minimax estimator of $f$ in the fast microscopic regime. In that Section we also define wavelet functions and Besov spaces that are used for the estimation and describe the law of the increments. Those results are also used in Section \[section slow\] where we give an adaptive minimax estimator of $f$ in the slow microscopic regime. In both Sections \[section fast\] and \[section slow\] we give upper bounds for the rate of convergence of the estimator of $f$ for the $L_p$ loss defined in , $p\geq1$, uniformly over Besov balls. In Section \[section num ex\], a numerical example illustrates the behavior of the estimators of $f$ introduced in Sections \[section fast\] and \[section slow\]. Finally Section \[Section proof\] is dedicated to the proofs.
Estimation of $f$ in the fast microscopic regime\[section fast\]
================================================================
Preliminary on Besov spaces and wavelet thresholding
----------------------------------------------------
For the estimation, we use wavelet threshold density estimators and study their performance uniformly over Besov balls. In this paragraph we reproduce some classical results on Besov spaces, wavelet bases and wavelet threshold estimators (see Cohen [@Cohen], Donoho *et al.* [@Donoho96] or Kerkyacharian and Picard [@KP00]) that we use in the next sections.
### Wavelets and Besov spaces {#wavelets-and-besov-spaces .unnumbered}
We describe the smoothness of a function with Besov spaces on $\mathcal{D}$. We recall here some well documented results on Besov spaces and their connection to wavelet bases (see Cohen [@Cohen], Donoho *et al.* [@Donoho96] or Kerkyacharian and Picard [@KP00]). Let $\big(\psi_{\lambda}\big)_\lambda$ be a regular wavelet basis adapted to the domain $\mathcal{D}$. The multi-index $\lambda$ concatenates the spatial index and the resolution level $j=|\lambda|$. Set $\Lambda_j:=\{\lambda,|\lambda|=j\}$ and $\Lambda=\cup_{j\geq -1}\Lambda_j$, for $f$ in $L_p({\mathbb{R}})$ we have $$\begin{aligned}
\label{eq fdecomp1}
f&=\sum_{j\geq-1}\sum_{\lambda\in\Lambda_j}\langle f,\psi_\lambda\rangle\psi_\lambda,\end{aligned}$$ where $j=-1$ incorporates the low frequency part of the decomposition and $\langle .,\rangle$ denotes the usual $L_2$ inner product. For $s>0$ and $\pi \in (0,\infty]$ a function $f$ belongs to the Besov space $\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})$ if the norm $$\begin{aligned}
\label{eq besov norm}\|f\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}:= \|f\|_{L_\pi(\mathcal{D})}+\|f^{(n)}\|_{L_\pi(\mathcal{D})}+\Big\|\frac{w_{\pi}^2(f^{(n)},t)}{t^a}\Big\|_{L_\infty(\mathcal{D})}\end{aligned}$$ is finite, where $s=n+a$, $n\in{\mathbb{N}}$ and $a\in (0,1]$, $w$ is the modulus of continuity defined by $$w_{\pi}^2(f,t)=\underset{|f|\leq t}{\sup}\big\|\mathbf{D} ^h\mathbf{D} ^h[f]\big\|_{L_\pi(\mathcal{D})}$$ and $\mathbf{D} ^h [f](x)=f(x-h)-f(x)$. Equivalently we can define Besov space in term of wavelet coefficients (see Härdle *et. al.* [@KerkPicTsyb] p. 123), $f$ belongs to the Besov space $\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})$ if the quantity $$\begin{aligned}
&\underset{j\geq-1}{\sup}2^{j(s+1/2-1/\pi)}\Big(\sum_{\lambda\in\Lambda_j}|\langle f,\psi_\lambda\rangle|^\pi\Big)^{1/\pi}\nonumber\end{aligned}$$ is finite, with usual modifications if $\pi=\infty$.
We need additional properties on the wavelet basis $\big(\psi_{\lambda}\big)_\lambda$, which are listed in the following assumption.
\[Ass\] For $p\geq1$,
- We have for some $\mathfrak{C}\geq 1$ $$\mathfrak{C}^{-1}2^{|\lambda|(p/2-1)}\leq \|\psi_\lambda\|_{L_p(\mathcal{D})}^p\leq \mathfrak{C} 2^{|\lambda|(p/2-1)}.$$
- For some $\mathfrak{C}>0$, $\sigma >0$ and for all $s\leq\sigma$, $J\geq0$, we have $$\begin{aligned}
\label{eq ass1} \big\|f-\sum_{j\leq J}\sum_{\lambda\in\Lambda_j} \langle f,\psi_\lambda\rangle\psi_\lambda\big\|_{L_p(\mathcal{D})}\leq \mathfrak{C}2^{-Js}\|f\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}.\end{aligned}$$
- If $p\geq 1$, for some $\mathfrak{C}\geq 1$ and for any sequence of coefficients $\big(u_\lambda\big)_{\lambda\in\Lambda}$, $$\begin{aligned}
\label{eq ass2}\mathfrak{C}^{-1}\Big\|\sum_{\lambda\in\Lambda}u_\lambda\psi_\lambda\Big\|_{L_p(\mathcal{D})}\leq \Big\|\Big( \sum_{\lambda\in\Lambda}|u_\lambda\psi_\lambda|^2\Big)^{1/2}\Big\|_{L_p(\mathcal{D})} \leq\mathfrak{C}\Big\|\sum_{\lambda\in\Lambda}u_\lambda\psi_\lambda\Big\|_{L_p(\mathcal{D})}.\end{aligned}$$
- For any subset $\Lambda_0\subset\Lambda$ and for some $\mathfrak{C}\geq 1$ $$\begin{aligned}
\label{eq ass3} \mathfrak{C}^{-1} \sum_{\lambda\in\Lambda_0}\|\psi_\lambda\|_{L_p(\mathcal{D})}^p\leq \int_{\mathcal{D}}\Big(\sum_{\lambda\in\Lambda_0}|\psi_\lambda(x)|^2\Big)^{p/2}\leq \mathfrak{C}\sum_{\lambda\in\Lambda_0}\|\psi_\lambda\|_{L_p(\mathcal{D})}^p .\end{aligned}$$
Property ensures that definition of Besov spaces matches the definition in terms of linear approximation. Property ensures that $\big(\psi_{\lambda}\big)_\lambda$ is an unconditional basis of $L_p$ and is a super-concentration inequality (see Kerkyacharian and Picard [@KP00] p. 304 and p. 306).
### Wavelet threshold estimator {#wavelet-threshold-estimator .unnumbered}
Let $(\phi,\psi)$ be a pair of scaling function and mother wavelet that generate a basis $\big(\psi_{\lambda}\big)_\lambda$ satisfying Assumption \[Ass\] for some $\sigma>0$. We rewrite $$\begin{aligned}
f&=\sum_{k\in \Lambda_0}\alpha_{0k}\phi_{0k}+\sum_{j\geq 1}\sum_{k\in\Lambda_j}\beta_{jk}\psi_{jk},\end{aligned}$$ where $\phi_{0k}(\bullet)=\phi(\bullet-k)$ and $\psi_{jk}(\bullet)=2^{j/2}\psi(2^j\bullet-k)$ and $$\begin{aligned}
\alpha_{0k}&=\int \phi_{0k}(x)f(x)dx\\ \beta_{jk}&=\int \psi_{jk}(x)f(x)dx.\end{aligned}$$ For every $j\geq 0$, the set $\Lambda_j$ has cardinality $2^j$ and incorporates boundary terms that we choose not to distinguish in the notation for simplicity. An estimator of a function $f$ is obtained when replacing the $(\alpha_{0k})$ and $(\beta_{jk})$ by estimated values. In the sequel we uses $(\gamma_{jk})$ to design either $(\alpha_{0k})$ or $(\beta_{jk})$ and $(g_{jk})$ for the wavelet functions $(\phi_{0k})$ or $(\psi_{jk})$.
We consider classical hard threshold estimators of the form $$\begin{aligned}
\widehat{f}(\bullet)&= \sum_{k\in \Lambda_0}\widehat{\alpha_{0k}}\phi_{0k}(\bullet)+\sum_{j= 1}^J\sum_{k\in\Lambda_j}\widehat{\beta_{jk}}\mathds{1}_{\big\{|\widehat{\beta_{jk}}|\geq\eta\big\}}\psi_{jk}(\bullet),\end{aligned}$$ where $\widehat{\alpha_{0k}}$ and $\widehat{\beta_{jk}}$ are estimators of $\alpha_{0k}$ and $\beta_{jk}$, $J$ and $\eta$ are respectively the resolution level and the threshold, possibly depending on the data. Thus to construct $\widehat{f}$ we have to specify estimators $(\widehat{\gamma_{jk}})$ of the $(\gamma_{jk})$ and the coefficients $J$ and $\eta$.
Construction of the estimator
-----------------------------
Assume that we have $\lfloor T\Delta^{-1}\rfloor$ discrete data at times $i\Delta$ for some $\Delta >0$ of the process $X$ $$\big(X_\Delta,\ldots,X_{\lfloor T\Delta^{-1}\rfloor\Delta}\big).$$ Introduce the increments $$\begin{aligned}
\mathbf{D}^\Delta X_i=X_{i\Delta}-X_{(i-1)\Delta},\ \ \ \mbox{ for } i=1,\dots,\lfloor T\Delta^{-1}\rfloor,\end{aligned}$$ where $X_0=0$. By Assumption \[ass f\], they are identically distributed but not independent.
\[PropDefOperator\] The distribution of the increment $\mathbf{D}^\Delta X_{1}$ is $$\big(1-p(\Delta)\big)\delta_0+p(\Delta)\mathbf{P}_\Delta[f]$$ where $\delta_0$ is the dirac delta function, $p(\Delta)={\mathbb{P}}(R_\Delta\ne 0)$ and $$\begin{aligned}
\label{DefOperator}\mathbf{P}_\Delta[f]&=\sum_{m=1}^\infty p_m(\Delta) f^{\star m},\end{aligned}$$ where $\star $ is the convolution product, $f^{\star m}$ is $f$ convoluted $m$ times and $$p_m(\Delta)={\mathbb{P}}\big(R_\Delta=m|R_\Delta\ne0\big).$$
It is straightforward to verify that the operator $\mathbf{P}_\Delta$ is a mapping from $\mathcal{F}({\mathbb{R}})$ to itself. The following Lemma gives a polynomial control of the coefficients $\big(p_m(\Delta)\big)$. It is widely used in Sections \[section fast\] and \[section slow\] and does not depend on the rate at which $\Delta_T$ decays to 0.
\[lem pmControl\] Assume $\tau(0)>0$ and let $\Delta_0$ be such that $$\int_0^{\Delta_0}\tau(t)dt\leq\frac{1}{2}\ \ \ \mbox{ and } \ \ \ \underset{t\in[0,\Delta_0]}{\sup}\tau(t)\leq 2\tau(0).$$ For all $\Delta\leq\Delta_0$ we have $$1- 2\tau(0)\Delta\leq p_1(\Delta)\leq1,$$ and for $m\geq2$ $$0\leq p_m(\Delta)\leq 2 \frac{\big(2\tau(0)\big)^{m-1}}{m!}\Delta^{m-1},$$ where the $\big(p_m(\Delta))$ are defined in Proposition \[PropDefOperator\].
The assumption $\tau(0)>0$ in Lemma \[lem pmControl\] ensures that the given inequalities are sharp. In the Poisson case it is always true since $\tau(0)$ is the positive intensity. In the renewal case we may have $\tau(0)=0$, if so two cases must be distinguished. The first one is when $\tau$ as infinitely many derivatives null at 0; it is the case if $\tau$ is bounded away from 0. Then straightforward computations give for any $K$ in ${\mathbb{N}}$: $p_1(\Delta)=1+O\big(\Delta^K\big),$ thus the procedure of Section \[section fast\] enables to achieve optimal rates even in slow microscopic regimes. It is not the purpose of this paper. The second case is $\tau(0)=0$ but there exists $l_0$ in ${\mathbb{N}}$ such that $\tau^{(l_0)}(0)>0$, then Lemma \[lem pmControl\] can be adapted replacing $\tau(0)$ by $\tau^{(l_0)}(0)$ and $\Delta$ by $\Delta^{l_0}$. In the sequel we assume that $\tau(0)>0$ and leave to the reader the changes to be made when $\tau(0)=0$.
In this Section we consider the regimes for which $\Delta=\Delta_T$ is such that $T\Delta_T^2=O(1),$ up to logarithmic factors in $T$. To estimate $f$, we use the approximation $\mathbf{P}_{\Delta_T}[f]\approx f.$ It is equivalent to consider that nonzero increments are realisations of $f$. We construct wavelet threshold density estimators of ${\mathbf{P}_\Delta}[f]$ from the observations $$\big(\mathbf{D}^\Delta X_i,i=1,\ldots,\lfloor T\Delta^{-1}\rfloor\big).$$ Define the wavelet coefficients $$\begin{aligned}
\label{eq est coeff NC}
\widehat{\gamma}_{jk}&=\frac{1}{\big(1-p(\Delta)\big)\lfloor T\Delta^{-1}\rfloor}\sum_{i=1}^{\lfloor T\Delta^{-1}\rfloor}g_{jk} \Big(\mathbf{D}^{\Delta} X_i\Big)\mathds{1}_{\big\{\mathbf{D}^{\Delta} X_i\ne 0\big\}},\end{aligned}$$ where $p(\Delta)$ is defined in Proposition \[PropDefOperator\]. Let $\eta>0$ and $J\in {\mathbb{N}}\setminus\{0\}$, the estimator $\widehat{P_{\Delta}}$ of $\mathbf{P}_\Delta[f]$ is for $x$ in $\mathcal{D}$ $$\begin{aligned}
\label{eq est P NC}
\widehat{P_{\Delta}}(x)&=\sum_{k}\widehat{\alpha}_{0k}\phi_{0k}(x)+\sum_{j= 0}^J\sum_{k}\widehat{\beta}_{jk}\mathds{1}_{\big\{|\widehat{\beta}_{jk}|\geq \eta\big\}}\psi_{jk}(x).\end{aligned}$$
\[def est f NC\]We define $\widehat{f}_{T,\Delta}$ an estimator of $f$ for $x$ in $\mathcal{D}$ as $$\begin{aligned}
\label{eq Est f Threshold NC}
\widehat{f}_{T,\Delta}(x)&=\widehat{P_{\Delta}}(x).\end{aligned}$$
Convergence rates
-----------------
We estimate densities $f$ which verify a smoothness property in term of Besov balls $$\mathcal{F}(s,{\pi},\mathfrak{M})=\big\{ f\in \mathcal{F}({\mathbb{R}}), \|f\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq\mathfrak{M}\big\},$$ where $\mathfrak{M}$ is a positive constant. We are interested in estimating $f$ on the compact interval $\mathcal{D}$, that is why we only impose that its restriction to $\mathcal{D}$ belongs to a Besov ball.
\[thm Renewal 1\] We work under Assumptions \[ass f\] and \[Ass\], let $\Delta_T$ be such that $T\Delta_T^2=O(1)$ up to logarithmic factors in $T$. Let $\pi>0$, $\sigma>s>1/\pi$, $p\geq1\wedge \pi$ and $\widehat{P_{\Delta_T}}$ be the wavelet threshold estimator of $\mathbf{P}_{\Delta_T}[f]$ on $\mathcal{D}$ constructed from $(\phi,\psi)$ and defined in . Take $J$ such that $$2^JT^{-1}\log\big(T^{1/2}\big)\leq 1,$$ and $$\eta=\kappa T^{-1/2}\sqrt{\log\big(T^{1/2}\big)},$$ for some $\kappa>0$. Let $$\begin{aligned}
\label{eq alpha}\alpha(s,p,\pi)=\min\Big\{\frac{s}{2s+1},\frac{s+1/p-1/{\pi}}{2\big(s+1/2-1/{\pi}\big)}\Big\}.\end{aligned}$$ 1) The estimator $\widehat{P_{\Delta_T}}$ verifies for large enough $T$ and sufficiently large $\kappa>0$ $$\begin{aligned}
\underset{\mathbf{P}_{\Delta_T}[f]\in\mathcal{F}(s,{\pi},\mathfrak{M})}{\sup}\big({\mathbb{E}}\big[\big\|\widehat{P_{\Delta_T}} -\mathbf{P}_{\Delta_T}[f] \big\|_{L_p(\mathcal{D})}^p\big] \big)^{1/p}&\leq \mathfrak{C} T^{-\alpha(s,p,\pi)},\end{aligned}$$ up to logarithmic factors in $T$ and where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi,\mu$.\
2)The estimator $\widehat{f}_{T,\Delta_T}$ defined in verifies for $T$ large enough, sufficiently large $\kappa>0$ and any positive constants $\underline{\mathfrak{a}}<\overline{\mathfrak{a}}$ $$\begin{aligned}
\underset{(\mu,\tau(0))\in[\underline{\mathfrak{a}},\overline{\mathfrak{a}}]^2}{\sup}\ \ \underset{f\in\mathcal{F}(s,{\pi},\mathfrak{M})}{\sup}\big({\mathbb{E}}\big[ \| \widehat{f}_{T,\Delta_T}-f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}&\leq \mathfrak{C} T^{-\alpha(s,p,\pi)},\end{aligned}$$ up to logarithmic factors in $T$, where $\mu=\int t\tau(t)dt$ and where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi,\underline{\mathfrak{a}}$ and $\overline{\mathfrak{a}}$.
The proof of Theorem \[thm Renewal 1\] is postponed to Section \[proof thm 1\]. Theorem \[thm Renewal 1\] guarantees that when $\Delta=\Delta_T$ tends rapidly to 0, namely $T\Delta_T^2=O(1)$, the approximation $f\approx \mathbf{P}_{\Delta_T}[f]$ enables to achieve minimax rates of convergence (see Section \[section discuss\]). The estimator does not depend on $\tau$.
Estimation of $f$ in the slow microscopic regime\[section slow\]
================================================================
In this Section we consider the regimes for which there exists $0<\delta< 1$ with $T\Delta_T^2=O(T^\delta),$ up to logarithmic factors in $T$.
Construction of the estimator
-----------------------------
We construct the estimator corrected at order $K$, following the estimation procedure described in Section \[section our res\].
### Construction of the inverse {#construction-of-the-inverse .unnumbered}
Define the space $$\begin{aligned}
\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})=\Big\{h, \|h\|_{L_1(\mathcal{D})}\leq \mathfrak{O}, \|h\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq\mathfrak{N}\Big\},\end{aligned}$$ where $\mathfrak{O}$ is any constant strictly greater than 1 and $\mathfrak{N}$ is a positive constant strictly greater than $\mathfrak{M}$. The space $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ is a subset of $\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})$ which is a Banach space if equipped with the Besov norm .
First we approach the inverse of $\mathbf{P}_\Delta$ with a fixed point method. Consider the mapping $\mathbf{H}_{\Delta,f}$ defined for $h$ in $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ by $$\begin{aligned}
\label{eq Op Point fixe}
\mathbf{H}_{\Delta,f}[h]:= \mathbf{P}_\Delta[f]+h-\mathbf{P}_\Delta[h]. \end{aligned}$$ We immediately verify that $f$ is a fixed point: $\mathbf{H}_{\Delta,f}[f]=f$. The constraints $1<\mathfrak{O}$ and $\mathfrak{M}<\mathfrak{N}$ ensure that if $f$ is in $\mathcal{F}(s,{\pi},\mathfrak{M})$, then $\mathbf{H}_{\Delta,f}[h]$ sends elements of $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ into itself (see Proposition \[prop complet contract\]). The following Proposition guarantee that the definition of the operator matches the assumptions of the Banach fixed point theorem.
\[prop complet contract\] The following properties hold.\
1) Let $\pi\geq 1$, the space $\big(\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N}),\|.\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\big)$ is a closed set of a Banach space and is then complete.\
2) The mapping $\mathbf{H}_{\Delta,f}$ sends elements of $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ into itself and is a contraction. For all $h_1,h_2 \in \mathcal{H}(s,{\pi},\mathfrak{O})$ we have that $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}[h_1]-\mathbf{H}_{\Delta,f}[h_2]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}&\leq \mathfrak{K}(\Delta)\|h_1-h_2\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})},\end{aligned}$$ where $$\begin{aligned}
\label{eq contract} \mathfrak{K}(\Delta)=
2\mathfrak{O}(e^{2\tau(0)\Delta}-1)+2\tau(0)\Delta.\end{aligned}$$ Moreover since $\Delta_T\rightarrow0$ we have $$\begin{aligned}
\label{eq contract 0}\mathfrak{K}(\Delta_T)\leq \mathfrak{C}\Delta_T<1\end{aligned}$$for some positive constant $\mathfrak{C}$ depending on $\tau(0)$ and $\mathfrak{O}$.
Proposition \[prop complet contract\] enables to apply the Banach fixed point theorem; we derive that $f$ is the unique fixed point of $\mathbf{H}_{\Delta,f}$ and from any initial point $h_0$ in $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ we have $$\big\|f-\mathbf{H}_{\Delta,f}^{\circ K} [h_0]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\rightarrow 0\ \ \ \mbox{ as } \ K\rightarrow\infty,$$ where $\circ$ stands for the composition product and $\mathbf{H}_{\Delta,f}^{\circ K}$ is $\mathbf{H}_{\Delta,f}\circ \ldots\circ \mathbf{H}_{\Delta,f}$, $K$ times. We choose $h_0=\mathbf{P}_\Delta[f]$ as a starting point (Lemma \[lem loicontY boule besov\] in Section \[Section proof\] ensures that $\mathbf{P}_\Delta[f]$ belongs to $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$).
\[prop inverseTronqueeK\] Let $\pi\geq 1$ and define the operator $\mathbf{L}_{\Delta,K}$ as the $K$th degree Taylor polynomial of $\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]$ in $\Delta$. It verifies for $p\geq1$$$\begin{aligned}
\label{eq Taylor Approx renewal}\Big\|\mathbf{H}_{\Delta,f}^{\circ K
}\big[\mathbf{P}_\Delta[f]\big]-\mathbf{L}_{\Delta,K} \Big\|_{L_p(\mathcal{D})}\leq\mathfrak{C}\Delta^{K+1}\end{aligned}$$ where $\mathfrak{C}$ is a positive constant depending on $\tau(0)$, $\mathfrak{M}$ and $\mathfrak{O}$. Moreover we have $$\begin{aligned}
\label{eq LfK linear}
\mathbf{L}_{\Delta,K}=\sum_{m=1}^{K+1} l_m(\Delta) {\mathbf{P}_\Delta[f]}^{\star m},\end{aligned}$$ where for $m=1,\ldots,K+1$ we have $|l_m(\Delta)|\leq \mathfrak{C}\Delta^{m-1}$ where $\mathfrak{C}$ is a positive constant that depends on $\tau(0)$ and $K$.
### Construction of estimators of the $\mathbf{P}_\Delta[f]^{\star m}$ {#construction-of-estimators-of-the-mathbfp_deltafstar-m .unnumbered}
Consider the increments $\big(\mathbf{D}^\Delta X_i=X_{i\Delta}-X_{(i-1)\Delta},i=1,\dots,\lfloor T\Delta^{-1}\rfloor\big)$ introduced earlier and define the nonzero ones using $$\begin{aligned}
S_1&=\inf\big\{j, \mathbf{D}^\Delta X_j\ne0\big\}\wedge \lfloor T\Delta^{-1}\rfloor\\S_i&=\inf\big\{j>S_{i-1}, \mathbf{D}^\Delta X_j\ne0\big\}\wedge \lfloor T\Delta^{-1}\rfloor\ \ \ \mbox{for }i\geq1,\end{aligned}$$ where $S_i$ is the random index of the $i$th jump. Let $$N_T=\sum_{i=1}^{\lfloor T\Delta^{-1}\rfloor}\mathds{1}_{\{\mathbf{D}^\Delta X_i\ne0\}}$$ the random number of nonzero increments observed over $[0,T]$. By Assumption \[ass f\], on the event $\{\mathbf{D}^\Delta X_i=0\},$ no jump occurred between $(i-1)\Delta$ and $i\Delta$. In the microscopic regime when $\Delta=\Delta_T\rightarrow 0$ as $T$ goes to infinity many increments are null and convey no information about $f$, hence for the estimation of $f$ we focus on the nonzero ones $$\big(\mathbf{D}^\Delta X_{S_1},\ldots,\mathbf{D}^\Delta X_{S_{N_T}}\big).$$ They are identically distributed of density given by ; Lemma \[lem pmControl\] still applies.
We construct wavelet threshold density estimators of the $K+1$ first convolution powers of $\mathbf{P}_\Delta[f]$; define the wavelet coefficients for $m\geq 1$ $$\begin{aligned}
\label{eq est Coeffconvol}
\widehat{\gamma}^{(m)}_{jk}&=\frac{1}{N_{T,m}}\sum_{i=1}^{N_{T,m}}g_{jk} \Big(\mathbf{D}^{\Delta}_m X_{S_i}\Big),\end{aligned}$$ where $N_{T,m}=\big\lfloor N_T/m \big\rfloor\geq 1$ for large enough $T$ and $$\mathbf{D}^{\Delta}_m X_{S_i}=\mathbf{D}^\Delta X_{S_i}+ \mathbf{D}^\Delta X_{S_{N_{T,m}+i}}+\dots+\mathbf{D}^\Delta X_{S_{(m-1)N_{T,m}+i}}.$$ Let $\eta>0$ and $J\in {\mathbb{N}}\setminus\{0\},$ define $\widehat{P_{\Delta,m}}$ the estimator of $\mathbf{P}_\Delta[f]^{\star m}$ over $\mathcal{D}$ for $m\geq 1$ $$\begin{aligned}
\label{eq est convol}
\widehat{P_{\Delta,m}}(x)&=\sum_{k}\widehat{\alpha}_{0k}^{(m)}\phi_{0k}(x)+\sum_{j= 0}^J\sum_{k}\widehat{\beta}_{jk}^{(m)}\mathds{1}_{\big\{|\widehat{\beta}_{jk}^{(m)}|\geq \eta\big\}}\psi_{jk}(x),\ \ \ x\in\mathcal{D}.\end{aligned}$$
As mentioned earlier $\tau$ is a nuisance that needs to be estimated. To simplify the problem, we make the following parametric assumption on $\tau$.
\[ass param tau\] Assume there exists $\vartheta$ in $\Theta$ a compact subset of ${\mathbb{R}}$ such that $$\begin{aligned}
\tau(x)&=\tau_1(x,\vartheta),\ \ \ \ \forall x\in [0,\infty),\end{aligned}$$ where $\tau_1$ is known, $\tau_1(0,\vartheta)>0$ and $\vartheta\rightarrow\tau_1(.,\vartheta)$ is $C^1$ . Assume there exists $q$ from $\Theta$ to $[0,1]$, invertible, such that $q(\vartheta)={\mathbb{P}}(R_\Delta\ne0)$ and whose inverse $q^{-1}$ is bounded.
Assumption \[ass param tau\] enables to estimate the unknown coefficients $\big(p_m(\Delta)\big)$ and $\big(l_m(\Delta)\big)$, and to compute the estimator of $f$ defined hereafter.
\[def est corr K\]Let $\widehat{f}^K_{T,\Delta}$ be the estimator corrected at order $K$ defined for $K$ in ${\mathbb{N}}$ and $x$ in $\mathcal{D}$ as $$\begin{aligned}
\label{eq Est f Threshold Corr renew}
\widehat{f}^K_{T,\Delta}(x)&=\sum_{m=1}^{K+1} l_m(\Delta,\widehat{\vartheta_T}) \widehat{P_{\Delta, m}}(x),\end{aligned}$$ where $$\widehat{\vartheta_T}=q^{-1}\Big(\frac{1}{\lfloor T \Delta^{-1}\rfloor}\sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor}\mathds{1}_{\mathbf{D}^\Delta X_i\ne0}\Big)$$ and the $l_m(\Delta,\vartheta)$ are defined in Proposition \[prop inverseTronqueeK\].
When $\Delta_T$ satisfies $T\Delta_T^2=O(1)$, $\widehat{f}^0_{T,\Delta}$ defined in with $K=0$ and $\widehat{f}_{T,\Delta}$ defined in coincides.
Convergence rates
-----------------
\[ass queue tau\] Assume that there exist $(\mathfrak{A},\mathfrak{a},\mathfrak{g})$ positive constants such that $$\begin{aligned}
\tau(x)\leq \mathfrak{A}\exp\big(-\mathfrak{a}x^\mathfrak{g}),\ \ \ \ \forall x\in [0,\infty).\end{aligned}$$
Assumption \[ass queue tau\] is a technical condition which ensures that $\tau$ has moments of all order. It is used in the proofs to replace $N_T/T$ by its asymptotic deterministic limit. Compactly supported densities and densities with subexponential queues satisfies Assumption \[ass queue tau\].
\[thm Renewal 2\] We work under Assumptions \[ass f\], \[Ass\], \[ass param tau\] and \[ass queue tau\] and assume that there exists $0<\delta< 1$ such that $$T\Delta_T^2=O(T^\delta),$$ up to logarithmic factors in $T$. Let $\pi\geq 1$, $\sigma>s>1/\pi$, $p\geq1$ and $\widehat{P_{\Delta_T,m}}$ be the threshold wavelet estimator of $\mathbf{P}_{\Delta_T}[f]^{\star m}$ on $\mathcal{D}$ constructed from $(\phi,\psi)$ and defined in . Take $J$ such that $$2^JT^{-1}\log\big(T^{1/2}\big)\leq 1,$$ and $$\eta=\kappa T^{-1/2}\sqrt{\log\big(T^{1/2}\big)},$$ for some $\kappa>0$.\
1) For $m\geq1$ the estimator $\widehat{P_{\Delta_T,m}}$ of $\mathbf{P}_{\Delta_T}[f]^{\star m}$ verifies for sufficiently large $\kappa>0$ $$\begin{aligned}
\underset{\mathbf{P}_{\Delta_T}[f]^{\star m}\in\mathcal{F}(s,{\pi},\mathfrak{M})}{\sup}\big({\mathbb{E}}\big[\big\|\widehat{P_{\Delta_T,m}} -\mathbf{P}_{\Delta_T}[f]^{\star m}\big\|_{L_p(\mathcal{D})}^p\big] \big)^{1/p}&\leq \mathfrak{C} T^{-\alpha(s,p,\pi)},\end{aligned}$$ up to logarithmic factors in $T$, where $\alpha(s,p,\pi)$ is defined in and where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi$ and $\vartheta$.\
2) The estimator corrected at order $K$ $\widehat{f}^K_{T,\Delta}$ for $K\in{\mathbb{N}}$ defined in verifies for $T$ large enough, sufficiently large $\kappa>0$ and any compact set $\Theta\subset{\mathbb{R}}$ $$\begin{aligned}
\underset{\vartheta\in\Theta}{\sup}\underset{f\in\mathcal{F}(s,{\pi},\mathfrak{M})}{\sup}\big({\mathbb{E}}\big[ \| \widehat{f}^K_{T,\Delta} -f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}&\leq \mathfrak{C} \max\big(T^{-\alpha(s,p,\pi)},\Delta_T^{K+1}\big),\end{aligned}$$up to logarithmic factors in $T$ and where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi$ and $K$.
The proof of Theorem \[thm Renewal 2\] is postponed to Section \[section proof thm2\]. Since $\alpha(s,p,\pi)\leq1/2$, Theorem \[thm Renewal 2\] ensures that whenever $\Delta_T$ and $T$ are polynomially related it is always possible to find $K_0$ such that the estimator corrected at order $K_0$ achieves the minimax rate of convergence (see Section \[section discuss\]). If $\Delta_T$ decays slower than any power of $1/T$, for instance if it decreases logarithmically with $T$, the estimator corrected at order $K$ still provide a consistent estimator of $f$.
A numerical example\[section num ex\]
=====================================
In this Section we illustrate the results of Theorems \[thm Renewal 1\] and \[thm Renewal 2\]. In both cases we compare the performances of our estimator with an oracle: the wavelet estimator we would compute in the idealised framework where all the jumps are observed $$\begin{aligned}
\widehat{f}^{Oracle}(x)=\sum_{k}\widehat{\alpha}_{0k}^{Oracle}\phi_{0k}(x)+\sum_{j= 0}^J\sum_{k}\widehat{\beta}_{jk}^{Oracle}\mathds{1}_{\big\{|\widehat{\beta}_{jk}^{Oracle}|\geq \eta\big\}}\psi_{jk}(x),\end{aligned}$$ where $$\widehat{\alpha}_{0k}^{Oracle}=\frac{1}{R_T}\sum_{i=1}^{R_T}\phi_{0k}(\xi_i)\ \ \ \mbox{and}\ \ \ \widehat{\beta}_{jk}^{Oracle}=\frac{1}{R_T}\sum_{i=1}^{R_T}\phi_{0k}(\xi_i),$$ $R_T$ being the value of the renewal process $R$ at time $T$ and $(\xi_i)$ the jumps. The parameters $J$ and $\eta$ as well as the wavelet bases $(\phi,\psi)$ are the same for all the estimators.
We consider a renewal process with a $Beta(1,\vartheta)$ interarrival density $\tau$. We have $\vartheta=3$, the first shape parameter is set to 1 to ensure the condition $0<\tau_1(0,\vartheta)<\infty$. We estimate the compound law given by $$f(x)=(1-a) f_1(x)+af_2(x),$$ where $f_1$ is the uniform distribution over $[-2,2]$ and $f_2$ is a Laplace with location parameter 1 and scale parameter 0.5, we take $a=0.5$. We estimate the mixture $f$ on $\mathcal{D}=[-10,10]$ with the estimator corrected at order $K$ for different values of $K$ and study the results with the $L_2$ error. We also compare them with the oracle $\widehat{f}^{Oracle}$. Wavelet estimators are based on the evaluation of the first wavelet coefficients, to perform those we use Symlets 4 wavelet functions and a resolution level $J=10$. Moreover we transform the data in an equispaced signal on a grid of length $2^L$ with $L=8$, it is the binning procedure (see Härdle *et al.* [@KerkPicTsyb] Chap. 12). The threshold is chosen as in Theorems \[thm Renewal 1\] and \[thm Renewal 2\]. The estimators we obtain take the form of a vector giving the estimated values of the density $f$ on the uniform grid $[-10,10]$ with mesh $0.01$. We use the wavelet toolbox of .
Illustration in the fast microscopic case
-----------------------------------------
In this case we choose $\Delta=T^{-1/2}$. Figure \[Fig Fast1\] represents the estimator $\widehat{f}_{T,\Delta}$ of Definition \[def est f NC\] and the oracle. The estimators are evaluated on the same trajectory. They are quite hard to distinguish, what is confirmed by the comparison of their $L_2$ losses.
![Estimators of the density $f$ (plain dark) for $T=10000$ and $\Delta=0.01$: the oracle (dotted red) and the estimator $\widehat{f}_{T,\Delta}$ (dashed green). []{data-label="Fig Fast1"}](EstFast)
We approximate the $L_2$ errors by Monte Carlo. For that we compute $M=1000$ times each estimator (for $T=10000$ and $\Delta=0.01$) and approximate the $L_2$ loss by$$\frac{1}{M}\sum_{i=1}^M\Big(\sum_{p=0}^{2000} \big(\widehat{f}(-10+0.01 p)-f(-10+0.01 p)\big)^2\times0.01\Big).$$ For each Monte Carlo iteration the estimators are evaluated on the same trajectory. The results are reproduced in the following table.
Estimator Oracle $\widehat{f}_{T,\Delta}$
-------------------- -------- -------------------------- -- -- --
$L_2$ error 0.1916 0.2040
Standard deviation 0.4519 0.4605
Illustration in the slow microscopic case
-----------------------------------------
We now study the behaviour of the estimator corrected at order $K$ for different values of $K$. We choose $T=10000$ and $\Delta=0.1$. In that case $T\Delta^2$ is large but $T\Delta^4$ is 1. According to Theorem \[thm Renewal 2\] we should observe that the estimator corrected at order 2, behaves as the oracle. Figure \[Fig Slow1\] represents the estimators $\widehat{f}^K_{T,\Delta}$ defined in Definition \[def est corr K\] for $K\in \{0,1,2,3\}$ and the oracle. The estimators are evaluated on the same trajectory. They all manage to reproduce the shape of the density $f$, and graphically apart from the estimator corrected at order 0 they are difficult to distinguish.
![Estimators of the density $f$ (plain dark) for $T=10000$ and $\Delta=0.1$: the oracle (dotted red) and the estimator $\widehat{f}^K_{T,\Delta}$ for $K=0,1,2,3$ (dashed light to dark grey). []{data-label="Fig Slow1"}](ESTSlow)
We compare their $L_2$ losses in the following tabular.
Estimator Oracle $K=0$ $K=1$ $K=2$ $K=3$
-------------------- -------- -------- -------- -------- --------
$L_2$ error 0.1896 0.5176 0.3037 0.2959 0.2946
Standard deviation 0.4348 0.7800 0.7533 0.7462 0.7466
This confirms that there is an actual gain in considering the estimator corrected at order 1 instead of the uncorrected one. In the following table we estimate the $\big(p_m(\Delta)\big)$ defined in Proposition \[PropDefOperator\].
Estimated quantity $\widehat{p_1}$ $\widehat{p_2}$ $\widehat{p_3}$
-------------------- ----------------- ----------------- -----------------
Estimation 0.8527 0.1327 0.0135
Standard deviation 0.9185 0.7388 0.1597
It turns out that making no correction is equivalent to estimate a density on a data set where $15\%$ of the observations are realisations of a law which is not target. This explains why it is relevant to take them into account when estimating $f$. Considering more than 1 or 2 corrections is unnecessary as the $L_2$ losses get stable afterwards. The $L_2$ loss of the oracle is strictly lower than the loss of the estimator corrected at order $K$, even for large $K$. That difference is explained by the fact that to estimate the $m$th convolution power we do not use $N_T$ data points but $N_{T,m}=\lfloor N_T/m\rfloor$. Therefore we do not loose in terms of rate of convergence, but we surely deteriorate the constants in comparison with the oracle.
Discussion and Conclusion\[section discuss\]
============================================
#### Attainable rates.
Without loss of generality, assuming $T$ is an integer if we observe $T$ independent realisations of the density $f$, it is possible to achieve the minimax rates of convergence $T^{-\alpha(s,\pi,p)}$ (see for instance Donoho *et al.* [@Donoho96]). When the process $X$ is continuously observed over $[0,T]$, we have $R_T$ independent and identically distributed realisations of $f$. Moreover for $T$ large enough, the elementary renewal theorem guarantees that $R_T$ is of the order of $T$ (see for instance Lindvall [@Lindvall]). It follows that the estimators of $f$ given in Sections \[section fast\] and \[section slow\] enables to attain the minimax rates of convergence of an experiment where $X$ is continuously observed.
#### Comparison with a previous work.
The results of this paper are the generalisation to the renewal reward case of Duval [@Du2]; a compound Poisson process is a particular renewal reward process and Theorems \[thm Renewal 1\] and \[thm Renewal 2\] enable to recover the results of [@Du2]. However in this paper we do not have an explicit formula for the estimator corrected at order $K$ but only a construction method. In the Poisson case it is much more simpler to apply the results of [@Du2].
#### Extension to the case where $\Delta$ is fixed.
We established Theorem \[thm Renewal 2\] for $\Delta_T$ vanishing to 0. Since the approximation of the inverse depends only the fact that $\mathbf{H}_{\Delta,f}$ is a contraction, the method remains valid for $\Delta$’s such that $\mathfrak{K}(\Delta)$ defined in is strictly lower than 1. Which means that we can expand the results to cases where $\Delta$ does not go to 0 but satisfies $\mathfrak{K}(\Delta)<1.$ The value of the maximum value $\Delta_1$ satisfying the former inequality depends on $\tau(0)$ and $\mathfrak{O}$ but is not only determined by . Another hidden condition on $\Delta$ have to be satisfied for $\mathbf{H}_{\Delta,f}$ to send elements of $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ into itself. Then to find $\Delta_1$ one has to solve an optimisation program with constraints to find $\Delta_1$ and $\mathfrak{O}$ giving the maximum coverage for $\Delta$. To get an idea of the value of $\Delta_1$ we use the function `NMaximize` of `Mathematica` and find that one should take $\mathfrak{O}=1.645$ and $\Delta_1=0.071/\tau(0)>0,$ which is positive. The results of Theorem \[thm Renewal 2\] should generalise in for all $\Delta_T\rightarrow\Delta_\infty$ such that $\Delta_\infty<\Delta_1$ and for $K\in {\mathbb{N}}$ the rate of convergence for the estimator corrected at order $K$ is bounded by $$\max\big\{T^{-\alpha(s,p,\pi)},\Delta^{K+1}_\infty\big\}.$$ However to achieve suitable rates theoretically one should consider larger $K$, therefore the dependency in $K$ in the constants need to be handled carefully. In practice for $T=10000$ and $\Delta=0.1$ considering $K=2$ appears sufficient to have $T^{-\alpha(s,p,\pi)}$ predominant in front of $\Delta^{K+1}$.
#### Discussion on Assumptions \[ass f\] and \[ass param tau\].
In the present paper we made two simplifying assumptions on the interarrival density $\tau$. First we assume that $J_1$ was distributed according to $\tau_0$ to work with a process with stationary increments. In fact if $\tau$ has finite expectation this assumption is not necessary since asymptotically the process has stationary increments (see Lindvall [@Lindvall]). The second assumption is that $\tau$ is described by a 1-dimensional parameter $\vartheta$. Generalising the result to a $d$-dimensional parameter should be possible at small cost, but removing all parametric assumption on $\tau$ would demand to solve a nonstandard nonparametric program for $\tau$ from the observations : observations only give access to truncated values of realisations of $\tau$ spaced of more than $\Delta$. Then the problem of estimating $\tau$ from should be considered separately.
#### Other generalisations.
We constructed in the microscopic regime an adaptive minimax estimator of the jump density of a renewal reward process. The methodology presented here should adapt to any process defined similarly to $X$ but whose counting process has stationary increments and manageable dependencies. We consider in the present paper a renewal counting measure since we are interested in expanding the methodology to other regimes of $\Delta$, namely when $\Delta=\Delta_T$ tends to a constant (intermediate regime) or to infinity (macroscopic regime). The macroscopic regime is of special interest since the observed process presents diffusive or anomalous asymptotic behaviour determined by the laws $f$ and $\tau$ (see for instance Meerschaert and Scheffler [@Meerschaert04; @Meerschaert05] or Kotulski [@kolt]) and many applications have a model based on a macroscopically observed renewal reward processes. For instance in physics where they are used to model particle motion (see Watkins and Credgington [@Watkins] or Cuppen *et al.* [@cuppen]), in biology to model the proliferation of tumor cells (see Fedotov and Iomin [@Fedotov]) or lipid granule motion (see Jeon *et al.* [@Jeon]), they are also used to model records (see Sabhapandit [@Sabhapandit]).
Proofs {#Section proof}
======
In the sequel $ \mathfrak{C}$ denotes a constant which may vary from line to line.
Proof of Theorem \[thm Renewal 1\]\[proof thm 1\] \[proof thm 1\]
-----------------------------------------------------------------
### Proof of part 1) of Theorem \[thm Renewal 1\] {#proof-of-part-1-of-theorem-thm-renewal-1 .unnumbered}
To prove part 1) of Theorem \[thm Renewal 1\] we apply the general results of Kerkyacharian and Picard [@KP00]. For that we establish some technical lemmas.
\[lem loicontY boule besov\] If $f$ belongs to $\mathcal{F}(s,{\pi},\mathfrak{M})$ then for $m\geq 1$, $\mathbf{P}_\Delta[f]^{\star m}$ also belongs to $\mathcal{F}(s,{\pi},\mathfrak{M})$.
To prove Theorem \[thm Renewal 1\], we use Lemma \[lem loicontY boule besov\] for $m=1$ only, but we take some advance on the proof of Theorem \[thm Renewal 2\].
It is straightforward to derive $\big\|\mathbf{P}_\Delta[f]^{\star m}\big\|_{L_1({\mathbb{R}})}=1$. The remainder of the proof is a consequence of the following result: Let $f\in {\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}$ and $g\in L_1$ we have $$\begin{aligned}
\|f\star g\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}\leq \|f\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}\|g\|_{L_1({\mathbb{R}})}.\label{eq etoile}\end{aligned}$$ To prove we use the definition of the Besov norm ; the result is a consequence of Young’s inequality and elementary properties of the convolution product. First Young’s inequality gives $$\begin{aligned}
\label{eq ine1}\|f_1\star f_2\|_{L_\pi({\mathbb{R}})}\leq \|f_1\|_{L_\pi({\mathbb{R}})}\|f_2\|_{L_1({\mathbb{R}})}.\end{aligned}$$ Then the differentiation property of the convolution product leads for $n\geq1$ to $$\begin{aligned}
\label{eq ine2}\Big\|\frac{d^n}{dx^n}(f_1\star f_2)\Big\|_{L_\pi(\mathcal{D})}=\Big\|\Big(\frac{d^n}{dx^n} f_1\Big)\star f_2\Big\|_{L_\pi({\mathbb{R}})}\leq \Big\|\frac{d^n}{dx^n} f_1\Big\|_{L_\pi(\mathcal{D})}\|f_2\|_{L_1({\mathbb{R}})}.\end{aligned}$$ Finally translation invariance of the convolution product enables to get $$\begin{aligned}
\big\|\mathbf{D} ^h\mathbf{D} ^h[(f_1\star f_2)^{(n)}]\big\|_{L_\pi(\mathcal{D})}&=\big\|(\mathbf{D} ^h\mathbf{D} ^h[f_1^{(n)}])\star f_2\big\|_{L_\pi(\mathcal{D})}\nonumber\\&\leq \big\|\mathbf{D} ^h\mathbf{D} ^h[f_1^{(n)}]\big\|_{L_\pi(\mathcal{D})}\|f_2\|_{L_1({\mathbb{R}})}.\label{eq ine3}\end{aligned}$$ Inequality is then obtained by bounding $\|f\star g\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}$ using , and . To complete the proof of Lemma \[lem loicontY boule besov\], we apply $m-1$ times which leads to $$\forall m\in {\mathbb{N}}\setminus\{0\},\ \ \ \ \ \big\|\mathbf{P}_\Delta[f]^{\star m}\big\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}\leq \big\|\mathbf{P}_\Delta[f]\big\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}.$$ The triangle inequality gives $\|\mathbf{P}_\Delta[f]^{\star m}\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}\leq \|f\|_{{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}}\leq\mathfrak{M}$ which concludes the proof.
\[lem Ros NC\] Let $2^{j}\leq T$, then for $p\geq 1$ we have $${\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}-\gamma_{jk}\big|^p\big]\leq \mathfrak{C}_{p
,\|g\|_{L_p({\mathbb{R}})},\mathfrak{M}} T^{-p/2},$$ where $\widehat{\gamma}_{jk}$ is defined in and $$\begin{aligned}
\label{eq coeff gam}\gamma_{jk}=\int g_{jk}(y) \mathbf{P}_\Delta[f](y)dy.\end{aligned}$$
The proof is obtained with Rosenthal’s inequality: let $p\geq 1$ and let $(Y_1,\ldots,Y_n)$ be independent random variables such that ${\mathbb{E}}[Y_i]=0$ and ${\mathbb{E}}\big[|Y_i|^p\big]<\infty$. Then there exists $\mathfrak{C}_p$ such that $$\begin{aligned}
\label{eq Ros}{\mathbb{E}}\bigg[\Big|\sum_{i=1}^nY_i \Big|^p\bigg]\leq \mathfrak{C}_p\bigg\{\sum_{i=1}^n{\mathbb{E}}\big[|Y_i|^p\big]+\Big(\sum_{i=1}^n{\mathbb{E}}\big[|Y_i|^2\big]\Big)^{p/2}\bigg\}.\end{aligned}$$
According to Proposition \[PropDefOperator\] the $\big(\mathbf{D}^{\Delta_T} X_i\big)$ have distribution $$f_{\mathbf{D}^{\Delta_T}(x) X_1}=p(\Delta_T)\delta_0(x)+\big(1-p(\Delta_T)\big)\mathbf{P}_{\Delta_T}[f](x),\ \ \ \ x\in\mathcal{D}$$ where $\delta_0$ is the Dirac delta function and $p(\Delta_T)={\mathbb{P}}(R_{\Delta_T}=0)$. We derive $$\begin{aligned}
{\mathbb{E}}\big[ \widehat{\gamma}_{jk}\big]&=\int g_{jk}(z)\mathds{1}_{\{z\ne0\}}\frac{f_{\mathbf{D}^{\Delta_T} X_1}(z)}{1-p(\Delta_T)}dz = \int g_{jk}(z)\mathbf{P}_{\Delta_T}[f](z)dz =\gamma_{jk}.\end{aligned}$$ Then $ \widehat{\gamma}_{jk}-\gamma_{jk}$ is a sum of centered and identically distributed random variables, define $$Z_i=\frac{1}{1-p(\Delta_T)}g_{jk}\big(\mathbf{D}^{\Delta_T} X_i\big)\mathds{1}_{\{\mathbf{D}^{\Delta_T} X_i\ne0\}}.$$ Since $x$ is a renewal reward process, nonzero and nonconsecutive $Z_i$ are independent, then if we separate the sum in two sums of nonzero and nonconsecutive indices we can apply Rosenthal’s inequality for independent variables to each sum, it wont affect the rates but the constant will modified. For $p\geq1$ we have by convex inequality $$\begin{aligned}
{\mathbb{E}}\big[\big|Z_i-{\mathbb{E}}[Z_i]\big|^p\big]&\leq 2^p {\mathbb{E}}\big[\big|Z_i\big|^p\big]\\&\leq\frac{2^p2^{jp/2}}{\big(1-p(\Delta_T)\big)^p}\int |g(2^jy-k)|^p\mathds{1}_{\{y\ne 0\}}f_{\mathbf{D}^{\Delta_T}}(y)dy\\
&=\frac{2^p2^{j(p/2-1)}}{\big(1-p(\Delta_T)\big)^{p-1}}\int |g(z)|^p \mathbf{P}_{\Delta_T}[f]\Big(\frac{z+k}{2^j}\Big)dz,\end{aligned}$$ where we made the substitution $z=2^jy-k$. Lemma \[lem loicontY boule besov\] and Sobolev embeddings (see [@Cohen; @Donoho96; @KerkPicTsyb]) $$\begin{aligned}
\label{eq Sobolev embed}\mathcal{B}^s_{{\pi} \infty}\hookrightarrow \mathcal{B}^{s'}_{p \infty}\ \ \ \mbox{ and }\ \ \ \mathcal{B}^{s'}_{{\pi} \infty}\hookrightarrow\mathcal{B}^s_{{\infty} \infty},\end{aligned}$$where $p>\pi$, $s\pi>1$ and $s'=s-1/\pi+1/p$, give $\big\|\mathbf{P}_{\Delta_T}[f]\big\|_\infty\leq \mathfrak{M}$. It follows that $${\mathbb{E}}\big[\big|Z_i-{\mathbb{E}}[Z_i]\big|^p\big]\leq 2^p2^{j(p/2-1)}\|g\|_{L_p({\mathbb{R}})}^p\mathfrak{M}/(1-p(\Delta_T))^{p-1}$$ and $${\mathbb{E}}\big[\big|Z_i-{\mathbb{E}}[Z_i]\big|^2\big]\leq \mathfrak{M}/(1-p(\Delta_T))$$ since $\|g\|_{L_2({\mathbb{R}})}=1$. Rosenthal’s inequality gives for $p\geq1$ $$\begin{aligned}
{\mathbb{E}}\big[\big| \widehat{\gamma}_{jk}-\gamma_{jk}\big|^p\big]&\leq \mathfrak{C}_p \Big\{2^p\Big(\frac{2^{j}}{ A_{T}}\Big)^{\frac{p}{2}-1}\|g\|_{L_p({\mathbb{R}})}^p\mathfrak{M}+\mathfrak{M}^{p/2}\Big\}A_{T}^{-\frac{p}{2}},
\end{aligned}$$ where $A_{T}=\lfloor T\Delta_T^{-1}\rfloor(1-p(\Delta_T))$. To conclude we use that $$1-p(\Delta_T)={\mathbb{P}}(J_1\geq\Delta_T)=\frac{1}{\mu}\int_0^{\Delta_T}\big(1-F(u)\big)du,$$ since $J_1$ has distribution , and derive that there exists $\Delta_1>0$ such that $ F(\Delta_1)\leq \frac{1}{2}$ and for all $\Delta_T\leq \Delta_1$ we have $$\begin{aligned}
\label{eq pcontrol Renew} \frac{\Delta_T}{2\mu}\leq 1-p(\Delta_T)\leq \frac{\Delta_T}{\mu}.\end{aligned}$$ It follows that $$\frac{T}{2\mu}\leq A_T \leq \frac{T}{\mu}$$ and then using $2^j\leq T$ $$\begin{aligned}
{\mathbb{E}}\big[\big| \widehat{\gamma}_{jk}-\gamma_{jk}\big|^p\big]&\leq\mathfrak{C}_{p,\|g\|_{L_p({\mathbb{R}})},\mathfrak{M},\mu}T^{-p/2}.
\end{aligned}$$ The proof the complete.
\[lem Bern NC\] Choose $j$ and $c$ such that $$2^j T^{-1}\log(T^{1/2})\leq1\mbox{ and }c^2\geq 32\mu\Big(\mathfrak{M}+\frac{c\|g\|_\infty}{6}\Big).$$ For all $r\geq 1$, let $\kappa_r=cr$. We have $${\mathbb{P}}\Big(\big|\widehat{\gamma}_{jk}-\gamma_{jk}\big|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\leq T^{-r/2},$$ where $\widehat{\gamma}_{jk}$ is defined in and $\gamma_{jk}$ in .
The proof is obtained with Bernstein’s inequality. Consider $Y_1,\ldots,Y_n$ independent random variables such that $|Y_i|\leq \mathfrak{A}$, ${\mathbb{E}}[Y_i]=0$ and $b_n^2=\sum_{i=1}^n{\mathbb{E}}[Y_i^2]$. Then for any $\lambda>0$, $$\begin{aligned}
\label{eq Bernstein}{\mathbb{P}}\Big(\Big|\sum_{i=1}^nY_i\Big| >\lambda\Big)\leq 2\exp\Big(-\frac{\lambda^2}{2(b_n^2+\frac{\lambda \mathfrak{A}}{3})}\Big).\end{aligned}$$ We keep notation $Z_i$ introduced in the proof of Lemma \[lem Ros NC\], $ \widehat{\gamma}_{jk}-\gamma_{jk}$ is a sum of centered and identically distributed random variables bounded by $2^{j/2}\|g\|_\infty/(1-p(\Delta_T))$ which verify $${\mathbb{E}}\big[ \big|Z_i-{\mathbb{E}}[Z_i]\big|^2\big]\leq\mathfrak{ M}/(1-p(\Delta_T)).$$ After separating the sum to get two sums of nonzero and nonconsecutive indices we apply Bernstein’s inequality for independent variables to each sum, which modify the constants. It follows that $$\begin{aligned}
{\mathbb{P}}\Big(|\widehat{\gamma}_{jk}-\gamma_{jk}|&\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\\
\leq&2\exp\Bigg(-\frac{\kappa_r^2T^{-1}\log(T^{1/2}) \lfloor T\Delta_T^{-1}\rfloor\big(1-p(\Delta_T)\big)} {16\Big( \mathfrak{M} +\frac{\kappa_r T^{-1/2}\sqrt{\log(T^{1/2})}2^{j/2}\|g\|_\infty}{6}\Big)}\Bigg). \end{aligned}$$ Using that $2^jT^{-1}\log(T^{1/2})\leq1$ and which gives $$T^{-1}\lfloor T\Delta_T^{-1}\rfloor(1-p(\Delta_T)) \geq \frac{1}{2\mu},$$ we have $$\begin{aligned}
{\mathbb{P}}\Big(|\widehat{\gamma}_{jk}-\gamma_{jk}|&\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\\
\leq&2\exp\bigg(-\frac{c^2r} {32\mu\big(\mathfrak{M}+\frac{\kappa_r\|g\|_\infty}{6}\big)}r\log(T^{1/2})\bigg)\leq T^{-r/2},
\end{aligned}$$ since $c^2\geq 32\mu\big(\mathfrak{M}+\frac{c\left\|g\right\|_\infty}{6}\big).$ The proof is complete.
It is a consequence of Lemma \[lem loicontY boule besov\], \[lem Ros NC\], \[lem Bern NC\] and of the general theory of wavelet threshold estimators of Kerkyacharian and Picard [@KP00]. It suffices to have conditions (5.1) and (5.2) of Theorem 5.1 of [@KP00], which are satisfied –Lemma \[lem Ros NC\] and \[lem Bern NC\]– with $c(T)=T^{-1/2}$ and $\Lambda_n=c(T)^{-1}$ (with the notation of [@KP00]). We can now apply Theorem 5.1, its Corollary 5.1 and Theorem 6.1 of [@KP00] to obtain the result.
### Completion of the proof of Theorem \[thm Renewal 1\] {#completion-of-the-proof-of-theorem-thm-renewal-1 .unnumbered}
To prove part 2) of Theorem \[thm Renewal 1\] we decompose the $L_p$ loss as follows $$\begin{aligned}
\big({\mathbb{E}}\big[ \| \widehat{f}_{T,\Delta_T}-&f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\\&\leq\big({\mathbb{E}}\big[\big\|\widehat{f}_{T,\Delta_T}-\mathbf{P}_{\Delta_T}[f] \|_{L_p(\mathcal{D})}^p\big]\big)^{1/p} +\big\|\mathbf{P}_{\Delta_T}[f] -f\big\|_{L_p(\mathcal{D})}.\end{aligned}$$ An upper bound for the first term is given by part 1) of Theorem \[thm Renewal 1\] $$\begin{aligned}
\label{eq prf thm11}\big({\mathbb{E}}\big[\big\|\widehat{f}_{T,\Delta_T}-\mathbf{P}_{\Delta_T}[f] \|_{L_p(\mathcal{D})}^p\big]\big)^{1/p} &\leq \mathfrak{C} T^{-\alpha(s,p,\pi)},\end{aligned}$$ where $\mathfrak{C}$ continuously depends on $\mu$, and on $s,\pi,p,\mathfrak{M},\phi,\psi$ and $\mu$. Since $$\begin{aligned}
\mathbf{P}_{\Delta_T}[f]-f=-(1-p_1(\Delta_T))f+\sum_{m=2}^{\infty}p_m(\Delta_T)\mathbf{P}_{\Delta_T}[f]^{\star m}\end{aligned}$$Lemma \[lem pmControl\], Young’s inequality, which gives $\|\mathbf{P}_{\Delta_T}[f]^{\star m}\|_{L_p(\mathcal{D})}\leq \|f\|_{L_p(\mathcal{D})}$ and Sobolev embeddings , which give $\|f\|_{L_p(\mathcal{D})}\leq \mathfrak{M}$, enable to get the bound $$\begin{aligned}
\label{eq prf thm12}\big\|\mathbf{P}_{\Delta_T}[f] -f\big\|_{L_p(\mathcal{D})}&\leq 2\tau(0)\Delta_T+2\mathfrak{M}\sum_{m=2}^\infty \frac{\big(2\tau(0)\Delta_T\big)^{m-1}}{m!}\leq \mathfrak{C}\Delta_T,\end{aligned}$$ where $\mathfrak{C}$ continuously depends on $\tau(0)$ and $\mathfrak{M}$. We finish the proof noticing that is predominant in front of since $\alpha(s,p,\pi)\leq 1/2$ and $T\Delta_T^2=O(1)$. Finally we take the supremum in $\mu$ and $\tau(0)$ over any compact of $(0,\infty)$ to render the constant independent of the unknown interarrival law $\tau$. The proof is now complete.
Proof of Proposition \[prop complet contract\]
----------------------------------------------
First we prove part 1) of Proposition \[prop complet contract\]. The set $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ is a subset of $$\big(\mathcal{B}^s_{{\pi} \infty}(\mathcal{D}),\|.\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\big)$$ which is a Banach space. We show that $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ is complete since it is a closed subset of a Banach space. For that we establish the following assertions; for all sequence $h_n\in \mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ such that there exists $h$ with $$\|h_n-h\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\rightarrow 0, \ \ \ \mbox{ as } n\rightarrow \infty,$$ we have $h\in \mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ *i.e.* $\|h\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq\mathfrak{N}$ and $\|h\|_{L_1(\mathcal{D})}\leq\mathfrak{O}$. The first inequality is immediate. The second one is a consequence of the compactness of $\mathcal{D}$. Indeed $$\|h_n-h\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\rightarrow 0, \ \ \ \mbox{ as } n\rightarrow \infty,$$ we have by definition of the Besov norm that $$\|h_n-h\|_{L_\pi(\mathcal{D})}\rightarrow 0, \ \ \ \mbox{ as } n\rightarrow \infty.$$ Since $\mathcal{D}$ is compact and $\pi\geq1$ we derive from Hölder’s inequality that $$\|h_n-h\|_{L_1(\mathcal{D})}\rightarrow 0, \ \ \ \mbox{ as } n\rightarrow \infty$$ and then $\|h\|_{L_1(\mathcal{D})}\leq \mathfrak{O}$ follows. The proof of part 1) of Proposition \[prop complet contract\] is now complete.
To prove part 2) of Proposition \[prop complet contract\] we show that $\mathbf{H}_{\Delta,f}$ sends elements of $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ into $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ and that it is a contraction. We start with the first assertion, the triangular inequality gives for $h\in\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}[h]\big\|_{L_1(\mathcal{D})}&\leq\big\|\mathbf{P}_\Delta[f]\big\|_{L_1(\mathcal{D})} +\big(1-p_1(\Delta)\big)\|h\|_1+\sum_{m=2}^\infty p_m(\Delta)\|h^{\star m}\|_{L_1(\mathcal{D})},\end{aligned}$$ where $\big\|\mathbf{P}_\Delta[f]\big\|_{L_1(\mathcal{D})}\leq \big\|\mathbf{P}_\Delta[f]\big\|_{L_1({\mathbb{R}})} =1$. Immediate induction on Young’s inequality leads to $$\|h^{\star m}\|_{L_1(\mathcal{D})}\leq \|h\|_{L_1(\mathcal{D})}^m\leq \mathfrak{O}^m$$ since $h\in \mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ and with Lemma \[lem pmControl\] we get $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}[h]\big\|_1&\leq 1+2\mathfrak{O}\tau(0)\Delta +\frac{1}{\tau(0)\Delta}(e^{2\mathfrak{O}\tau(0)\Delta}-1-2\mathfrak{O}\tau(0)\Delta)\leq\mathfrak{O}\end{aligned}$$ for $\Delta$ small enough since $\mathfrak{O}>1$. Similar computations and give $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}[h]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq&\big\|\mathbf{P}_\Delta[f]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}+\big(1-p_1(\Delta)\big)\|h\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\\&+\sum_{m=2}^\infty p_m(\Delta)\|h^{\star m}\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\\
\leq& \mathfrak{M}+2\tau(0)\Delta\mathfrak{N}+\frac{\mathfrak{N}}{\tau(0)\Delta\mathfrak{O}} (e^{2\mathfrak{O}\tau(0)\Delta}-1-2\mathfrak{O}\tau(0)\Delta) \leq\mathfrak{N},
\end{aligned}$$ for $\Delta$ small enough since $\mathfrak{M}<\mathfrak{N}$. Then if $h$ is in $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$, $\mathbf{H}_{\Delta,f}$ belongs to $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$.
For the contraction property, we have for all $h_1,h_2 \in \mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$$$\begin{aligned}
\mathbf{H}_{\Delta,f}[h_1]-\mathbf{H}_{\Delta,f}[h_2]=&(1-p_1(\Delta))(h_1-h_2)\nonumber\\&- (h_1-h_2)\star \sum_{m=2}^{\infty}p_m(\Delta)\sum_{q=0}^{m-1}h_1^{\star q}\star h_2^{\star m-1-q}\label{eq prof cont1}\end{aligned}$$ Lemma \[lem pmControl\] gives $$\begin{aligned}
\label{eq prof cont2}0\leq1-p_1(\Delta)\leq 2\tau(0)\Delta,\end{aligned}$$ and with Young’s inequality and since $h_1$ and $h_2$ belong to $\mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ we get $$\begin{aligned}
\label{eq prof cont3}
\Big\|\sum_{m=2}^{\infty}p_m(\Delta)\sum_{q=0}^{m-1}h_1^{\star q}\star h_2^{\star m-1-q}\Big\|_{L_1(\mathcal{D})}&\leq 2\mathfrak{O}\big(e^{2\tau(0)\mathfrak{O}\Delta}-1\big)\end{aligned}$$ for $\Delta$ small enough. Finally injecting and into leads to the contraction property for all $h_1,h_2 \in \mathcal{H}(s,{\pi},\mathfrak{O},\mathfrak{N})$ $$\big\|\mathbf{H}_{\Delta,f}[h_1]-\mathbf{H}_{\Delta,f}[h_2]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})} \leq \big(2\tau(0)\Delta+2\mathfrak{O}\big(e^{2\tau(0)\mathfrak{O}\Delta}-1\big)\big)\|h_1-h_2\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})},$$ which concludes the proof.
Proof of Theorem \[thm Renewal 2\]\[section proof thm2\]
--------------------------------------------------------
### Preliminary {#preliminary .unnumbered}
The estimators of the convolution powers of $\mathbf{P}_\Delta[f]$ depend on $N_T$ which is random and depends on the $\mathbf{D}^\Delta_m X_i$.
\[lem Nconcentre Renewal\] Work under Assumption \[ass queue tau\] and let $p(\Delta)={\mathbb{P}}(R_{\Delta}\ne0)$ and $\Delta_1$ be such that $\int_0^{\Delta_1}\tau(x)dx\leq \frac{1}{2}.$ Then for all $\lambda>0$ and $\Delta\leq\Delta_1$ we have $$\begin{aligned}
{\mathbb{P}}\Big(\Big|\frac{N_T}{\lfloor T\Delta^{-1}\rfloor}-p(\Delta)\Big|>\lambda \Delta\Big)\leq \exp\Big(-\mathfrak{C}\sqrt{T}\Delta\Big),
\end{aligned}$$ where $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\mathfrak{g},\mu,\lambda$.
We have $$\frac{N_T}{\lfloor T\Delta^{-1}\rfloor}-p(\Delta)=\frac{1}{\lfloor T\Delta^{-1}\rfloor}\sum_{i=1} ^{\lfloor T\Delta^{-1}\rfloor}Y_i,$$ where $$Y_i=\mathds{1}_{\{\mathbf{D}^{\Delta}X_i\ne0\}}-p(\Delta),\ \ \ \ i=1,\ldots,\lfloor T\Delta^{-1}\rfloor$$ are centered random variables, bounded by $M=1-p(\Delta)$ and such that $${\mathbb{E}}[Y_i^2]\leq p(\Delta).$$ To show the result, we apply Theorem 4.5 of Dedecker *et. al.* [@Doukhan] which is a Bernstein-type inequality for dependent data. We have to verify conditions (4.4.16) and (4.4.17) of Theorem 4.5 of [@Doukhan]. With their notation, condition (4.4.16) ensures that for all $u$-tuples $(s_1,\ldots,s_u)$ and all $v$-tuples $(t_1,\ldots,t_v)$ such that $$1\leq s_1\leq\ldots\leq s_u\leq t_1\leq\ldots \leq t_v\leq\lfloor T\Delta^{-1}\rfloor$$ we have $$\big|Cov(Y_{s_1}\ldots Y_{s_u},Y_{t_1}\ldots Y_{t_v})\big|\leq K^2M^{u+v-2}uv \rho(t_1-s_u),$$ for some positive constant $K$ and a nonincreasing function $\rho$ satisfying (4.4.17) namely $$\sum_{s=0}^\infty (s+1)^k\rho(s)\leq L_1L_2^k(k!)^\nu,\ \ \ \forall k\geq0,$$ where $L_1$, $L_2$ and $\nu$ are positive constants.
Since $X$ is a renewal process, $Y_{s_1}\ldots Y_{s_u}$ and $Y_{t_1}\ldots Y_{t_v}$ are independent if there exists $r$ such that $s_u< r <t_1$ and $Y_r=1-p(\Delta)$ *i.e* there is a jump between $Y_{s_u}$ and $Y_{t_1}$. For the covariance to be nonzero it is necessary that no jump occurred between $s_u\Delta$ and $(t_1-1)\Delta$. Let $s=t_1-s_u-1$ using that $R$ is stationary we get an upper bound for $\rho$ $$\begin{aligned}
\label{eq upp rho}\rho(t_1-s_u)\leq{\mathbb{P}}\big(R_{(t_1-1)\Delta}-R_{s_u\Delta}=0\big)={\mathbb{P}}\big(R_{s\Delta}=0\big)=\int_{s\Delta}^\infty \tau_0(x)dx\end{aligned}$$ which decreases with $s$. Moreover since the $Y_i$ are centered and bounded by $M\leq 1$ we have by Cauchy-Schwarz and ${\mathbb{E}}[Y_i^2]\leq p(\Delta)$ $$\begin{aligned}
\big|Cov(Y_{s_1}\ldots Y_{s_u},Y_{t_1}\ldots Y_{t_v})\big|&\leq \big|Cov(Y_{s_1},Y_{t_v})\big|\leq \sqrt{{\mathbb{E}}\big[Y_{s_1}^2\big]{\mathbb{E}}\big[Y_{t_v}^2\big]}\leq p(\Delta).\end{aligned}$$ We deduce that condition (4.4.16) is fulfilled with $K=p(\Delta)^{1/2}$ and the nonincreasing sequence $\rho$.
Next we show that $\rho$ satisfies (4.4.17), using Assumption \[ass queue tau\] and we get for $s\geq1$ $$\begin{aligned}
\rho(s)&\leq \frac{1}{\mu}\int_{s\Delta_T}^\infty \big(1-\int_0^x \tau(t)dt\big)dx\leq \mathfrak{C}\exp\big(-\mathfrak{a}(s\Delta)^\mathfrak{g'}\big),\end{aligned}$$ where $\mathfrak{g}<\mathfrak{g'}$ and $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\mu,\mathfrak{g}$. Which leads to for $k\geq0$ $$\begin{aligned}
\sum_{s=1}^\infty s^k\rho(s)&\leq \sum_{s=1}^{1/\Delta}s^k+\mathfrak{C}\sum_{s=1/\Delta}^\infty s^k\exp\big(-\mathfrak{a}(s\Delta)^\mathfrak{g'}\big)\nonumber\\&\leq \Delta^{-(k+1)}+\mathfrak{C}\Delta^{-k}\sum_{s'=1}^\infty s'^{k}\exp\big(-\mathfrak{a}(s')^\mathfrak{g'}\big)\nonumber\\&\leq \mathfrak{C}\Delta^{-(k+1)}\label{eq prf bernDep}\end{aligned}$$ where $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\mathfrak{g},\mu$, condition (4.4.17) follows with $L_1=\mathfrak{C}\Delta^{-1}$, $L_2=\Delta^{-1}$ and $\nu=0$.
We can now apply Theorem 4.5 which gives for all $\lambda>0$ $$\begin{aligned}
{\mathbb{P}}\Big(\Big|&\frac{N_T}{\lfloor T\Delta^{-1}\rfloor}-p(\Delta)\Big|>\lambda \Delta\Big)\\&\leq 2\exp\bigg(-\frac{\lfloor T\Delta^{-1}\rfloor^2\Delta^2\lambda^2}{2\big(\lfloor T\Delta^{-1}\rfloor p(\Delta)+(\lfloor T\Delta^{-1}\rfloor\Delta\lambda)^{3/2}\sqrt{2^5\Delta^{-2}}\big)}\bigg).\end{aligned}$$ Using we derive for $\Delta\leq \Delta_1$ $$\begin{aligned}
{\mathbb{P}}\Big(\Big|\frac{N_T}{\lfloor T\Delta^{-1}\rfloor}-p(\Delta)\Big|>\lambda \Delta\Big)&\leq \exp\Big(-\mathfrak{C}\sqrt{T}\Delta\Big),\end{aligned}$$ where $\mathfrak{C}$ depends on $\lambda,\mathfrak{A},\mathfrak{a},\mathfrak{g},\mu$. The proof is now complete.
### Proof of part 1) of Theorem \[thm Renewal 2\] {#proof-of-part-1-of-theorem-thm-renewal-2 .unnumbered}
As for the proof of part 1) of Theorem \[thm Renewal 1\] we apply the general results of Kerkyacharian and Picard [@KP00] and first establish some technical lemmas.
\[lem Ros C\] Let $2^{j}\leq T$, then for $p\geq 1$ we have for all $m\geq1$ $${\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p\big]\leq \mathfrak{C}_{p,m,\|g\|_{L_p({\mathbb{R}})},\mathfrak{M},\mu,\tau} T^{-p/2},$$ where $\widehat{\gamma}_{jk}^{(m)}$ is defined in and $$\begin{aligned}
\label{eq coeff gamma}\gamma_{jk}^{(m)}=\int g_{jk}(y) \mathbf{P}_\Delta[f]^{\star m}(y)dy.\end{aligned}$$
For $m\geq 1$, $\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}$ is the sum of $\lfloor N_T/m\rfloor$ identically distributed random variables, where $N_T$ is random. First we replace $N_T$ by its deterministic asymptotic limit using the following decomposition $$\begin{aligned}
{\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p\big]= & {\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p\mathds{1}_{\big\{\big|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\big|\geq\lambda\big\}}\big]\\&+{\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p \mathds{1}_{\big\{|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\big|<\lambda\big\}}\big].\end{aligned}$$ Take $\lambda=1/4\mu$ and denote $n_m=\big\lfloor T/m\mu\big\rfloor$ and $n'_m=\big\lfloor T/(4m\mu)\big\rfloor$, we have with that that$$\begin{aligned}
{\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p\big]
\leq& 2^{jp/2}\|g\|_\infty {\mathbb{P}}\Big(\Big|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\Big|\geq\frac{\Delta_T}{4\mu}\Big)\\
&+{\mathbb{E}}\Big[\Big|\frac{1}{n'_m}\sum_{i=1}^{n_m}\big(\mathbf{D}^{\Delta}_m X_{S_i}-{\mathbb{E}}\big[\mathbf{D}^{\Delta}_m X_{S_i}\big] \big)\Big|^p\Big].\end{aligned}$$ For the first term of the right hand part, $T\Delta_T=O(T^\delta)$, Lemma \[lem Nconcentre Renewal\] and $2^{j}\leq T$ leads to $$\begin{aligned}
{\mathbb{P}}\Big(\big|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\big|\geq\frac{\Delta_T}{4\mu}\Big)& \leq \mathfrak{C}T^{p/2}\exp\Big(-\mathfrak{C}T^\delta\Big)\leq \mathfrak{C}\exp(-T^{\delta'_p})\label{eq prf Ros C1},\end{aligned}$$ for some $\delta'_p<\delta$ and where $\mathfrak{C}$ depends on $p,\|g\|_\infty,\mathfrak{A},\mathfrak{a},\mu$. For the second term we apply Rosenthal’s inequality . Since $X$ is a renewal process the variables $\big(\mathbf{D}^{\Delta}_m X_{S_{2i}}\big)_i$ are independent but dependent of the variables $\big(\mathbf{D}^{\Delta}_m X_{S_{2i+1}}\big)_i$ which are independent. It ensures that the variables $\big(\mathbf{D}^{\Delta}_m X_{S_i}\big)$ are distributed according to $\mathbf{P}_{\Delta_T}[f]^{\star m}$. Moreover if we separate the sum $\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}$ between odd and even indices we can apply Rosenthal’s inequality for independent variables to each sum. For $p\geq1$ we have by convex inequality $$\begin{aligned}
{\mathbb{E}}\big[\big|g_{jk}(\mathbf{D}^{\Delta}_m X_{S_i})-\gamma_{jk}\big|^p\big]&\leq 2^p {\mathbb{E}}\big[\big|g_{jk}(\mathbf{D}^{\Delta}_m X_{S_i})\big|^p\big]\\&\leq 2^p2^{jp/2}\int |g(2^jy-k)|^p\mathbf{P}_{\Delta_T}[f]^{\star m}(y)dy \\&\leq 2^p2^{j(p/2-1)}\int |g(z)|^p\mathbf{P}_{\Delta_T}[f]^{\star m}\Big(\frac{z+k}{2^j}\Big)dz,\end{aligned}$$ where we made the substitution $z=2^jy-k$. Lemma \[lem loicontY boule besov\] and Sobolev embeddings give $\big\|\mathbf{P}_{\Delta_T}[f]^{\star m}\big\|_\infty\leq \mathfrak{M}$. It follows that $${\mathbb{E}}\big[\big|g_{jk}(\mathbf{D}^{\Delta}_m X_{S_i})-\gamma_{jk}\big|^p\big]\leq 2^p2^{j(p/2-1)}\|g\|_{L_p({\mathbb{R}})}^p\mathfrak{M}$$ and $${\mathbb{E}}\big[\big|g_{jk}(\mathbf{D}^{\Delta}_m X_{S_i})-\gamma_{jk}\big|^2\big]\leq \mathfrak{M}$$ since $\|g\|_{L_2({\mathbb{R}})}=1$. We derive for $p\geq1$ $$\begin{aligned}
{\mathbb{E}}\Big[\Big|\frac{1}{n'_m}\sum_{i=1}^{n_m}&\big(\mathbf{D}^{\Delta}_m X_{S_i}-{\mathbb{E}}\big[\mathbf{D}^{\Delta}_m X_{S_i}\big] \big)\Big|^p\Big]\nonumber\\ &\leq \mathfrak{C}_p \Big\{2^p\Big(\frac{2^{j}}{ n_m}\Big)^{\frac{p}{2}-1}\|g\|_{L_p({\mathbb{R}})}^p\mathfrak{M}+\mathfrak{M}^{p/2}\Big\}{n_m'}^{-p/2}\nonumber\\
&\leq \mathfrak{C}_{p,m,\|g\|_{L_p({\mathbb{R}})},\mathfrak{M},\mu}T^{-p/2}.\label{eq prf Ros C2}\end{aligned}$$ It follows from and that$$\begin{aligned}
{\mathbb{E}}\big[\big|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}\big|^p\big]\leq & \mathfrak{C}\exp(-T^{\delta'_p})+\mathfrak{C}T^{-\frac{p}{2}}\leq \mathfrak{C}T^{-\frac{p}{2}},\end{aligned}$$ since the first term is negligible in front of the second as $\delta'>0$ where $\mathfrak{C}$ depends on $p,m,\|g\|_{L_p({\mathbb{R}})},\|g\|_\infty,\mathfrak{A},\mathfrak{a},\mathfrak{M},\mu$. It concludes the proof.
\[lem Bern C\] Choose $j$ and $c$ such that $$2^jT^{-1}\log(T^{1/2})\leq1\mbox{ and }c^2\geq 256m\mu\Big(\mathfrak{M}+\frac{c\|g\|_\infty}{24}\Big).$$ For all $r\geq 1$ let $\kappa_r=cr$. We have for all $m\geq1$ $${\mathbb{P}}\Big(|\widehat{\gamma}_{jk}^{(m)}-\gamma_{jk}^{(m)}|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\leq T^{-r/2},$$ where $\widehat{\gamma}_{jk}^{(m)}$ is defined in and $\gamma_{jk}^{(m)}$ in .
As for the proof of Lemma \[lem Ros C\] we decompose as follow for $m\geq1$ $$\begin{aligned}
{\mathbb{P}}\Big(|\widehat{\gamma}_{jk}^{(m)}&-\gamma_{jk}^{(m)}|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\\&\leq {\mathbb{P}}\Big(\Big|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\Big|\geq\frac{\Delta_T}{4\mu}\Big)\\ &+ {\mathbb{P}}\Big(\big|\frac{1}{n'_m}\sum_{i=1}^{n_m}\big(\mathbf{D}^{\Delta}_m X_{S_i}-{\mathbb{E}}\big[\mathbf{D}^{\Delta}_m X_{S_i}\big] \big)\big|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big),\end{aligned}$$ where $n_m=\big\lfloor T/m\mu\big\rfloor$ and $n'_m=\big\lfloor T/(4m\mu)\big\rfloor$. From $T\Delta_T^2=O(T^\delta)$ and Lemma \[lem Nconcentre Renewal\] we derive $$\begin{aligned}
\label{eq prf Bern C1}{\mathbb{P}}\Big(\Big|\frac{N_T}{\lfloor T\Delta_T^{-1}\rfloor}-p(\Delta_T)\Big|\geq\frac{\Delta_T}{4\mu}\Big)& \leq \exp\Big(-\mathfrak{C}T^\delta\Big),\end{aligned}$$ where $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\mu$. For the second term we apply Bernstein’s inequality and as in the proof of Lemma \[lem Ros C\] we separate the sum between odd and even indices to work with independent variables. We get $$\begin{aligned}
{\mathbb{P}}\Big(\Big|\frac{1}{n'_m}\sum_{i=1}^{n_m}&\big(\mathbf{D}^{\Delta}_m X_{S_i}-{\mathbb{E}}\big[\mathbf{D}^{\Delta}_m X_{S_i}\big] \big)\Big|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)\\
\leq&2\exp\Bigg(-\frac{\kappa_r^2{n'_m}^2T^{-1}\log(T^{1/2})} {16\Big(n_m\mathfrak{M}+\frac{\kappa_rn'_mT^{-1/2}\sqrt{\log(T^{1/2})}2^{j/2}\|g\|_\infty}{6}\Big)}\Bigg)\\
\leq&2\exp\Bigg(-\frac{c^2r} {128m\mu\Big(\mathfrak{M}+\frac{\kappa_rT^{-1/2}\sqrt{\log(T^{1/2})}2^{j/2}\|g\|_\infty}{24}\Big)}r\log(T^{1/2})\Bigg).
\end{aligned}$$ With $2^jT^{-1}\log(T^{1/2})\leq1$ and $c^2\geq 256m\mu\big(\mathfrak{M}+\frac{c\|g\|_\infty}{24}\big)$ we have for $r\geq1$ $$\begin{aligned}
\label{eq prf Bern C2}
{\mathbb{P}}\Big(\Big|\frac{1}{n'_m}\sum_{i=1}^{n_m}&\big(\mathbf{D}^{\Delta}_m X_{S_i}-{\mathbb{E}}\big[\mathbf{D}^{\Delta}_m X_{S_i}\big] \big)\Big|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)
\leq v(T)^r. \end{aligned}$$It follows from and that $$\begin{aligned}
{\mathbb{P}}\Big(|\widehat{\gamma}_{jk}^{(m)}&-\gamma_{jk}^{(m)}|\geq \frac{\kappa_r}{2}T^{-1/2}\sqrt{\log(T^{1/2})}\Big)&\leq \exp\Big(-\mathfrak{C}T^\delta\Big) +T^{-r/2}\leq T^{-r/2}\end{aligned}$$ since the first term is negligible in front of the second since $\delta>0$. It concludes the proof.
It is a consequence of Lemma \[lem loicontY boule besov\], \[lem Ros C\], \[lem Bern C\] and of the general theory of wavelet threshold estimators of Kerkyacharian and Picard [@KP00]. It suffices to have conditions (5.1) and (5.2) of Theorem 5.1 of [@KP00], which are satisfied –Lemma \[lem Ros C\] and \[lem Bern C\]– with $c(T)=T^{-1/2}$ and $\Lambda_n=c(T)^{-1}$ (with the notation of [@KP00]). We can now apply Theorem 5.1, its Corollary 5.1 and Theorem 6.1 of [@KP00] to obtain the result.
### Completion of the proof of Theorem \[thm Renewal 2\] {#completion-of-the-proof-of-theorem-thm-renewal-2 .unnumbered}
To prove Theorem \[thm Renewal 2\] we define for $K$ in ${\mathbb{N}}$ and $x$ in $\mathcal{D}$ the quantity $$\begin{aligned}
\widetilde{f}^K_{T,\Delta}(x)&=\sum_{m=1}^{K+1} l_m(\Delta,\vartheta) \widehat{P_{\Delta, m}}(x).
\end{aligned}$$ It is the estimator of $f$ one would compute if $\tau$ were known. We decompose the $L_p$ error as follows $$\begin{aligned}
\big({\mathbb{E}}\big[\|\widehat{f}^K_{T,\Delta_T}-f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\leq& \big({\mathbb{E}}\big [\|\widehat{f}^K_{T,\Delta_T}-\widetilde{f}^K_{T,\Delta_T}\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\\ &+\big({\mathbb{E}}\big[\|\widetilde{f}^K_{T,\Delta_T}-f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p},\end{aligned}$$ and control each term separately.
First we look at the second term $$\begin{aligned}
\big({\mathbb{E}}\big[ \| \widetilde{f}^K_{T,\Delta_T} -f\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\leq&\big({\mathbb{E}}\big[\big\|\widetilde{f}^K_{T,\Delta_T}-\mathbf{L}_{\Delta_T,K} \|_{L_p(\mathcal{D})}^p\big]\big)^{1/p} \nonumber\\&+\big\|\mathbf{L}_{\Delta_T,K} -\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big\|_{L_p(\mathcal{D})}\nonumber\\&+\big\|\mathbf{H}_{\Delta,f}^{ \circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] -f\big\|_{L_p(\mathcal{D})}.\label{eq prf thm2 1}\end{aligned}$$ An upper bound for the first term is given by part 1) of Theorem \[thm Renewal 2\], given the definition of $\mathbf{L}_{\Delta_T,K}$ and Triangular’s inequality we derive $$\begin{aligned}
\label{eq prf thm2 2}\big({\mathbb{E}}\big[\big\|\widetilde{f}^K_{T,\Delta_T}-\mathbf{L}_{\Delta_T,K} \|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\leq \mathfrak{C}T^{-\alpha(s,p,\pi)},\end{aligned}$$ where $\mathfrak{C}$ depends on $\vartheta,s,\pi,p,\mathfrak{M},\phi,\psi,$ and $K$. By , we have $$\begin{aligned}
\label{eq prf thm2 3}\big\|\mathbf{L}_{\Delta_T,K} -\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big\|_{L_p(\mathcal{D})}\leq\mathfrak{C}\Delta_T^{K+1},\end{aligned}$$ where $\mathfrak{C}$ depends on $\vartheta$, $\mathfrak{O}$ and $\mathfrak{M}$. For the last term we use the fixed point theorem’s approximation, first we have to relate the $L_p$ norm with the Sobolev one. Triangular’s inequality ensures that if $f$ is in $\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})$ then $\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] -f$ is in $\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})$. It follows using Sobolev embeddings that $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] -f\big\|_{L_p(\mathcal{D})}&\leq\big\|\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] -f\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}.\end{aligned}$$ We now use the approximation given by the Banach fixed point theorem $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] &-f\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq\mathfrak{K}(\Delta_T)^{K} \big\|\mathbf{H}_{\Delta,f}\big[\mathbf{P}_{\Delta_T}[f]\big] -\mathbf{P}_{\Delta_T}[f]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}.\end{aligned}$$ After replacing $\mathbf{H}_{\Delta,f}\big[\mathbf{P}_{\Delta_T}[f]\big] $ by its expression and using triangular’s inequality we have $$\begin{aligned}
\big\|\mathbf{H}_{\Delta,f}\big[\mathbf{P}_{\Delta_T}[f]\big] -\mathbf{P}_{\Delta_T}[f]\big\|_{\mathcal{B}^s_{{\pi} \infty}(\mathcal{D})}\leq \mathfrak{C}\Delta_T,\end{aligned}$$which leads to $$\begin{aligned}
\label{eq prf thm2 4}\big\|\mathbf{H}_{\Delta,f}^{ \circ K}\big[\mathbf{P}_{\Delta_T}[f]\big] -f\big\|_{L_p(\mathcal{D})}\leq \mathfrak{C}\Delta_T\mathfrak{K}(\Delta_T)^{K},\end{aligned}$$ $\mathfrak{C}$ depends on $\vartheta,\mathfrak{M},\mathfrak{O},K$. We conclude by injecting , , and in and taking the supremum in $\vartheta$ over the compact set $\Theta$.
We now control ${\mathbb{E}}\big [\|\widehat{f}^K_{T,\Delta_T}-\widetilde{f}^K_{T,\Delta_T}\|_{L_p(\mathcal{D})}^p\big] $, the triangle inequality leads to $$\begin{aligned}
\big({\mathbb{E}}\big[\|\widehat{f}^K_{T,\Delta_T}-&\widetilde{f}^K_{T,\Delta_T}\|_{L_p(\mathcal{D})}^p \big]\big)^{1/p}\\&\leq \sum_{m=1}^{K+1}\big({\mathbb{E}}\big[\|\big(l_m(\Delta_T,\widehat{\vartheta_T})-l_m(\Delta_T,\vartheta)\big)\widehat{P_{\Delta_T,m}}\|_{L_p(\mathcal{D})}^p \big]\big)^{1/p},\end{aligned}$$ where $\widehat{P_{\Delta_T,m}}$ does not depend on $\vartheta$ (see ). Cauchy-Schwarz inequality leads to $$\begin{aligned}
{\mathbb{E}}\big[\|\big(l_m(\Delta_T,\widehat{\vartheta_T})-&l_m(\Delta_T,\vartheta)\big)\widehat{P_{\Delta_T,m}}\|_{L_p(\mathcal{D})}^p \big]^2\\&\leq {\mathbb{E}}\Big[\big|l_m(\Delta_T,\widehat{\vartheta_T})-l_m(\Delta_T,\vartheta)\big|^{2p}\Big] {\mathbb{E}}\Big[\big\|\widehat{P_{\Delta_T,m}}\big\|_{L_{p}(\mathcal{D})}^{2p}\Big] ,\end{aligned}$$ where using part 1) of Theorem \[thm Renewal 2\], the triangle inequality and that $T\geq1$ we have $$\begin{aligned}
{\mathbb{E}}\Big[\big\|\widehat{P_{\Delta_T,m}}\big\|_{L_{p}(\mathcal{D})}^{2p}\Big]& \leq {\mathbb{E}}\Big[\|\widehat{P_{\Delta_T,m}}-\mathbf{P}_{\Delta_T}[f]^{\star m}\|_{L_{p}(\mathcal{D})}^{2p}\Big]+\|\mathbf{P}_{\Delta_T}[f]^{\star m}\|_{L_{p}(\mathcal{D})}^{2p}\nonumber\\&\leq \mathfrak{C}T^{-2{\alpha(s,p,\pi)} p}+\mathfrak{M}^{2p}\leq \mathfrak{C}\label{eq thm bound}\end{aligned}$$ where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi,\vartheta$. We conclude the proof with the following Lemma, proof of which is given in the Appendix.
\[lem lm ros\] Work under Assumptions \[ass param tau\] and \[ass queue tau\]. We have for all $r\geq2$ $$\begin{aligned}
{\mathbb{E}}\big[|l_m(\Delta_T,\widehat{\vartheta_T})-l_m(\Delta_T,\vartheta)&|^{r}\big]\leq \mathfrak{C}\big(T^{1-r}+T^{-r/2}\big)\end{aligned}$$ where $\mathfrak{C}$ depends on $r,\mathfrak{A},\mathfrak{a},\vartheta.$
It follows from and Lemma \[lem lm ros\] applied with $r=2p$ that $$\begin{aligned}
{\mathbb{E}}\big[\|\widehat{f}^K_{T,\Delta_T}-\widetilde{f}^K_{T,\Delta_T}&\|_{L_p(\mathcal{D})}\big]^{1/p}\leq \mathfrak{C}\big(T^{1-1/(2p)}+T^{-1/2}\big) ,\end{aligned}$$ where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi,\mathfrak{A},\mathfrak{a},\vartheta.$ We deduce for $p\geq 1$ $$\begin{aligned}
\underset{\vartheta\in \Theta }{\sup}\underset{f\in\mathcal{F}(s,{\pi},\mathfrak{M})}{\sup}\big({\mathbb{E}}\big[\|\widehat{f}^K_{T,\Delta_T} \big(\widehat{\vartheta}\big)-\widehat{f}^K_{T,\Delta_T}&\|_{L_p(\mathcal{D})}^p\big]\big)^{1/p}\\&\leq \mathfrak{C}\big(T^{-(1-1/(2p))}+T^{-1/2}\big)\end{aligned}$$ where $\mathfrak{C}$ depends on $s,\pi,p,\mathfrak{M},\phi,\psi,\mathfrak{A},\mathfrak{a},K$. It is negligible compared to $T^{-{\alpha(s,p,\pi)}}$ since ${\alpha(s,p,\pi)}\leq1/2$. The proof of Theorem \[thm Renewal 2\] is now complete.
Appendix {#appendix .unnumbered}
========
Proof of Proposition \[PropDefOperator\] {#proof-of-proposition-propdefoperator .unnumbered}
----------------------------------------
Let $x\in{\mathbb{R}}$, we have by stationarity $$\begin{aligned}
{\mathbb{P}}(\mathbf{D}^{\Delta} X_{S_1}\leq x)&={\mathbb{P}}(X_\Delta\leq x|X_\Delta\ne0)\\&=\sum_{m=0}^\infty{\mathbb{P}}(X_\Delta\leq x|R_\Delta=m,R_\Delta\ne0) {\mathbb{P}}(R_\Delta=m) \\
&=\sum_{m=1}^\infty p_m(\Delta) {\mathbb{P}}(X_\Delta\leq x|R_\Delta=m) \end{aligned}$$ where $
{\mathbb{P}}(X_\Delta\leq x|R_\Delta=m)=\int_{-\infty}^xf^{\star m}(y)dy$ for $m\geq1$. It follows $$\begin{aligned}
{\mathbb{P}}(\mathbf{D}^{\Delta} X_{S_1}\leq x)&=\int_{-\infty}^x \mathbf{P}_\Delta[f](y)dy.\end{aligned}$$
Proof of Lemma \[lem pmControl\] {#proof-of-lemma-lem-pmcontrol .unnumbered}
--------------------------------
We start with the second assertion. For $m\geq 1$ we have $$p_m(\Delta)=\frac{{\mathbb{P}}(R_\Delta=m)}{1-{\mathbb{P}}(R_\Delta=0)}.$$ First we derive the lower bound $$\begin{aligned}
\label{eq pmC1}1-{\mathbb{P}}(R_\Delta=0)&=1-{\mathbb{P}}(J_1\geq \Delta) \geq\frac{1}{\mu}\int_0^\Delta1-F(\Delta)dx\geq \frac{\Delta}{2\mu}
\end{aligned}$$ since $F$ is a cumulative distribution function; it is positive, increasing and continuous with $F(0)=0$. Then there exists $\Delta_1$ such that for all $\Delta\leq\Delta_1$ we have $F(\Delta)\leq\frac{1}{2}$. Second we have for all $m\geq1$ $$\begin{aligned}
{\mathbb{P}}(R_\Delta=m)
&\leq {\mathbb{P}}(J_1+\ldots+J_{m}\leq \Delta)=\int_0^\Delta \tau_0\star \tau^{\star m-1}(x)dx,\end{aligned}$$ where for all $x\in [0,\Delta]$ $$\begin{aligned}
\tau_0&\star \tau^{\star m-1}(x)=x^{m-1}\int_0^1\tau_0(xt_1)\int_0^{1-t_1}\tau(xt_2)\ldots \\ &\int_0^{1-t_1-\ldots-t_{m-2}}\tau(xt_{m-1})\tau(x(1-t_1-\ldots-t_{m-2}-t_{m-1}))dt_1\ldots dt_{m-1}.\end{aligned}$$ We derive $$\begin{aligned}
\tau_0&\star \tau^{\star m-1}(x)\nonumber\\&\leq x^{m-1} \underset{t\in[0,x]}{\sup}\tau_0(t)\big(\underset{t\in[0,x]}{\sup}\tau(t)\big)^{m-1}\int_0^1\int_0^{1-t_1}\ldots \int_0^{1-t_1-\ldots-t_{m-2}}dt_1\ldots dt_{m-1}\nonumber\\
&\leq \frac{1}{\mu}\big(\underset{t\in[0,\Delta]}{\sup}\tau(t)\big)^{m-1} \frac{x^{m-1}}{(m-1)!},$$ since $$\int_0^1\int_0^{1-t_1}\ldots \int_0^{1-t_1-\ldots-t_{m-2}}dt_1\ldots dt_{m-1}=\frac{1}{(m-1)!}.$$ It follows that $$\begin{aligned}
\label{eq pmC2}
{\mathbb{P}}(R_\Delta=m)&\leq \frac{1}{\mu}\big(\underset{t\in[0,\Delta]}{\sup}\tau(t)\big)^{m-1}\frac{\Delta^{m}}{m!}.\end{aligned}$$ Since $\tau$ is continuous, there exists $\Delta_2$ such that $$\underset{t\in[0,\Delta_2]}{\sup}\tau(t)\leq 2\tau(0).$$ Taking $\Delta_0=\Delta_1 \wedge \Delta_2$, and lead to the second assertion. The first one is straightforward from the previous computations.
Proof of Proposition \[prop inverseTronqueeK\] {#proof-of-proposition-prop-inversetronqueek .unnumbered}
----------------------------------------------
According to the definition of $\mathbf{L}_{\Delta,K}$ inequality is immediate. The dependency in $\tau(0)$ and $\mathfrak{M}$ of the constant is a consequence of Lemma \[lem pmControl\], part 2) of Proposition \[prop complet contract\] and Lemma \[lem loicontY boule besov\]. A rearrangement of the terms enables to write $\mathbf{L}_{\Delta,K}$ as a sum of increasing powers of ${\mathbf{P}_\Delta[f]}^{\star m}$. Thus we have to prove that only the $K+1$ first convolution powers of ${\mathbf{P}_\Delta[f]}$ intervene and that the coefficient $l_m(\Delta)$ in front of ${\mathbf{P}_\Delta[f]}^{\star m}$ in the rearrangement satisfies $$\big|l_m(\Delta)\big|\leq \mathfrak{C}_{\tau(0)}\Delta^{m-1}.$$
For that we show that for all $L\geq 1$ the Taylor expansion of order $L$ in $\Delta$ of $\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]$, that we denote $\widetilde{\mathbf{L}}_{\Delta,K,L}$, only depends on ${\mathbf{P}_\Delta[f]}^{\star m}$, $m=1,\ldots,L+1$ with coefficients such that $\widetilde l_{m,K}(\Delta)\leq \mathfrak{C}_{\tau(0)}\Delta^{m-1}.$ We prove the result by induction on $K$. For $K=1$ we immediately have the result by Lemma \[lem pmControl\] since $$\begin{aligned}
\mathbf{H}_{\Delta,f}\big[\mathbf{P}_\Delta[f]\big]&= 2\mathbf{P}_\Delta[f]-\sum_{m=1}^\infty p_m(\Delta)\mathbf{P}_\Delta[f]^{\star m},
\end{aligned}$$ it follows that $$\widetilde{\mathbf{L}}_{\Delta,L,1}=(2-p_1(\Delta)\mathbf{P}_\Delta[f]-\sum_{m=2}^{L+1}p_m(\Delta)\mathbf{P}_\Delta[f]^{\star m}$$ with $\widetilde l_{1,1}(\Delta)=(2-p_1(\Delta))\leq 2$ and $\widetilde l_{m,1}(\Delta)=p_m(\Delta)\leq \mathfrak{C}_{\tau(0)}\Delta^{m-1}$. Then using the definition of $\mathbf{H}_{\Delta,f}$ we have $$\begin{aligned}
\mathbf{H}_{\Delta,f}^{\circ (K+1)}\big[\mathbf{P}_\Delta[f]\big]=\mathbf{P}_\Delta[f]+\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]-\sum_{m=1}^\infty p_m(\Delta)\Big(\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]\Big)^{\star m}.\end{aligned}$$ The induction hypothesis and Lemma \[lem pmControl\], with part 2) of Proposition \[prop complet contract\] which ensures that $\mathbf{H}_{\Delta,f}^{\circ K}\big[\mathbf{P}_\Delta[f]\big]\in\mathcal{H}(s,\pi,\mathfrak{O},\mathfrak{N}) $, lead to $$\begin{aligned}
\widetilde{\mathbf{L}}_{\Delta,L,K+1}&=\mathbf{P}_\Delta[f]+ \widetilde{\mathbf{L}}_{\Delta,L,K}-\sum_{m=1}^{L+1} p_m(\Delta)\Big(\widetilde{\mathbf{L}}_{\Delta,L,K}\Big)^{\star m}\\
&=\mathbf{P}_\Delta[f]+ \sum_{m=1}^{L+1}\widetilde l_{m,L}(\Delta)\mathbf{P}_\Delta[f]^{\star m}-\sum_{m=1}^{L+1} p_m(\Delta)\Big(\sum_{m'=1}^{L+1}\widetilde l_{m',L}(\Delta)\mathbf{P}_\Delta[f]^{\star m'}\Big)^{\star m}\\&=
\sum_{m=1}^{L+1}\widetilde l_{m,L+1}(\Delta)\mathbf{P}_\Delta[f]^{\star m},
\end{aligned}$$ where $ \widetilde l_{1,L+1}(\Delta)=1$ and $$\begin{aligned}
\widetilde l_{m,L+1}(\Delta)&=\widetilde l_{m,L}(\Delta)-\sum_{k=1}^mp_k(\Delta)\sum_{n_1+\ldots+n_k=m}\widetilde l_{n_1,L}(\Delta)\ldots \widetilde l_{n_k,L}(\Delta)
\end{aligned}$$ which we bound with Lemma \[lem pmControl\] and the induction hypothesis by $$\begin{aligned}
\big|\widetilde l_{m,L+1}(\Delta)\big|&\leq \mathfrak{C}\Big(\Delta^{m-1}+\sum_{k=1}^m\Delta^{k-1}\sum_{n_1+\ldots+n_k=m}\widetilde \Delta^{n_1-1}\ldots \Delta^{n_k-1}\Big)\\
&=\mathfrak{C}\Big(\Delta^{m-1}+m\Delta^{m-1}\Big)\leq\mathfrak{C}\Delta^{m-1},
\end{aligned}$$ where $\mathfrak{C}$ is a positive constant depending on $\tau(0)$ and $K$. We conclude the proof having $L=K$ and $l_m(\Delta)=\widetilde l_{m,K}$ for $m=1,\ldots,K+1$.
Proof of Lemma \[lem lm ros\] {#proof-of-lemma-lem-lm-ros .unnumbered}
-----------------------------
### Preliminary {#preliminary-1 .unnumbered}
\[lem theta ros\]Work under assumptions \[ass queue tau\] and \[ass param tau\], for all $r\geq2$ $$\begin{aligned}
{\mathbb{E}}\big[|\widehat{\vartheta_T}-\vartheta|^r\big]&\leq \mathfrak{C}\big(T^{1-r}+T^{-r/2}\big),\end{aligned}$$ where $\mathfrak{C}$ depends on $r,\mathfrak{A},\mathfrak{a},\vartheta$ and $\widehat{\vartheta_T}$ is defined in Definition \[def est corr K\].
Let $r>2$, the proof is a consequence of Proposition 5.5 of Dedecker *et al.* [@Doukhan] which is a Rosenthal type inequality for dependent data. Define $$S_T=\sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor}Y_1$$ where $S_0=X_0=0$ and the $Y_i=\mathds{1}_{\mathbf{D}^\Delta X_i\ne0}-q(\vartheta)$ are centered identically distributed random variables bounded by 1. To apply Proposition 5.5 of [@Doukhan] we have to verify that $(Y_i)$ is a sequence of $\theta_{1,\infty}-$dependent random variables. For that Proposition 2.3 of [@Doukhan] ensures that it is sufficient to have a $\theta-$dependent sequence which is defined as follows with notation of [@Doukhan]; Let $\Gamma(u,v,k)$ be the set of $(i,j)$ in ${\mathbb{Z}}^u\times{\mathbb{Z}}^v$ such that $$i_1<\ldots<i_u\leq i_u+k<j_1<\ldots<j_v,$$ we have to show that for all $f \in\mathcal{F}_u$ the set of bounded function from ${\mathbb{R}}^u$ to ${\mathbb{R}}$ and for all $g\in \mathcal{G}_v$ the set of Lipschitz function from ${\mathbb{R}}^v$ to ${\mathbb{R}}$ with Lipschitz coefficient denoted $\mbox{Lip}g$ the sequence $\theta(k)$ defined as $$\theta(k)=\underset{u,v}{\sup}\underset{(i,j)\in\Gamma(u,v,k)}{\sup}\underset{f\in \mathcal{F}_u,g\in\mathcal{G}_v}{\sup} \frac{\big|Cov\big(f(Y_{i_1},\ldots,Y_{i_u}),g(Y_{j_1},\ldots,Y_{j_v})\big)\big|}{v\|f\|_\infty\mbox{Lip}g}$$ tends to 0. We denote as $Y_i$ and $Y_j$ respectively $(Y_{i_1},\ldots,Y_{i_u})$ and $(Y_{j_1},\ldots,Y_{j_v})$, and due to the fact that $X$ is a renewal process $Y_i$ and $Y_j$ are independent if there exists $r$ such that $i_u< r <j_1$ and $Y_r=1-p(\Delta)$ *i.e* there is a jump between $Y_{i_u}$ and $Y_{j_1}$. We denote by $A$ the event ”there exists $r$ such that $i_u< r <j_1$ and $Y_r=1-p(\Delta)$. It follows that $$\begin{aligned}
\big| Cov\big(f(X_i),g(X_j)\big)\big|&= \big|{\mathbb{E}}\big[\big(f(X_i)-{\mathbb{E}}[f(X_i)]\big)\big(g(X_j)-g(0_j)\big)\mathds{1}_{\{A\}}\big]\big|\\
&\leq 2\|f\|_\infty \mbox{Lip}g{\mathbb{E}}\big[\|X_j\|\mathds{1}_{\{A\}}\big]\\&\leq 2v\|f\|_\infty \mbox{Lip}g{\mathbb{P}}(R_{k\Delta}\ne 0),\end{aligned}$$ since $\|X_j\|\leq v$ as the $Y_i$ are bounded by 1, for every $L_p$ norm $p\geq0$, and ${\mathbb{E}}[\mathds{1}_{\{A\}}]$ is bounded by ${\mathbb{P}}(R_{k\mathfrak{T}}\ne 0)$. We immediately derive that $\theta(k)\leq 2{\mathbb{P}}(R_{k\Delta}\ne 0)$ and by Assumption \[ass queue tau\] we derive $$\begin{aligned}
\label{eq prf lem th01}\theta(k)\leq \mathfrak{C}\exp(-\mathfrak{a} (k\Delta)^\mathfrak{g'}),\end{aligned}$$ where $\mathfrak{g}<\mathfrak{g'}$, it tends to 0. We verify the hypothesis of Proposition 5.5 of [@Doukhan] and get for all $r>2$ $$\begin{aligned}
{\mathbb{E}}\big[|S_T|^r\big]&\leq \mathfrak{C}\Big(\lfloor T \Delta^{-1}\rfloor\sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor} i^{r-2}\theta(i)+\big(\lfloor T \Delta^{-1}\rfloor\sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor}\theta(i)\big)^{r/2}\Big)\end{aligned}$$ where $\mathfrak{C}$ depends on $r$. Since we have the upper bound , we derive applying with $k=0$ and $k=r-2$ $$\sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor}\theta(i)\leq \mathfrak{C}\Delta\ \ \ \mbox{ and }\ \ \ \sum_{i=1}^{\lfloor T \Delta^{-1}\rfloor} i^{r-2}\theta(i)\theta(i)\leq \mathfrak{C}\Delta^{r-1}$$ where $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\vartheta.$ It follows $$\frac{1}{\lfloor T \Delta^{-1}\rfloor^r}{\mathbb{E}}\big[|S_T|^r\big]\leq \mathfrak{C} \big(T^{1-r}+T^{-r/2}\big),$$ where $\mathfrak{C}$ depends on $r,\mathfrak{A},\mathfrak{a},\vartheta.$ The case $r=2$ is a consequence of $${\mathbb{E}}\big[|S_T|^2\big]=\frac{1}{\lfloor T \Delta^{-1}\rfloor}{\mathbb{V}}(Y_1)+\frac{2}{\lfloor T \Delta^{-1}\rfloor^2}\sum_{1\leq i<j\leq T}Cov(Y_i,Y_j)$$ and the upper bounds ${\mathbb{V}}(Y_1)\leq\mathfrak{C}\Delta$ where $\mathfrak{C}$ depends on $\vartheta$ and $$|Cov(Y_i,Y_{i+k})|\leq \mathfrak{C}\exp(-\mathfrak{a} k\Delta).$$ We derive $$\frac{1}{\lfloor T \Delta^{-1}\rfloor^2}{\mathbb{E}}\big[|S_T|^2\big]=\mathfrak{C}T^{-1}$$ where $\mathfrak{C}$ depends on $\mathfrak{A},\mathfrak{a},\vartheta.$ We conclude the proof using Assumption \[ass param tau\], for all $r\geq2$ $$\begin{aligned}
{\mathbb{E}}\big[|\widehat{\vartheta_T}-\vartheta|^r\big]&={\mathbb{E}}\big[|q^{-1}(q(\widehat{\vartheta_T}))-q^{-1}(q(\vartheta))|^r\big]\\
&\leq \|q^{-1}\|_\infty\ {\mathbb{E}}\big[|q(\widehat{\vartheta_T})-q(\vartheta)|^r\big]\leq \mathfrak{C}\big(T^{1-r}+T^{-r/2}\big),\end{aligned}$$where $\mathfrak{C}$ depends on $r,\mathfrak{A},\mathfrak{a},\vartheta.$
### Completion of the proof of Lemma \[lem lm ros\] {#completion-of-the-proof-of-lemma-lem-lm-ros .unnumbered}
The remaining of the proof is now based on the fact that under Assumption \[ass param tau\] the functions $\vartheta\rightarrow p_m(.,\vartheta)$ are Lipschitz continuous. We show that their derivative with respect to $\vartheta$ is bounded, we have for $m\geq1$ that $$\begin{aligned}
\partial_\vartheta[p_m(\Delta,\vartheta)]=&\frac{1}{\int_0^\Delta\tau_2(z,\vartheta)dz}\Big(\int_0^\Delta \partial_\vartheta[\tau_2(.,\vartheta)\star \tau_1^{\star m-1}(.,\vartheta)](z) dz\nonumber\\&-\int_0^\Delta \partial_\vartheta[\tau_2(.,\vartheta)\star \tau_1^{\star m}(.,\vartheta)](z)dz\Big)\nonumber\\&- \frac{\int_0^\Delta\partial_\vartheta[\tau_2(z,\vartheta)]dz}{\Big(\int_0^\Delta\tau_2(z,\vartheta)dz\Big)^2}\Big(\int_0^\Delta \tau_2(x,\vartheta)\star \tau_1^{\star m-1}(.,\vartheta)(z) dz\nonumber\\&-\int_0^\Delta \tau_2(x,\vartheta)\star \tau_1^{\star m}(.,\vartheta)(z)dz\Big)\label{eq prf lem 10 03}\end{aligned}$$ where $\tau_2(.,\vartheta)/\mu$ is the density of $J_1$. Immediate induction gives for $m\geq1$ $$\begin{aligned}
\partial_\vartheta[\tau_2(.,\vartheta)\star \tau_1^{\star m}(.,\vartheta)](z)=&\partial_\vartheta[\tau_2(.,\vartheta)]\star \tau_1^{\star m}(.,\vartheta)(z)\nonumber\\&+m\tau_2(.,\vartheta)\star \partial_\vartheta[\tau_1(.,\vartheta)]\star \tau_1^{\star m-1}(.,\vartheta)(z)\label{eq prf lem 10 04}\end{aligned}$$ and $$\begin{aligned}
\label{eq prf lem 10 05} \int_0^\Delta g^{\star m}(x)dx\leq \mathfrak{C}\Delta^{m-1}\end{aligned}$$ for some constant $\mathfrak{C}$ and any bounded function $g$ supported by $(0,\infty)$. Moreover we have $$\begin{aligned}
\partial_\vartheta[\tau_2(z,\vartheta)]=&-\int_0^z \partial_\vartheta \tau_1(x,\vartheta)dx,\end{aligned}$$ and it follows from Assumption \[ass param tau\] that for $\Delta$ small enough we have $\forall z\leq \Delta$ $$\begin{aligned}
\label{eq prf lem 10 06}
0<\frac{\tau_1(0,\vartheta)}{2}&\leq \tau_1(z,\vartheta)\leq 2\tau_1(0,\vartheta),\end{aligned}$$ and that its derivative is bounded over $[0,\Delta]$. Finally we bound , using and , we get $$\begin{aligned}
\big|\partial_\vartheta[p_m(\Delta,\vartheta)]\big|&\leq \mathfrak{C} \Delta^{m-1},\end{aligned}$$ where $\mathfrak{C}$ continuously depends on $\vartheta$. Then taking the supremum in $\vartheta$ over the compact set $\Theta$ we derive $$\begin{aligned}
\big|\partial_\vartheta[p_m(\Delta,\vartheta)]\big|&\leq \mathfrak{C} \Delta^{m-1},\end{aligned}$$ where $\mathfrak{C}$ is a positive constant independent of $\vartheta$. It follows that for $m\geq 1$, the functions $\vartheta\rightarrow p_m(.,\vartheta)$ are Lipschitz continuous and with Lemma \[lem theta ros\] we derive $$\begin{aligned}
{\mathbb{E}}\big[\big|p_m(\Delta,\widehat{\vartheta_T}) -p_m(\Delta,\vartheta)\big|^r\big]&\leq \mathfrak{C}\Delta^{m-1}{\mathbb{E}}\big[\big|\widehat{\vartheta_T}-\vartheta\big|^r\big]\\&\leq \mathfrak{C} \big(T^{1-r}+T^{-r/2}\big),\end{aligned}$$ where $\mathfrak{C}$ is a positive constant depending on $r,\mathfrak{A},\mathfrak{a},\vartheta.$ We conclude the proof using that $l_m(\Delta,\vartheta)=l(p_1(\Delta,\vartheta),\ldots,p_m(\Delta,\vartheta))$ where $l$ is Lipschitz in every argument and the argument are bounded by 1.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is a part of the author’s Ph.D thesis under the supervision of Marc Hoffmann whom I would like to thanks for his valuable remarks on this paper. The author’s research is supported by a PhD GIS Grant.
[99]{}
[^1]: GIS-CREST and CNRS-UMR 8050, 3, avenue Pierre Larousse, 92245 Malakoff Cedex, France.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $n$ be a positive integer. We show that a unit rational space vector whose multiple by $n$ is an integer vector can be extended to a rational orthonormal basis whose all members have the same property.'
author:
- 'Masanori Kobayashi, Chikara Nakayama'
title: On rational points of orthogonal group
---
Main results
============
A system $S : v_1,\ldots,v_k$ of vectors is called **orthoregular** if $(i)$ $v_i$ and $v_j$ are perpendicular to each other if $i \neq j$, and $(ii)$ $|v_i|=|v_j|\not=0$. We call $|v_i|$ the **length** of the system.
\[odd\] If $n$ is odd, the length of orthoregular $n$ vectors in $\Z^n$ is an integer.
Let $v_1, \ldots,v_n$ be orthoregular integer vectors and $l$ be its length. Then the determinant $|v_1, \ldots, v_n| = |v_1|\cdots|v_n|=l^n$ is an integer since all the entries are integers. Since $l^2$ is an integer, we have $l \in \Q$, and $l \in \Z$ by the normality of $\Z$.
Let $k$ be a positive integer less than $n$. We denote the maximum number of orthoregular integer vectors extending an orthoregular system $S \subset \Z^n$ by $E(S)$, and the minimum of those $E(S)$ for all orthoregular $S$ with $\sharp S=k$ by $E(n,k)$. For a subset $K$ of $\R$, we also use $E_K(n,k)$ for all orthoregular sets of $k$ integer vectors of length in $K$. We mainly use the cases $K=\R$ or $\Z$.
\[p:det\] If $n$ is even or $K=\Z$, $E_K(n,n-1)=n$.
Let $v_1,\ldots, v_{n-1}$ be an orthoregular system in $\Z^n$ of length $l$. Take a square matrix $(v_1, \ldots, v_{n-1}, v)$, where $v$ is an arbitrary vector. Let $v_n$ be the vector consists of the cofactors of the last column. Then $v_n$ is orthogonal to $v_1,\ldots,v_{n-1}$, thus $l^{n-1}|v_n|=|v_n|^2$ by cofactor expansion. Hence $|v_n|=l^{n-1}$.
We will now show that $(1/l^{n-2})v_n \in \Z^n$, which will complete the proof. It is enough to show that the last component of $v_n$ is in $l^{n-2}\Z$ since the others are shown similarly. We put $v_j =
\begin{pmatrix}
v'_j \\ x_j
\end{pmatrix}$ for $1 \leq j \leq n-1$, where $v'_j$ has $(n-1)$-rows and $x_j$ is a scalar. We show that $D := |v'_1, \ldots, v'_{n-1}|$ belongs to $l^{n-2}\Z$. Since $l^{n-2} \in \Z$ by the assumption that $n$ is even or $l \in \Z$, it is enough to prove that $D^2 \in l^{2(n-2)}\Z$. $$D^2 = |(v'_i,v'_j)_{1\leq i,j\leq n-1}|=|((v_i,v_j)-x_ix_j)_{i,j}|,$$ which is nothing but $\varphi(l^2)$, where $\varphi(t)$ is the characteristic polynomial for the matrix $(x_ix_j)_{i,j}$. Since the matrix has rank $\leq 1$, $\varphi(t)$ has a form $t^{n-2}(t-a)$ $(a \in \Z)$. Thus $D^2 \in l^{2(n-2)}\Z$.
The next is the main result in this paper.
\[t:main\] A nonzero integer vector $v$ can be extended to an integral orthoregular basis if and only if $|v|$ is an integer. In particular, $E_{\Z}(3,1)=3$.
Examples. Here is the list of smallest primitive nonnegative integral space vectors whose norms are integers.
$\t(0,0,1),
\t(1,2,2),
\t(0,3,4),
\t(2,3,6),
\t(1,4,8), \t(4,4,7),
\t(2,6,9), \t(6,6,7),....$
By starting from each vector in the list, you can always complete it to an orthoregular basis consisting of integral vectors. Dividing by its norm, you get a rational orthonormal basis, thus a rational point of $O(3)$.
Proof of main theorem
=====================
In this section, we prove \[t:main\]. Our proof is constructive and gives an algorithm for all solutions.
First we prove two lemmas.
\[l:ortho\] Let $v,w \in \Z^3\smallsetminus \{\bold o\}$. Assume $|v| \in \Z$ and $(v,w)=0$. Then $v_p(|w|^2)$ is even for any prime $p$ with $p\equiv3 \bmod 4$. Here $(-,-)$ is the standard inner product and $v_p$ is the normalized valuation with respect to $p$.
Let $v=
\begin{pmatrix}
a \\ b \\ c
\end{pmatrix},
w=
\begin{pmatrix}
x \\ y \\ z
\end{pmatrix}$ and let $l=|v|$. We may assume $a\not=0$. Since $ax+by+cz=0$, we have $x=\frac{-by-cz}a$. Hence $$|w|^2=\left(\frac{by+cz}a\right)^2+y^2+z^2,$$ $$a^2|w|^2=(a^2+b^2)y^2+2bcyz+(a^2+c^2)z^2.$$ Multiplying $a^2+b^2$, we deduce $$a^2(a^2+b^2)|w|^2=((a^2+b^2)y+bcz)^2+((a^2+c^2)(a^2+b^2)-b^2c^2)z^2.$$ But the coefficient of $z^2$ in the right-hand-side coincides with $a^2l^2$, so that the both terms of the right-hand-side are perfect squares. Hence $v_p$ of the right-hand-side is even for any $p$ with $p\equiv3 \bmod 4$. Since $v_p(a^2(a^2+b^2))$ is also even, $v_p(|w|^2)$ is even, too.
\[l:bilin\] Let $a,b,c,l \in \Z$ satisfying $ac-b^2=l^2$. Then there exist $x,y \in \Z$ such that $ax^2+2bxy+cy^2=l^2$ if and only if $a\geq0$ and, when $a>0$, $v_p(a)$ is even for any prime $p$ with $p\equiv3 \bmod 4$.
It is probably well-known. Our proof here uses the prime factorization in the integer ring $\Z[i]$, where $i=\sqrt{-1}$.
We prove the only if part. By multiplying $a$ with the given equation, we have $(ax+by)^2+(ly)^2=al^2$. Hence $l$ divides $ax+by$. Letting $u=\frac{ax+by}l$, we have $u^2+y^2=a$, which implies that the desired condition holds.
We prove the if part. The case $a=0$ is trivial. Assume $a>0$. If there are $u$ and $y$ with $u^2+y^2=a$ such that $a$ divides $ul-by$, then $x=\frac{ul-by}a$ and $y$ give an integer solution for $ax^2+2bxy+cy^2=l^2$. We prove the existence of $u$ and $y$ as above under the assumption that $v_p(a)$ is even for any prime $p$ with $p\equiv3 \bmod 4$. First factorize $a$ into $$a=2^m
\prod_{p_j\equiv1 \bmod 4}p_j^{m_j}
\prod_{q_k\equiv3 \bmod 4}q_k^{n_k},$$ where $p_j$’s are primes with $p_j\equiv1 \bmod 4$ and $q_k$’s are primes with $q_k\equiv3 \bmod 4$. Here, each $n_k$ is even by the assumption. Since $l^2+b^2=ac$, we can write $l+ib$ as a product $$l+ib=u_0(1+i)^{m'}
\prod p_j^{s_j}\pi_j^{t_j}
\prod q_k^{n'_k/2}$$ in $\Z[i]$, where $u_0$ is a unit of $\Z[i]$, $m' \geq m$, $p_j=\pi_j\overline \pi_j$ in $\Z[i]$, where $\pi_j$ and $\overline \pi_j$ are conjugate prime elements, $2s_j + t_j \geq m_j$, and $n'_k \geq n_k$ is an even.
Claim. There exist $a_j \geq0, b_j \geq0$ such that $s_j + \min(t_j, b_j) \geq a_j + b_j$ and $2a_j +b_j = m_j$.
Proof of Claim. If $2s_j \leq m_j$, let $a_j=s_j$ and $b_j = m_j -2s_j$. Then, we have $t_j \geq m_j -2s_j=b_j$ and $a_j$, $b_j$ satisfy the condition.
If $2s_j > m_j$, let $a_j := \lfloor \frac{m_j}2\rfloor$ and $b_j=m_j-2\lfloor \frac{m_j}2\rfloor$. Then, we have $b_j \leq 1$ and $s_j \geq a_j +1 \geq a_j + b_j$ and $a_j$, $b_j$ satisfy the condition. This completes the proof of Claim.
Using these $a_j$ and $b_j$, define $$u+iy:=(1+i)^m\prod p_j^{a_j}\overline \pi_j^{b_j} \prod q_k^{n_k/2}.$$ Then, since $2a_j+b_j=m_j$, we have $u^2+y^2=a$. Further, $ul-by$, the real part of $(u+iy)(l+ib)$, is divided by $2^m\prod p_j^{a_j+s_j+\min(b_j,t_j)}\prod q_k^{n_k}$. Since $a_j+s_j+\min(b_j,t_j) \geq 2a_j + b_j = m_j$, this is divided by $a$.
Now we prove \[t:main\]. The only if part is by \[odd\]. We prove the if part. By \[p:det\], it is enough to show $E_{\Z}(3,1) \geq 2$. Let $v_1=\begin{pmatrix}
a_1 \\ b_1 \\ c_1
\end{pmatrix}\in \Z^3\smallsetminus\{\bold o\}$ such that $l:=|v_1|$ is an integer. It suffices to show $E_{\Z}(\{v_1) \geq 2$. Since $m:=(a_1,b_1,c_1)$ divides $l$, replacing $v_1$ with $\frac {v_1}m$, we may and will assume that $m=1$. Then, the homomorphism $(v_1,-)\colon \Z^3 \to \Z$ is surjective. The kernel $K$ of this surjection is free of rank two. Let $Q$ be the positive definite bilinear form on $K$ obtained by restricting the standard metric on $\Z^3$. It is enough to show that $Q$ represents $l^2$. Take a base $w_1, w_2$ of $K$. Let $A:=$Gram($w_1,w_2$), the symmetic matrix which represents $Q$ with respect to $w_1, w_2$.
Claim. $\det(A)=l^2$.
Proof of Claim. Let $S$ be the area of the parallelogram spanned by $w_1, w_2$. What we have to see is $S=l$. Since there are isomorphisms $\Z^3/\langle v_1,w_1,w_2\rangle \cong
\Z/(v_1,v_1)\Z\cong
\Z/l^2\Z$, the volume of the parallelepiped spanned by $v_1, w_1, w_2$ is $l^2$. On the other hand, since $v_1$ is perpendicular to $K$, this volume is $|v_1|S$, too. Hence $S=l$. This completes the proof of Claim.
Let $a:=(w_1,w_1)$. By \[l:ortho\] with $v=v_1$ and $w=w_1$, $v_p(a)$ is even for any prime $p$ with $p\equiv3 \bmod 4$. Hence by \[l:bilin\] with $a=(w_1,w_1)$, $b=(w_1,w_2)$, $c=(w_1,w_2)$, we conclude that $Q$ represents $l^2$, which completes the proof of \[t:main\].
More results
============
Here we gather more results on $E_K(n,k)$.
For $0 < k < n$ and let $K$ be a subset of $\R$.
1. $E_{\Z}(n+1,k+1) \leq E_{\Z}(n,k)+1$.
2. If $n$ is odd, then $E(n,k) \leq n-1$.
3. If $n$ is odd and each coordinate of $v \in \Z^n$ is odd, then $E(v)=1$.
4. If $n$ is odd, then $E(n,1)=1$.
5. If $n \equiv 1 \bmod 8$, then $E_{\Z}(n,1)=1$.
6. If $n$ is the sum of positive integers $n_i$, $E_K(n,1) \geq \min_i E(n_i,1)$.
7. For a positive integer $m$, $E_K(mn,1) \geq E(n,1)$, $E_K(2m,1)\geq 2$, $E_K(4m,1)\geq 4$ and $E_K(8m,1) \geq 8$.
Question. Is $E_K(16m,1) \geq 16$ valid?
$(1)$ Take an orthoregular system $S:v_1, \ldots, v_k$ of vectors which attains $E_\Z(n,k)=E(S)$. Let $l$ be the length of $S$. Then $E \left(
\begin{pmatrix}
l \\ 0 \\
\end{pmatrix},
\begin{pmatrix}
0 \\ v_1 \\
\end{pmatrix}, \ldots,
\begin{pmatrix}
0 \\ v_k \\
\end{pmatrix} \right)
= E(v_1,\ldots, v_k)+1$.
$(2)$ It is enough to show that $E(v_1)\leq n-1$ for $v_1 = \t (1 \ 1 \ 0 \ \cdots \ 0) \in \Z^n$. This follows from $\ref{odd}$. (In fact, $E(v_1)=n-1$. This is seen by taking $v_2= \t (1 \ -1 \ 0 \ \cdots \ 0 ), \ldots, v_{2k-1}=\t (0 \ \cdots \ 0 \ 1 \ 1 \ 0 \ \cdots \ 0 ), v_{2k}=\t (0 \ \cdots \ 0 \ 1 \ -1 \ 0 \ \cdots \ 0),\ldots$ $(1 \leq k \leq (n-1)/2)$. Here the $i$-th components of $v_{2k-1}$ and of $v_{2k}$ are zero unless $i=2k-1,2k$.)
$(3)$ Suppose that an integer vector $v' = \t ( a_1 \ \cdots \ a_n)$ satisfies $(v,v')=0$ and $|v|=|v'|$. By taking modulo two, we have $|v|^2=|v'|^2=a_1^2+\cdots+a_n^2 \equiv a_1 + \cdots + a_n$. Since the components of $v$ are odd, the last integer is congruent to $(v,v')=0$, that is, $|v|^2 \equiv0 \bmod 2$. On the other hand, we have $|v|^2 \equiv n \equiv 1 \bmod 2$. A contradiction.
$(4)$ Take $v = \t(1 \cdots 1)$ and apply $(3)$.
$(5)$ Use $(3)$ and the following Claim. (The part $(b) \Rightarrow (a)$ suffices.)
Claim. For an odd integer $n>0$, the followings are equivalent.
$(a)$ There exist $n$ odd integers whose square sum is again square.
$(b)$ $n \equiv 1 \mod 8$.
Proof of Claim. $(a) \Rightarrow (b)$. Assume that there are $x_1,\ldots,x_n,y \in \Z$ such that $x_1^2+\cdots+ x_n^2=y^2$. Since the square of any odd number is congruent to $1$ modulo $8$, we have $n \equiv x_1^2+\cdots+ x_n^2=y^2 \bmod 8$. Since $y$ is also odd, this implies $n \equiv 1 \bmod 8$.
$(b) \Rightarrow (a)$. Take $m \geq1$ such that $(2m-1)^2 \leq n \leq (2m+1)^2$. If we prove $n \geq \frac {(2m+1)^2-n}8$, then, $(x_1,\ldots, x_n)=(3,\ldots,3,1\ldots,1)$ will do, where the number of $3$ is $\frac {(2m+1)^2-n}8$, for $x_1^2+\cdots+ x_n^2=n+(9-1)\frac {(2m+1)^2-n}8=(2m+1)^2$.
We prove the above inequality. This is equivalent to $9n \geq (2m+1)^2$, which reduces to $9(2m-1)^2 \geq (2m+1)^2$, $3(2m-1) \geq 2m+1$, $m\geq1$, which completes the proof of Claim.
$(6)$ Let $v \in \Z^n\smallsetminus\{\bold o\}$. By the isomorphism $\Z^n \cong \Z^{n_1} \oplus \Z^{n_2}\oplus \cdots$, we regrad $v$ as a collection $v_1, v_2,\ldots$ with $v_i \in \Z^{n_i}$. Let $m:=\min_iE(n_i,1)$. For each $i$, if $v_i$ is not zero, apply $E(n_i,1)$ to $v_i$, and we get an orthoregular set $v_{i1}=v_i,v_{i2},\ldots, v_{im}$. If $v_i$ is zero, let $v_{i1}=v_{i2}=\cdots=v_{im}$ be zero. For each $j=2,\ldots, m$, let $v_j \in \Z^n$ be the vector corresponding to $(v_{ij})_i$. Then, $v_1,v_2,\ldots, v_m$ is the desired system.
$(7)$ By $(6)$, we may assume that $m=1$. Then this is by the next proposition with $K=\R$.
Let $K$ be an arbitrary subset of $\R$ satisfying $-K=K$. If $n=2,4$ or $8$, then $E_K(n,1)=n$ for any $K$.
Let $v_1 \in K^n$. Via a standard matrix representation of complex number field, quarternion algebra or Cayley algebra as an $\R$-algebra, the rotations multiplied by the unit imaginary numbers gives the desired $v_2,\ldots, v_n$. Explicitly, for $v_1=
\begin{pmatrix}
a_0 \\ a_1 \\
\end{pmatrix}$, we give $v_2=iv_1=
\begin{pmatrix}
-a_1 \\ a_0 \\
\end{pmatrix};$ for $v_1=
\begin{pmatrix}
a_0 \\ a_1 \\ a_2 \\ a_3
\end{pmatrix}$, we give $(v_2,v_3,v_4)=(iv_1,jv_1,kv_1)=
\left(
\begin{pmatrix}
-a_1 \\ a_0 \\ -a_3 \\ a_2
\end{pmatrix},
\begin{pmatrix}
-a_2 \\ a_3 \\ a_0 \\ -a_1
\end{pmatrix},
\begin{pmatrix}
-a_3 \\ -a_2 \\ a_1 \\ a_0
\end{pmatrix}
\right);$ for $v_1=
\begin{pmatrix}
a_0 \\ a_1 \\ a_2 \\ a_3 \\ a_4 \\ a_5 \\ a_6 \\ a_7
\end{pmatrix}$, we give $$(v_2,v_3,v_4,v_5,v_6,v_7,v_8)=
\left(
\begin{pmatrix}
-a_1 \\ a_0 \\ -a_3 \\ a_2 \\ -a_5 \\ a_4 \\ a_7 \\ -a_6
\end{pmatrix},
\begin{pmatrix}
-a_2 \\ a_3 \\ a_0 \\ -a_1 \\ -a_6 \\ -a_7 \\ a_4 \\ a_5
\end{pmatrix},
\begin{pmatrix}
-a_3 \\ -a_2 \\ a_1 \\ a_0 \\ -a_7 \\ a_6 \\ -a_5 \\ a_4
\end{pmatrix},
\begin{pmatrix}
-a_4 \\ a_5 \\ a_6 \\ a_7 \\ a_0 \\ -a_1 \\ -a_2 \\ -a_3
\end{pmatrix},
\begin{pmatrix}
-a_5 \\ -a_4 \\ a_7 \\ -a_6 \\ a_1 \\ a_0 \\ a_3 \\ -a_2
\end{pmatrix},
\begin{pmatrix}
-a_6 \\ -a_7 \\ -a_4 \\ a_5 \\ a_2 \\ -a_3 \\ a_0 \\ a_1
\end{pmatrix},
\begin{pmatrix}
-a_7 \\ a_6 \\ -a_5 \\ -a_4 \\ a_3 \\ a_2 \\ -a_1 \\ a_0
\end{pmatrix}
\right).$$
$E(6,k) \leq 4$ for $1 \leq k \leq 4$.
Let $v_1=\t (1,1,1,1,1,1)$. It is enough to show $E(6,v_1)\leq 4$. First observe that any integer vector who has the same length as $v_1$ satisfies either (a) the absolute value of any its component is $1$ or (b) the absolute values of its components are $2,1,1,0,0,0$.
If an integer vector satisfying (a) is perpendicular to $v_1$, then it has exactly $3$ components whose value is $-1$. From this, we see that there is no orthoregular set $v_1, v_2, v_3$ such that both $v_2$ and $v_3$ satisfy (a).
On the other hand, if an integer vector satisfying (b) is perpendicular to $v_1$, then it satisfies (b${}'$) its components are $2,-1,-1,0,0,0$ or $-2,1,1,0,0,0$. Further, if both $v_2$ and $v_3$ satisfy (b${}'$) and they are perpendicular, their supports should be disjoint. From this, we see that there is no orthoregular set $v_1, v_2, v_3, v_4$ such that three of $v_2$, $v_3$, and $v_4$ satisfy (b).
Therefore there are no orthoregular set $v_1, v_2, v_3, v_4, v_5$, that is, $E(6,v_1)\leq 4$, as desired.
Remark: In fact, $E(6,v_1)=4$ as seen by $v_2=\t(1,1,1,-1,-1,-1),
v_3=\t(2,-1,-1,0,0,0),
v_4=\t(0,0,0,2,-1,-1)$.
Masanori Kobayashi
Department of Mathematics and Information Sciences\
Tokyo Metropolitan University\
1-1 Minami-Ohsawa, Hachioji, Tokyo, 192-0397\
Japan
[email protected]
Chikara Nakayama Department of Economics\
Hitotsubashi University\
2-1 Naka, Kunitachi, Tokyo 186-8601\
Japan
[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss phase coexistence of polydisperse colloidal suspensions in the presence of adhesion forces. The combined effect of polydispersity and Baxter’s sticky-hard-sphere (SHS) potential, describing hard spheres interacting via strong and very short-ranged attractive forces, give rise, within the Percus-Yevick (PY) approximation, to a system of coupled quadratic equations which, in general, cannot be solved either analytically or numerically. We review and compare two recent alternative proposals, which we have attempted to by-pass this difficulty. In the first one, truncating the density expansion of the direct correlation functions, we have considered approximations simpler than the PY one. These $C_{n}$ approximations can be systematically improved. We have been able to provide a complete analytical description of polydisperse SHS fluids by using the simplest two orders $C_{0}$ and $C_{1}$, respectively. Such a simplification comes at the price of a lower accuracy in the phase diagram, but has the advantage of providing an analytical description of various new phenomena associated with the onset of polydispersity in phase equilibria (e.g. fractionation). The second approach is based on a perturbative expansion of the polydisperse PY solution around its monodisperse counterpart. This approach provides a sound approximation to the real phase behavior, at the cost of considering only weak polydispersity. Although a final seattlement on the soundness of the latter method would require numerical simulations for the polydisperse Baxter model, we argue that this approach is expected to keep correctly into account the effects of polydispersity, at least qualitatively.'
author:
- |
Domenico Gazzillo, Riccardo Fantoni & Achille Giacometti [^1]$^\ast$\
Istituto Nazionale per la Fisica della Materia and Dipartimento di Chimica Fisica, Università di Venezia, S. Marta DD 2137, I-30123 Venezia, Italy
title: |
Phase behavior of polydisperse sticky hard spheres:\
analytical solutions and perturbation theory
---
Introduction
============
New technological advances in physico-chemical manipulation of colloidal mixtures have brought up again the issue of theoretically understanding the phase behaviour of polydisperse systems [@Sollich02]. ‘Polydispersity’ in colloidal solutions means that, due to their production process, suspended macroparticles with the same chemical composition cannot be exactly identical to each other, but in general have different sizes, and possibly different surface charges, shapes, etc. In practice, a polydisperse system can be reckoned as a mixture with very large – or essentially infinite – number $M$ of different species or components, identified by one or several parameters ($M$ large but finite refers to *discrete polydispersity*, whereas $M\rightarrow \infty $ with a continuous distribution of polydisperse parameters corresponds to *continuous polydispersity*). The present paper will consider discrete polydispersity of spherical colloidal particles, with their diameter being the only polydisperse attribute (size-polydispersity).
When polydispersity is not negligible, the phase behaviour becomes much richer, but the determination of phase transition boundaries requires a much more involved formalism, compared to the monodisperse counterpart. In fact, the coexistence condition in terms of intensive variables requires that all phases must have equal temperature, pressure and chemical potentials of the $%
M$ components. In the presence of polydispersity, one should thus solve a number of equations of the order of $M^{2}$, a task which is practically impossible for $M$ large or infinite.
However, the study of phase equilibria can conveniently start from the appropriate thermodynamic potential, which is the Helmholtz free energy $A$ when the experimentally controlled variables are temperature, volume and numbers of different colloidal species. In the one-component case, the coexistence condition of equal pressure and chemical potential has a simple geometrical interpretation in terms of free energy density $a$: the densities of two coexisting phases are determined by constructing a double-tangent to $a$ plotted versus particle density. This recipe leads to the well known Maxwell construction, which connects suitably selected points along a van der Waals (vdW) subcritical isotherm, in order to ’reduce’ its unphysical loop to a constant-pressure line characteristic of a first-order phase transition.
In the polydisperse case, a significant progress in the very difficult problem of predicting phase equilibria can be obtained for models with *truncatable* free energies [@Sollich02]. Here ‘truncatable’ means that the excess free energy of the polydisperse system turns out to depend only on a *finite* number of moments of the distribution corresponding to the polydisperse attribute (the diameter $\sigma $ in the simplest cases). For spherical colloids, the excess free energy of the vdW model extended to polydisperse fluids has such a truncatable structure. Due to this property, this vdW theory has often been employed as the simplest model to investigate the effects of polydispersity on the gas-liquid transition [@Sollich02; @Bellier-Castella00]. On the other hand, the influence of polydispersity on freezing has been addressed by using the hard-sphere (HS) mixture model, which also admits a truncatable free energy [@Sollich02] (for the fluid phase, the Boublik-Mansoori-Carnahan-Starling-Leland (BMCSL) [@Mansoori71] expression was employed). It is worth recalling that it is currently believed that size-polydispersity might destabilize crystallization, eventually inhibiting freezing above a certain ‘terminal’ value of polydispersity [Sollich02]{}.
The present paper focuses on – and reviews – a number of recent attempts to investigate polydisperse phase equilibria, at least within some approximations, for another prototype model useful for studying colloidal suspensions, namely Baxter’s sticky-hard-spheres (SHS) [@Baxter68]. Here the particles are hard spheres with a surface adhesion, and the corresponding potential can be obtained as a limit of an attractive square-well which becomes infinitely deep and narrow, according to a particular prescription which ensures a finite non-zero contribution of adhesion to the second virial coefficient (‘sticky limit’). For the one-component version of this model, Baxter and collaborators [@Baxter68] solved the Ornstein-Zernike (OZ) integral equation coupled with the Percus-Yevick (PY) approximation (‘closure’). This *fully analytical* solution allows to determine all structural and thermodynamical properties of the SHS fluid. On the other hand, the multi-component PY solution, which soon followed Baxter’s work [@Perram75; @Barboy79], is practically inapplicable in the presence of significant polydispersity. In fact, it requires the computation of a set of parameters $\left\{ \Lambda
_{ij}\right\} $ determined by a system of $M(M+1)/2$ quadratic equations, which – in general – cannot be solved even numerically for a mixture with a large number of components. Moreover, even in the presence of a general solution for this non-linear algebraic system, the problem of phase coexistence would still remain out of reach in view of the previous remarks.
In a series of recent papers [Gazzillo00,Gazzillo02,Gazzillo03,Fantoni05,Fantoni06,Gazzillo06]{}, we have attempted to make some progress along two different lines.
First, starting from the *density* expansion of the cavity function at contact, we have considered a sequence of simpler approximations (compared to the PY one) [Gazzillo00,Gazzillo02,Gazzillo03,Gazzillo04,Fantoni05]{}. Within the two simplest ones among these approximations, denoted as $C_{0}$ and $C_{1}$ (for reasons which will become clear in the following), we have been able to derive analytically all relevant information regarding structure and thermodynamics, including the phase coexistence, in view of the fact that the corresponding free energy turns out to be truncatable [@Fantoni05]. Due to the simplicity of $C_{0}$ and $C_{1}$, it is however reasonable to expect these approximations to fail at high packing fractions, with a consequently incomplete or even incorrect description of the effects of polydispersity on the phase diagram.
Therefore, in collaboration with Peter Sollich, we have recently explored a second approach [@Fantoni06], where the expansion variable (which must be small) is an appropriate polydispersity index. In such a way, we have tried to solve the non-linear algebraic system – involved in the PY result – *perturbatively in polydispersity*, starting from the monodisperse PY solution.
Baxter’s SHS model and PY solution
==================================
The SHS model is defined as limiting case of a particular square-well (SW) model [@Baxter68], based upon a potential including steeply repulsive core and short-ranged attractive tail, i.e.
$$\phi _{ij}^{\mathrm{Baxter\ SW}}\left( r\right) =\left\{
\begin{array}{ccc}
+\infty & & 0<r<\sigma _{ij}\equiv (\sigma _{i}+\sigma _{j})/2 \\
-\epsilon _{ij}^{\mathrm{Baxter~SW}} & & \sigma _{ij}\leq r\leq
R_{ij}\equiv \sigma _{ij}+w_{ij} \\
0 & & r>R_{ij}\text{ ,}%
\end{array}%
\right. \label{eq0}$$
with$$\epsilon _{ij}^{\mathrm{Baxter~SW}}=k_{B}T\ \ln \left( 1+t_{ij}\frac{\sigma
_{ij}}{w_{ij}}\right) \text{ ,}$$where $\sigma _{i}$ is the HS diameter of species $i$, $\epsilon _{ij}^{%
\mathrm{Baxter~SW}} >0$ and $w_{ij}$ are the depth and the width of the well, respectively, $k_{B}$ denotes Boltzmann’s constant, $T$ the temperature. Moreover, $$t_{ij}=\frac{1}{12\tau _{ij}}\geq 0,$$where the conventional Baxter parameter $\tau _{ij}$ is an unspecified increasing function of $T$, while $\tau _{ij}^{-1}$ measures the strength of surface adhesion or ‘stickiness’ between particles of species $i$ and $j$.
The ‘sticky limit’ of $\phi _{ij}^{\mathrm{Baxter~SW}}\left( r\right) $ corresponds to taking $w_{ij}\rightarrow 0$. While the SW width goes to zero, its depth $\epsilon _{ij}^{\mathrm{Baxter~SW}}$ diverges, giving rise to a Dirac delta function in the Boltzmann factor [@Baxter68], i. e.$$e^{-\beta \phi _{ij}^{\mathrm{SHS}}\left( r\right) }=\theta \left( r-\sigma
_{ij}\right) +t_{ij}\ \sigma _{ij}\delta \left( r-\sigma _{ij}\right)
\label{eq1}$$where $\beta =\left( k_{B}T\right) ^{-1}$, while $\theta $ and $\delta $ are the Heaviside step function and the Dirac delta function, respectively.
The advantage of the sticky limit is that one effectively deals with a single parameter $\tau _{ij}$ for each pair, rather than a combination of energy and length scales (as occurs in the square-well model, for which no analytical solution is known). On the one hand, this particular limit has the disadvantage of introducing some pathologies in the model, notably in the one-component case [@Stell91]. On the other hand, Baxter’s model represents the simplest paradigmatic example of a combination of steep repulsion and short-range attraction which entails a complete analytical solution in the one-component case, within a robust approximation such as the PY closure.
In the multicomponent case, the PY solution of the OZ equation in terms of Baxter’s factor correlation function reads [@Perram75; @Barboy79]$$q_{ij}(r)=\left\{
\begin{array}{l}
\frac{1}{2}a_{i}(r-\sigma _{ij})^{2}+(b_{i}+a_{i}\ \sigma _{ij})(r-\sigma
_{ij})+\Lambda _{ij},\qquad (\sigma _{i}-\sigma _{j})/2\leq r\leq \sigma
_{ij} \\
\qquad 0,\qquad \qquad \qquad \qquad \text{\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ elsewhere,}%
\end{array}%
\right. \label{eq2}$$where the expressions for the parameters $a_{i}$ and $b_{i}$ may be found in [@Gazzillo00], while the quantity $$\Lambda _{ij}=t_{ij}\ y_{ij}(\sigma _{ij})\ \sigma _{ij}^{2}\ , \label{eq3}$$which depends on the cavity function at contact $y_{ij}(\sigma _{ij})$, must be solution of the following system of quadratic equations $$\Lambda _{ij}=\alpha _{ij}+\beta _{ij}\sum_{m}x_{m}\left[ \Lambda
_{im}\Lambda _{jm}-\frac{1}{2}\left( \Lambda _{im}\Gamma _{mj}+\Lambda
_{jm}\Gamma _{mi}\right) \right] ~,~~~i,j=1,2,\ldots ,M \label{eq4}$$Here, $x_{m}$ is the molar fraction of the $m$-th species ($m=1,\ldots ,M$), while $\alpha _{ij}=t_{ij}\ y_{ij}^{\mathrm{HS-PY}}(\sigma _{ij})\ \sigma
_{ij}^{2}$, $\beta _{ij}=12\rho \ t_{ij}\sigma _{ij}$ ($\rho $ is the total number density), and $\Gamma _{ij}=\sigma _{i}\sigma _{j}/(1-\eta )$, with $\eta $ being the packing fraction [@Fantoni06]. The solution of these equations for $\left\{ \Lambda _{ij}\right\} $ is the real bottleneck of the multi-component PY result, as mentioned in the Introduction: for large $M$ (and in particular for $M\rightarrow \infty $ ) this calculation is next to impossible, neither analytically nor numerically.
As a consequence, although the PY closure is commonly believed to be very sound for short-range potentials (for one-component SHS fluids this was confirmed by recent numerical simulations [@Miller03]), one has to conclude that, in the multi-component (polydisperse) case, the PY solution has a very limited practical usefulness, since its solution scheme cannot be fully accomplished. This is the reason why other possible routes have been attempted, as we discuss next.
Simplified closures: the class of $C_{n}$ approximations
========================================================
A ‘closure’ is a relationship, added to the OZ equation, between the direct correlation function $c_{ij}(r)$ and $h_{ij}(r)=g_{ij}(r)-1$ or $\gamma
_{ij}(r)=h_{ij}(r)-c_{ij}(r)$ ($g_{ij}(r)$ being the radial distribution function) [@Barrat03].
Let us go back to Baxter’s SW model given by Eq. (\[eq0\]) (i.e. *before* the ‘sticky limit’), and consider the following general class of ‘closures’ [@Gazzillo04] $$c_{ij}(r)=\left\{
\begin{array}{cc}
-1-\gamma _{ij}\left( r\right) & 0<r<\sigma _{ij} \\
c_{ij}^{\mathrm{shrink}}(r) & \text{ \ \ \ \ }\sigma _{ij}\leq r\leq
R_{ij}\qquad \\
0 & \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }r>R_{ij}\text{ .}\qquad
\end{array}%
\right. \label{eq6}$$The expression $c_{ij}(r)=-1-\gamma _{ij}\left( r\right) $ inside the core ($%
r<\sigma _{ij}$) is exact and dictated by the HS potential. The form outside the well ($r>R_{ij}$) may then be identified with the PY approximation, $$c_{ij}^{\mathrm{PY}}(r)=f_{ij}(r)\left[ 1+\gamma _{ij}\left( r\right) \right]
,$$since, for Baxter’s potential, the Mayer function, $f_{ij}(r)=\exp \left[
-\beta \phi _{ij}(r)\right] -1$, vanishes for $r>R_{ij}$.
The choice of $c_{ij}^{\mathrm{shrink}}(r)$ inside the well (region which ’shrinks’ in the sticky limit) defines one particular closure within the proposed class. Of course, $c_{ij}^{\mathrm{shrink}}(r)=c_{ij}^{\mathrm{PY}%
}(r)$ corresponds to the PY approximation. On the other hand, when $c_{ij}^{%
\mathrm{shrink}}(r)\neq c_{ij}^{\mathrm{PY}}(r)$, we are in the presence of *mixed closures,* which have frequently appeared in the literature [@Herrera91]. In order to define mixed closures simpler than the PY approximation, we consider the *density* expansion of the exact direct correlation function [@Barrat03], and denote as $C_{n}$ approximation a truncation of this series to order $O(\rho ^{n})$. The simplest two approximations are
$$\begin{array}{ccc}
\begin{array}{c}
c_{ij}^{\mathrm{shrink}}(r)=f_{ij}(r) \\
\end{array}
& &
\begin{array}{c}
\text{(}C_{0}\text{ closure)} \\
\end{array}
\\
c_{ij}^{\mathrm{shrink}}(r)=f_{ij}\left( r\right) \ \left[ 1+\left( \
\sum_{k}\rho _{k\ }f_{ik}\ast f_{kj}\right) \left( r\right) \right] & &
\text{(}C_{1}\text{ closure),}%
\end{array}%$$
where $\rho _{k\ }$ is the number density of species $k$, while $\ast $ denotes convolutive integration [@Gazzillo04].
In the ‘sticky limit’ $R_{ij}\rightarrow \sigma _{ij}^{+}$ the well region shrinks, but a ‘memory’ of the approximation chosen for $c_{ij}^{\mathrm{%
shrink}}$ remains in the solution of the OZ integral equation. In fact, although all solutions $q_{ij}(r)$ corresponding to closures belonging to the class given by Eq. (\[eq6\]) have the same functional form as the PY solution – Eq. (\[eq2\]) – , each closure is characterized by its own approximation to $y_{ij}(\sigma _{ij})$, which is involved in the expressions of the parameters $a_{i}$, $b_{i}$, $\Lambda _{ij}$. For instance, the $C_{0}$ and $C_{1}$ approximations correspond to
$$\begin{array}{ccc}
\begin{array}{c}
y_{ij}(\sigma _{ij})=1 \\
\end{array}
& &
\begin{array}{c}
\text{(}C_{0}\text{ closure)} \\
\end{array}
\\
y_{ij}(\sigma _{ij})=1+y_{ij}^{(1)}(\sigma _{ij})\ \eta & & \text{(}C_{1}%
\text{ closure),}%
\end{array}%$$
which are, respectively, the zeroth– and first–order truncations of the density expansion for the exact cavity function at contact (see Ref. [Gazzillo04]{} for details).
While a brute-force truncation of the above-mentioned density expansions leads to analytical expressions simple enough to be applied to the multi-component (polydisperse) case, one should reasonably expect less accuracy, expecially in the high-density regime. In the one-component case, we can carefully check this point.
In Figure \[fig1\] coexistence curves obtained from the $C_0$ and $C_{1}$ approximations are compared with the PY ones (using both compressibility and energy routes), and with Monte Carlo simulations from Ref. [@Miller03]. It is apparent how the PY energy route (PYE) yields a rather precise representation of the MC results, unlike the compressibility route (PYC). It is worth noting that the results stemming from the $C_{1}$ approximation, although rather close to the MC data in the low-density branch, clearly fail to accurately reproduce them for higher densities, as expected.
In spite of their lack of accuracy, the $C_{0}$ and $C_{1}$ approximations provide however a rather sound basis for getting some insight into phase equilibria of polydisperse SHS fluids, since they allow simple analytical, or semi-analytical, treatments.
A first important feature of the $C_{0}$ and $C_{1}$ approximations for polydisperse SHS is that the corresponding free energy has a *truncatable* structure, that is it depends upon few ($4$ at most) moments of the (discrete) size distribution, $\xi _{\nu
}=(\pi /6)\rho \sum_{j}x_{j}\sigma _{j}^{\nu }$ with $\nu =0,1,2,3$.
A second remarkable fact is that the $C_{0}$ and $C_{1}$ approximations are able to describe the so-called *fractionation* phenomena characteristic in phase equilibria of polydisperse systems. While we refer to a recent review [@Sollich02] for detailed description of the increased complexity in the polydisperse phase diagrams, here we just mention the two important points. First, fractionation means that daughter phases, obtained from demixing of a parent homogeneous phase, need not have the same composition of the parent phase. As a consequence, there is no a single coexistence line (‘binodal’) as in the one-component case, but one rather finds a *cloud curve*, representing the temperature-density dependence line of the low-density majority phase (’gas’), and a *shadow curve* representing the temperature-density dependence of the high-density minority phase (incipient ’liquid’). While for one-component systems these two curves are identical, for polydisperse systems in general they are not, with the exception of the critical point where they coincide by definition.
However, in order to apply the $C_{0}$ and $C_{1}$ approximations to the multi-component SHS model, we have to tackle a further important problem, that is the definition of the stickiness parameters $\tau _{ij}$.
Size-dependence of stickiness parameters
========================================
In mixtures, $\tau _{ij}$ will depend on the particular pair $i,j$ considered, and reasonably should be expected to be related to the particles sizes. Assuming that we are dealing only with size-polydispersity, we can always decouple temperature and adhesion as $$\frac{1}{\tau _{ij}}=\frac{1}{\tau }\ \epsilon _{ij}=\frac{1}{\tau }\mathcal{%
F}\left( \sigma _{i},\sigma _{j}\right)$$where the last equality stems from the assumption of size-polydispersity and of a purely pairwise potential. Unfortunately, the exact form of the size-dependence of these stickiness parameters is still an open problem, due to the lack of experimental and theoretical insights on this [@Gazzillo06]. On the other hand, few guidelines – based on arguments discussed in Refs. [@Fantoni05; @Fantoni06] – provide, as reasonable and plausible, the following dependences $$\epsilon _{ij}=\mathcal{F}(\sigma _{i},\sigma _{j})=\left\{
\begin{array}{ll}
\sigma _{0}^{2}/\sigma _{ij}^{2} & \mbox{Case I}~, \\
\sigma _{i}\sigma _{j}/\sigma _{ij}^{2} & \mbox{Case II}~, \\
1 & \mbox{Case IV}~, \\
\sigma _{0}/\sigma _{ij} & \mbox{Case V}~.%
\end{array}%
\right. \label{eq9}$$Here $\sigma _{0}$ is a characteristic reference length (e.g. the parental mean diameter) and the numbering of the various cases follows the convention of previous work [@Fantoni05; @Fantoni06].
Figure \[fig2\] reports the results of the calculation of the cloud and shadow curves for polydisperse SHS within the simple $C_{0}$ approximation. Here and below, polydispersity is measured by an index $s$, which is the normalized standard deviation of the size distribution. Hence $s=0$ corresponds to a mono-disperse case, whereas $s=0.1$ and $s=0.3$ indicate moderate and significant polydispersity, respectively. The top panel of Fig. \[fig2\] depicts the results for case I of size-dependence of stickiness parameters. As $s$ increases, the coexistence region shrinks, thus suggesting that polydispersity *disfavours* the phase transition. On the other hand, this trend is markedly case-dependent, as illustrated in the bottom panel of Fig. \[fig2\], where the cloud-shadow pairs with polydispersity $s=0.3$ are displayed for different cases of size-dependence. It can be clearly seen that while, for cases I and V the same trend is observed, case IV seems to suggest a *widening* of the phase coexistence region (and hence a favouring of the phase transition).
In view of the lack of numerical simulation for polydisperse SHS to compare with, we have no way, at the present stage, to check how realistic those results are. On the other hand, we might suspect, based on the comparison in the one-component case, $C_{0}$ to fail to provide accurate representation in the region of low temperatures and high densities. This is the reason why other possible approaches have been recently tested. We now illustrate a different perturbative approach, which has proved to be promising in this context.
Perturbative treatment of the SHS-PY solution
=============================================
The main difficulty in dealing with the PY solution for polydisperse SHS stems from the solution of the coupled quadratic system of equations ([eq4]{}), as remarked. As the one-component case has a well-defined solution, one might then suspect that – for weak polydispersity – a perturbative expansion around this solution might include the main effects of polydispersity. This is in fact what happens, as recently shown [Fantoni06]{} by exploiting a general perturbation theory due to Evans [Evans01]{}. The main idea is that, for weak polydispersity, size-distributions are narrowly peaked around a mean reference value ($%
\sigma _{0}$ in the present case), and hence all quantities such as$$\delta _{i}=\frac{\sigma _{i}-\sigma _{0}}{\sigma _{0}}\ll 1\text{,}
\label{eq10}$$are small. Therefore one might expand both $\epsilon _{ij},$ and all quantities appearing in $\Lambda _{ij},$ in powers of $\delta _{i}$. Similar expansion can be performed in the free energy, and hence all thermodynamic quantities can be computed. The entire procedure is described in detail in Refs. [@Sollich02; @Fantoni06; @Evans01]. Here we just summarize the main results.
The approximate range of validity of the perturbation expansion can be envisaged by considering the polydisperse HS case where the ‘exact’ BMCSL approximation [@Mansoori71] can be compared with the corresponding perturbative solution based on the one-component ($s=0$) counterpart. This is reported in the top-left part of Figure \[fig3\], where the quantity $\beta P v_0$ ($v_0=\pi \sigma_0^3/6$) is plotted against the packing fraction $\eta$ for increasing values of polydispersity. It is apparent how the perturbative solution remains rather close to the ’exact’ polydisperse BMCSL solution even for moderate polydispersity $s \le 0.3$, which is the maximum value considered in the remaining part of the work. The remaining plots in Figure \[fig3\] display the effect of polydispersity on the PY pressure equation of state as obtained from the energy route and for decreasing values of the temperature $\tau$. In the one component case $s=0$, a van der Waals loop starts to appear when we cross the critical temperature $\tau_c \sim 0.1185$ coming from the high $\tau$ regime. Obviously this signals the onset of a liquid-gas phase transition, and the corresponding phase diagram can be obtained by a standard Maxwell construction by connecting appropriate points with the same pressure. In the presence of polydispersity (here represented by choice IV for size-dependence of the stickiness parameters) the same procedure [*cannot*]{} be applied due to fractionation, as already discussed. Nevertheless, we can clearly see that as $s$ increases, the van der Waals loop region (when present) expands, thus suggesting that phase transition is favored by the presence of polydispersity. A similar feature occurs for the polydisperse van der Waals model [@Bellier-Castella00] and for the numerical results of the PY compressibility equation of state [@Robertus89] (note that in the latter a [*gap*]{} rather that a [*loop*]{} is signalling the onset of the transition). A somewhat surprising feature is that, at fixed packing fraction $\eta$, the pressure decreases with increasing polydispersity [*less*]{} in the presence of adhesion rather than in its absence (i.e. for the HS case). An intuitive plausible interpretation of this feature can be found in Ref. [@Fantoni06].
The same perturbative approach allows the determination of the cloud and shadow curves for the various cases of size-dependence of $\ \tau _{ij}^{-1}$. This is reported in Figure \[fig4\] for cases II, IV (top panel) and I, V (bottom panel). In the first case the cloud and shadow lines collapse into a single curve, and this can be understood on the basis of the particular scaling properties of the free energy to this order in perturbation theory [@Fantoni06]. In all cases there is a breakdown of the perturbation theory on approaching the critical point, and this is a known general drawback of Evans’ perturbative scheme. Nevertheless, in all cases and to this order in perturbation theory, there is a tendency of the phase coexistence region to *increase* with polydispersity, in qualitative agreement with the intuitive picture obtained from Figure \[fig3\].
It is worth stressing the difference with previous non-perturbative results stemming from the $C_{0}$ solution, where all different cases (with the notable exception of IV) were predicting a reduction of the phase coexistence region. While in the $C_{0}$ description we have provided a careful treatment of polydispersity at the expenses of accuracy of the exploited approximation, in the perturbative description of the PY solution, polydispersity is assumed to be small and hence one might suspect that solutions with large polydispersities cannot fit within this picture. On a balance, nevertheless, we would favour the latter rather than the former description. An almost correct representation of the one-component counterpart, is a necessary requirement to check the effect of polydispersity on it, and we are not aware of any physical or experimental system where the effects of polydispersity are so strong that they could not kept into account, at least at the simplest qualitative level, by the perturbative scheme proposed here. Along this line, some further proposals have been put forward in Ref. [@Fantoni06] to derive a phenomenological BMCSL-like approximation for SHS, which might be regarded as our ’best and simplest guess’ to the *exact* phase behavior of polydisperse SHS. Even on the size-dependence of $\tau _{ij}^{-1}$ some possible support of the proposed forms may be argued [@Fantoni06; @Gazzillo06].
Conclusions
===========
In this work we have summarized recent advances in predicting theoretically the phase diagram for polydisperse suspensions of colloidal particles with surface adhesion, within the simple description of Baxter’s model. Emphasis was put on the crucial – unsolved – step required to get the multi-component SHS-PY solution, and the proposed recipies to deal with this problem. This first one is based on a simplification of the closure. It has the advantage of allowing a complete analytical analysis on the effects of polydispersity, including fractionation, but has the disadvantage of a very questionable accuracy. The second one is based on a perturbative method, starting from the energy PY one-component solution, which is known to provide an accurate description of the phase diagram. The drawback of this scheme is that it works for mild polydispersity, and that cannot describe the changes of the critical point region. Notwithstanding these limitations, this novel approach is expected to find practical application in the interpretation of all those phenomena where Baxter’s model and polydispersity both play a privileged role.
**Ackowledgments**\
Part of the work appearing here has been obtained in collaboration with Peter Sollich.\
[99]{} P. Sollich, [*J. Phys.: Condens. Matter*]{} [**14**]{} R79 (2002).
L. Bellier-Castella, H. Xu and M. Baus, [*J. Chem. Phys.*]{} [**113**]{}, 8337 (2000)
T. Boublík, [*J. Chem. Phys.*]{} [**53**]{}, 471 (1970); G. A. Mansoori, N. F. Carnahan, K. E. Starling and T. W. Leland Jr, [*J. Chem. Phys.*]{} [**54**]{}, 1523 (1971); J. J. Salacuse and G. Stell, [*J. Chem. Phys.*]{} [**77**]{}, 3714 (1982).
R. J. Baxter, [*J. Chem. Phys.*]{} [**49**]{}, 2270 (1968). R. J. Baxter, in: *Physical Chemistry, an Advanced Treatise,* Vol. 8A, ed. D. Henderson (Academic Press, New York, 1971) ch. 4. R. O. Watts, D. Henderson and R. J. Baxter, [*Advan. Chem. Phys.*]{} [**21**]{}, 421 (1971).
J. W. Perram and E. R. Smith, [*Chem. Phys. Lett.*]{} [**35**]{}, 138 (1975).
B. Barboy and R. Tenne, [*Chem. Phys.*]{} [**38**]{}, 369 (1979).
D. Gazzillo, and A. Giacometti, [*J. Chem. Phys.*]{} [**113**]{}, 9837 (2000); ibidem [*Physica A*]{}, [**304**]{}, 202 (2002).
D. Gazzillo, and A. Giacometti, [*Mol. Phys.*]{} [**100**]{}, 3307 (2002).
D. Gazzillo, and A. Giacometti, [*Mol. Phys.*]{} [**101**]{}, 2171 (2003)
D. Gazzillo, and A. Giacometti, [*J. Chem. Phys.*]{} [**120**]{}, 4742 (2004).
R. Fantoni, D. Gazzillo, and A. Giacometti, [*J. Chem. Phys.*]{} [**122**]{}, 034901 (2005); ibidem [*Phys. Rev. E*]{} [**72**]{}, 011503 (2005).
R. Fantoni, D, Gazzillo, A. Giacometti, and P. Sollich, [*J. Chem. Phys.*]{} (2006), in press.
D, Gazzillo, A. Giacometti, R. Fantoni, and P. Sollich, [*Phys. Rev. E.*]{} (2006), in press.
G. Stell, [*J. Stat. Phys.*]{} [**63**]{}, 1203 (1991).
M. A. Miller, and D. Frenkel, [*Phys. Rev. Lett.*]{} [*90*]{}, 135702 (2003); *ibidem*, [*J. Chem. Phys.*]{} [**121**]{} , 535 (2004). *ibidem*, [*J. Phys. Cond. Mat.*]{} [**16**]{}, S4901 (2004).
J. L. Barrat, and J. P. Hansen, [*Basic Concepts for Simple and Complex Liquids*]{} (Cambridge University Press, Cambridge) (2003); J. P. Hansen and I. R. Mc Donald, [*The Theory of Simple Liquids*]{} (Academic, London) (1986).
M. L. Evans, [*J. Chem. Phys.*]{} [**114**]{}, 1915 (2001)
C. Robertus, W.H. Philipse, J.G. Joosten, and Y. K Levine, [*J. Chem. Phys.*]{} [**90**]{}, 4482 (1989)
J. N. Herrera and L. Blum, [*J. Chem. Phys.*]{} [**94**]{}, 5077 (1991)
0.5cm ![Coexistence (binodal) curves for the one-component Baxter model. Both compressibility (PYC) and energy (PYE) equation of state as obtained from the Percus-Yevick approximation (see Ref [@Baxter68]) are reported and compared with $C_0$ and $C_1$ approximations from Ref. [@Fantoni05] and with Monte Carlo simulation (MC) from Ref. [@Miller03]. In the MC case the continuous line is simply a guide to the eye.[]{data-label="fig1"}](coexistence.eps "fig:"){width="15cm"}
0.5cm ![(Top) Cloud and shadow curve for model I within the $C_0$ approximation at increasing polydispersity: $s=0$, $s=0.1$ and $s=0.3$ . (Bottom) Same as above for fixed value of polydispersity $s=0.3$ and different choice of the stickiness adhesion (model I, IV and V) []{data-label="fig2"}](cs_modI.eps "fig:"){width="15cm"}\
2.0cm ![(Top) Cloud and shadow curve for model I within the $C_0$ approximation at increasing polydispersity: $s=0$, $s=0.1$ and $s=0.3$ . (Bottom) Same as above for fixed value of polydispersity $s=0.3$ and different choice of the stickiness adhesion (model I, IV and V) []{data-label="fig2"}](cs_s.3.eps "fig:"){width="15cm"}
0.5cm ![Behavior of the energy equation of state within our perturbative scheme. In all cases the quantity $\beta P v_0$ is plotted against the packing fraction $\eta$. In clockwise order, the first curve (left, top) reports a comparison of the perturbative versus the ‘exact’ BMCSL solution in the equation of state for polydisperse HS ($\tau=+\infty$). The other curves report the perturbative solution for the energy equation of state within the PY approximation for SHS Baxter model. Results are depicted for three values of temperature $\tau=0.15> \tau_c$, $\tau=0.1185\sim
\tau_c$ and $\tau=0.1<\tau_c$ and for different degrees of polydispersity. The choice for the size-dependence of stickiness parameters corresponds to model IV. []{data-label="fig3"}](Fig3a.eps "fig:"){height="9.0cm" width="7.5cm"} ![Behavior of the energy equation of state within our perturbative scheme. In all cases the quantity $\beta P v_0$ is plotted against the packing fraction $\eta$. In clockwise order, the first curve (left, top) reports a comparison of the perturbative versus the ‘exact’ BMCSL solution in the equation of state for polydisperse HS ($\tau=+\infty$). The other curves report the perturbative solution for the energy equation of state within the PY approximation for SHS Baxter model. Results are depicted for three values of temperature $\tau=0.15> \tau_c$, $\tau=0.1185\sim
\tau_c$ and $\tau=0.1<\tau_c$ and for different degrees of polydispersity. The choice for the size-dependence of stickiness parameters corresponds to model IV. []{data-label="fig3"}](Fig3b.eps "fig:"){height="9.0cm" width="7.5cm"}\
1.5cm ![Behavior of the energy equation of state within our perturbative scheme. In all cases the quantity $\beta P v_0$ is plotted against the packing fraction $\eta$. In clockwise order, the first curve (left, top) reports a comparison of the perturbative versus the ‘exact’ BMCSL solution in the equation of state for polydisperse HS ($\tau=+\infty$). The other curves report the perturbative solution for the energy equation of state within the PY approximation for SHS Baxter model. Results are depicted for three values of temperature $\tau=0.15> \tau_c$, $\tau=0.1185\sim
\tau_c$ and $\tau=0.1<\tau_c$ and for different degrees of polydispersity. The choice for the size-dependence of stickiness parameters corresponds to model IV. []{data-label="fig3"}](Fig3d.eps "fig:"){height="9.0cm" width="7.5cm"} ![Behavior of the energy equation of state within our perturbative scheme. In all cases the quantity $\beta P v_0$ is plotted against the packing fraction $\eta$. In clockwise order, the first curve (left, top) reports a comparison of the perturbative versus the ‘exact’ BMCSL solution in the equation of state for polydisperse HS ($\tau=+\infty$). The other curves report the perturbative solution for the energy equation of state within the PY approximation for SHS Baxter model. Results are depicted for three values of temperature $\tau=0.15> \tau_c$, $\tau=0.1185\sim
\tau_c$ and $\tau=0.1<\tau_c$ and for different degrees of polydispersity. The choice for the size-dependence of stickiness parameters corresponds to model IV. []{data-label="fig3"}](Fig3c.eps "fig:"){height="9.0cm" width="7.5cm"}
0.5cm ![Cloud/Shadow pairs from the perturbative results for the PY solution of SHS Baxter model. In clockwise order the results for choices II, IV (top) and I,V (bottom) are depicted. In the two top panels, the cloud and shadow curves coincide to this order in perturbation, whereas in the bottom panels they are different. In order to have all pictures on the same scale, the selected value for polydispersity is $s=0.3$ for models II, IV (top) and $s=0.1$ for models I, V (bottom). In all cases the continuous curve represents the monodisperse ($s=0$) result. []{data-label="fig4"}](Fig4a.eps "fig:"){height="9.0cm" width="7.5cm"} 0.5cm ![Cloud/Shadow pairs from the perturbative results for the PY solution of SHS Baxter model. In clockwise order the results for choices II, IV (top) and I,V (bottom) are depicted. In the two top panels, the cloud and shadow curves coincide to this order in perturbation, whereas in the bottom panels they are different. In order to have all pictures on the same scale, the selected value for polydispersity is $s=0.3$ for models II, IV (top) and $s=0.1$ for models I, V (bottom). In all cases the continuous curve represents the monodisperse ($s=0$) result. []{data-label="fig4"}](Fig4b.eps "fig:"){height="9.0cm" width="7.5cm"}\
1.5cm ![Cloud/Shadow pairs from the perturbative results for the PY solution of SHS Baxter model. In clockwise order the results for choices II, IV (top) and I,V (bottom) are depicted. In the two top panels, the cloud and shadow curves coincide to this order in perturbation, whereas in the bottom panels they are different. In order to have all pictures on the same scale, the selected value for polydispersity is $s=0.3$ for models II, IV (top) and $s=0.1$ for models I, V (bottom). In all cases the continuous curve represents the monodisperse ($s=0$) result. []{data-label="fig4"}](Fig4d.eps "fig:"){height="9.0cm" width="7.5cm"} 0.5cm ![Cloud/Shadow pairs from the perturbative results for the PY solution of SHS Baxter model. In clockwise order the results for choices II, IV (top) and I,V (bottom) are depicted. In the two top panels, the cloud and shadow curves coincide to this order in perturbation, whereas in the bottom panels they are different. In order to have all pictures on the same scale, the selected value for polydispersity is $s=0.3$ for models II, IV (top) and $s=0.1$ for models I, V (bottom). In all cases the continuous curve represents the monodisperse ($s=0$) result. []{data-label="fig4"}](Fig4c.eps "fig:"){height="9.0cm" width="7.5cm"}
[^1]: $^\ast$Corresponding author. Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $S$ be a minimal complex surface of general type with $q(S)=0$. We prove the following statements concerning the algebraic fundamental group ${\pi_1^{\mathrm{alg}}}(S)$:
- Assume that $K^2_S\le3\chi(S)$. Then $S$ has an irregular étale cover if and only if $S$ has a free pencil of hyperelliptic curves of genus $3$ with at least $4$ double fibres.
- If $K^2_S=3$ and $\chi(S)=1$, then $S$ has no irregular étale cover.
- If $K^2_S<3\chi(S)$ and $S$ does not have any irregular étale cover, then $|{\pi_1^{\mathrm{alg}}}(S)|\le9$. If $|{\pi_1^{\mathrm{alg}}}(S)|= 9$, then $K^2_S=2$, $\chi(S)=1$.
[*2000 Mathematics Subject Classification:*]{} 14J29, 14F35.
author:
- Margarida Mendes Lopes
- Rita Pardini
title: 'On the algebraic fundamental group of surfaces with $K^2\le3\chi$ '
---
Introduction
=============
Every minimal surface $S$ of general type satisfies the Noether inequality: $$K^2_S\ge 2\chi(S)-6.$$
It has been clear for a long time that the closer a surface is to the Noether line $K^2=2\chi-6$, the simpler its algebraic fundamental group is. In fact, Reid has conjectured that for $K^2<4\chi$ the algebraic fundamental group of $S$ is either finite or it coincides, up to finite group extensions, with the fundamental group of a curve of genus $g\geq 1$, i. e. it is [*commensurable*]{} with the fundamental group of a curve, (, see also [@bpv], pp. 294).
In the case of irregular surfaces or of regular surfaces having an irregular étale cover, Reid’s conjecture follows from the Severi inequality, recently proved in [@severi], which states that the Albanese map of an irregular surface with $K^2<4\chi$ is a pencil.
Indeed, let $S$ be an irregular surface satisfying $K^2<4\chi$, let $a\colon S\to B$ be the Albanese pencil of $S$ and $F$ a general fibre of $a$. The inclusion $F{\hookrightarrow}S$ induces a map $\psi\colon {\pi_1^{\mathrm{alg}}}(F)\to{\pi_1^{\mathrm{alg}}}(S)$. By [@xiaoslope Theorem 1] the image $H$ of $\psi$ is either 0 or ${\mathbb Z}_2$, and $H={\mathbb Z}_2$ is possible only if $F$ is hyperelliptic. The cokernel of $\psi$ is the so-called [*orbifold fundamental group*]{} of the fibration $a$ (cf. [@cko], [@catanese Lemma 4.2]). If $a$ has no multiple fibres, then we have an exact sequence: $$\label{pisequence}
1\to H\to{\pi_1^{\mathrm{alg}}}(S)\to {\pi_1^{\mathrm{alg}}}(B)\to 1.$$ If $a$ has multiple fibres, then it is possible to find a Galois cover $B'\to B$ such that the fibration $a'\colon S'\to B'$ obtained from $a$ by base change and normalization has no multiple fibres and the map $S'\to S$ is étale. Since ${\pi_1^{\mathrm{alg}}}(S')$ is a normal subgroup of ${\pi_1^{\mathrm{alg}}}(S)$ of finite index, it follows that in any case the algebraic fundamental group of an irregular surface satisfying $K^2<4\chi$ is commensurable with the fundamental group of a curve. Of course the same is true for a regular surface satisfying $K^2<4\chi$ and having an irregular étale cover.
Reid’s conjecture is still open for surfaces not having an irregular cover. However for surfaces satisfying $K^2<3\chi$ not only Reid’s conjecture is true ([@miles1] and [@horikawaV]) but work by several authors gives more precise results on the algebraic fundamental group (cf. [@bombieri], [@horikawaV], [@miles1], [@milesk2], [@xiaohyp], [@xiaoslope]). The picture that emerges from their work is the following:
- if $K^2_S<2\chi(S)$, then $S$ is regular and ${\pi_1^{\mathrm{alg}}}(S)$ is finite.
- if $K^2_S<\frac{8}{3}\chi(S)$ and $S$ is irregular, then the Albanese map of $S$ is a pencil of curves of genus 2. If $K^2_S<\frac{8}{3}\chi(S)$ and $S$ is regular, then ${\pi_1^{\mathrm{alg}}}(S)$ is finite.
- if $K^2_S<3\chi(S)$ and $S$ is irregular, then the Albanese map of $S$ is a pencil of hyperelliptic curves of genus 2 or 3. If $S$ is regular, then either ${\pi_1^{\mathrm{alg}}}(S)$ is finite or there exists an irregular étale cover $X\to S$. The Albanese map of $X$ is a pencil of hyperelliptic curves of genus 3, which induces on $S$ a free pencil of hyperelliptic curves of genus 3 with at least $4$ double fibres. Conversely, if $S$ has such a pencil, then it admits an irregular étale cover.
These results give a good understanding of the algebraic fundamental group of a surface $S$ with $K^2<3\chi$ and infinite ${\pi_1^{\mathrm{alg}}}(S)$.
In fact, if $S$ is irregular and the Albanese map $a\colon S\to B$ has multiple fibres, then by statement (III) and by the adjunction formula we have $g=3$ and the multiple fibres are double fibres. Then there is a Galois cover $B'\to B$ with Galois group $G$ such that the $G$-cover $S'\to S$ obtained by base change and normalization is étale and the induced fibration $a'\colon S'\to B'$ has no multiple fibres. One can show that $G$ can be chosen to be a quotient of the dihedral group of order 8. So we have an exact sequence: $$1\to{\pi_1^{\mathrm{alg}}}(S')\to{\pi_1^{\mathrm{alg}}}(S)\to G\to 1.$$ and the group ${\pi_1^{\mathrm{alg}}}(S')$ is described by sequence (\[pisequence\]).
If $S$ is a regular surface such that $K^2_S<3\chi(S)$ and ${\pi_1^{\mathrm{alg}}}(S)$ is infinite, then using (III), one constructs an irregular étale Galois cover $X\to S$ with Galois group ${\mathbb Z}_2$ or ${\mathbb Z}_2^2$ whose Albanese map is a pencil of curves of genus 3 without multiple fibres (more precisely, we have ${\mathbb Z}_2$ if the number $k$ of double fibres of $a$ is even and ${\mathbb Z}_2^2$ if $k$ is odd). Then the group ${\pi_1^{\mathrm{alg}}}(X)$ is a normal subgroup of ${\pi_1^{\mathrm{alg}}}(S)$ of index $2$ or $4$ which can be described as explained above.
However if the algebraic fundamental group of $S$ is finite then the above results give no additional information.
In this paper we give two improvements of the above results.
We first extend part of (III) to surfaces on the line $K^2= 3\chi$:
\[main\] Let $S$ be a minimal complex surface of general type with $q(S)=0$ and $K^2_S\le3\chi(S)$.
Then $S$ has an irregular étale cover if and only if there exists a fibration $f\colon S\to{\mathbb P}^1$ such that:
1. the general fibre $F$ of $f$ is hyperelliptic of genus $3$;
2. $f$ has at least $4$ double fibres.
This improvement is made possible by the Severi inequality.
In the case $p_g(S)=0$, Theorem \[main\] can be made more precise:
\[K3\] Let $S$ be a smooth minimal surface of general type with $p_g(S)=0$, $K^2_S=3$.
Then $S$ has no irregular étale cover.
Theorem \[K3\] is sharp in a sense, since there are examples, due to Keum and Naie (cf. [@naie]), of surfaces with $K^2=4$ and $p_g=0$ that have an irregular cover.
On the other hand, it remains an open question whether the algebraic fundamental group of a surface with $K^2=3$ and $p_g=0$ is finite or more generally whether the algebraic fundamental group of a surface with $K^2=3\chi$ that has no étale irregular cover is finite.
In even greater generality one would like to know whether the algebraic fundamental group of a surface with $K^2<4\chi$ that has no étale irregular cover is finite, deciding thus Reid’s conjecture. This is a very challenging problem, which however does not seem possible to resolve with the methods of the present paper.
Finally, we bound the cardinality of ${\pi_1^{\mathrm{alg}}}(S)$ in the case when it is a finite group:
\[T2\] Let $S$ be a minimal surface of general type such that $K^2_S<3\chi(S)$. If $S$ has no irregular étale cover, then ${\pi_1^{\mathrm{alg}}}(S)$ is a finite group of order $\le9$.
Moreover, if ${\pi_1^{\mathrm{alg}}}(S)$ has order 9, then $\chi(S)=1$ and $K^2_S=2$, namely $S$ is a numerical Campedelli surface.
This bound is sharp, since there are examples of surfaces with $p_g=0$, $K^2=2$ and ${\pi_1^{\mathrm{alg}}}(S)={\mathbb Z}_9$, ${\mathbb Z}_3^2$ (cf. [@xiaog2 Ex. 4.11], [@pi9]).
By this theorem only a very short list of finite groups can occur as the algebraic fundamental groups of surfaces with $K^2\leq 3\chi-1$. The list is even more restricted if $K^2\leq 3\chi-2$: in [@3chi-2] it is shown that in this case $|{\pi_1^{\mathrm{alg}}}(S)|\le 5$, with equality holding only for surfaces with $K^2_S=1$ and $p_g(S)=0$. Moreover $|{\pi_1^{\mathrm{alg}}}(S)|=3$ is possible only for $2\le \chi(S)\le 4$ and $K^2=3\chi-3$.
[**Notation and conventions.**]{} We work over the complex numbers. All varieties are projective algebraic. We denote by $\chi$ or $\chi(S)$ the holomorphic Euler characteristic of the structure sheaf of the surface $S$.
The proof of Theorem \[main\]
=============================
In this section we assume that $S$ is a minimal complex surface of general type with $q(S)=0$ and $K^2_S\le3\chi(S)$. In order to prove Theorem \[main\] we need some intermediate steps.
\[fibre3\] Let $\rho\colon Z\to S$ be an étale cover such that $q(Z)>0$.
Then the Albanese pencil $a\colon Z\to A$ induces a fibration $f\colon S\to{\mathbb P}^1$ such that:
1. the general fibre $F$ of $f$ is a curve of genus $3$;
2. $f$ has at least 4 double fibres.
Moreover, all irregular étale covers of $S$ induce the same fibration $f\colon S\to{\mathbb P}^1$.
If $\rho \colon Z\to S$ is an irregular étale cover, then the Galois closure of $\rho$ is an irregular Galois étale cover. We denote by $\pi\colon Y\to S$ a minimal element of the set of irregular Galois étale covers of $S$.
Denote by $d$ the degree of $\pi$. The surface $S$ is minimal of general type with $K^2_Y=dK^2_S$, $\chi(Y)=d\chi(S)$. Hence $K^2_Y\le3\chi(Y)<4\chi(Y)$ and therefore, by the Severi inequality ([@severi]), the image of the Albanese map of $Y$ is a curve. Write $a\colon Y\to B$ for the Albanese pencil, and let $b$ be the genus of $B$ and $g$ the genus of the general fibre $F$ of $a$. The Galois group $G$ of $\pi$ acts on the curve $B$. This action is effective by the assumption that $\pi$ is minimal among the irregular étale covers of $Y$. Hence we have a commutative diagram: $$\label{diagram}
\begin{CD}
Y @>{\pi}>>S \\
@V{a}VV @VV{f}V\\
B @>{\bar{\pi}}>> {\mathbb P}^1
\end{CD}$$
The map $\bar{\pi}$ is a Galois cover with group $G$ and the general fibre of $f$ is also equal to $F$. Since the map $\pi$ is obtained from $f$ by taking base change with $\bar{\pi}$ and normalizing, the fibre of $f$ over a point $x$ of ${\mathbb P}^1$ has multiplicity equal to the ramification order of $\bar{\pi}$ over $x$. Notice that, since ${\mathbb P}^1$ is simply connected, the branch divisor of $\bar{\pi}$ is nonempty and therefore the fibration $f$ always has multiple fibres. Notice also that, since $S$ is of general type, the existence of multiple fibres implies $g\ge 3$.
We remark that the fibration $a$ is not smooth and isotrivial. In fact, if this were the case then $Y$ would be a free quotient of a product of curves, hence it would satisfy $K^2_Y=8\chi(Y)$. Hence we may define the slope of $a$ (cf. [@xiaoslope]): $$\lambda(a):=\frac{K^2_Y-8(b-1)(g-1)}{\chi(Y)-(b-1)(g-1)}.$$ The slope inequality ([@xiaoslope], cf. also [@ch], [@stoppino]) gives $$\label{slope}
4(g-1)/g\le \lambda(a)\le K^2_Y/\chi(Y)=K^2_S/\chi(S)\le3$$ where the second inequality is a consequence of $b>0$. Hence we get $g=3$ or $g=4$.
Assume $g=4$. In this case (\[slope\]) becomes: $$3\le \lambda(a)\le K^2_S/\chi(S)\le3.$$ It follows that the slope inequality is sharp in this case and $K^2_S=3\chi(S)$. By [@konnoslope Prop. 2.6], this implies that $F$ is hyperelliptic. Let ${\sigma}$ be the involution of $S$ induced by the hyperelliptic involution on the fibres of $f$. The divisorial part $R$ of the fixed locus of ${\sigma}$ satisfies $FR=10$. As remarked above, $f$ has at least a fibre of multiplicity $m>1$, that we denote by $mA$. Since $g=4$, by the adjunction formula $\frac{6}{m}$ is divisible by $2$, yielding $m=3$. Hence $3AR=10$, a contradiction. So we have proved $g=3$.
Using the adjunction formula again, we see that the multiple fibres of $f$ are double fibres, hence all the branch points of $\bar{\pi}$ have ramification order equal to $2$. Let $k$ be the number of branch points of $\bar{\pi}$. By applying the Hurwitz formula to $\bar{\pi}$, we get $k\ge 4$.
Given an irregular étale cover $\rho\colon Z\to S$, we can always find an étale cover $W\to S$ which dominates both $Z$ and $Y$. The Albanese pencil of $W$ is a pullback both from $Y$ and from $Z$, hence the fibrations induced on $S$ by the Albanese pencils of $Z$, $W$ and $Y$ are the same.
We introduce some more notation. Assume that $f\colon S\to {\mathbb P}^1$ is the fibration defined in Lemma \[fibre3\]. Let $\bar{\pi}\colon B\to {\mathbb P}^1$ be the double cover branched on 4 points corresponding to double fibres $2F_1,\dots,2F_4$ of $f$ and $\pi\colon Y\to S$ the étale double cover obtained by base change with $\bar{\pi}$ and normalization, as in diagram (\[diagram\]). Then $K^2_Y=2K^2_S$, $\chi(Y)=2\chi(S)$ and $q(Y)=1$. We write $\eta:=F_1+F_2-F_3-F_4$. Clearly, $\eta$ has order 2 in $\operatorname{Pic}(S)$ and $\pi$ is the étale double cover corresponding to $\eta$.
\[hyp\] The general fibre $F$ of $f$ is hyperelliptic.
Assume by contradiction that $F$ is not hyperelliptic and consider the pencil $a\colon Y\to B$, whose general fibre is also equal to $F$. Set $\mathcal E:=a_*\omega_Y$ and denote by $\psi\colon Y\to {\mathbb P}(\mathcal E)$ the relative canonical map, which is a morphism by Remark 2.4 of [@konnoslope]. Let $V$ be the image of $\psi$. The surface $V$ is a relative quartic in ${\mathbb P}(\mathcal E)$ and, by Lemma 3.1 and Theorem 3.2 of [@konnoslope], its singularities are at most rational double points. The map $\psi$ is birational and it contracts precisely the nodal curves of $Y$, which are all vertical since $B$ has genus 1. Hence $V$ is the canonical model of $Y$.
Let $\iota$ be the involution associated to the cover $Y\to S$. This involution induces automorphisms of $B$, $\mathcal E$, ${\mathbb P}(E)$ and $V$ (that we denote again by $\iota$) compatible with $a$, $\psi$ and the inclusion $V\subset {\mathbb P}(\mathcal E)$. Given $b\in B$, write ${\mathbb P}^2_b$ for the fiber of ${\mathbb P}(\mathcal E)$ over $b$ and $V_b:=V\cap {\mathbb P}^2_b$. The curve $V_b$ is a plane quartic inside ${\mathbb P}^2_b$. For every $b\in B$, the map $\iota$ induces a projective isomorphism between ${\mathbb P}^2_b$ and ${\mathbb P}^2_{\iota(b)}$ that restricts to an isomorphism of $V_b$ with $V_{\iota(b)}$. In particular, if $b$ is one of the four fixed points of $\iota$ on $B$, then $\iota$ induces an involution of ${\mathbb P}^2_b$ that preserves the quartic $V_b$. Since the fixed locus of an involution of the plane contains a line, it follows that $\iota$ has at least a fixed point on $V_b$. In particular, the action of $\iota$ on $V$ is not free.
On the other hand, one checks that a fixed point free automorphism of a minimal surface of general type induces a fixed point free automorphism of the canonical model. So we have a contradiction.
We can now give:
The “if” part is a consequence of Lemma \[fibre3\] and Lemma \[hyp\]. Conversely, if $S$ has a fibration with 4 double fibres $2F_1,\dots,2F_4$ then the étale double cover associated with $\eta:=F_1+F_2-F_3-F_4$ has irregularity equal to 1.
The proof of Theorem \[K3\]
===========================
In this section we let $S$ denote a smooth minimal surface of general type with $p_g(S)=0$, $K^2_S=3$. To prove Theorem \[K3\] we argue by contradiction.
Thus assume that $S$ has an irregular étale cover. Then by Theorem \[main\] there exists a fibration $f\colon S\to{\mathbb P}^1$ whose general fibre is hyperelliptic of genus 3 and with at least 4 double fibres $2F_1,\dots,2F_4$. As before, denote by $\pi\colon Y\to S$ the étale double cover given by $\eta=F_1+F_2-F_3-F_4$ and by $\iota$ the involution associated with $\pi$. The invariants of $Y$ are: $q(Y)=1$, $p_g(Y)=2$, $K^2_Y=6$.
The hyperelliptic involution on the fibres of $a\colon Y\to B$ and $f\colon S\to Y$ induces involutions $\tau$ of $Y$ and ${\sigma}$ of $S$. By construction, these involutions are compatible with the map $\pi\colon Y\to S$, namely we have $\pi\circ \tau={\sigma}\circ\pi$. We denote by $p\colon S\to {\Sigma}:=S/{\sigma}$ the quotient map.
\[commute\] The involutions $\tau$ and $\iota$ of $Y$ commute.
Denote by $h$ the composite map $Y\to S\to{\Sigma}$. By construction, both $\iota$ and $\tau$ belong to the Galois group $G$ of $h$. Since $h$ has degree 4 and $\iota$ and $\tau$ are involutions, the group $G$ is isomorphic to ${\mathbb Z}_2\times {\mathbb Z}_2$ and $\iota$ and $\tau$ commute.
\[16\] The involution $\iota\tau$ has at least $16$ isolated fixed points on $Y$.
Let $q\colon Y\to Z:=Y/\iota\tau$ be the quotient map. The surface $Z$ is nodal. The regular $1$-forms and $2$-forms of $Z$ correspond to the elements of $H^0(Y, \Omega^1_Y)$, respectively $H^0(Y,\omega_Y)$, that are invariant under the action of $\iota\tau$. By the same argument, since $p_g(S)=p_g(Y/\tau)=0$, both $\iota$ and $\tau$ act on $H^0(Y,\omega_Y)$ as multiplication by $-1$. It follows that $\iota\tau$ acts trivially on $H^0(Y,\omega_Y)$ and $p_g(Z)=2$. Since $\iota$ acts on $B$ as an involution with quotient ${\mathbb P}^1$ and $\tau$ acts trivially on $B$, it follows that the action of $\iota\tau$ on $B$ is equal to the action of $\iota$ and that $q(Z)=0$.
Let $D$ be the divisorial part of the fixed locus of $\iota\tau$ on $Y$ and let $k$ be the number of isolated fixed points of $\iota\tau$. We recall the Holomorphic Fixed Point formula (see [@as], p.566): $$\sum_i(-1)^i \operatorname{Tr}(\iota\tau|H^i(Y,{\mathcal{O}}_Y))= (k-K_YD)/4.$$ By the above considerations, this can be rewritten as: $$k=16+K_YD.$$ The statement now follows from the fact that $K_Y$ is nef.
By Lemma \[commute\], the involution $\iota\tau$ of $Y$ induces ${\sigma}$ on $S$. By Lemma \[16\], $\iota\tau$ has at least 16 isolated fixed points. Since the images on $S$ of these points are isolated fixed points of ${\sigma}$, the involution ${\sigma}$ has at least 8 isolated fixed points. On the other hand, by [@CCM Prop. 3.3] there are at most $K^2_S+4=7$ isolated fixed points of ${\sigma}$. So we have a contradiction, and thus $S$ has no irregular étale cover.
The proof of Theorem \[T2\]
===========================
To prove Theorem \[T2\] we will use the following two results proved in [@beauville Cor. 5.8], although not stated explicitly.
\[be\] Let $Y$ be a surface of general type such that the canonical map of $Y$ has degree $2$ onto a rational surface. If $G$ is a group that acts freely on $Y$, then $G={\mathbb Z}_2^r$, for some $r$.
The group $G$ is finite, since a surface of general type has finitely many automorphisms.
Let $T$ be the quotient of $Y$ by the canonical involution. The surface $T$ is rational, with canonical singularities, and $G$ acts on $T$.
Since $T$ is rational, each element $g\in G$ acts with fixed points. The argument in the proof of [@beauville Cor. 5.8] shows that each $g$ has order $2$, hence $G={\mathbb Z}_2^r$.
\[Z2\] Let $S$ be a minimal surface of general type such that $K^2_S<3\chi(S)$, and $S$ has no irregular étale cover. If $Y\to S$ is an étale G-cover, then either $|G|\leq 10$ or $G={\mathbb Z}_2^r$, for some $r\geq 4$.
Let $\pi\colon Y\to S$ be an étale G-cover of degree $d>10$. By assumption we have $q(Y)=0$ and $K^2_Y<3p_g(Y)-7$, and therefore the canonical map of $Y$ is 2-to-1 onto a rational surface by [@beauville Theorem 5.5]. Hence $G={\mathbb Z}_2^r$ for some $r\ge 4$ by Proposition \[be\].
For related statements see the results of [@xiaohyp] on hyperelliptic surfaces and the results of [@ak] and [@konnopg].
We remark that the next result is well known for the cases $\chi(S)=1$ and $K^2_S=1$ or $2$ ([@milesk2]).
\[pi9\] Let $S$ be a minimal surface of general type with $K^2_S<3\chi(S)$. If $S$ has no irregular étale cover, then $|{\pi_1^{\mathrm{alg}}}(S)|\le9$.
Let $Y\to S$ be an étale G-cover. By Corollary \[Z2\], it is enough to exclude the following possibilities: a) $G={\mathbb Z}_2^r$ for some $r\ge 4$, and b) $|G|=10$.
Consider case a) and assume by contradiction that $\pi\colon Y\to S$ is a Galois étale cover with Galois group $G={\mathbb Z}_2^4$. By [@miyaoka1], $\chi(S)\geq 2$. We have $\chi(Y)=16\chi(S)\ge 32$ and $K^2_Y<3(\chi(Y)-5)$. Notice that, since $K^2_Y<3\chi(Y)-10$, by [@beauville Theorem 5.5] the surface $Y$ has a pencil of hyperelliptic curves. Hence $Y$ satisfies the assumptions of [@xiaohyp Theorem 1] and there exists a unique free pencil $|F|$ of hyperelliptic curves of genus $g\le3$ on $Y$. The action of $G$ preserves $|F|$ by the uniqueness of $|F|$. Since $\operatorname{Aut}({\mathbb P}^1)$ does not contain a subgroup isomorphic to ${\mathbb Z}_2^3$, there is a subgroup $H<G$ of order $\ge 4$ that maps every curve of $|F|$ to itself. Since the action of $G$ on $Y$ is free, this implies that $g-1$ is divisible by 4, contradicting $g\le3$ and $S$ of general type.
Consider now case b) and assume by contradiction that $\pi \colon Y\to S$ is a Galois cover with Galois group $G$ of order 10. For $K^2_S<3\chi(S)-1$, we have $K^2_Y<3\chi(Y)-10$ and, as in the proof of Corollary \[Z2\], $G$ is of the form ${\mathbb Z}_2^a$, a contradiction. So we have $K^2_S=3\chi(S)-1$, $K^2_Y=3\chi(Y)-10$, $q(Y)=0$ and so, by [@ak], the canonical map of $Y$ is either birational or 2-to-1 onto a rational surface. By Proposition \[be\], the last possibility does not occur, since $G$ has order 10.
The surface $Y$ satisfies $p_g(Y)=10\chi(S)-1\ge 9$. Surfaces on the Castelnuovo line $K^2=3\chi-10$ with birational canonical map are classified (cf. [@harris], [@rick] and [@ak]): for $p_g(Y)\ge 8$, the canonical model $V$ of $Y$ is a relative quartic inside a ${\mathbb P}^2$-bundle $${\mathbb P}:=\operatorname{Proj}({\mathcal{O}}_{{\mathbb P}^1}(a)\oplus {\mathcal{O}}_{{\mathbb P}^1}(b)\oplus{\mathcal{O}}_{{\mathbb P}^1}(c)),$$ where $0\le a\le b\le c$ and $a+b+c=p_g(Y)+3$.
If the Galois group $G$ preserves the fibration $f\colon V\to{\mathbb P}^1$ induced by the projection ${\mathbb P}\to{\mathbb P}^1$, then, as in Lemma \[hyp\], we obtain a contradiction by considering the action on $V$ of an element of order 2 of $G$.
So, to conclude the proof we just have to show that $G$ preserves $f$. Let $W$ be the image of ${\mathbb P}$ via the tautological linear system. By the results of [@ak], [@harris], [@rick], the threefold $W$ is the intersection of all the quadrics that contain the canonical image of $Y$ and therefore it is preserved by the automorphisms of $V$. One checks that $W$ has a unique ruling by planes which induces the fibration $f$ on $V$. Therefore every automorphism of $V$ preserves the fibration $f$.
To obtain the statement of Theorem \[T2\] we now show the following:
\[campedelli\] Let $S$ be a minimal surface of general type with $K^2_S<3\chi(S)$. If $|{\pi_1^{\mathrm{alg}}}(S)|=9$, then $\chi(S)=1$ and $K^2_S=2$, namely $S$ is a numerical Campedelli surface.
Suppose that $|{\pi_1^{\mathrm{alg}}}(S)|=9$ and $\chi(S)\geq 2$. The argument in the proof of Proposition \[pi9\] shows that $K^2_S=3\chi(S)-1$. Let $\pi\colon Y\to S$ be the universal cover. We have $K^2_Y=3p_g(Y)-6$, $p_g(Y)=9\chi(Y)-1\ge 17$. By [@konnopg Lem. 2.2] the bicanonical map of $Y$ has degree 1 or 2. Arguing as in the proof of Proposition \[pi9\], one shows that the bicanonical map of $Y$ is birational. Then, since $p_g(Y)\ge 11$, by the results of [@konnopg] the situation is analogous to the case of a surface with $K^2=3p_g-7$ and birational canonical map. Namely, the intersection of all the quadrics through the canonical image of $Y$ is a threefold $W$, which is the image of a ${\mathbb P}^2$-bundle ${\mathbb P}:=\operatorname{Proj}({\mathcal{O}}_{{\mathbb P}^1}(a)\oplus{\mathcal{O}}_{{\mathbb P}^1}(b)\oplus{\mathcal{O}}_{{\mathbb P}^1}(c))$ via the tautological linear system, and $Y$ is birational to a relative quartic of ${\mathbb P}$. In particular, there is a fibration $f\colon Y\to{\mathbb P}^1$ with general fibre a nonhyperelliptic curve of genus 3. One can show as above that the Galois group $G={\pi_1^{\mathrm{alg}}}(S)$ of $\pi$ preserves $f$. Then we obtain a contradiction, since the multiple fibres of a genus 3 fibration are double fibres and a smooth genus 3 curve does not admit a free action of a group of order 9.
[*Remark.*]{} Numerical Campedelli surfaces with fundamental group ${\mathbb Z}_9$ and ${\mathbb Z}_3^2$ do exist (cf. [@xiaog2 Ex. 4.11], [@pi9]).
[ABC]{} T. Ashikaga, K. Konno, [*Algebraic surfaces of general type with $c\sp 2\sb 1=3p\sb g-7$*]{}, Tohoku Math. J. (2) [**42**]{} (1990), no. 4, 517–536. MR1076174, Zbl 0735.14026. M.F. Atiyah, I.M. Singer, [*The index of elliptic operators: III*]{}, Ann. of Math. [**87**]{} (1968), 546–604. MR0236952, Zbl 0164.24301.
W. Barth, K. Hulek, C. Peters, A. Van de Ven, [*Compact complex surfaces*]{}, 2nd edition, Ergebnisse der Mathematik und ihrer Grenzgebiete, [**3.**]{} Folge, Band [**4**]{}, Springer 2004. MR2030225, Zbl 1036.14016.
A. Beauville, [*L’application canonique pour les surfaces de type général*]{}. Inv. Math. [**55**]{} (1979), 121–140. MR0553705, Zbl 0403.14006. E. Bombieri, [*Canonical models of surfaces of general type*]{}, Inst. Hautes Études Sci. Publ. Math. No. 42 (1973), 171–219. MR0318163, Zbl 0259.14005. A. Calabri, C. Ciliberto, M. Mendes Lopes, [*Numerical Godeaux surfaces with an involution*]{}, to appear in Trans. A.M.S.. F. Catanese, [*Fibred Kähler and quasi-projective groups*]{}, Adv. Geom., Special issue dedicated to Adriano Barlotti suppl. (2003), S13–S27. MR2028385, Zbl 1051.32013. F. Catanese, J. Keum, K. Oguiso, [*Some remarks on the universal cover of an open $K3$ surface*]{}, Math. Ann. 325 (2003), no. 2, 279–286. MR1962049, Zbl 1073.14535. M. Cornalba, J. Harris, [*Divisor classes associated to families of stable varieties, with applications to the moduli space of curves*]{}, Ann. Sci. École Norm. Sup., (4) [**21**]{} (1988), no. 3, 455–475. MR0974412, Zbl 0674.14006. J. Harris, [*A bound on the geometric genus of projective varieties*]{}, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) [**8**]{} (1981), no. 1, 35–68. MR0616900, Zbl 0467.14005. E. Horikawa, [*Algebraic surfaces of general type with small $c\sp{2}\sb{1}$. V*]{}, J. Fac. Sci. Univ. Tokyo Sect. IA Math. [**28**]{} 3 (1981), 745–755 (1982). MR0656051, Zbl 0505.14028. K. Konno, [*Algebraic surfaces of general type with $c\sp 2\sb 1=3p\sb g-6$*]{}, Math. Ann. [**290**]{} (1991), no. 1, 77–107. MR1107664, Zbl 0711.14021. K. Konno, [*Non-hyperelliptic fibrations of small genus and certain irregular canonical surfaces*]{}, Ann. Sc. Norm. Sup. Pisa Cl. Sci. (4) [**20**]{} (1993), 575–595. MR1267600, Zbl 0822.14009. Y. Miyaoka, [*On numerical Campedelli surfaces*]{}, Complex Anal. Algebr. Geom., Collect. Pap. dedic. K. Kodaira (1977), 113–118. MR0447258, Zbl 0365.14007 . M. Mendes Lopes, R. Pardini, [*Numerical Campedelli surfaces with fundamental group of order 9*]{}, J.E>M.S. (to appear). math.AG/0602633. M. Mendes Lopes, R. Pardini, [*The order of finite algebraic fundamental groups of surfaces with $K^2\le3\chi-2$*]{}, in “Algebraic geometry and Topology” Suurikaiseki kenkyusho Koukyuuroku, No. 1490 (2006), 69–75, math.AG/0605733. R. Miranda, [*On canonical surfaces of general type with $K\sp 2=3\chi-10$*]{}, Math. Z. [**198**]{} (1988), no. 1, 83–93. MR0938031, Zbl 0622.14028. D. Naie, [*Surfaces d’Enriques et une construction de surfaces de type général avec $p_g=0$*]{}, Math. Z. [**215**]{} (2) (1994), 269–280. MR1259462, Zbl 0791.14016. R. Pardini, [*The Severi inequality $K^2\ge 4\chi$ for surfaces of maximal Albanese dimension*]{}, Invent. math. [**159**]{} (2005), no. 3, 669 –672. MR2125737, Zbl 1082.14041. M. Reid, [*$\pi \sb{1}$ for surfaces with small $K\sp{2}$*]{}, Algebraic geometry (Proc. Summer Meeting, Univ. Copenhagen, Copenhagen, 1978), 534–544, Lecture Notes in Math., 732, Springer-Verlag, , Berlin, (1979). MR0555716, Zbl 0423.14021. M. Reid, [*Surfaces with $p_g=0$, $K^2_S=2$*]{}, preprint available at http://www.maths.warwick.ac.uk/$\!\sim$miles/surf/ L. Stoppino, [*A remark on the slope inequality for fibred surfaces*]{}, preprint math.AG/0411639 G. Xiao, [*Surfaces fibrées en courbes de genre deux*]{}, Lecture Notes in Mathematics, 1137, Springer-Verlag, Berlin, (1985). MR0872271, Zbl 0579.14028. G. Xiao, [*Hyperelliptic surfaces of general type with $K^2<4\chi$*]{}, Manuscripta Math. [**57**]{} (1987), 125–148. MR0871627, Zbl 0615.14022. G. Xiao, [*Fibered algebraic surfaces with low slope*]{}, Math. Ann. [**276**]{} (1987), no. 3, 449–466. MR0875340, Zbl 0596.14028.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, the structural controllability of switched linear systems is investigated. In particular, switched linear systems with independent subsystem models are studied. The structural controllability is a generalization of the traditional controllability concept for dynamical systems, and purely based on the graphic topologies among state and input vertices. First, two kinds of graphic representations of switched linear systems are proposed. Second, graph theory based necessary and sufficient characterizations of the structural controllability for switched linear systems are presented. Finally, the paper concludes with one illustrative example and discussions on the results and future work.'
author:
- 'Xiaomeng Liu$^{a}$, Hai Lin$^{b}$$^{\ast}$, and Ben M. Chen$^{a}$, [^1] [^2] [^3]'
title: 'Structural Controllability of Switched Linear Systems\*'
---
[Xiaomeng Liu : Structural Controllability of Switched Linear Systems]{}
Structural controllability, switched linear system, graphic interpretation.
Introduction
============
a special class of hybrid control systems, a switched linear system consists of several linear subsystems and a rule that orchestrates the switching among them. Switching between different subsystems or different controllers can greatly enrich the control strategies and may accomplish certain control objective which can not be achieved by conventional dynamical systems. For example, it provided an effective mechanism to cope with highly complex systems and/or systems with large uncertainties [@AWV]. References [@KJ] presented good examples that switched controllers could provide a performance improvement over a fixed controller. Besides, switched linear systems also have promising applications in control of mechanical systems, aircrafts, satellites and swarming robots. Driven by its importance in both theoretical research and practical applications, switched linear system has attracted considerable attention during the last decade [@linhai]-[@D2].
Much work has been done on the controllability of switched linear systems. For example, the controllability and reachability for low-order switched linear systems have been presented in [@KJI]. Under the assumption that the switching sequence is fixed, references [@JA][@F] introduced some sufficient conditions and necessary conditions for controllability of switched linear systems. Complete geometric criteria for controllability and reachability were established in [@sunzd][@GL]. Up to now, all the previous work mentioned above has been based on the traditional controllability concept of switched linear systems. In this paper, we propose a new notion for the controllability of switched linear system: structural controllability, which may be more reasonable in face of uncertainties. Actually, it is more often than not that most of system parameter values are difficult to identify and only known to certain approximations. On the other hand, we are usually pretty sure where zero elements are either by coordination or by the absence of physical connections among components of the system. Thus structural properties that are independent of a specific value of unknown parameters, e.g., the structural controllability studied here, are of particular interest. It is therefore assumed here that all the elements of matrices of switched linear systems are fixed zeros or free parameters. Furthermore, the switched linear system is said to be structurally controllable if one can find a set of values for the free parameters such that the corresponding switched linear system is controllable in the classical sense. For linear structured systems, generic properties including structural controllability have been studied extensively and it turns out that generic properties including structural controllability are true for almost all values of the parameters [@lin]-[@LD]. That is also true for switched linear systems studied here and presents one of the reasons why this kind of structural controllability is of interest.
Graphic conditions can help to understand how the graphic topologies of dynamical systems influence the corresponding generic properties, here especially for the structural controllability. This would be helpful in many practical applications. For example, in multi-agent systems, graphic interpretations for structural controllability help us to understand the necessary information exchange among agents to make the whole team well-behaved, e.g., controllable. Therefore, this motivates our pursuit on illuminating the structural controllability of switched linear systems from a graph theoretical point of view. In this paper, we propose two graphic representations of switched linear systems and finally, it turns out that the structural controllability of switched linear systems only depends on the graphic topologies of the corresponding systems.
The organization of this paper is as follows: In Section II, we introduce some basic preliminaries and the problem formualtion, followed by structural controllability study of switched linear systems in Section III, where several graphic necessary and sufficient conditions for the structural controllability are given. One illustrative example together with discussions on a more general case are also presented. Finally, some concluding remarks are drawn in Section IV.
Preliminaries and Problem Formulation
=====================================
Graph Theory Preliminaries
--------------------------
First of all, definition and example of a structured matrix are introduced as follows:
$P$ is said to be a structured matrix if its entries are either fixed zeros or independent free parameters. $\tilde{P}$ is called admissible (with respect to $P$) if it can be obtained by fixing the free parameters of $P$ at some particular values. In addition $P_{ij}$ is adopted to represent the element of $P$ from row $i$ and column $j$.
$P$=$\left[
\begin{array}{ccc}
0&\lambda_1\\
\lambda_2&\lambda_3\\
\end{array}\right]$ is a structured matrix, where $\lambda_1$, $\lambda_2$ and $\lambda_3$ are free parameters, and $\tilde{P}$=$\left[
\begin{array}{ccc}
0&1\\
2&0\\
\end{array}\right]$ is admissible with respect to $P$.
Now consider a linear control system: $$\label{l1}
\dot{x}=Ax(t)+Bu(t),$$ where $x(t)\in \mathbb{R}^n$ and $u(t)\in \mathbb{R}^r$. The matrices $A$ and $B$ are structured matrices, which means that their elements are either fixed zeros or free parameters. This structured system given by matrix pair $(A, B)$ can be described by a directed graph [@lin]:
The representation graph of structured system $(A, B)$ is a directed graph $\mathcal{G}$, with vertex set $\mathcal{V}=\mathcal{X}\cup
\mathcal{U}$, where $\mathcal{X}=\{x_1,x_2,\ldots,x_n\}$, which is called $state~vertex~ set$ and $\mathcal{U}=\{u_1,u_2,\dots,u_r\}$, which is called $input~vertex~ set$, and edge set $\mathcal{I}=\mathcal{I}_{UX}\cup\mathcal{I}_{XX}$, where $\mathcal{I}_{UX}=\{(u_i,x_j)|B_{ji}\neq 0, 1\leq i \leq r, 1\leq
j\leq n\}$ and $\mathcal{I}_{XX}=\{(x_i,x_j)|A_{ji}\neq 0, 1\leq i
\leq n, 1\leq j\leq n\}$ are the oriented edges between inputs and states and between states defined by the interconnection matrices $A$ and $B$ above. This directed graph (for notational simplicity, we will use digraph to refer to directed graph) $\mathcal{G}$ is also called the graph of matrix pair $(A,B)$ and denoted by $\mathcal{G}(A,B)$.
Note that the total number of vertices in $\mathcal{G}(A,B)$ equals to the summation of states number and inputs number. One important graphic definition is needed before we proceed forward:
\[def14\](*Stem* [@lin]) An alternating sequence of distinct vertices and oriented edges is called a directed path, in which the terminal node of any edge never coincide to its initial node or the initial or the terminal nodes of the former edges. A stem is a directed path in the state vertex set $\mathcal{X}$, that begins in the input vertex set $\mathcal{U}$.
Two graphic properties ‘accessibility’ and ‘dilation’ were proposed by [@lin], which will serve as the basis of following discussion. We state them as follows:
(*Accessibility* [@lin]) A vertex (other than the input vertices) is called $nonaccessible$ if and only if there is no possibility of reaching this vertex through any stem of the graph $\mathcal{G}$.
\[def17\](*Dilation* [@lin]) Consider one vertex set $S$ formed by the vertices from the state vertices set $\mathcal{X}$ and determine another vertex set $T(S)$, which contains all the vertices $v$ with the property that there exists an oriented edge from $v$ to one vertex in $S$. Then the graph $\mathcal{G}$ contains a ‘$dilation$’ if and only if there exist at least a set $S$ of $k$ vertices in the vertex set of the graph such that there are no more than $k-1$ vertices in $T(S)$.
Switched Linear System, Controllability and Structural Controllability
----------------------------------------------------------------------
In general, a switched linear system is composed of a family of subsystems and a rule that governs the switching among them, and is mathematically described by $$\begin{aligned}
\label{eq2}
\dot x(t)=&A_{\sigma(t)} x(t)+B_{\sigma(t)} u(t) ,
%y(t)=&C_\sigma x(t) .\end{aligned}$$ where $x(t)\in \mathbb{R}^n$ are the states, $u(t)\in
\mathbb{R}^{r}$ are piecewise continuous input, $\sigma:[0,\infty)\rightarrow M \triangleq \{1,\ldots,m\}$, where time slot 0 is system initial time, is the switching signal. System (\[eq2\]) contains $m$ subsystems $(A_i,B_i),$ $i\in
\{1,\ldots,m\}$ and $\sigma(t)$= $i$ implies that the $i$th subsystem $(A_i,B_i)$ is active at time instance $t$. In the sequel, the following definition of controllability of system (\[eq2\]) will be adopted (Definition 7 in [@sunzd]):
\[def1\] Switched linear system (\[eq2\]) is said to be (completely) controllable if for any initial state $x_0$ and final state $x_f$, there exist a time instance $t_f>0$, a switching signal $\sigma:[0,t_f)\rightarrow
M$ and an input $u:[0,t_f)\rightarrow \mathbb{R}^{r}$ such that $x(0)=x_0$ and $x(t_f)=x_f$.
For the controllability of switched linear systems, a well-known matrix rank condition was given in [@ZS]:
\[lem1\]([@ZS]) If matrix:$$\begin{split}\label{eq3} &[B_1,B_2,\ldots,B_m,
A_1B_1,A_2B_1,\ldots,A_mB_1,A_1B_2,A_2B_2,\ldots,A_mB_2,\ldots,
A_1B_m,A_2B_m,\ldots,A_mB_m,\\&A_1^2B_1,A_2A_1B_1,\ldots,A_mA_1B_1,A_1A_2B_1,A_2^2B_1,\ldots,A_mA_2B_1,\ldots,A_1A_mB_m,A_2A_mB_m,\ldots,A_m^2B_m,
\\&,\ldots,\\&A_1^{n-1}B_1,A_2A_1^{n-2}B_1,\ldots,A_mA_1^{n-2}B_1,A_1A_2A_1^{n-3}B_1,A_2^2A_1^{n-3}B_1,\ldots,A_mA_2A_1^{n-3}B_1\ldots,A_1A_m^{n-2}B_m,
\\&A_2A_m^{n-2}B_m\ldots,
A_m^{n-1}B_m] \end{split}$$has full row rank $n$, then switched linear system (\[eq2\]) is controllable, and vice versa.
This matrix is called controllability matrix of switched linear system (\[eq2\]) and for simplicity, we will use $\mathcal{C}$$(A_1,\ldots,A_m,B_1,\ldots,B_m)$ to represent it. If we use $Im$$P$ to represent the range space of arbitrary matrix $P$, actually, $Im$$\mathcal{C}$$(A_1,\ldots,A_m,B_1,\ldots,B_m)$ is the controllable subspace of switched linear system (\[eq2\])[@sunzd][@ZS]. The above lemma implies that system (\[eq2\]) is controllable if and only if $Im$$\mathcal{C}$$(A_1,\ldots,A_m,B_1,\ldots,B_m)=\mathcal{R}^n$. Besides, controllable subspace can be expressed as $\langle
A_1,\ldots,A_m|B_1,\ldots,B_m \rangle$, which is the smallest subspace containing $Im$$B_i$, $i=1,\ldots,m$ and invariant under the transformations $A_1,\ldots,A_m$ [@D2].
In view of structural controllability, system (\[eq2\]) will be treated as structured switched linear system defined as:
For structured system (\[eq2\]), the elements of all the matrices $(A_1, B_1,\ldots, A_m, B_m)$ are either fixed zero or free parameters and free parameters in different subsystems $(A_i,B_i),
i\in M$ are independent. A numerically given matrices $(\tilde{A}_1,
\tilde{B}_1,\ldots, \tilde{A}_m, \tilde{B}_m)$ is called an admissible numerical realization (with respect to $(A_1, B_1,\ldots,
A_m, B_m)$) if it can be obtained by fixing all free parameter entries of $(A_1, B_1,\ldots, A_m, B_m)$ at some particular values.
Similar with the definition of structural controllability of linear system in [@K], we have the following definition for structural controllability of switched linear system (\[eq2\]):
\[def2\] Switched linear system (\[eq2\]) given by its structured matrices $(A_1, B_1,\ldots, A_m, B_m)$ is said to be structurally controllable if and only if there exists at least one admissible realization $(\tilde{A}_1,
\tilde{B}_1,\ldots, \tilde{A}_m, \tilde{B}_m)$ such that the corresponding switched linear system is controllable in the usual numerical sense.
Before proceeding further, we need to introduce the definition of $g$-rank of one matrix:
The generic rank ($g$-rank) of a structured matrix $P$ is defined to be the maximal rank that $P$ achieves as a function of its free parameters.
Then, we have the following algebraic condition for structural controllability of system (\[eq2\]):
\[lem2\] Switched linear system (\[eq2\]) is structurally controllable if and only if
$g$-rank $\mathcal{C}$$(A_1,\ldots,A_m,B_1,\ldots,B_m)$ = $n$.
Structural Controllability of Switched Linear Systems
=====================================================
Criteria Based on Union Graph
-----------------------------
For switched linear system (\[eq2\]), let’s use $\mathcal {G}_i$ with vertex set $\mathcal{V}_i$ and edge set $\mathcal{I}_i$ to represent the underlying graph of subsystem $(A_i,B_i)$, $i \in
\{1,\ldots,m\}.$ As to the whole switched system, one kind of representation graph, which is called union graph, is described in the following definition:
\[def3\] Switched linear system (\[eq2\]) can be represented by a union digraph $\mathcal{G}$ (sometimes named union graph without leading to confusion). Mathematically, $\mathcal{G}$ is defined as $$\begin{aligned}
\mathcal{G}_1\cup \mathcal{G}_2\cup\ldots\cup
\mathcal{G}_m=\{\mathcal{V}_1\cup \mathcal{V}_2\cup \ldots \cup
\mathcal{V}_m;\mathcal{I}_1\cup \mathcal{I}_2\cup \ldots\cup
\mathcal{I}_m\}\nonumber\end{aligned}$$ For the union graph $\mathcal{G}$, the vertex set is the same as the vertex set of every subgraph $\mathcal{G}_i$. The edge set of $\mathcal{G}$ equals to the union of the edge sets of the subgraphs. Note that there are no multiple edges between any two vertices in $\mathcal{G}$.
\[rem3\]It turns out that union graph $\mathcal{G}$ is the representation of linear structured system: $(A_1+A_2+\ldots+A_m,B_1+B_2+\ldots+B_m)$. The reason is that: If the element at position $a_{ji}(b_{ji})$ in $[A_1+A_2+\ldots+A_m, B_1+B_2+\ldots+B_m]$ is a free parameter, this implies that there exist some matrices $[A_p, B_p]$, $p=1,\ldots,m$ such that the element at position $a_{ji}(b_{ji})$ is also a free parameter and in the corresponding subgraph $\mathcal{G}_p$, there is an edge from vertex $i$ to vertex $j$. According to the definition of union graph, it follows that there is also an edge from vertex $i$ to vertex $j$ in union graph $\mathcal{G}$. If the element at position $a_{ji}(b_{ji})$ in $[A_1+A_2+\ldots+A_m,
B_1+B_2+\ldots+A_m]$ is zero, this implies that for every matrices $[A_p, B_p]$, $p=1,\ldots,m$, the element at position $a_{ji}(b_{ji})$ is zero and in the corresponding subgraph $\mathcal{G}_p$, there is no edge from vertex $i$ to vertex $j$. It follows that there is also no edge in union graph $\mathcal{G}$ from vertex $i$ to vertex $j$.
Before proceeding further, we need to introduce two definitions which were proposed in [@lin] for linear system (\[l1\]) first:
\[def5\]([@lin]) The matrix pair $(A,B)$ is said to be reducible or of form I if there exists a permutation matrix $P$ such that they can be written in the following form: $$\begin{aligned}
PAP^{-1}=\left[
\begin{array}{ccc}
A_{11}&0\\
A_{21}&A_{22}\\
\end{array}\right],PB=\left[
\begin{array}{cc}
0\\
B_{22}\\
\end{array}\right],\end{aligned}$$\
where $ A_{11}\in \mathbb{R}^{p \times p}$ , $A_{21 }
\in \mathbb{R}^{(n - p) \times p}$,$ $ $ A_{22} \in \mathbb{R}^{(n
- p) \times (n- p)} $ and $ B_{22} \in \mathbb{R}^{(n- p) \times
r}$.
\[rem4\] Whenever the matrix pair $(A,B)$ is of form I, the system is structurally uncontrollable [@lin] and meanwhile, the controllability matrix $C\triangleq\left[B,AB,\ldots,A^{n-1}B\right]$ will have at least one row which is identically zero for all parameter values [@KL]. If there is no such permutation matrix $P$, we say that the matrix pair $(A,B)$ is irreducible.
\[def6\] ([@lin]) The matrix pair $(A,B)$ is said to be of form II if there exists a permutation matrix $P$ such that they can be written in the following form: $$\begin{aligned}
\left[PAP^{-1},PB\right]=\left[
\begin{array}{ccc}
P_1\\
P_2
\end{array}\right],\end{aligned}$$\
where $P_2\in \mathbb{R}^{(n-k)\times (n+r)}$ , $P_1 \in \mathbb{R}^{k \times (n+r)}$ with no more than $k-1$ nonzero columns (all the other columns of $P_1$ have only fixed zero entries).
The following lemma, which will underpin the following analysis on switched linear systems, details the criteria for evaluating structural controllability of linear system $(A,B)$ [@lin]-[@K]:
\[lem3\]([@lin]-[@K]) For a linear system $(A,B)$, the following statements are equivalent:
1. the pair $(A,B)$ is structurally controllable;
2. i)$[A,B]$ is irreducible or not of form I,\
ii)$[A,B]$ has $g$-rank$[A,B]=n$ or is not of form II;
3. i)there is no nonaccessible vertex in $\mathcal{G}(A,B)$,\
ii)there is no ‘dilation’ in $\mathcal{G}(A,B)$.
This lemma proposed interesting graphic conditions for structural controllability of linear systems and revealed that the structural controllability is totally determined by the underlying graph topology. However, for switched linear systems, to the best of our knowledge, proper graphic representations which can determine the structural controllability properties of switched linear systems are still lacking in the literature.
With the previous lemmas and definitions, we are in the position to present the first main result of the paper, which is actually a graphic sufficient condition for structural controllability of switched linear systems:
\[the1\]Switched linear system (\[eq2\]) with graphic topologies $\mathcal{G}_i$, $i\in \{1,\ldots,m\}$, is structurally controllable if its union graph $\mathcal{G}$ satisfies:
1. there is no nonaccessible vertex in $\mathcal{G}$,
2. there is no ‘dilation’ in $\mathcal{G}$.
Assume the two conditions in this theorem are satisfied. According to remark \[rem3\] and lemma \[lem3\], the corresponding linear system $(A_1+A_2+\ldots +A_m, B_1+B_2+\ldots +B_m)$ is structurally controllable. It follows that there exist some scalars for the free parameters in matrices $(A_i, B_i),i=1,2,\ldots,m$ such that controllability matrix $$\begin{split}
&[B_1+B_2+\ldots+B_m,(A_1+A_2+\ldots+A_m)( B_1+B_2+\ldots+B_m),\\&
(A_1+A_2+\ldots+A_m)^2(B_1+B_2+\ldots+B_m),~\ldots,(A_1+A_2+\ldots+A_m)^{n-1}(
B_1+B_2+\ldots+B_m)]\nonumber\end{split}$$ has full row rank $n$. Expanding the matrix, it follows that matrix $$\begin{split}[&B_1+B_2+\ldots+B_m, A_1B_1+A_2B_1+\ldots+A_mB_1+A_1B_2+A_2B_2\\&
+\ldots+A_mB_2+\ldots
+A_1B_m+A_2B_m\ldots+A_mB_m,~\ldots,~A_1^{n-1}B_1
+A_2A_1^{n-2}B_1+\ldots+A_m^{n-1}B_m]\nonumber\end{split}$$ has full rank $n$.\
The following matrix can be got after adding some column vectors to the above matrix: $$\begin{split}
&[B_1+B_2+\ldots+B_m,
B_2,\ldots,B_m,A_1B_1+A_2B_1+\ldots+A_mB_1+A_1B_2+A_2B_2\\&+\ldots+A_mB_2+\ldots+
A_1B_m+A_2B_m+\ldots+A_mB_m,A_2B_1,\ldots,A_mB_m,~\ldots ,\\&
A_1^{n-1}B_1
+A_2A_1^{n-2}B_1+\ldots+A_1A_m^{n-2}B_1+\ldots+A_m^{n-1}B_m,
A_2A_1^{n-2}B_1,\ldots, A_1A_m^{n-2}B_1,\ldots,
A_m^{n-1}B_m].\nonumber\end{split}$$ Since this matrix still has $n$ linear independent column vectors, it follows that it has full row rank $n$. Next, subtracting $B_2,\ldots,B_m$ from $
B_1+B_2+\ldots+B_m$; subtracting $A_2B_1,\ldots,A_mB_m$ from $
A_1B_1+A_2B_1+\ldots+A_mB_1+\ldots+A_1B_m+\ldots+A_mB_m $ and subtracting $A_2A_1^{n-2}B_1,\ldots,A_1A_m^{n-2}B_1,\ldots,
A_m^{n-1}B_m$ from $A_1^{n-1}B_1
+A_2A_1^{n-2}B_1+\ldots+A_1A_m^{n-2}B_1+\ldots+ A_m^{n-1}B_m$, we can get the following matrix:$$\begin{split}
&[B_1,B_2,\ldots,B_m,A_1B_1,A_2B_1,\ldots,A_mB_m,~\ldots,
\\&A_1^{n-1}B_1,A_2A_1^{n-2}B_1,\ldots,A_1A_m^{n-2}B_1,\ldots,
A_m^{n-1}B_m],\nonumber \end{split}$$ which is the controllability matrix for switched linear systems (\[eq2\]). Since column fundamental transformation does not change the matrix rank, this matrix still has full row rank $n$. Hence, the switched linear system (\[eq2\]) is structurally controllable.
Actually, from the proof, we can see that full rank of controllability matrix of linear system $(A_1+A_2+\ldots+A_m,B_1+B_2+\ldots+B_m)$ in Remark \[rem3\] implies the full rank of controllability matrix of system (\[eq2\]), which means that the structural controllability of this linear system implies structural controllability of system (\[eq2\]). It turns out that this criterion is not necessary for system (\[eq2\]) to be structurally controllable (see the example in subsection $C$). This implies that the union graph does not contain enough information for determining structural controllability. This is because edges from different subsystems are not differentiated in union graph. In the following subsection, another graphic representation of switched linear systems is proposed, from which necessary and sufficient conditions for structural controllability arise.
Criteria Based on Colored Union Graph
-------------------------------------
Another graphic representation:*‘colored union graph’* is defined as follows:
\[def4\] Switched linear system (\[eq2\]) can be represented by a colored union digraph $\mathcal{\tilde{G}}(\mathcal{\tilde{V}},\mathcal{\tilde{I}})$ (sometimes named colored union graph without leading to confusion), where vertex set $\mathcal{\tilde{V}}=\{\mathcal{V}_1\cup
\mathcal{V}_2\cup \ldots \cup \mathcal{V}_m$} and edge set $\mathcal{\tilde{I}}=\{e|e\in \mathcal{I}_i, i=1,2,\ldots,m\}$, i.e., for $i\in \{1,\ldots, m\}$, to each edge $e$ we associate index i in $\mathcal{\tilde{G}}$, if this edge is associated to the subsystem $i$ (subgraph $\mathcal{G}_i$). Note that we associate several indexes (several different colors) to an edge $e$ if it belongs to several subsystems.
With this colored union graph, several graphic properties are introduced in the following lemmas.
\[lem4\] There is no nonaccessible vertex in the colored union graph $\mathcal{\tilde{G}}$ of switched linear system (\[eq2\]) if and only if the matrix $$\label{r7} [A_1+A_2+\cdots+A_m,
B_1+B_2+\cdots+B_m]$$ is irreducible or not of form I.
One vertex is accessible if and only if it can be reached by a stem. From definitions \[def3\] and \[def4\], it follows that there is no nonaccessible vertex in the colored union graph if and only if there is no nonaccessible vertex in the union graph. Besides, from remark \[rem3\], it is clear that the matrix representation of the union graph is $[A_1+A_2+\cdots+A_m, B_1+B_2+\cdots+B_m]. $ According to lemma \[lem3\], there is no nonaccessible vertex in the union graph if and only if matrix (\[r7\]) is irreducible or not of form I. Consequently the equivalence between accessibility of colored union graph and irreducibility of matrix (\[r7\]) gets proved.
A new graphic property ‘$S$-$dilation$’ in colored union graph needs to be introduced here:
\[def7\] In colored union graph $\mathcal{\tilde{G}}$, which is composed of subgraphs $\mathcal{G}_i$, $i=1,2,\ldots,m$, consider one vertex set $S$ formed by the vertices from the state vertex set $\mathcal{X}$ and determine another vertex set $T(S)=\{v|v\in T_i(S),
i=1,2,\ldots,m\}$, where $T_i(S)$ is a vertex set in $\mathcal{G}_i$ which contains all the vertices $w$ with the property that there exists an oriented edge from $w$ to one vertex in $S$. Then $|T(S)|=\sum_{i=1}^m |T_i(S)|$. If $|T(S)|<|S|$, we say that there is a $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$.
Based on this new graphic property, the following lemma can be given:
\[lem5\] There is $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$ of switched linear system (\[eq2\]) if and only if matrix $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$ is of form II. It means that this matrix can be written into: $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$=$\left[
\begin{array}{ccc}
P_1\\
P_2
\end{array}\right]$, where $P_1 \in \mathbb{R}^{p \times k}$ with no more than $p-1$ nonzero columns (all the other columns of $P_1$ have only fixed zero entries).
From [@lin][@hm] or lemma \[lem3\], it is known that in linear systems, there is no ‘dilation’ in the corresponding graph if and only if the matrix pair $[A,B]$ can not be of form II or have $g$-rank $n$. From the explanation of this result in [@lin] and definition \[def6\], $P_1$ in $[A,B]$ has $p$ rows, which actually represents the $p$ vertices of vertex set $S$ (defined for dilation), and each nonzero element of each row of $P_1$ represents that there is one vertex pointing to the vertex presented by this row. Therefore, the number of nonzero columns in $P_1$ is the number of vertices pointing to some vertex in $S$, and actually equals to $|T(S)|$. Furthermore, by the definition of $S$-$dilation$, $|T(S)|$ is now the summation of $|T_i(S)|$, $i\in
\{1,\ldots,m\}$, in every subgraph. Consequently, it follows that there is $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$ if and only if matrix $[A_1,A_2,\ldots,A_m,
B_1,B_2,\ldots,B_m]$ in of form II.
Before going further to give another algebraic explanation of $S$-$dilation$, one definition and lemma proposed in [@RJ] must be introduced first:
\[def8\]([@RJ]) A structured $n\times m^\prime$ $(n\leq m^\prime)$ matrix $A$ is of form $(t)$ for some $t$, $1\leq t\leq n$, if for some $k$ in the range $m^\prime-t<k\leq m^\prime$, $A$ contains a zero submatrix of order $(n+m^\prime-t-k+1)\times k$.
\[lem6\]([@RJ]) $g$-rank of $A=t$
1. for $t=n$ if and only if $A$ is not of form $(n)$;
2. for $1\leq t< n$ if and only if $A$ is of form $(t+1)$ but not of form $(t)$.
From the above definition and lemma, another lemma is proposed here:
\[lem7\] There is no $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$ of switched linear system (\[eq2\]) if and only if the following matrix $$\label{eq9}
[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$$has g-rank $n$.
* Necessity:* If matrix (\[eq9\]) has $g$-rank $< n$, from lemma \[lem6\], it follows that matrix (\[eq9\]) is of form $(n)$. Then referring to definition \[def8\], matrix (\[eq9\]) must have a zero submatrix of order $(n+m^\prime-t-k+1)\times k$. Here, $t$ can be chosen as $n$, then (\[eq9\]) has a zero submatrix of order $(m^\prime-k+1)\times k$. For this $(m^\prime-k+1)$ rows, there are only $(m^\prime-k)$ nonzero columns. Consequently, matrix (\[eq9\]) is of form II and by lemma \[lem5\], there is $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$ of switched linear system (\[eq2\]).
*Sufficiency*: If there is $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$, by lemma \[lem5\], matrix (\[eq9\]) is of form II, then obviously $P_1$ in (\[eq9\]) can not have row rank equal to $k$ and furthermore, matrix (\[eq9\]) can not have $g$-rank = $n$.
With the above definitions and lemmas, a graphic necessary and sufficient condition for switched linear system to be structurally controllable can be proposed here:
\[the2\] Switched linear system (\[eq2\]) with graphic representations $\mathcal{G}_i$, $i\in \{1,\ldots,m\}$, is structurally controllable if and only if its colored union graph $\mathcal{\tilde{G}}$ satisfies the following two conditions:
1. there is no nonaccessible vertex in the colored union graph $\mathcal{\tilde{G}}$,
2. there is no $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$.
* Necessity:* *(i)* If there exist nonaccessible vertices in $\mathcal{\tilde{G}}$, by Lemma \[lem4\], the matrix $[A_1+A_2+\cdots+A_m, B_1+B_2+\cdots+B_m]$ is reducible or of form I. It follows that the controllability matrix $$\begin{split}[
&B_1+B_2+\ldots+B_m,(A_1+A_2+\ldots +A_m)(
B_1+B_2+\ldots+B_m),\\&(A_1+A_2+\ldots +A_m)^2(
B_1+B_2+\ldots+B_m),~\cdots,~(A_1+A_2+\ldots+A_m)^{n-1}(
B_1+B_2+\ldots+B_m)]\nonumber
\end{split}$$ always has at least one row that is identically zero (remark \[rem4\]). It is clear that every component of the matrix, such as $B_i,A_iB_j~and ~A_i^pA_j^qB_r$ has the same row always to be zero. As a result, the controllability matrix $$\begin{split} &[B_1,\ldots,B_m,
A_1B_1,\ldots,A_mB_1,\ldots,A_mB_m,A_1^2B_1,\ldots,A_mA_1B_1,\ldots,A_1^2B_m,\ldots,
A_mA_1B_m,\ldots,\\&A_1^{n-1}B_1,\ldots,A_mA_1^{n-2}B_1,\ldots,A_1A_m^{n-2}B_m,\ldots,
A_m^{n-1}B_m] \end{split}\nonumber$$ always has one zero row and can not be of full rank $n$. Therefore, switched linear system (\[eq2\]) is not structurally controllable.
*(ii)* Suppose that the switched linear system (\[eq2\]) is structurally controllable, i.e., the controllability matrix satisfies $g$-rank $\mathcal{C}$$(A_1,\ldots,A_m,B_1,\ldots,B_m)=n$. Specifically, $\textit{Im}[B_1,\ldots,B_m,A_1B_1,\ldots,\\A_mB_m,A_1^2B_1,\ldots,A_m^{n-1}B_m]=\mathbb{R}^n.$ Since $\forall P\in \mathbb{R}^{n\times r}$, $\textit{Im}(A_iP)\subseteq\textit{Im}(A_i)$, we have that $\textit{Im}[B_1,\ldots,B_m,\\A_1B_1,\ldots,A_mB_m,A_1^2B_1,\ldots,A_m^{n-1}B_m]\subseteq\textit{Im}[A_1,A_2,\ldots,A_m,
B_1,B_2,\ldots,B_m]\subseteq\mathbb{R}^n$. Thus $g$-rank $\mathcal{C}$$(A_1,\\\ldots,A_m,B_1,\ldots,B_m)=n$ requires that $\textit{Im}[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]=\mathbb{R}^n$ and therefore $g$-rank $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]=n$. However, if there is $S$-$dilation$ in the colored union graph $\mathcal{\tilde{G}}$, by Lemma \[lem5\], $g$-rank $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]<n$. Consequently, the switched linear system (\[eq2\]) is not structurally controllable.
*Sufficiency*: The general idea in the sufficiency proof is that we will assume that the two graphical conditions in the theorem hold. Then a contradiction will be found such that it is impossible that switched linear system (\[eq2\]) is structurally uncontrollable.
Before proceeding to the switched linear system (\[eq2\]), firstly, consider a structured linear system: $$\label{r1}
\dot{x}(t)=Ax(t)+Bu(t)$$ It is well known that system (\[r1\]) is structurally controllable if and only if there exists a numerical realization $(\tilde{A},\tilde{B})$, such that rank $(sI-\tilde{A},\tilde{B})=n,\forall s\in \mathbb{C}$. Otherwise, the PBH test [@TK] states that system (\[r1\]) is uncontrollable if and only if for every numerical realization, there exists a row vector $q\neq0$ such that $q\tilde{A}=s_0q,s_0\in \mathbb{C}$ and $q\tilde{B}=0$, where rank $(s_0I-\tilde{A},\tilde{B})<n$.
On one hand, if for every numerical realization rank $(sI-\tilde{A},\tilde{B})=n,\forall s\in \mathbb{C}\setminus \{0\}$, then the uncontrollability of system (\[r1\]) implies necessarily that for every numerical realization there exists a vector $q\neq0$ such that $q\tilde{A}=0$ and $q\tilde{B}=0$.
On the other hand, Lemma 14.1 of [@K] states that, if in the digraph associated to (\[r1\]), every state vertex is an end vertex of a stem (accessible), then $g$-rank $(sI-A,B)=n,\forall
s\in \mathbb{C}\setminus \{0\}$, which means that for almost all numerical realization $(\tilde{A},\tilde{B})$, rank $(sI-\tilde{A},\tilde{B})=n,\forall s\in \mathbb{C}\setminus \{0\}$.
Now considering the switched linear system (\[eq2\]), assume that the two conditions in Theorem \[the2\] are satisfied. Due to Lemma 14.1 of [@K], as all the parameters of matrices $A_1,\ldots,A_m,B_1,\ldots,B_m$ are assumed to be free, the condition *(i)* of Theorem \[the2\] implies that, for almost all vector values $\bar{u}=(\bar{u}_1,\ldots,\bar{u}_m)$, we have $g$-rank $(sI-(\bar{u}_1A_1+\ldots+\bar{u}_mA_m),(\bar{u}_1B_1+\ldots+\bar{u}_mB_m))=n,
\forall s\neq 0$. On the other hand, if the switched linear system (\[eq2\]) is structurally uncontrollable, then for all constant values, $\bar{u}=(\bar{u}_1,\ldots,\bar{u}_m)$, linear systems defined by matrices $(\bar{A},\bar{B})$ are also uncontrollable, where $\bar{A}=\sum^m_{i=1}\bar{u}_iA_i$ and $\bar{B}=\sum^m_{i=1}\bar{u}_iB_i$. We write the numerical realization of $(\bar{A},\bar{B})$ as $(\tilde{\bar{A}},\tilde{\bar{B}})$. This is due to the fact that for all constant values $\bar{u}$, $\textit{Im}(\mathcal{C}(\bar{A},\bar{B})\subseteq
\textit{Im}(\mathcal{C}(A_1,\ldots,A_m,B_1,\ldots,B_m)).$ Therefore, if the switched linear system is structurally uncontrollable, since for almost all $\bar{u}=(\bar{u}_1,\ldots,\bar{u}_m)$, $g$-rank $(sI-(\bar{u}_1A_1+\ldots+\bar{u}_mA_m),(\bar{u}_1B_1+\ldots+\bar{u}_mB_m))=n,
\forall s\neq 0$, we have that for every numerical realization matrix pair $(\tilde{\bar{A}},\tilde{\bar{B}})$, there exists a nonzero vector $q$ such that $q\tilde{\bar{A}}=0$ and $q\tilde{\bar{B}}=0$. Since this statement is true for almost all the values $\bar{u}=(\bar{u}_1,\ldots,\bar{u}_m)$, we have that for almost all $n\cdot m$-tuple values $\bar{u}^j=(\bar{u}^j_1,\ldots,\bar{u}^j_m), j=1,\ldots,n\cdot m$, we can find nonzero vectors $q_j$ such that the following holds: $$\label{r2}
\left\{\begin{array}{clc}
\sum^m_{i=1}\bar{u}^j_iq_j\tilde{A}_i=0, \hfill j=1,\ldots,n\cdot
m\\
\sum^m_{i=1}\bar{u}^j_iq_j\tilde{B}_i=0. \hfill j=1,\ldots,n\cdot m\\
\end{array}\right.$$ Obviously, there can not exist more than $n$ linear independent vectors $q_j$. Let us denote $q_1,q_2,\ldots,q_n$ the vectors such that $span$ $(q_1,q_2,\ldots,q_{n\cdot m})\subseteq$ $span$ $(q_1,q_2,\ldots,q_n)$ (we can renumber the vectors if necessary). All the vectors $q_j,j=n+1,\ldots,n\cdot m$ are linear combinations of $q_1,q_2,\ldots,q_n$. Therefore, system (\[r2\]) contains the following equations: $$\label{r3}
\left\{\begin{array}{clc}
\sum^n_{k=1}\sum^m_{i=1}a^j_{i,k}(\bar{u})q_k\tilde{A}_i=0 \hfill ~~~~j=1,\ldots,n\cdot m \\
\sum^n_{k=1}\sum^m_{i=1}a^j_{i,k}(\bar{u})q_k\tilde{B}_i=0 \hfill ~~~~j=1,\ldots,n\cdot m \\
\end{array}\right.$$ where $a^j_{i,k}(\bar{u})$ are linear functions of $\bar{u}^j,j=1,\ldots,n\cdot m$. Since system (\[r2\]) is satisfied for almost all the values, we can find $\bar{u}^j,j=1,\ldots,n\cdot m$ such that $$\label{r4}
det\left[ {\begin{array}{*{20}c}
a^1_{1,1}(\bar{u})& a^1_{1,2}(\bar{u})&\ldots&a^1_{m,n}(\bar{u}) \\
a^2_{1,1}(\bar{u})& a^2_{1,2}(\bar{u})&\ldots&a^2_{m,n}(\bar{u}) \\
\vdots&\vdots&\vdots&\vdots\\
a^{n\cdot m}_{1,1}(\bar{u})& a^{n\cdot m}_{1,2}(\bar{u})&\ldots&a^{n\cdot m}_{m,n}(\bar{u})
\end{array}} \right]\neq 0.\nonumber%~~det\left[ {\begin{array}{*{20}c}
% b^1_{0,1}(\bar{u})& b^1_{0,2}(\bar{u})&\ldots&b^1_{m,n}(\bar{u}) \\
% b^2_{0,1}(\bar{u})& b^2_{0,2}(\bar{u})&\ldots&b^2_{m,n}(\bar{u}) \\
% \vdots&\vdots&\vdots&\vdots\\
% b^{n\cdot m}_{0,1}(\bar{u})& b^{n\cdot m}_{0,2}(\bar{u})&\ldots&b^{n\cdot m}_{m,n}(\bar{u})
% \end{array}} \right]\neq 0$$ In this case, the only solution of (\[r3\]) is $q_k\tilde{A}_1=\ldots=q_k\tilde{A}_m=q_k\tilde{B}_1=\dots=q_k\tilde{B}_m=0,$ $k=1,\ldots,n$. Obviously, if the switched linear system is structurally uncontrollable, then vector $q_k,k=1,\ldots,n$ is nonzero. Consequently, the switched linear system (\[eq2\]) is structurally uncontrollable only if for every numerical realization there exists at least one nonzero vector $q$ such that $qA_1=\ldots=qA_m=qB_1=\dots=qB_m=0$. However, if condition *ii* of Theorem \[the2\] is satisfied, then $g$-rank $[A_1,\ldots,A_m,B_1,\ldots,B_m]=n$ and therefore, for at least one numerical realization, there does not exist a vector $q\neq 0$ such that $qA_1=\ldots=qA_m=qB_1=\dots=qB_m=0$. Hence, the two conditions are sufficient to ensure the structural controllability of switched linear system (\[eq2\]).
Actually, using the terminologies $`dilation'$ and $`S$-$dilation'$ as graphic criteria is not so numerically efficient. For example, to check the second condition of Theorem \[the2\], we need to test for all possible vertex subsets to see whether there exist $S$-$dilation$ in the colored union graph or not. Consequently, we will adopt another notion $`S$-$disjoint$ $edges$’ to form a more numerically efficient graphic interpretation of structural controllability.
\[r5\] In the colored union graph $\mathcal{\tilde{G}}$, consider $k$ edges $e_1=(v_1,v'_1),e_2=(v_2,v'_2),\ldots,e_k=(v_k,v'_k)$. We define for $i=1,\ldots,k,$ $S_i$ as the set of integers $j$ such that $v_j=v_i$, i.e., $S_i=\{1\leq j\leq k|v_j=v_i\}$. $e_1,e_2,\ldots,e_k$ are $S$-$disjoint$ if the following two conditions are satisfied:
1. edges $e_1,e_2,\ldots,e_k$ have distinct end vertices,
2. for $i=1,\ldots,k$, $S_i=\{i\}$ or there exist $r$ distinct integers $i_1,i_2,\ldots,i_r$ such that $e_{j_1}\in
\mathcal{I}_{i_1},e_{j_2}\in \mathcal{I}_{i_2},\ldots,e_{j_r}\in
\mathcal{I}_{i_r}$, where $j_1,j_2,\ldots,j_r$ are all the elements of $S_i$.
Roughly speaking, $k$ edges are $S$-$disjoint$ if their end vertices are all distinct and if all the edges which have the same begin vertex can be associated to distinct indexes $i$. For this new graphic property, the following lemma can be given:
\[10\] Considering switched linear system (\[eq2\]), there exist $n$ $S$-$disjoint$ edges in associated colored union graph $\mathcal{\tilde{G}}$ if and only if $[A_1,A_2,\ldots,A_m,
B_1,B_2,\ldots,B_m]$ has $g$-rank = $n$.
* Necessity:* If there exist $n$ $S$-$disjoint$ edges in $\mathcal{\tilde{G}}$, matrix $[A_1,A_2,\ldots,A_m,
B_1,B_2,\ldots,B_m]$ contains at least $n$ free parameters. Since the $n$ $S$-$disjoint$ edges have distinct end vertices, the corresponding $n$ free parameters lie on $n$ different rows. Besides, the $n$ $S$-$disjoint$ edges have distinct begin vertices or have same begin vertex that can be associated to distinct indexes $i$. This implies that these $n$ free parameters lie on $n$ different columns. keep these $n$ free parameters and set all the other free parameters to be zero. We can see that matrix $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$ has the following form: $\left[ {\begin{array}{*{20}c}
0& \lambda_1&0&0&\ldots&0 \\
0& 0&0& \lambda_2&\ldots&0\\
\vdots&\vdots&\vdots&\vdots\\
\lambda_n&0&0&0&\ldots&0
\end{array}}. \right]$, which has $g$-rank $n$.
*Sufficiency*: From the definition 12.3 and the following discussions of [@K], for a structured matrix $Q$, $g$-rank $Q$ = $s$-rank $Q$. where $s$-rank of $Q$ is defined as the maximal number of free parameters that no two of which lie on the same row or column. If matrix $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$ has $g$-rank = $n$, it follows that there exists $n$ free parameters from $n$ different rows, which implies that the corresponding $n$ edges have different end vertices, from $n$ different columns, which implies that these $n$ edges start from different vertices or start from same vertices but can be associated to different indexes. Hence the condition that the matrix has $g$-rank $n$ is sufficient to ensure the existence of $n$ $S$-$disjoint$ edges.
We can get another necessary and sufficient condition for structural controllability of system (\[eq2\]).
\[the6\] Switched linear system (\[eq2\]) with graphic representations $\mathcal{G}_i$, $i\in \{1,\ldots,m\}$, is structurally controllable if and only if its colored union graph $\mathcal{\tilde{G}}$ satisfies the following two conditions:
1. there is no nonaccessible vertex in the colored union graph $\mathcal{\tilde{G}}$,
2. there exist $n$ $S$-$disjoint$ edges in the colored union graph $\mathcal{\tilde{G}}$.
Lemma \[lem5\] and Lemma \[10\] show that there exist $n$ $S$-$disjoint$ edges in the colored union graph $\mathcal{\tilde{G}}$ if and only if there is no $S$-$dilation$ in $\mathcal{\tilde{G}}$. Then this theorem follows immediately.
\[rem5\] Compared with condition using $`S$-$dilation'$, this condition using ‘$S$-$disjoint$ edges’ does not require to check all the vertex subsets, which is a more efficient criterion. The maximal number of ‘$S$-$disjoint$ edges’ can be calculated using bipartite graphs. For example, we can use the algorithm in [@SV], which allows to compute the cardinality of maximum matching into a bipartite graph. A bipartite graph is a graph whose vertices can be divided into two disjoint sets $\mathcal{U}$ and $\mathcal{W}$ such that every edge connects a vertex in $\mathcal{U}$ to one in $\mathcal{W}$. To build a bipartite graph in directed subgraph $\mathcal{G}_i(\mathcal{V}_i,
\mathcal{I}_i)$, what we need to do is adding some vertices and make $\mathcal{U}_i=\{v\in \mathcal{V}_i|\exists (v,v')\in
\mathcal{I}_i\}$, which implies that cardinality $|\mathcal{U}_i|$ equals to the number of nonzero columns in matrix $[A_i, B_i]$. Besides, $\mathcal{W}_i=\mathcal{X}_i$, i.e., the state vertex set. Then it follows that the maximum matching in this bipartite graph is the same as the maximal $S$-$disjoint$ edge set in $\mathcal{G}_i(\mathcal{V}_i, \mathcal{I}_i)$. According to definition of $S$-$disjoint$ edges, the beginning vertex from different subgraphs should be differentiated when building the bipartite graph for colored union graph $\mathcal{\tilde{G}}$. Therefore for the bipartite graph of $\mathcal{\tilde{G}}$, $\mathcal{U}=\{v|\exists (v,v')\in \mathcal{I}_i, i=1,2,\ldots,m\}$, which implies that cardinality $|\mathcal{U}|$ equals to the number of nonzero columns in matrix $[A_1,A_2,\ldots,A_m,
B_1,B_2,\ldots,B_m]$. And $\mathcal{W}=\mathcal{X}$, i.e., the state vertex set. Similarly, the maximum matching in this bipartite graph is the same as the maximal $S$-$disjoint$ edge set in colored union graph. Therefore the complexity order of algorithm using method in [@SV] is $O(\sqrt{p+n}\cdot q)$, where $q$ is the number of edges in colored union graph, i.e., the number of free parameters in all system matrices, $p$ is the number of nonzero columns in matrix $[A_1,A_2,\ldots,A_m, B_1,B_2,\ldots,B_m]$ and $n$ is number of state variables. Compared with condition (ii), condition (i) is easier to check. We have to look for paths which connect each state vertex with one of the input vertex. This is a standard task of algorithmic graph theory. For example, depth-first search or breadth-first search algorithm for traversing a graph can be adopted and the complexity order is $O(|V|+|E|)$, where $|V|$ and $|E|$ are cardinalities of vertex set and edge set in union graph.
Numerical Example
-----------------
Consider a switched linear system with two subsystems as depicted by the graphic topologies in Fig. 1(a)-(b). In colored union graph $\mathcal{\tilde{G}}$ (Fig. 1(d)), thin lines represent edges from subgraph (a) and thick lines represent the edges from subgraph (b). It turns out that the colored union graph $\mathcal{\tilde{G}}$ has no nonaccessible vertex and no $S$-$dilation$. Besides, the three edges are $S$-$disjoint$ edges since they have different end vertices and one edge begins at vertex 3 and two edges begin at vertex 0 but they come from different subsystems.
(80,35)
(45,20)(75,20)(75,20) (95,20)(125,20)(75,20) (145,20)(175,20)(75,20) (45,35) (95,35) (145,35)
(45,50)(95,50)(145,50)
(173,20)[(-1,0)[27]{}]{} (173,21)[(-1,1)[28]{}]{}(145,21)[(0,1)[13.5]{}]{} (73,20)[(-1,0)[27]{}]{} (95,21)[(0,1)[13.5]{}]{} (123,21)[(-1,1)[28]{}]{} (42,20)[(0,0)\[c\][$3$]{}]{}(92,20)[(0,0)\[c\][$3$]{}]{}(142,20)[(0,0)\[c\][$3$]{}]{} (79,20)[(0,0)\[c\][$0$]{}]{}(129,20)[(0,0)\[c\][$0$]{}]{}(179,20)[(0,0)\[c\][$0$]{}]{} (42,50)[(0,0)\[c\][$1$]{}]{}(92,50)[(0,0)\[c\][$1$]{}]{}(142,50)[(0,0)\[c\][$1$]{}]{} (42,35)[(0,0)\[c\][$2$]{}]{}(92,35)[(0,0)\[c\][$2$]{}]{}(142,35)[(0,0)\[c\][$2$]{}]{}
(60,10)[(-2,0)[$(a)$]{}]{} (110,10)[(-2,0)[$(b)$]{}]{} (160,10)[(-2,0)[$(c)$]{}]{} (210,10)[(-2,0)[$(d)$]{}]{} (195,20)(225,20)(195,35)(195,50) (192,20)[(0,0)\[c\][$3$]{}]{}(229,20)[(0,0)\[c\][$0$]{}]{}(192,35)[(0,0)\[c\][$2$]{}]{}
(224,20)[(-1,0)[27]{}]{} (224,20)[(-1,1)[29]{}]{} (195,20)[(0,1)[13.5]{}]{} (192,50)[(0,0)\[c\][$1$]{}]{} (130,-5)[(15,0)\[c\][[Fig. 1. Switched linear system with two subsystems ]{}]{}]{}
According to Theorem \[the2\] or \[the6\], the switched linear system is structurally controllable. On the other hand, the system matrices of each subsystem of corresponding subgraph are: $$\label{124}
A_1= \left[ {\begin{array}{*{20}c}
0&0 &0 \\
0&0 & 0 \\
0&0 &0
\end{array}} \right],~~B_1=\left[ {\begin{array}{*{20}c}
0 \\
0 \\
\lambda_1
\end{array}} \right];
A_2= \left[ {\begin{array}{*{20}c}
0&0 & 0 \\
0&0 & \lambda_2 \\
0&0 & 0
\end{array}} \right],~~B_2=\left[ {\begin{array}{*{20}c}
\lambda_3 \\
0 \\
0
\end{array}} \right].\nonumber$$ controllability matrix (\[eq3\]) can be calculated and can be shown to have $g$-$rank$=3. In addition, there exist a $dilation$ in union graph Fig. 1(c), which shows that the condition in Theorem \[the1\] is not necessary for structural controllability.
For further extension purpose, next we will use examples to illustrate that the dependence among system parameters will make some edges ‘useless’ or ‘excessive’ in judging the structural controllability. See the following switched linear system first $$\label{124}
A_1= \left[ {\begin{array}{*{20}c}
0&0 \\
0&0 \\
\end{array}} \right],~~B_1=\left[ {\begin{array}{*{20}c}
\lambda_1\\
\lambda_2\\
\end{array}} \right];
A_2= \left[ {\begin{array}{*{20}c}
0&0 \\
0&0 \\
\end{array}} \right],~~B_2=\left[ {\begin{array}{*{20}c}
\lambda_3\\
\lambda_4\\
\end{array}} \right].\nonumber$$ According to Theorem 2 or 3, this system is structurally controllable. However, if dependent parameters are considered, see the following switched linear system (a linear system actually) $$\label{124}
A_1= \left[ {\begin{array}{*{20}c}
0&0 \\
0&0 \\
\end{array}} \right],~~B_1=\left[ {\begin{array}{*{20}c}
\lambda_1\\
\lambda_2\\
\end{array}} \right];
A_2= \left[ {\begin{array}{*{20}c}
0&0 \\
0&0 \\
\end{array}} \right],~~B_2=\left[ {\begin{array}{*{20}c}
\lambda_1\\
\lambda_2\\
\end{array}} \right].\nonumber$$ The dependence of all the parameters in matrix $B_1$ and $B_2$ makes this system not structurally controllable and the results in Theorem 2 and 3 not hold, even though it would be structurally controllable if the parameters in $B_2$ are replaced with $\lambda_3$ and $\lambda_4$ or simply remove $\lambda_1$ or $\lambda_2$ in the second subsystem.
Conclusions and Future Work
===========================
In this paper, structural controllability for switched linear systems has been investigated. Combining the knowledge in the literature of switched linear systems and graph theory, several graphic necessary and sufficient conditions for the structurally controllability of switched linear systems have been proposed. These graphic interpretations provide us a better understanding on how the graphic topologies of switched linear systems will influence or determine the structural controllability of switched linear systems. This shows us a new perspective that we can design the switching algorithm to make the switched linear system structurally controllable conveniently just having to make sure some properties of the corresponding graph (union or colored union graph) are kept during the switching process. In this paper, the parameters in different subsystem models are assumed to be independent. A more general assumption is that some free parameters remain the same among different subsystems switching, i.e., dependence among subsystems. It turns out that our necessary and sufficient condition derived here would be a necessary condition under this dependence assumption. To obtain a full characterization for the dependent case needs further investigation.
[99]{}
A. Leonessa, W. M. Haddad, and V. S. Chellaboina, “Nonlinear system stabilization via hierarchical switching control,” *IEEE Trans. Autom. Control,* vol. 46, no. 1, pp. 17–28, Jan. 2001.
K. S. Narendra and J. Balakrishnan, “Adaptive control using multiple models,” *IEEE Trans. Autom. Control,* vol. 42, no. 2, pp. 171 – 187, Feb. 1997.
H. Lin and P. J. Antsakis, “Switching stabilizability for continuous-time uncertain switched linear systems,” *IEEE Trans. Autom. Control,* vol. 52, no. 4, pp. 633–646, Apr. 2007.
K. A. Loparo, J. T. Aslanis, and O. Hajek, “Analysis of switching linear systems in the plain, part 2, golbal behavior of trajectories, controllability and attainablility,” *J. Optim. Theory Appl.,* vol. 52, no. 3, pp. 395–427, Mar. 1987. J. Ezzine and A. H. Haddad, “Controllability and observability of hybrid systems,” *Int. J. Control,* vol. 49, no. 6, pp. 2045–2055, Jun. 1989.
F. Szigeti, “A differential-algebraic condition for controllability and observability of time varying linear systems,” *in Proc. 31st IEEE Conf. Decision Control,* 1992, pp. 3088–3090.
Z. Sun, S. S. Ge, and T. H. Lee, “Controllability and reachability criteria for switched linear systems,” *Automatica,* vol. 38, no. 5, pp. 775–786, May 2002.
Z. Sun and S. S. Ge, “Switched Linear Systems–Control and Design,” Springer, New York, 2004.
G. Xie and L. Wang, “Controllability and stabilizability of switched linear systems,” *Systems & Control Letters* vol. 48, no. 2, pp. 135–155, Feb. 2003.
Y. Qiao and D. Cheng, “On partinioned controllability of switched linear systems,” *Automatica,* vol. 45, pp.225–229,2009. C. T. Lin, “Structural controllability,” *IEEE Trans. Autom. Control,* vol. 19, no. 3, pp. 201–208, Jun. 1974.
H. Mayeda, “On structural controllability theorem,” *IEEE Trans. Autom. Control,* vol. 26, no. 3, pp. 795–798, Jun. 1981.
R. W. Shields and J. B. Pearson, “Structural controllability of multi-input linear systems,” *IEEE Trans. Autom. Control,* vol. 21, no. 2, pp. 203–212, Apr. 1976.
K. Glover and L. M. Silverman, “Characterization of structural controllability,” *IEEE Trans. Autom. Control,* vol. 21, no. 4, pp. 534–537, Aug. 1976.
J. M. Dion, C. Commault, and J. van der Woude, “Generic properties and control of linear structured systems,” *Automatica,* vol. 39, no. 7, pp. 1125–1144, Jul. 2003.
J. W. van der Woude, “A graph theoretic characterization for the rank of the transfer matrix of a structured system,” *Mathematics of Control, Signals and Systems,* vol. 4, no. 1, pp. 33–40, May 1991. K. Murota, “System Analysis by Graphs and Matroids,” Springer-Verlag, New York, U.S.A., 1987. K. J. Reinschke, “Multivariable Control A Graph Theoretic Approach,” Springer-Verlag, New York, U.S.A., 1988. L. Blackhall and D. J. Hill, “On the structural controllability of networks of linear systems,” *in Proc. 2nd IFAC Workshop Distributed Estimation and Control in Networked System* 2010, pp. 245-250.
T. Kailath, “Linear systems,” Prentice Hall Information and system science series. Prentice Hall, Englewood Cliffs, 1980.
S. Micali and V. V. Vazirani, “An $O(\sqrt{|v|}\cdot|E|)$ algorithm for finding maximum matching in general graphs,” *in Proc. the 21st Annual Symposium on the Foundations of Computer Science,* 1980, pp. 17–27.
X. M. Liu, H. Lin, and B. M. Chen, “Graphic interpretations of structural controllability for switched linear systems,”*in Proc. 11th International Conf. Control, Automation, Robotics and Vision,* 2010, pp. 549–554.
[^1]: $^\ast$Corresponding author. Email: [email protected] tel 574-6313177 fax 574-6314393
[^2]: $^{a}$[*[Dept. of ECE, National University of Singapore, Singapore]{}*]{} $^{b}$[*[Dept of EE, Univ of Notre Dame, Notre Dame, IN 46556, USA]{}*]{}
[^3]: \*Manuscript titled “Structural Controllability of Switched Linear Systems” is submitted to IEEE Transactions on Automatic Control, 2011
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Passive spiral galaxies, despite their spiral morphological appearance, do not have any emission lines indicative of ongoing star formation in their optical spectra. Previous studies have suggested that passive spiral galaxies preferentially exist in infall regions of galaxy clusters, suggesting that the cluster environment is likely to be responsible for creating these galaxies. By carrying out spatially resolved long-slit spectroscopy on four nearby passive spiral galaxies with the Apache Point Observatory 3.5-m telescope, we investigated the stellar populations of passive spiral galaxies separately for their inner and outer regions. In the two unambiguously passive spiral galaxies among the four observed galaxies, H$\delta$ absorption lines are more prominent in the outer regions of the galaxies, whereas the 4000-[Å]{} breaks (D$_{4000}$) are strongest in the inner regions of the galaxies. A comparison with a simple stellar population model for the two passive spiral galaxies indicates that the outer regions of the galaxies tend to harbour younger populations of stars. The strong H$\delta$ absorption observed in the outer regions of the sample galaxies is consistent with that of galaxies whose star formation ceased a few Gyrs ago. Because of the large uncertainty in the absorption indices in our samples, further observations are needed in order to place constraints on the mechanisms that quench star formation in passive spiral galaxies.'
author:
- |
Miho Ishigaki$^{1,2}$[^1][^2],Tomotsugu Goto$^{1}$ and Hideo Matsuhara$^{1,2}$\
$^{1}$Institute of Space and Astronautical Science, Japan Aerospace Exploration Agency, 3-1-1 Yoshinodai,\
Sagamihara, Kanagawa 229-8510, Japan\
$^{2}$Tokyo Institute of Technology, 2-12-1 Ookayama, Meguro-ku,\
Tokyo, 152-8550, Japan
title: Spatially Resolved Spectroscopy of Passive Spiral Galaxies
---
\[firstpage\]
[**galaxies: evolution, galaxies: stellar content, galaxies: clusters: general** ]{}
Introduction {#sec:section1}
============
Passive spiral galaxies have peculiar spectroscopic characteristics among the galaxy populations having spiral morphologies. They show few or neither of the emission lines in H$\alpha$ nor \[O II\] in optical spectra that would be indicative of ongoing star formation (Couch et al. 1998; Dressler et al. 1999; Poggianti et al. 1999; Goto et al. 2003b). The optical g $-$ r colours observed in such galaxies are found to be significantly redder than those of spiral galaxies with emission lines, which confirms the lower star formation rate among passive spiral galaxies (Poggianti et al. 1999).
A similar population of galaxies, known as ’anemic’ spiral galaxies, has been found in the Virgo Cluster (van den Bergh 1976). They have smoothed spiral arms showing less prominent star formation activity than in other galaxies of the same Hubble type. Elmegreen et al. (2002) observed anemic spirals in the Virgo Cluster and found lower gas surface densities than for normal spirals for these galaxies. They showed that the gas surface density in these galaxies is below the threshold for star formation (Kennicutt 1989), suggesting that the lack of star formation is caused by the stripping of gas in the environment of the cluster (Elmegreen et al. 2002).
In order to investigate the effect of environment on the formation of such galaxies, Goto et al. (2003b) searched for passive spiral galaxies in all environments, including in dense cluster cores and field regions, using the large samples in the Sloan Digital Sky Survey (SDSS) data (Strauss et al. 2002). They found that passive spiral galaxies preferentially exist in the environment where a local galaxy density is intermediate between that of the cluster cores and the field regions (Goto et al. 2003b). This characteristic environment corresponds to the region where a significant decline in the star formation rate has been observed (Lewis et al. 2002; Gomez et al. 2003; Tanaka et al. 2004). These studies imply that cluster-related phenomena could be the main factors responsible for the formation of passive spiral galaxies.
Recent observations have revealed that the cluster environment may bring about a transformation in the galaxy population from star-forming spirals to passive galaxies, and thus passive spiral galaxies have been suggested to be a population in a transitional phase between these two populations (Poggianti et al. 1999). For instance, the fraction of blue galaxies is found to be larger in distant clusters than in local clusters, a phenomenon known as the Butcher-Oemler effect (Butcher & Oemler 1978; Couch et al. 1998; Kodama & Bower 2001; Goto et al. 2003d, 2004). The Butcher-Oemler effect implies that the environment of a cluster may affect the star formation activity of its member galaxies. Furthermore, morphological type (Goto et al. 2003c; Treu et al. 2003; Postman et al. 2005) is found to be correlated with various functions of cluster environment, such as local density or cluster centric radius (Goto et al. 2004; Tanaka et al. 2004). Numerical simulations (Bekki et al. 2002) have shown that gas-stripped galaxies may finally become S0 galaxies if no further accretion onto the disc occurs after the stripping. UV observations suggest that the cessation of star formation could take place before this morphological transformation into S0 galaxies (Moran, Ellis & Treu 2006). Indeed, Goto et al. (2003b) found that the Petrosian radius of S0s is smaller than that of spirals, which is consistent with the above scenario. Although these studies have provided significant broad implications concerning the origin of passive spiral galaxies, the details of the underlying physical mechanisms, especially the time-scale on which in-falling cluster galaxies terminate their star formation, are still uncertain.
Various mechanisms with different time-scales have been proposed. Because star formation could be sustained by cold gas accreted onto the disc, the time-scale is closely related to the time at which the reservoir of cold gas is removed by means of galaxy-intracluster medium (ICM) interactions (Treu et al. 2003; Diaferio et al. 2001; Kauffmann et al. 1999).
A possible mechanism that occurs on a relatively short time-scale is ram-pressure stripping (Gunn & Gott 1972; Fujita 2001; Fujita & Nagashima 1999; Fujita & Goto 2004; Abadi, Moore & Bower 1999; Quilis, Moore & Bower 2000). When a galaxy falls into a dense region of the cluster, ram pressure caused by the motion of the galaxy relative to the dense ICM removes the cold interstellar gas in the disc that is the fuel for the star formation (Gunn & Gott 1972). Vollmer et al. (2006) modelled the heavily ram-pressure-stripped galaxy NGC 4522 in the Virgo Cluster (Kenney, van Gorkom & Vollmer 2004). The ram-pressure-stripping model successfully reproduces the observed H I gas deficiency and the truncated gas disc of the galaxy (Vollmer et al. 2006). The time-scale for the ram-pressure stripping to terminate star formation was estimated to be 0.01-0.1 Gyr from numerical simulations (Abadi et al. 1999; Quilis et al. 2000; Fujita & Nagashima 1999).
Some authors, however, argue that ram pressure alone cannot explain the observed decline in star formation in cluster galaxies (Treu et al. 2003; Balogh et al. 2002; Kodama & Bower 2001). Treu et al. (2003) reported the mild decline in star formation at the periphery of the cluster Cl 0024$+$16, where the ram pressure may not be effective in stripping cold disc gas.
Other possible mechanisms have been proposed that require relatively longer time-scales to strip the galactic halo gas (Larson, Tinsley & Caldwell 1980), and these have been termed ’strangulation’ (Fujita 2004; Tanaka et al. 2004) or ’starvation’ (Treu et al. 2003; Boselli et al. 2006; Bekki et al. 2002). ’Strangulation’ involves the stripping of warm halo gas through the interaction with the ICM in a situation where further supply of halo gases to the disc is disrupted. Because this can occur even in less dense environments, this may explain the observed decline of star formation among galaxies in the outer regions of clusters.
Mechanisms such as mergers cannot be responsible for creating passive spiral galaxies because they may disturb the spiral arms. Similarly, ’harassment’, high-speed gravitational encounters between galaxies (Moore et al. 1996), may also lead to changes in morphology and thus cannot be a dominant mechanism for transforming normal spiral galaxies into passive ones.
In this paper, we perform spatially resolved spectroscopy on four of the passive spiral galaxies identified in our previous paper (Goto et al. 2003b) in order to obtain further constraints on the mechanism responsible for halting the star formation. The strength of the H$\delta$ absorption line and 4000-[Å]{} break are compared with the simple stellar population (SSP) model constructed by Bruzual & Charlot (2003) in order to estimate the age of the stellar population. We then attempt to estimate the time-scale for the mechanism to create passive spiral galaxies.
The method of data reduction and analysis are described in Section 2, the results are in Section 3, the discussion based on the comparison with the SSP model is presented in Section 4, and conclusions are given in Section 5. Unless otherwise stated, we adopt the best-fitting WMAP cosmology: (h, $\Omega_{m}$, $\Omega_{L}$) = (0.71, 0.27, 0.73) (Bennett et al. 2003).
\[tab1:1\]
---------------------------- ------------- --------------- -------- ------------------------------ -------------------- ----------------- ------------------ -----------
Name Ra(J2000) Dec(J2000) $z$ $\sigma_{V}$ km s$^{-1}$[^3] $r_{\rm P}$ arcsec ${\rm AP}_{in}$ ${\rm AP}_{out}$ $M_r$[^4]
SDSS J021534.35$-$090537.0 02 15 34.35 $-$09 05 37.0 0.0687 112$\pm$7 15.62 19.7 9.8 $-$22.58
SDSS J024732.02$-$065137.5 02 47 32.02 $-$06 51 37.5 0.0705 131$\pm$9 10.31 17.2 12.0 $-$21.75
SDSS J033322.66$-$000907.5 03 33 22.66 $-$00 09 07.5 0.0838 - 6.615 11.1 7.0 $-$20.89
SDSS J074452.51$+$373852.7 07 44 52.51 $+$37 38 52.7 0.0743 94$\pm$9 8.329 9.4 4.6 $-$20.80
---------------------------- ------------- --------------- -------- ------------------------------ -------------------- ----------------- ------------------ -----------
{width="60mm"} {width="60mm"}
The Method {#sec:section2}
==========
Sample selection
----------------
The target galaxies, SDSS J021534.35$-$090537, J024732.02$-$065137.5, J033322.65$-$000907.5 (the left panel of Fig. 1) and J074452.52$+$373852.7 (the right panel of Fig. 1), are a subset of the passive spiral galaxies selected in Goto et al. (2003b) from the volume-limited sample of the SDSS data. Galaxies are selected based on the following criteria: (1) the inverse concentration parameter, which is defined as the ratio of the Petrosian 50 per cent radius to the Petrosian 90 per cent radius, is less than 0.5 (details of the use of this parameter for classifying morphological type are given in Shimasaku et al. 2001) in order to select galaxies having spiral morphology; (2) the absence of \[O II\] and H$\alpha$ emission lines (the measured values of equivalent width are less than 1 $\sigma$ error, Goto et al. 2003b), which are indicative of ongoing star formation activity.
Observation and Data reduction {#subsec:2}
------------------------------
The observations were carried out using the Dual Imaging Spectrograph (DIS) installed on the Apache Point Observatory 3.5-m telescope on 2004 October 18. Both the blue and red cameras on the DIS were used in a medium-dispersion mode, with the dispersion covering the wavelength range 3000-9000 [Å]{}. The pixel scales of the spatial axis for the blue and the red cameras were 0.42 and 0.40 arcsec pix$^{-1}$, respectively. To perform spatially resolved spectroscopy, a long slit with a slit width of 1.5 arcsec was used. Each sample galaxy was observed three times with an exposure time of 1000-1500 s. After bias-subtraction and flat-fielding had been applied, the three frames were combined into one frame. Spectra of standard stars, HR 718 and Hilt 600, were taken with an exposure time of 20 and 1 s, respectively, and used for flux calibration. The seeing size measured using the full width at half maximum (FWHM) of the point spread function of Hilt 600, which was monitored during the observation of the target galaxies, was $\sim$1.7 and $\sim$1.3 arcsec for the red and the blue camera, respectively.
To perform the spatially resolved analysis, we refer to the Petrosian radius in r’ band ($r_{\rm P}$), which is a measure of the surface brightness profile of galaxies, obtained through the SDSS (Blanton et al. 2001). The values are presented in Table 1. The long-slit data were divided into 11 spatial bins using the iraf routine apall, which outputs 11 spectra for all bins. Then, the three bins around the centre and the remaining eight bins sampling the spectra of two sides of the galaxies were summed, respectively, to increase the signal-to-noise ratio. Hereafter, these summed data for the inner and outer parts of the galaxies are designated as AP$_{in}$ and AP$_{out}$, respectively. AP$_{in}$ samples approximately $r$ $\leq$ 0.27$r_{\rm P}$, whereas AP$_{out}$ samples 0.27$r_{\rm P}$ $<$ r $\leq$ 1$r_{\rm P}$, where $r$ represents the distance from the galaxy centre. The diameter of the inner part, $\sim$3.7 arcsec, is much larger than the seeing size ($<$1.7 arcsec). Wavelength calibration was performed using observations of a HeNeAr lamp. The sensitivity function was obtained using the spectra of the standard stars HR 718 and Hilt 600. Because it was difficult to fit the whole wavelength range with a single sensitivity function, we excluded the data points at the edges of each red and blue spectrum. The resulting spectral coverage, included in the fitting of the sensitivity correction, is 3700-5650 [Å]{} for the blue spectrum and 6300-9500 [Å]{} for the red spectrum.
The wavelength resolution was measured using the FWHM of the emission lines of the HeNeAr lamp spectra. The result was $\sim 8.5 \pm 0.5$ [Å]{} FWHM at approximately 5000 [Å]{}.
Measurement of spectral features {#subsec:3}
--------------------------------
\[fig:ha\] {width="45mm"} {width="45mm"} {width="45mm"} {width="45mm"}
Before we measured the spectral features, the redshift for each galaxy was computed from the Ca H line at 3970 [Å]{}, one of the most prominent features in the obtained spectra, by fitting a Gaussian for this line with the iraf splot routine. The wavelengths of the observed spectrum were shifted to the rest-frame based on the derived redshifts. The derived redshifts are shown in Table 1.
The Lick/IDS absorption-line index system (Trager et al. 1998; Worthey 1994) was used to measure the strengths of absorption features in the spectra. The wavelength definitions for the index measurement were taken from Worthey & Ottaviani (1997) for the H$\gamma$ and H$\delta$ indices, and from Trager et al. (1998) for the other 21 indices. The equivalent width or magnitudes for each index were computed as defined in Trager et al. (1998). We also measured the 4000-[Å]{} break (D$_{4000}$) to estimate the age of the stellar population, as it is a broad feature and can be measured with a greater signal-to-noise ratio than can individual lines. D$_{4000}$ is widely used as a diagnostic of the age and metallicity of stellar populations and can be compared with values in the literature (Bruzual 1983; Gorgas et al. 1999; Kauffmann et al. 2003). D$_{4000}$ is calculated as defined in Balogh et al. (1999), who take a narrower spectral region for the red and blue continua than do Bruzual (1983). The reason for using this narrower definition is that it is less affected by the uncertainty in the sensitivity correction, and, more importantly, by the reddening effect (Balogh et al. 1999).
In order to confirm the absence of star formation activity for the observed candidates of passive spiral galaxies, equivalent widths of H$\alpha$ and \[O II\] emission line were measured. Fig. 2 shows the spectral regions around the H$\alpha$ emission lines. Two of the four target galaxies (SDSS J021534.35$-$090537 and SDSS J024732.02$-$065137.5) are found to show detectable H$\alpha$ emission (H$\alpha$ $-$ 1$\sigma$ error $\geq$ 10 [Å]{}) at AP$_{out}$, and thus do not meet the criteria for passive spiral galaxies defined in Goto et al. (2003b). The \[O II\] emission lines, whose equivalent widths are less than 7 [Å]{}, were also detected at AP out, indicating the presence of current star formation activity in the exterior regions of these galaxies. The non-detection of \[O II\] in the SDSS data was presumably the result of the aperture of the SDSS spectroscopic fiber (diameter of 3 arcsec), which samples only the inner parts of the galaxies, where the observed light may be dominated by the bulge component over the disc component (Abazajian et al. 2005). The results indicate the importance of using the whole light, including that from the outer regions of the galaxies, when identifying and investigating passive spiral galaxies.
We restrict our discussion to galaxies with no prominent star formation activity over the whole galaxy. The remaining two galaxies, SDSS J033322.65$-$000907.5 and SDSS J074452.52$+$373852.7, are hereafter denoted as SDSSJ0333$-$0009 and SDSSJ0744$+$3738. The velocity dispersion (available only for SDSSJ0744$+$3738), Petrosian radius, absolute magnitude in the r’ band obtained by the SDSS, and the measured values of redshift and signal-to-noise ratio are shown in Table 1. The g’, r’, i’-composite SDSS images (Fukugita et al. 1996) of SDSSJ0333$-$0009 and SDSSJ0744$+$3738 are shown in Fig. 1.
Results: Spectral characteristic {#sec:section3}
================================
Figs 3 and 4 show the spectra for AP$_{in}$ (upper figure) and AP$_{out}$ (lower figure) for each target galaxy over the rest-frame wavelength range of 3900-4900 [Å]{}. These spectra are smoothed by a boxcar function over two data points and normalized to the flux at 4400 [Å]{} to clarify the comparison of the absorption strengths. In the wavelength bands of the major spectral indices, H$\delta$, Ca, G4300, H$\gamma$, Fe and H$\beta$ are marked. Smoothed spectra around the D$_{4000}$ and the H$\delta$ absorption are shown in Fig. 5. To enable visualization of the strength of D$_{4000}$, average fluxes over the red (4000-4100 [Å]{}) and blue (3850-3950 [Å]{}) bandpasses are represented by horizontal lines. Table 2 summarizes the measured equivalent widths of H$\delta_{\rm A}$, H$\delta_{\rm F}$ (the subscripts A and F refer to a wider and narrower index definition, respectively) and D$_{4000}$. It also shows these indices measured for the spectrum obtained in SDSS DR5 (Adelman-McCarthy et al. 2006) using the same bandpasses and methodology as are applied to our samples. The errors in the indices are computed by propagating the standard errors of the pseudo-continuum bandpasses. Errors for D$_{4000}$ are estimated in such a way that the errors of each pixel propagate according to the definition of D$_{4000}$ (Balogh et al. 1999).
The following subsections describe the characteristics of the spectra for each passive spiral galaxy.
SDSS J033322.65$-$000907.5
--------------------------
![The rest-frame spectra for AP$_{out}$ and AP$_{in}$ for SDSS J033322.65$-$000907.5. The vertical axis shows the flux normalized to the flux at 4400 [Å]{}. The horizontal axis shows the wavelength in the rest-frame. The spectra were smoothed with a boxcar function of width two data points. Shaded areas show wavebands used for the measurement of the spectral indices as defined in Worthey & Ottaviani (1997) and Trager et al. (1998).[]{data-label="fig:spec1"}](spec033_final_final.ps){width="75mm"}
As shown in Fig. 3, Balmer absorption lines, Ca and Fe absorption lines and deep absorption in G4300 bands are observed for both AP$_{out}$ and AP$_{in}$ in SDSSJ0333$-$0009. The \[O II\] emission line was not detected for either AP$_{out}$ or AP$_{in}$, which ensures the passive nature of this galaxy. The H$\alpha$ emission line, which is also an indicator of current star formation was, however, weakly detected for AP$_{out}$ with an equivalent width of $\sim -3$ [Å]{}. H$\delta_{\rm A}$ absorption is stronger for AP$_{out}$ than that for AP$_{in}$ by $\sim$1.6 [Å]{} (Table 2). By contrast, the D$_{4000}$ break is larger for AP$_{in}$ than it is for AP$_{out}$ by $\sim$ 0.07 (see Fig. 5). These results suggest that there is a higher proportion of old stars in the inner than there is in the outer regions for these galaxies.
SDSS J074452.52$+$373852.7
--------------------------
As shown in Fig. 4, the spectrum of SDSSJ0744$+$3738 is dominated by the Balmar absorptions. The metal absorptions and the deep absorption in the G4300 band look similar to those of SDSSJ0333$-$0009, but the difference in the equivalent widths of observed features between AP$_{in}$ and AP$_{out}$ is larger: H$\delta_{\rm F}$ absorption is stronger for AP$_{out}$ than for AP$_{in}$ by $\sim$4.0 [Å]{} as shown in Table 2, and the H$\beta$ absorption is stronger for AP$_{out}$ by 3.5 $\pm$ 0.5 [Å]{}. D$_{4000}$ is larger for AP$_{in}$ than for AP$_{out}$ by $\sim$0.5 (Table 2). These results indicate that the stellar population is younger in the outer regions, similar to the situation for the previous galaxy. The equivalent width of the H$\alpha$ emission line is found to be weak: $\sim -2.3 \pm 0.5$ [Å]{} for AP$_{in}$ and $\sim -3.4 \pm 0.6$ [Å]{} for AP$_{out}$. The weakness of the H$\alpha$ emission lines suggests the lack of prominent star formation activity in both the inner and the outer regions of this galaxy.
![The rest-frame spectra for AP$_{out}$ and AP$_{in}$ for SDSS J074452.52$+$373852.7. The notation is the same as in Fig. 3[]{data-label="fig:spec2"}](spec074_final_final.ps){width="75mm"}
{width="55mm"} {width="55mm"}
Target Spatial bin ${\rm H}\delta_{A}$ ${\rm H}\delta_{F}$ $D_{4000}$ ${\rm H}\delta_{A}$(SDSS) H$\delta_{F}$(SDSS) $D_{4000}$(SDSS)
---------------------------- ------------------ --------------------- --------------------- ---------------- --------------------------- --------------------- ------------------
SDSS J033322.65$-$000907.5 ${\rm AP}_{in}$ 2.0$\pm$1.8 2.6$\pm$1.1 1.68$\pm$ 0.03 -1.9$\pm$2.4 0.2$\pm$1.6 1.66
${\rm AP}_{out}$ 3.6$\pm$4.3 4.4$\pm$ 2.6 1.61$\pm$0.03
SDSS J074452.52$+$373852.7 ${\rm AP}_{in}$ 1.0 $\pm$1.7 1.6$\pm$1.2 1.80$\pm$ 0.03 -0.1$\pm$1.2 1.1$\pm$0.7 1.72
${\rm AP}_{out}$ 22.2$\pm$8.1 5.5$\pm$11.0 1.29$\pm$0.02
Discussion {#sec:section4}
==========
In the following subsection, we first try to estimate the metallicity and the effects of $\alpha$-enhancement on the absorption strengths for the observed galaxies using the measured Lick indices. The super-solar $\alpha$/Fe ratio could affect the strength of the H$\delta_{\rm A}$ and the H$\delta_{\rm F}$ indices, as discussed in Thomas, Maraston & Bender (2003), which leads to an underestimate of the age of stellar populations. Considering the estimated metallicity and the effect of $\alpha$-enhancement, Section 4.2 discusses the light-averaged ages of the stellar populations using a H$\delta_{\rm F}$-D$_{4000}$ plane.
The Lick indices measured with $>$1$\sigma$ for both the inner and outer regions were Fe4383, Fe4668, H$\beta$, Fe5270 and H$\delta_{\rm F}$ for SDSSJ0333$-$0009. For SDSSJ0744$+$3738, G4300, Fe4531 and Fe4668 were detected with $>$1$\sigma$. Unfortunately, the H$\delta_{\rm F}$ index, which is used as an age indicator, does not have a good enough signal-to-noise ratio for SDSSJ0744$+$3738. Furthermore, neither the inner nor the outer region of this galaxy has an Mgb index, which is used as an indicator of metallicity and $\alpha$-enhancement, with a signal-to-noise ratio $>$1. We therefore focus our discussion of the metallicity and $\alpha$-enhancement on SDSSJ0333$-$0009, for which the iron indices and the Mgb indices were measured with a relatively high signal-to-noise ratio.
Metallicity and $\alpha$-enhancement {#sec:metal}
------------------------------------
The measured Lick indices (Fe4383, Fe4668, H$\beta$, Fe5270 and H$\delta_{\rm F}$) for SDSSJ0333$-$0009 indicate that the inner region is more metal-rich than the outer region. Fig. 6 shows the strength of these indices for AP$_{in}$ (open triangles) and AP$_{out}$ (filled triangles) as a function of the Mgb index. The SSP models of Thomas et al. (2003) with age = 1.0, 5.0, 10.0, 15.0 Gyr and metallicity Z = 0.67, 0.00(Z$_{\odot}$), -2.25, assuming \[$\alpha$/Fe\] = 0.0, are overlaid. A comparison of the strength of the observed indices with that of the SSP models suggests that the observed indices for both the inner and outer regions are consistent with the model with super-solar metallicity. The iron indices (Fe4383, Fe4668 and Fe5270), especially, suggest that there is a radial gradient in the metallicity; the observed indices for the inner region are marginally consistent with \[Z/H\] = 0.67 models, whereas for the outer region the observed indices are reproduced by the model with \[Z/H\] = 0.00 in the Fe4668-Mgb plane and \[Z/H\] = 0.00-0.67 in the Fe5270-Mgb plane. Therefore, although the index strengths are associated with large uncertainties, they suggest that the inner region is more metal-rich than the outer region. It should be noted that H$\delta_{\rm F}$ for both regions and Fe4383 for the outer region were not reproduced by the models for \[$\alpha$/Fe\] = 0.0. Rather, they seem to be consistent with models with \[$\alpha$/Fe\] = 0.5, as discussed in the next section.
Comparisons of the Fe4383 indices, which are especially sensitive to the $\alpha$-enhancement (Thomas et al. 2003), with the model suggest that the outer region of this galaxy possibly has a super-solar $\alpha$/Fe ratio. Fig. 7 shows the strength of the Fe4383 index as a function of the Mgb index. Models with \[$\alpha$/Fe\] = 0.0 and 0.5 for ages and metallicities the same as those shown in Fig. 6 are overlaid. Open and filled triangles show the observed index strength for AP$_{in}$ and AP$_{out}$, respectively. This figure indicates that the strength of Fe4383 for AP$_{out}$ is more consistent with the models with \[$\alpha$/Fe\] = 0.5 than it is with those with \[$\alpha$/Fe\] = 0.0.
In order to elucidate the dependence of the H$\delta$ absorption strength on \[$\alpha$/Fe\], Fig. 8 shows the model predictions for the H$\delta_{\rm F}$ index as a function of the Mgb index for \[$\alpha$/Fe\] = 0.0 and 0.5. This illustrates that the difference in H$\delta_{\rm F}$ absorption owing to the $\alpha$-enhancement is expected to be $\sim$ 0.8-1.2 [Å]{} , depending on the age and metallicity of the stellar population.
The inner region (open triangles) is roughly consistent with the model with age $\sim$ 1.0-1.5 Gyr, metallicity Z $\sim$ 0.67, and \[$\alpha$/Fe\] = 0.5. The outer region (filled triangles), by contrast, shows stronger H$\delta_{\rm F}$ absorption, which suggests that the stellar population in the outer region is, on average, younger than that in the inner region. The age and the metallicity implied from the plot for H$\delta_{\rm F}$ against Mgb are $<$ 1.0 Gyr and Z $\sim$ 0.00-0.67, respectively. Although the metallicity may be difficult to determine quantitatively from Fig. 8 because of the quite large uncertainty in the Mgb index, Z $\sim$ 0.00-0.67 is consistent with the value implied from the strength of the Fe4668 and Fe5270 indices against the Mgb index as shown in Fig. 6.
-- -- --
-- -- --
--------------------------------------- ------------------------------------
{width="55mm"} {width="55mm"}
--------------------------------------- ------------------------------------
Comparison with the SSP models {#sec:age}
------------------------------
In the following discussions, we use the H$\delta_{\rm F}$ index and the D$_{4000}$ feature as age indicators of the stellar population. The H$\delta_{\rm F}$-D$_{4000}$ plot is widely used as a diagnostic tool to estimate the age of a stellar population in galaxies (Kauffmann et al. 2003; Poggianti & Barbaro 1997; Balogh et al. 1999). The great advantage in using the H$\delta_{\rm F}$ and D$_{4000}$ indices as age indicators is that they are less affected by the reddening caused by interstellar dust (Kauffmann et al. 2003). We note that the ages discussed in this section refer to the light-averaged ages of stellar populations, and thus we aimed to estimate the time at which star formation ceased in the samples qualitatively.
In Fig. 9, the measured H$\delta_{\rm F}$ index and the values of D$_{4000}$ are compared with predictions from the simple stellar population (SSP) models in the galaxev package of Bruzual & Charlot (2003). We used the model with the Salpeter initial mass function with lower and upper cut-offs of stellar masses of 0.1 ${\rm M}_{\odot}$ and 100 ${\rm M}_{\odot}$, respectively. The model assumes an instantaneous burst of star formation at age = 0 for the various metallicities (Bruzual & Charlot 2003).
The spectra of the SSP model used in Fig. 9 were broadened to the instrumental resolution of the observations (8.5 [Å]{} FWHM), which corresponds to a velocity dispersion of $\sigma$ $\sim$ 197 km s$^{-1}$. The spectral broadening was applied using the program vel\_disp included in the galaxev package (Bruzual & Charlot 2003). The resulting SSP model tracks for metallicities \[Fe/H\] = $-1.6464$, $-0.6392$, $+0.0932$, $+0.5595$ are overlaid in Fig. 9. Points with the same ages (1.0, 1.6, 2.5, 4.0, 6.2, 10.0 Gyr) are joined with dotted lines.
Based on the Fe4383 index, the outer regions of SDSSJ0333$-$0009 have \[$\alpha$/Fe\] $>$ 0.0, as discussed in the previous subsection. Because a stellar population which has \[$\alpha$/Fe\] $>$ 0.0 is reported to be enhanced in H$\delta$ absorption strength compared to that having solar $\alpha$/Fe ratios with similar age (Thomas et al. 2004), we have to take this effect into account to estimate the stellar population ages using H$\delta_{\rm F}$ indices. According to the stellar population model of Thomas et al. (2003), the enhanced strength in H$\delta_{F}$ indices is up to $\sim$ 1.2 [Å]{} for the model with \[$\alpha$/Fe\] = 0.5. Therefore, the data points should be shifted downwards from those actually plotted in Fig. 6 by up to $\sim$ 1.2 [Å]{} following the correction for the $\alpha$-enhancement. For SDSSJ0744$-$3738, \[$\alpha$/Fe\] could not be determined from the observed spectra, because the Mg indices were not measured with a high enough signal-to-noise ratio. Therefore, the additional uncertainty arising from the $\alpha$-enhancement should be taken into account when considering H$\delta_{\rm F}$ versus D$_{4000}$ diagnostics.
Another point of caution in using the H$\delta_{\rm F}$-D$_{4000}$ plot is that the stellar absorption of H$\delta_{\rm F}$ could be contaminated by nebular emission (Kauffmann et al. 2003; Goto et al. 2003a). The nebular emission of H$\alpha$ is weak for SDSSJ0333$-$0009 and SDSSJ0744$+$3738 ($<-4$ [Å]{}), and, furthermore, H$\beta$ was detected in absorption. The effect of emission filling on the H$\delta_{\rm F}$ absorption is expected to be negligible ($\sim$ 0.3 [Å]{} assuming the Case B recombination value for the H$\alpha$/H$\delta$ nebulae emission line ratio without dust extinction). In the following subsection, we estimate the light-averaged age of the stellar populations in the inner and outer regions for the observed galaxies.
### SDSS J033322.65$-$000907.5
The measured H$\delta_{\rm F}$ and D$_{4000}$ for AP$_{in}$ and AP$_{out}$ are shown in Fig. 9 by the open and filled triangles, respectively. If we assume a metallicity \[Fe/H\] $> + 0.0932$ and a correction for the $\alpha$-enhancement of $\sim$1.2 [Å]{} as suggested in Section 4.1, the data for APout are broadly consistent with a model with age older than $\sim$ 1.0 Gyr. By contrast, for the inner region (AP$_{in}$) the observed H$\delta_{\rm F}$ index seems stronger than the value that can be reproduced by the model with super-solar metallicity (\[Fe/H\] = $+0.5595$). However, the observed strength of D$_{4000}$ for AP$_{in}$ suggests an age older than $\sim$ 1.6 Gyr. Overall, the H$\delta_{\rm F}$-D$_{4000}$ plane suggests that the average age of the stellar population is younger in the outer region than in the inner region for this galaxy.
### SDSS J074452.52$+$373852.7
The open and filled squares in \[fig:4\] correspond to the data for AP$_{in}$ and AP$_{out}$, respectively, for SDSSJ0744$+$3738. For this galaxy, the metallicity and \[$\alpha$/Fe\] are not available because of the large errors in the Lick indices. Although the determination of the age of the stellar population for SDSSJ0744$+$3738 is quite uncertain, the observed large difference in D$_{4000}$ between the core and the exterior of the galaxy imply that the two regions have distinct stellar populations in terms of their age and/or metallicity. For APin, the large value of D$_{4000}$ is marginally consistent with the model with metallicity \[Fe/H\] = $+0.5595$ in the H$\delta_{\rm F}$-D$_{4000}$ plane. Furthermore, in the outer region (AP$_{out}$), the H$\delta_{\rm F}$ index is quite strong (5.5$\pm$11.1[Å]{}), although we note that the index strength is associated with large errors.
Galaxies with strong H$\delta$ absorption are classified as ’H$\delta$-strong’ (HDS) galaxies [@b57; @b20; @b9; @2004MNRAS.348..515G; @b61]. As strong H$\delta$ absorption could arise from A-type stars in the main-sequence phase [@b2], HDS galaxies probably terminated their starburst activity $\sim$0.1-1.5 Gyr ago [@b32; @b9]. Therefore, the strong H$\delta_{\rm F}$ absorption in the outer regions of the observed galaxies could be clear evidence that they terminated their star formation a few Gyrs ago. Nevertheless, because of the large uncertainty associated with H$\delta_{F}$, we need further observations with higher signal-to-noise ratios to detect a clear signature of quenched star formation.
![Equivalent width of H$\delta_{\rm F}$ as a function of D$_{4000}$. The open and filled triangles represent the data for AP$_{in}$ and AP$_{out}$, respectively, for SDSSJ0333$-$0009. The open and filled squares represent the data for APin and APout, respectively, for SDSSJ0744$+$3738. Overlaid lines show predictions of SSP models broadened to the instrumental resolution ($\sigma$ = 197 km s$^{-1}$) with various metallicities \[Fe/H\] = $-1.6464$, $-0.6392$, $+0.0932$, $+0.5595$ (Bruzual & Charlot 2003). Dotted lines connect points of the same age (from bottom to top, 10.0, 6.2, 4.0, 2.5, 1.6 and 1.0 Gyr, respectively).[]{data-label="fig:4"}](inst_burst.ps){width="80mm"}
### Implications
It is interesting that through the spectral diagnostics of spatially separated regions of passive spiral galaxies we can estimate the history of these galaxies. Even our spectra with relatively low signal-to-noise ratios suggest that passive spiral galaxies were in a star-forming phase for several gigayears before ceasing activity. Moreover, one of the passive spirals, SDSSJ0744$+$3738, probably terminated its star formation approximately 1-2 Gyr ago. If confirmed, this result will provide us with an important constraint in uncovering the underlying (cluster-related) physical mechanism responsible for the creation of passive spiral galaxies; for example, ram-pressure stripping [@b24; @b37; @b38; @b47; @b48; @b25] stops star formation much quicker than the strangulation scenario [@b52; @b46]. To reach a firm conclusion on the subject, however, further observations are needed: it is well known that the age/metallicity estimates are degenerate, for example the spectra of an old ($>$ 2 Gyr) stellar population looks almost identical when the age is doubled and the total metallicity reduced by a factor of 3 [@b62]. It is therefore important to break this degeneracy by obtaining spectra at a higher signal-to-noise ratio to measure indices sensitive only to age [e.g. Balmer lines; @b4] and only to metallicity [e.g. @b19]. Furthermore, we only had enough observing time for two passive spiral galaxies. As larger samples of passive spiral galaxies are now becoming available [e.g. @b1], it is important to investigate the statistical proportion of passive spiral galaxies that do not suffer from shot-noise to draw firm conclusions on this subject.
Summary {#sec:section5}
=======
We have performed spatially resolved long-slit spectroscopy of four candidate passive spiral galaxies using the APO 3.5-m telescope. Two of the observed galaxies show detectable \[O II\] and H$\alpha$ emission lines in the exterior regions of the disc, emphasizing the importance of investigating spatially large regions including the exterior regions of passive spiral galaxies. For the other two galaxies, for which the emission lines were not detected or were weak, we found radial gradients in the H$\delta_{\rm F}$ absorption and the strength of the 4000 [Å]{} break: Balmer absorption is more prominent and $D_{4000}$ is smaller in the outer regions of these galaxies. Taking into account the metallicity and the \[$\alpha$/Fe\] ratio roughly estimated for one of the sample galaxies, SDSSJ0333$-$0009, the comparison with the stellar population model suggests that the outer regions of the samples harbour younger populations of stars. The other observed passive spiral galaxy, SDSSJ0744$+$3738, also shows a younger population of stars in the outer regions, which presumably experienced the quenching of star formation a few Gyrs ago.
Our results have opened a door to understanding the history of cluster infall galaxies through spatially resolved spectroscopy. The results are, however, based on relatively low signal-to-noise spectra of only two passive spiral galaxies. To draw firm conclusions on this subject it is therefore essential to obtain the spectra of a larger sample of passive spiral galaxies at higher signal-to-noise ratios.
Acknowledgement {#acknowledgement .unnumbered}
===============
We are grateful to Youichi Ohyama for valuable advice on the data reduction. We wish to thank Chris Pearson for much advice and for comments that improved the paper. We also thank the anonymous referee for many insightful comments. This work is based on observations obtained with the Apache Point Observatory 3.5-m telescope, which is owned and operated by the Astrophysical Research Consortium. This research was partially supported by the Japan Society for the Promotion of Science through Grant-in-Aids for Scientific Research (nos 16204013 and 18840047) and by a Sasakawa Scientific Research Grant from the Japan Science Society.
[99]{} Abadi M. G., Moore B. and Bower R. G., 1999, MNRAS, 308, 947 Abazajian K., Adelman-McCarthy J. K., Agueros M. A., et al., 2005, AJ, 129, 1755 Adelman-McCarthy J. K., Agueros M. A., Allam S. S., et al., 2006, ApJS, 162, 38 Abraham R. G., Smecker-Hane T. A., Hutchings J. B., et al., 1996, ApJ, 471, 694 Balogh M. L., Morris S. L., Yee H. K. C., et al., 1999, ApJ, 527, 54 Balogh M. L., Bower R. G., Smail I., et al., 2002, MNRAS, 337, 256 Bekki K., Couch W. J., Shioya Y., et al., 2002, ApJ, 577, 651 Bennett C. L., Halpern M., Hinshaw G., et al., 2003, ApJS, 148, 1 Blanton, M. R., Dalcanton, J., Eisenstein, D., Loveday, J., Strauss, M. A., SubbaRao, M., Weinberg, D. H., 2001, AJ, 121, 2358 Boselli A., Boissier S., Cortese L., et al., 2006, ApJ, 651, 811 Bruzual A., G., 1983, ApJ, 273, 105 Bruzual G. and Charlot S., 2003, MNRAS, 344, 1000 Butcher H., Oemler A. Jr., 1978, ApJ, 226, 559 Couch W. J., Sharples R. M., 1987, MNRAS, 229, 423 Couch W. J., Barger A. J., Smail I., et al., 1998, ApJ, 497, 188 Diaferio A., Kauffmann G., Balogh M. L., et al., 2001, MNRAS, 323, 999 Dressler A., Smail I., Poggianti B. M., et al., 1999, ApJS, 122, 51 Elmergreen D. M., Elmegreen B. G., Frogel J. A., et al., 2002, AJ, 124, 777 Fujita Y., & Nagashima M., 1999, ApJ, 516, 619 Fujita Y., 2001, ApJ, 550, 612 Fujita Y., 2004, PASJ, 56, 29 Fujita Y.& Goto T., 2004, PASJ, 56, 621 Fukugita M., Ichikawa T., Gunn J. E., et al., 1996, AJ, 111, 1748 Gomez P. L., Nichol R. C., Miller C. J., et al., 2003, ApJ, 584, 210 Gorgas J., Cardiel N., Pedraz S., et al., 1999, A&A, 139, 29 Goto T., 2005, MNRAS, 357, 937G Goto T., 2006, MNRAS, 371, 769G Goto T., Nichol R. C., Okamura S., et al., 2003a, PASJ, 55, 771 Goto T., Okamura S., Sekiguchi M., et al., 2003b, PASJ, 55, 757 Goto T., Yamauchi C., Fujita Y., et al., 2003c, MNRAS, 346, 601 Goto T., Okamura S., Yagi M., et al., 2003d, PASJ, 55, 739 Goto T., Yagi M., Tanaka M., et al., 2004, MNRAS, 348, 515 Gunn J. E., Gott J. R., 1972, ApJ, 176, 1 Kauffmann G., Colberg J. M., Diaferio A., et al., 1999, MNRAS, 303, 188 Kauffmann G., Heckman T. M., White S. D. M., et al., 2003, MNRAS, 341, 33 Kennicutt R. C. Jr., 1983, AJ, 88, 483 Kennicutt R. C. Jr. 1989, ApJ, 344, 685 Kenney J. P. D., van Gorkom J. and Vollmer B., 2004, AJ, 127, 3361 Kodama T.& Bower R. T., 2001, MNRAS, 321, 18 Larson R. B., Tinsley B. M. and Caldwell C. N., 1980, 237, 692 Lewis I., Balogh M., De Propris R., et al., 2002, MNRAS, 334, 673 Millar N. A. and Owen F. N., 2002, AJ, 124, 2453 Moore B., Lake G., Dressler A., et al., 1996, Nat, 379,613 Moran S. M., Ellis R. S., Treu T., 2006, ApJ, 641L,97 Morris S. L., Hutchings J. B., Carlberg R. G., et al., 1998, ApJ, 507, 84 Poggianti B. M. and Barbaro G., 1996, A&A, 314, 379 Poggianti B. M. and Barbaro G., 1997, A&A, 325, 1025 Poggianti B. M., Smail I., Dressler A., et al., 1999, ApJ, 518, 576 Postman M., Franx M., Cross N. J. G., et al., 2005, ApJ, 623, 721 Quilis V., Moore B., Bower R., 2000, Sci, 288, 1617 Savege B. D. and Mathis J. S., 1979, ARA&A, 17, 73 Shimasaku K., Fukugita M., Doi M., et al., 2001, AJ, 122, 1238 Strauss M. A., Weinberg D. H., Lupton R. H., et al., 2002, AJ, 124, 1810 Tanaka M., Goto T., Okamura S., et al., 2004, AJ, 128, 2677 Thomas D., Maraston C. and Bender R., 2003, MNRAS, 339, 897 Trager S. C., Worthey G., Faber S. M., et al., 1998, ApJS, 116, 1 Treu T., Ellis R. S., Kneib J., et al., 2003, ApJ, 591, 53 van den Bergh S., 1976, ApJ, 206, 883 Vollmer B., Otmianowska-Mazur K., Kenney J. D. P., et al., 2006, A& A, 453, 883 Wilkerson M. S., 1980, ApJ, 240L, 115 Worthey G., 1994, ApJS, 95, 107 Worthey G. and Ottaviani D. L., 1997, ApJS, 111, 377
\[lastpage\]
[^1]: E-mail: [email protected] (M.I.); [email protected] (T.G.); [email protected] (H.M.)
[^2]: Astronomical Institute, Tohoku University, Sendai, 980-8578, Japan
[^3]: Velocity dispersion $\sigma_{V}$ and Petrosian radius $r_{\rm P}$ in $r'$ band are taken from SDSS Data Release 5 [@b3; @Adelman]
[^4]: Absolute magnitude in $r'$ band
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Lingyao Kong, Daniele Malafarina and Cosimo Bambi$^1$'
title: 'Can we observationally test the weak cosmic censorship conjecture?'
---
Introduction
============
One of the most important open problems in gravitational physics is that of the final fate of a heavy star after exhausting its nuclear fuel. For normal stars, the object contracts up to when the quantum pressure of electrons or neutrons stops the collapse and the outcome is either a white dwarf or a neutron star. However, if the star is very massive, there is no known mechanism capable of compensating the inward push of its own gravitational force, and the body will undergo a complete gravitational collapse. According to the theory of general relativity, the final product of gravitational collapse must be a spacetime singularity [@hp; @he]. In principle, the singularity may either be hidden behind a horizon, and in this case the result of the collapse is a black hole, or be naked, and thus visible to distant observers. While the weak cosmic censorship conjecture asserts that singularities created in gravitational collapse must be hidden within black holes [@wccc], today we know many physically relevant counterexamples in which naked singularities are formed from regular initial data (for a recent review, see e.g. Ref. [@jm] and Ref. [@Joshibook] for a detailed treatment). The possibility of detecting radiation from the high curvature region where classically we would expect the formation of a singularity would represent a unique opportunity to investigate strong gravity and observationally test the region where quantum gravity phenomena are supposed to show up [@z2; @bmm1; @bmm2].
The predictions of general relativity have been confirmed by experiments in Earth’s gravitational field [@llr; @gpb], by spacecraft missions in the Solar System [@cassini], and by accurate radio observations of binary pulsars [@wt05; @pulsar] (for a general review, see e.g. Ref. [@will]). In all these environments, the gravitational field is weak, in the sense that one can write $g_{tt} = 1 + \phi$ with $|\phi| \ll 1$. In the last few years, there have been significant efforts and progresses to test the theory in the strong field regime, where the approximation $|\phi| \ll 1$ breaks down. The ideal laboratory to test strong gravitational fields is the spacetime around astrophysical black hole candidates (see e.g. Refs. [@image; @jet1; @jet2; @qpo; @cfm; @iron; @wh; @cfmiron; @torres; @jp1; @jp2]; for a review see [@rev1; @rev2]). These studies have shown that the properties of the electromagnetic radiation emitted by the gas in the accretion disk can provide useful details about the spacetime geometry around the compact object and available radio and X-ray data can already be used to constrain new physics.
While deviations from the predictions of general relativity in the spacetime around astrophysical black hole candidates are definitively possible, since current observations can put only weak constraints, from purely theoretical arguments new physics is not strictly necessary here (see however [@mathur; @gd1; @gd2; @gd3]). The black hole’s event horizon has indeed no special properties for a freely falling observer. On the contrary, the existence of spacetime singularities, where observer-independent quantities like the scalar curvature or the Kretschmann scalar may diverge, is very likely a symptom of the break down of general relativity and new physics, presumably a quantum theory of gravity, is mandatory. Some observational tests have already been proposed in the literature [@ns1; @ns2; @ns3; @ns4; @ns5]. In this paper we study the question of principle whether the high density region close to the formation of the singularity can affect the outside universe by exploring a toy model describing the radiation emitted from the high curvature region of astrophysical collapsing bodies, where classically we would expect the formation of a singularity. More specifically, we want to figure out if – at least in principle – we can observationally distinguish the case in which the classical singularity that forms at the end of the collapse is not covered by the horizon from the case in which the horizon forms before the singularity. If this were to be possible, we would in principle be able to experimentally test the weak cosmic censorship conjecture.
During the collapse, the density and the temperature of the object increase. Subnuclear reactions, otherwise strongly suppressed, become important and the collapsing star can emit a large amount of energy in several forms of radiation. The luminosity curve of this radiation clearly depends on the evolution of the gravitational collapse, setting the evolution of the increase in density and temperature at any layer of the body. For instance, the detection of neutrinos from supernovae may be used to probe the equation of state of matter at supernuclear densities [@naka1; @naka2]. Experiments to detect neutrinos from supernovae already exist and they are simply waiting for the explosion of a nearby supernova. In 1987, neutrinos from the supernova SN1987A in the Large Magellanic Cloud were detected by experiments searching the proton decay (Kamiokane II, IMB, Baksan detected together 24 events). With the sensitivity of present experiments, a supernova explosion in our Galaxy could produce thousands of events in a detector like Super-Kamiokande and even millions of events in one like IceCube. In the same way, some weakly interacting radiation may be used to track the gravitational collapse of an astrophysical body and observationally test if the collapse follows the pattern expected for the formation of a black hole, for the creation of a naked singularity, or another one.
The simplest exact solution for gravitational collapse in which the outcome can be either a black hole or a naked singularity is the Leimatre-Tolman-Bondi (LTB) dust model [@lem; @tol; @bon]. The system has spherical symmetry and, depending on the initial density and velocity profile, it may behave in two different ways. Either the horizon develops first and the subsequent singularity is always covered or, vice versa, a singularity visible to distant observers forms before the formation of the horizon [@dust1; @dust2; @dust3; @dust4; @dust5]. If we consider a distant observer and we integrate backwards in time the photons’ trajectories, from the observer to the collapsing object, we can obtain the luminosity image at any time. Integrating over the whole image, we can find the curve luminosity produced by the collapsing object. As the evolution of the emissivity of the matter in the star depends on the gravitational collapse, the curve luminosity provide information on the collapse itself. In particular, one may expect that the formation of a event horizon and of a naked singularity can be characterized by qualitatively different light curves.
Unfortunately, this does not seem to be the case. Assuming two different emissivity functions, we find that both the black hole and the naked singularity case show very similar luminosity spectrum. This would leave the distinction undetermined even once a more realistic scenario is considered. While the results shown here are based on a very simple analytical toy model, they definitely suggest that any observational test of the weak cosmic censorship conjecture through the analysis of the emitted spectrum of a collapsing astrophysical body may be extremely challenging unless some effects coming from new physics at high densities do not intervene to completely change the picture.
The content of our manuscript is organized as follows. In Section \[s-1\], we briefly review the LTB dust collapse model. In Section \[s-2\], we describe how our code computes the luminosity of the collapsing object seen by a distant observer. In Section \[s-3\], we present the results of our numerical calculations. At first we consider the homogeneous case, the well-known Oppenheimer and Snyder model [@OS], in which the final product of the collapse is always a black hole. Then we extend the study to the inhomogeneous case, in which the collapse can create either a black hole or a naked singularity, depending on the initial matter density profile. We then compare the curve luminosity of the two scenarios. Summary and conclusions are reported in Section \[s-c\]. Throughout the paper, we use units in which $G_{\rm N} = c = 1$.
LTB dust collapse model \[s-1\]
===============================
The LTB model describes a spherically symmetric system composed of non interacting particles (dust) that undergoes complete gravitational collapse. The most general spherically symmetric line element can be written in comoving coordinates (namely coordinates attached to the infalling particles) as $$\begin{aligned}
ds^2 = -e^{2\nu}dt^2 + \frac{\rho'^2}{G}dr^2
+ \rho^2 \left(d\theta^2 + \sin^2\theta d\phi^2\right) \, ,\end{aligned}$$ where $\nu$, $\rho$, and $G$ are functions of the comoving time $t$ and radius $r$. Here and in what follows, the prime $(')$ denotes a derivative with respect to $r$. If we impose that $\nu$, $\rho$, and $G$ are independent of the $t$ coordinate, we find the class of static interior Schwarzschild solutions that was originally studied by Tolman [@Tolman]. As we are using comoving coordinates, the energy momentum tensor of a relativistic fluid is diagonal and can be written as $$\begin{aligned}
T^\mu_\nu = {\rm diag}\{\epsilon(r,t), p_r(r,t), p_\theta(r,t), p_\theta(r,t) \} \, ,\end{aligned}$$ where $\epsilon$ is the energy density and $p_r$ and $p_\theta$ are, respectively, the radial and tangential pressure of the cloud. Einstein’s equations then take the form $$\begin{aligned}
\epsilon &=& \frac{F'}{\rho^2 \rho'} \, , \label{ee-e} \\
p_r &=& - \frac{\dot{F}}{\rho^2 \dot{\rho}} \, , \label{ee-pr} \\
\nu' &=& 2 \frac{p_\theta - p_r}{\epsilon + p_r} \frac{\rho'}{\rho}
- \frac{p_r'}{\epsilon+p_r} \, , \label{ee-nu} \\
\dot{G} &=& \frac{2 \nu'}{\rho'} \dot{\rho} G \label{ee-g} \, ,\end{aligned}$$ where the dot $(\dot{})$ denotes a derivative with respect to $t$ and $F$ is the Misner-Sharp mass, defined by $$\begin{aligned}
\label{misner}
F = \rho (1 - G + e^{-2\nu} \dot{\rho}^2) \, .\end{aligned}$$ $F$ is proportional to the “gravitational mass” enclosed within the shell of radial coordinate $r$ at the time $t$. Eventually, we have a set of five coupled first order differential equations in seven unknown functions of $r$ and $t$. In general, it may be hard or impossible to solve the system. The usual approach is to look for simplifications that, while preserving the physical features of interest, allow us to solve the equations and say something about the nature of the collapse.
The dust case is obtained for particles that carry no self-interacting energy and can be described imposing that $p_r = p_\theta = 0$. It turns out that this case is a particularly simple model where the set of Einstein’s equations can be solved explicitly. From Eq. , we get $\nu = \nu(t)$ and we can change the time gauge in order to have a $t$ coordinate for which $\nu = 0$. Then Eq. reduces to $\dot{F} = 0$, which implies $F=F(r)$; that is, the amount of matter enclosed within the shell labeled by $r$ is conserved during the collapse. This, in turn, means that there is no inflow or outflow of matter/energy during the collapse and therefore, given the spherical symmetry of the system, the matching with the exterior geometry can be done with the simple and well-known Schwarzschild spacetime [@matching]. Furthermore, given the absence of pressures, the boundary radius $r_b$, which corresponds to the shrinking boundary area-radius $\rho_b(t)=\rho(r_b,t)$, can be chosen at will. From the matching conditions, it is easy to see that $F(r_b) = 2M_{\rm Sch}$, where $M_{\rm Sch}$ is the mass parameter in the exterior Schwarzschild metric. From Eq. , that for dust reads $\dot{G}=0$, we can obtain $G=G(r)$ as a free function, which is convenient to write in the form $G(r)=1+f(r)$. Finally, Eq. becomes the equation of motion of the system $$\begin{aligned}
\label{e-rhodot}
\dot{\rho}=-\sqrt{\frac{F}{\rho} + f} \, ,\end{aligned}$$ with the minus sign necessary to describe collapse. Given a certain mass profile $F$, after choosing the free function $f$, we can integrate Eq. to get $\rho(t,r)$. Plugging this solution into Eq. , we obtain $\epsilon(r,t)$, thus completely solving the system.
The free function $f$ coming from the integration of Eq. is related to the velocity of the infalling particles. The collapse is said to be bound if $f < 0$, marginally bound if $f = 0$, and unbound if $f > 0$. In the rest of the manuscript, we will restrict our attention on the marginally bound case $f = 0$, which represents particles that would have zero initial velocity at spatial infinity. The line element for the collapsing interior reduces to $$\begin{aligned}
ds^2_{\rm int} = - dt^2 + \rho'^2dr^2
+ \rho^2 \left(d\theta^2 + \sin^2\theta d\phi^2\right) \, .\end{aligned}$$ The collapse process leads eventually to the formation of a black hole when all the matter passes the threshold of trapping surfaces located at the event horizon in the Schwarzschild exterior. The condition for the formation of trapped surfaces for the collapsing cloud is given by $1 - F/\rho = 0$, and it reduces to $1 - 2 M_{\rm Sch}/R = 0$, where $R$ is the Schwarzschild radial coordinate, in the static case in vacuum. All the matter falls into the spacetime singularity that forms at $r=0$ and it is easy to check, for instance by evaluating the Kretschmann scalar, that this is a true curvature singularity. A factor of crucial importance for black hole physics is to determine whether the singularity always forms at a later time with respect to the formation of the horizon, thus being hidden to far away observers at all times, or if other possibilities are allowed.
The whole system has a gauge degree of freedom that can be fixed by setting the scale at a certain time. In models of collapse, one usually sets the area radius $\rho(t,r)$ equal to the comoving radius $r$ at the initial time $t_{i}$; that is, $\rho(t_{i},r)=r$. We can then introduce a scale function $a$ as $$\begin{aligned}
\rho(t,r)=ra(t,r) \, .\end{aligned}$$ with the initial condition $a(t_{i},r)=1$. Further one wishes to impose certain regularity conditions to ensure the physical validity of the model. For example, one wishes to have a density profile that is regular at the center at the initial time and that presents no cusps in $r=0$ at all times. In order to have such regularity conditions, we can impose that the mass function near the center behaves in a suitable way. Therefore we can define a function $M(r)$ such that $$F(r)=r^3M(r) \, .$$ We can rewrite the whole system of equations in terms of $M$ and $a$ and it is immediately found that the form of Eq. with $f=0$ is invariant under the substitution of $F$ with $M$ and $\rho$ with $a$. The energy density can now be written as $$\label{dens}
\epsilon(r,t)=\frac{3M+rM'}{a^2(a+ra')} \, ,$$ and it is easy to check that the singularity occurs for $a=0$, while values of the central shell $r=0$ with $a\neq 0$ are regular. This solves the problem of the divergence of $\epsilon$ at $\rho=0$ at all times coming from Eq. (\[ee-e\]).
Note from Eq. that the density diverges also when $\rho'=a+ra'$ goes to zero. This is related to the occurrence of so called “shell crossing” singularities [@cross1; @cross2; @cross3]. These are singularities that arise from a breakdown of the coordinate system rather than true gravitational singularities and they can generally be avoided by a suitable change of coordinates. Nevertheless it is important to check under what circumstances they can arise in the model in order to impose some prescription to rule them out. It is not difficult to see that for dust collapse, if one imposes a decreasing energy profile, then no shell crossing occurs. With the new scaling, together with the requirement of avoidance of shell crossing, the density diverges only when the singularity is achieved.
Oppenheimer-Snyder collapse
---------------------------
If, for simplicity, we want to describe homogeneous collapse where $\epsilon=\epsilon(t)$, we need to take $M(r)=M_0$, which implies $a=a(t)$. Then $\epsilon=3M_0/a^3$ and the equation of motion reduces to $$\label{motion}
\dot{a}=-\sqrt{\frac{M_0}{a}} \, .$$ Eq. (\[motion\]) can be easily integrated. The solution is given by $$a(t)=\left(1-\frac{3}{2}\sqrt{M_0}t\right)^{2/3} \, ,$$ with initial time $t_{i}=0$. The singularity forms at the time $t_s=2/3\sqrt{M_0}$. Here all the shells become singular at the same time and the horizon forms at the boundary before the formation of the singularity, therefore leaving the singularity always covered.
The boundary of the cloud collapses along the curve $\rho_b(t)=\rho(r_b,t)=r_b a(t)$ and the whole cloud becomes trapped inside the event horizon at the time $t_{tr}<t_s$ for which $\rho_b(t_{tr})=2M_{\rm Sch}=r_b^3M_0$, so $$t_{tr}=t_s-\frac{4M_{\rm Sch}}{3} \, .$$ For the homogeneous dust collapse we thus have the formation of a spacelike simultaneous singularity that is always covered by the horizon. This means that the region of extremely high density at the center of the collapsing cloud close to the formation of the singularity is causally disconnected from the outside universe.
Inhomogeneous collapse
----------------------
If we wish to analyze a more general case, still within the dust scenario, we can take $M(r)$ to be written as an expansion close to the center as $$\begin{aligned}
M(r)&=&M_0+M_1r+M_2r^2+... \end{aligned}$$ In this case, the density profile $\epsilon(r,t)$ is not homogeneous any more and the mass profile $M(r)$ will determine its form. It is easy to see that the behavior of trapped surfaces and the structure of formation of the singularity can change drastically (see for example [@ns1b; @ns2b; @ns2b-2; @ns3b] and references therein). It turns out that it is the sign of $M_1$ that will determine the character of the singularity curve and the apparent horizon near the center. If we require $M_1=0$ (as it is often done in astrophysical scenarios, where one desires to have only quadratic terms in the density and pressures), the behavior of the apparent horizon and of the singularity curve near the center will be determined by the value of $M_2$. In the following we will therefore consider $M_1=0$ and $M_2<0$. The case $M_2>0$ is not physically relevant, as it implies a density increasing with the radius.
Mathematically, since in the dust collapse there are no pressures, the matching with the outside region can be done at any radius and therefore, if one shows that the singularity is locally naked (meaning that there are outgoing geodesics originating at the singularity and reaching a finite radius without being trapped), then one can choose $r_b$ for the matching in order to make it globally naked (meaning that such geodesics can reach observers at infinity). In a realistic scenario, things might be different (see for example [@MJH]) and when pressures are considered it is preferable to perform the matching with the exterior region at the radius where the pressure vanishes. Nevertheless, the possibility remains that the central singularity be visible to far away observers (see for example [@JMS]).
The above formalism is enough to obtain the necessary information about the behavior of the dust cloud close to $r=0$ and close to the formation of the singularity. Two scenarios are possible:
1. In the black hole case, the trapped surfaces form at an outer shell before the formation of the singularity. Essentially each shell $r>0$ becomes trapped at a time smaller than the time of formation of the central singularity, $t_s(0)$.
2. In the naked singularity case, the trapped surfaces form at the center at the time of formation of the singularity. This means that the shell $r=0$ becomes trapped at the time $t_s(0)$, and light rays escaping from the central singularity can reach far away observers. At later times the trapped surface expands, thus covering the singularity.
The equation of motion is given by Eq. with $M(r)$ in place of $M_0$ and the solution takes again the form $$\label{a}
a(r,t)=\left(1-\frac{3}{2}\sqrt{M(r)}t\right)^{2/3} \: .$$ The singularity curve is described by the condition that $a(t_s(r),r)=0$ which gives $$\begin{aligned}
\label{ts2}
t_{s}(r)=\frac{2}{3\sqrt{M(r)}} \, ,\end{aligned}$$ from which we can see that different shells become singular at different times, with the central shell becoming singular first in the case that $M_2<0$. The apparent horizon curve is given by $$\begin{aligned}
\label{ts2}
t_{ah}(r)= t_{s}(r) - \frac{2}{3}r^3M(r)\, ,\end{aligned}$$ and it is easy to check that $t_{ah}$ is also increasing from the center and that $t_{ah}(0)=t_{s}(0)$. Therefore, in the inhomogeneous dust case with $M_2<0$ the central singularity is not trapped at the time of its formation and it may become trapped only afterwards. This means that for certain choices of the boundary radius the high density region that develops close to the singularity is causally connected to the outside universe (see for example [@ns1b; @ns2b; @ns2b-2] and [@ortiz] for the complete conformal structure of the model). Such a collapsing cloud can potentially bear an observational signature different from that of the black hole case discussed above. Not every negative value of $M_2$ is allowed. From the condition that the energy density is positive throughout the cloud we get the constraint $$M_2 > - \frac{3}{5} \frac{M_0}{r_b^2} \, .$$
The gravitational collapse of a dust cloud is just a simple toy model that has the advantage that it can be treated analytically. Obviously, if one wishes to describe a star, pressures are important[^1]. Of course here we are investigating a mathematical toy model describing a simple light spectrum emitted from the vicinity of a naked singularity that has no resemblance to the real spectrum emitted by a realistic collapsing object. However this investigation is important in that it helps us answer the question of principle of whether the visibility of the region surrounding the singularity could have long range effects in such a way as to make it distinguishable from the formation of a black hole. Therefore these models can constitute a first step to investigate what could possibly be observationally detectable if such naked singularities happened in realistic star collapse or in the formation of supermassive compact objects. In fact, if one thinks about supermassive compact objects, the formalism is exactly the same, but the time scales are much longer. Actually, in this case the dust model could be a better approximation than in the star collapse case, since supermassive objects are less dense than stars and tracing of geodesics inside a supermassive collapsing object in more meaningful.
Exterior geometry
-----------------
The whole spacetime can be described by a collapsing interior given by the LTB metric (in comoving coordinates $\{t,r\}$) that matches at the boundary $r_b$ to a vacuum exterior given by the Schwarzschild metric (in Schwarzschild coordinates $\{T,R\}$) $$ds^2_{\rm out}=-\left(1-\frac{2M_{\rm Sch}}{R}\right)dT^2+\left(1-\frac{2M_{\rm Sch}}{R}\right)^{-1}dR^2+R^2d\Omega^2 \; .$$ The two manifolds match across the 3-surface $\Sigma$ given by $r=r_b$, which corresponds to $\rho=\rho_b(t)=\rho(r_b,t)$, in the interior and $R=R_b(T)$ in the exterior. The matching of the two manifolds is achieved imposing the continuity of the first and second fundamental forms across the surface. Continuity of the first fundamental form reduces to continuity of the metric across the surface. The induced metric on the surface can be written as $$\begin{aligned}
ds^2_\Sigma&=&-dt^2+\rho_b(t)^2d\Omega^2=\nonumber\\
&=& - \left(1-\frac{2M_{\rm Sch}}{R_b}\right)dT^2+\left(1-\frac{2M_{\rm Sch}}{R_b}\right)^{-1}
\left(\frac{dR_b}{dT}\right)^2 dT^2+R_b^2d\Omega^2 \, ,\end{aligned}$$ and the coordinate transformation on the boundary of the cloud $\Sigma$ that relates $t$ to $T$ is given by $$\label{time}
\frac{dt}{dT}=\sqrt{\left(1-\frac{2M_{{\rm Sch}}}{R_b}\right)-\left(1-\frac{2M_{{\rm Sch}}}{R_b}\right)^{-1}\left(\frac{dR_b}{dT}\right)^2} \, .$$ The metric components are continuos across $\Sigma$ and we can express the trajectory of the boundary as $\rho_b(t)=R_b(T(t))$.
Tracing photons \[s-2\]
=======================
In this section we discuss the theoretical aspects of the procedure by which we intend to trace light rays from the collapsing cloud to far away observers. We consider the geodesics starting at some far away initial radius $R_*$ at the time $T=\bar{T}$ so that $R(\bar{T})=R_*$. Then we follow $R(T)$ tracing the photon backwards in time along the path from the observer to the collapsing cloud. Three scenarios are possible:
1. The photon escapes to infinity never hitting either the cloud or the event horizon.
2. The photon hits the event horizon.
3. The photon hits the collapsing cloud, thus reaching the boundary $R_b(T)$ at a time antecedent the formation of the horizon. In this case, the photon can either escape from the cloud, thus crossing again the boundary at a later time, or hit the event horizon.
First of all, we consider the Schwarzschild solution to construct the image of the object for a far away observer. This is the image seen after that all the photons coming from the LTB region reached the observer. Then we will consider the LTB region describing the collapsing cloud. This evolves from the initial time $T_i=T(t_i)$ until the formation of the event horizon at $T_{tr}=T(t_{tr})$ after which, from the perspective of external observers, we are left with a black hole. Therefore the image seen by the far away observer will change in time from the initial moment until the formation of the horizon.
Given the spherical symmetry of the spacetime, motion happens always on a plane and we can restrict our analysis to the equatorial plane without any loss of generality. Null geodesics are then described by the equation $$\left(\frac{dR}{d\lambda}\right)^2=E^2-\frac{L^2}{R^2}\left(1-\frac{2M}{R}\right) \; ,$$ for the Schwarzschild exterior and by $$\begin{aligned}
\label{ng}
&&\left(\frac{dt}{d\lambda}\right)^2
-\rho'\left(\frac{dr}{d\lambda}\right)^2-\frac{L^2}{\rho^2}=0 \; , \nonumber\\
&&\frac{d^2r}{d\lambda^2}
+ 2\frac{\dot{\rho}'}{\rho'}\frac{dr}{d\lambda}
\frac{dt}{d\lambda}
+ \frac{\rho''}{\rho'}\left(\frac{dr}{d\lambda}\right)^2
- \frac{L^2}{\rho^3\rho'} = 0 \; ,\end{aligned}$$ for the LTB interior. Here $\lambda$ is an affine parameter, while $E$ and $L$ are the conserved energy and angular momentum related to the killing vectors as defined below. Once the photon hits the boundary of the cloud, as in the case (3) above, to trace it inside the cloud we have to change coordinates from $T$ to $t$ via Eq. and follow the geodesic $r(t)$ (using the comoving time $t$ as the affine parameter) obtained by solving equation with the same value for $L$ as the one used for the outer part of the trajectory [^2].
Geodesics in the Schwarzschild spacetime
----------------------------------------
The Schwarzschild spacetime is static and spherically symmetric. We can thus define two quantities conserved along geodesics. They are related to the killing vectors associated to time translations and rotations. These quantities are the energy $E$ and angular moment $L$ and are given by $$\begin{aligned}
E=\left(1-\frac{2M_{{\rm Sch}}}{R}\right)\frac{dT}{d\lambda} \; ,
\quad L=R^{2}\frac{d\phi}{d\lambda} \; .\end{aligned}$$ Since the trajectory of a photon is independent of its energy $E$, for the study of null geodesics it is more convenient to use the “impact parameter” $b = L/E$ instead of $E$ and $L$. All the equations depend now on $b$ and $E$ and $L$ never appear. From the expression of the Schwarzschild metric, we can write the equation for null geodesics as $$\left(\frac{dR}{dT}\right)^{2}=\left(1-\frac{2M_{{\rm Sch}}}{R}\right)^{2}-\left(1-\frac{2M_{{\rm Sch}}}{R}\right)^{3}\frac{b^{2}}{R^{2}} \; .$$ which, once integrated with the initial condition $R(\bar{T})=R_*$, gives the trajectory $R(T)$ of the photon in the exterior spacetime.
The other ingredient necessary to trace the photon backward in time from the observer to the cloud is the trajectory of the boundary of the cloud as given by $R_{b}\left(T\right)$. This allows us to determine whether and when the photon hits the boundary of the collapsing object. From Eq. we use the equation of motion written at the boundary as $$\frac{d\rho_{b}}{dt}=-\sqrt{\frac{F}{\rho_{b}}}=-\sqrt{\frac{2M_{{\rm Sch}}}{R_{b}}}\; ,$$ and, noting that at the boundary of the collapsing object is $\rho_{b}(t)=R_{b}(T(t))$, we get $$\frac{dR_{b}}{dT}=\frac{dR_{b}}{dt}\frac{dt}{dT}=\frac{d\rho_{b}}{dt}\frac{dt}{dT} \; .$$ Now making use of equation we obtain $$\frac{dR_{b}}{dT}=-\sqrt{\frac{2M_{{\rm Sch}}}{R_{b}}}\left(1-\frac{2M_{{\rm Sch}}}{R_{b}}\right)\label{eq:RbT}\; ,$$ that, once integrated, gives $$\begin{aligned}
T (R_{b}) & = & T_{0}
-\frac{2}{3} \frac{R_{b}^{3/2}}{\sqrt{2M_{{\rm Sch}}}}
-2\sqrt{2M_{{\rm Sch}} R_{b}} + 2M_{{\rm Sch}} \ln
\left(\sqrt{R_b} +\sqrt{2M_{{\rm Sch}}}\right)
+ \nonumber\\ &&
-2M_{{\rm Sch}}\ln\left(\sqrt{R_b}-\sqrt{2M_{{\rm Sch}}}\right)\; ,\end{aligned}$$ that can be inverted to obtain $R_{b}(T)$. The intersection of the null geodesic $R(T)$ with the boundary curve $R_{b}(T)$ then gives the time $\bar{\bar{T}}$ at which the photon traveling along the geodesics hits the boundary.
Geodesics in the LTB spacetime
------------------------------
Some of the photons that are traced back in time from the screen will hit the boundary of the cloud and propagate in the interior. These are the actual photons that are coming from the collapsing object and we need to evaluate geodesics in the LTB interior in order to trace them. These were first studied in [@nakao]. The null geodesics in the interior can be obtained from $$\begin{aligned}
\frac{dt}{d\lambda} &=& \sqrt{\rho'^2
\left(\frac{dr}{d\lambda}\right)^2
+ \frac{b^2}{\rho^2}} \; , \label{eq:geo_ltb_t} \\
\frac{d^2r}{d\lambda^2} &=&
-2\frac{\dot{\rho}'}{\rho'}
\sqrt{\rho'^2 \left(\frac{dr}{d\lambda}\right)^2 + \frac{b^2}{\rho^2}}\frac{dr}{d\lambda}
- \frac{\rho''}{\rho'}\left(\frac{dr}{d\lambda}\right)^2
+ \frac{b^2}{\rho^3\rho'} \; .
\label{eq:geo_ltb_r}\end{aligned}$$ The photon now propagates inside the cloud following the null geodesics with initial conditions at the boundary, namely the photon motion in the LTB interior starts with position $\rho_b=R_b(\bar{\bar{T}})$ at the time $\bar{\bar{t}}=t(\bar{\bar{T}})$. From the expression for $\rho=ra$ with $a$ given by equation , we can find the time at which the photon coming from the observer’s screen is at the boundary in the interior coordinates. We have $$\begin{aligned}
\frac{dR_b}{d\lambda} & = & -\sqrt{\frac{2M_{{\rm Sch}}}{R_b}}\frac{dt}{d\lambda}
+R_b'\frac{dr}{d\lambda}\label{eq:dR} \; , \\
\left(\frac{dT}{d\lambda}\right)^{2} & = &
\left(1-\frac{2M_{{\rm Sch}}}{R_b}\right)^{-2}
\left[\left(\frac{dR_b}{d\lambda}\right)^{2}
+\left(1-\frac{2M_{{\rm Sch}}}{R_b}\right)\frac{b^{2}}{R_b^{2}}\right] \; .\end{aligned}$$ Now, with the help of Eq. (\[ng\]), we get $$\begin{aligned}
\frac{dT}{d\lambda} =
\left( 1-\frac{2M_{{\rm Sch}}}{R} \right)^{-1}
\left(\frac{dt}{d\lambda}-R'\sqrt{\frac{2M_{{\rm Sch}}}{R}}
\frac{dr}{d\lambda}\right)\label{eq:dT} \; ,\end{aligned}$$ and eventually we arrive at the initial conditions at the boundary for the first derivatives of $t(\lambda)$ and $r(\lambda)$, namely $$\begin{aligned}
\frac{dt}{d\lambda} & = & \frac{dT}{d\lambda}+\frac{\sqrt{\frac{2M_{{\rm Sch}}}{R}}\frac{dR}{d\lambda}}{\left(1-\frac{2M_{{\rm Sch}}}{R}\right)} \; ,
\label{eq:dt_init} \\
\frac{dr}{d\lambda} & = & \frac{1}{\rho'}\sqrt{\left(\frac{dt}{d\lambda}\right)^{2}-\frac{b^{2}}{R^{2}}}
\; . \label{eq:dr_init}\end{aligned}$$ With the above equations, it is straightforward to numerically calculate all the photon trajectories.
Observed spectrum
-----------------
In the following, we consider two examples, namely the homogenous collapse model, where we set $M=M_0$, and the inhomogeneous one, where we have $M=M_0+M_2r^2$. The spectrum at the time $T$ measured by the distant observer is given by [@worm] $$I\left(T,\nu_{{\rm obs}}\right)=\int2\pi b \, db\int_{\gamma}g^{3}jdl \; ,$$ where $\gamma$ is the photon’s path, $j$ is the emissivity per unit volume in the rest frame of the emitter and $g$ is the gravitational redshift $$g=\frac{\nu_{{\rm obs}}}{\nu_{e}}=
\frac{k_{\mu}v_{{\rm obs}}^{\mu}}{k_{\nu}v_{e}^{\nu}}
=\frac{E}{\frac{dt}{d\lambda}} \; ,$$ $\nu_{{\rm obs}}$ is the photon frequency as measured by the distant observer, $\nu_{e}$ is the photon frequency with respect to the emitter, $v_{{\rm obs}}^{\mu}=\left(1,0,0,0\right)$ is the 4-velocity of the distant observer, $v_{e}^{\mu}=\left(1,0,0,0\right)$ is the 4-velocity of the emitter, and $k^{\mu}$ is the 4-momentum of the photon. Also $dl$ is the proper length in the rest-frame of the emitter and in our model it turns out to be equal to $dt$ $$dl=\sqrt{^{3}g_{ij}\frac{dx^{i}}{d\lambda}\frac{dx^{j}}{d\lambda}}d\lambda=dt\,.$$ In the next section, for the sake of clarity, we will use two simple emissivity functions. In the first model, we assume that the emission is monochromatic with rest-frame frequency $\nu_{\star}$ and proportional to the square of the energy density (as we may expect in a two-body collision) $$\label{eq-delta}
j=\epsilon^{2} \delta\left(\nu_{e}-\nu_{\star}\right) \; .$$ In the second example, we replace the monochromatic emission with an exponential function that could somehow mimic a thermal emission (even if, strictly speaking, our object is made of dust and therefore the temperature is zero) $$\label{eq-exp}
j=\epsilon^{2}\nu_{e}^{2}\exp\left(-\frac{\nu_{e}}{\epsilon^{1/4}}\right)\,.$$
Results and discussion \[s-3\]
==============================
Here we report the results for the spectrum measured by observers at infinity for the two cases discussed above with a specific choice of the parameters involved. The first model is the homogeneous dust collapse (Oppenheimer-Snyder model) which terminates with the creation of a black hole. In this case the high density region that develops close to the formation of the singularity is always covered by the horizon. The second example is the inhomogeneous dust collapse model in which we find the formation of a singularity that can be visible to far away observers. The singular point from which null geodesics can propagate to infinity is that of $t_s(0)$ as given in equation . All the other points in the singularity curve $t_s(r)$ are covered by the apparent horizon, nevertheless it can be shown that once there exist one radial null geodesic emanating from $t_s(0)$ that can reach observers at infinity then there exists a whole family of null geodesics and also non radial and non null geodesics can escape (see for example Refs. [@geodesics; @geodesics2]). Still, the amount of radiation that can escape from the high density region close to the singularity could in principle be very small and thus negligible in comparison with the spectrum emitted from the low density region which behaves similarly to the homogeneous cloud. This in fact turns out to be the case as the qualitative features of both spectrums result to be similar.
For the numerical calculations we have therefore set a scale by fixing the total mass of the object in both cases. More specifically, we have set the total mass $2M_{{\rm Sch}}=1$. What changes is the way the mass is distributed within the cloud. Having fixed the value of the parameter $M_{0}=0.01$ to be the same in both models, in order to have the same total mass we must retrieve a different boundary in the two cases [^3]. Then for the inhomogeneous collapse model we have set $M_{2}=-0.00015$. The parameter $M_2$ of the second model has been chosen, among all the allowed values for which the singularity is globally visible, in such a way to have the maximal effect in the difference between the two light curves. From the relation $2M_{{\rm Sch}}=r_b^{3}M(r_b)$, we find that the radius of the boundary $r_{b}$ is $\approx 4.64$ in the homogeneous model and $\approx 6.06$ in the inhomogeneous one. Let us notice that, for the above choice of $M_0$ and $M_2$, the maximum boundary radius for the inhomogeneous collapse model is $\approx 6.32$ (the density at larger radii becomes negative and thus unphysical), and this also maximizes the difference in the light curves between the homogeneous and inhomogeneous collapse. The calculations have been performed from the distant observer (locate at the radius $R_* = 10^7$ in units $2M_{{\rm Sch}}=1$) backward in time to the collapsing object. At any time as measured by the distant observer, the photon trajectory is characterized only by the impact parameter $b$. Geodesics in the exterior Schwarzschild spacetime have been computed with a 4-th order Runge-Kutta method. As in the interior LTB spacetime we have a second order differential equation, we decided to use the Runge-Kutta-Nystrom method [@Lund2009] for the calculation of the trajectories.
The evolution of the light curve luminosity as a function of the proper time of the distant observer $T$ for the homogeneous/black hole and inhomogeneous/naked singularity scenarios are reported in Figs. \[f-delta\] and \[f-exp\], respectively for the monochromatic emissivity of Eq. (\[eq-delta\]) and the thermal-like one of Eq. (\[eq-exp\]). Figs. \[f-s-delta\] and \[f-s-exp\] show instead some spectra at specific times. The blueshift ($\nu_{\rm obs} > \nu_*$) shown in some panels of Fig. \[f-s-delta\] is an effect experienced by those photons that can propagate for a long time inside the collapsing object, so that the blueshift gained in the interior solution exceed the redshift in the exterior Schwarzschild spacetime from the surface of the object to the distant observer. The phenomenon is explained with some details in Ref. [@kmb].
As we can see from both the light curve luminosity and the spectra, there are no qualitatively significant features that can differentiate between the two scenarios. While in the inhomogeneous/naked singularity case we could have expected a much higher luminosity originating from the central region with high energy density just before the formation of the naked singularity, it turns out the that size of this region is too small, and the time scale too short, to produce a significant emission of radiation. On the one hand, this result may suggest that the formation of a spacetime singularity visible to distant observers in our Universe is not a catastrophic event incompatible with observations. It would seem that even within this simple and extremely idealized collapse model it is impossible to observationally distinguish the birth of a black hole from that of a naked singularity. Therefore, provided that the scenario is not drastically altered by some other effects (like for example those induced in the strong field regime by some theory of quantum gravity), the creation of a spacetime naked singularity as the endstate of collapse might not have significant direct observational consequences for far away observers. While we cannot at present exclude the possibility of testing the weak cosmic censorship conjecture and/or probe the high densities region where classically we would expect the formation of a spacetime singularity with astrophysical observations, it is clear that such possibilities are at least challenging, even in principle, and are likely to remain challenging also in the case of more sophisticated theoretical models that account for more realistic scenarios and astrophysical effects.
![Left panel: light curve luminosity of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Right panel: zoom of the left panel at the peak of the luminosity. $T$ in units $2M_{{\rm Sch}}=1$. Luminosity in arbitrary units. \[f-delta\]](delta-T0-logI "fig:"){width="7.8cm"}![Left panel: light curve luminosity of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Right panel: zoom of the left panel at the peak of the luminosity. $T$ in units $2M_{{\rm Sch}}=1$. Luminosity in arbitrary units. \[f-delta\]](delta-T0-logI-d "fig:"){width="7.8cm"}
![As in Fig. \[f-delta\] for the emissivity function in Eq. (\[eq-exp\]). $T$ in units $2M_{{\rm Sch}}=1$. Luminosity in arbitrary units. \[f-exp\]](exp-T0-logI "fig:"){width="7.8cm"}![As in Fig. \[f-delta\] for the emissivity function in Eq. (\[eq-exp\]). $T$ in units $2M_{{\rm Sch}}=1$. Luminosity in arbitrary units. \[f-exp\]](exp-T0-logI-d "fig:"){width="7.8cm"}
![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T0-spec "fig:"){width="7.8cm"}![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T20-spec "fig:"){width="7.8cm"}
![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T40-spec "fig:"){width="7.8cm"}![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T50-spec "fig:"){width="7.8cm"}
![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T53-spec "fig:"){width="7.8cm"}![Spectra of an LTB collapsing object with the emissivity function described by Eq. (\[eq-delta\]), for the homogeneous/black hole (red solid curve) and the inhomogeneous/naked singularity case (blue dashed curve). Luminosity in arbitrary units. \[f-s-delta\]](delta-T55-spec "fig:"){width="7.8cm"}
![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T0-spec "fig:"){width="7.8cm"}![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T20-spec "fig:"){width="7.8cm"}
![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T40-spec "fig:"){width="7.8cm"}![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T50-spec "fig:"){width="7.8cm"}
![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T53-spec "fig:"){width="7.8cm"}![As in Fig. \[f-s-delta\] for the emissivity function in Eq. (\[eq-exp\]). Luminosity and $\nu_{\rm obs}$ in arbitrary units. \[f-s-exp\]](exp-T55-spec "fig:"){width="7.8cm"}
Summary and conclusions \[s-c\]
===============================
In general relativity, gravitational collapse of type I matter fields satisfying basic energy conditions ends with the formation of a singularity of the spacetime, where the matter density diverges and standard physics breaks down. In particular this is the case for dust collapse, where, in the absence of pressures, a spacetime singularity is the only allowed outcome of collapse under the basic assumption of the positivity of mass and energy density. Spacetime singularities may either be hidden behind a horizon, as in the case of black holes, or be naked and thus visible to distant observers. We now know many physically meaningful examples of naked singularities created as the endstate of collapse of matter fields that respect the standard energy conditions, starting with regular initial data. On the other hand, the weak cosmic censorship conjecture asserts that singularities produced in any generic gravitational collapse scenario must be hidden within black holes and cannot be seen by distant observers. The validity of this conjecture is still an open and controversial problem, but it is a key-assumption in black hole thermodynamics and it is of crucial importance for astrophysics where observed massive compact objects that exceed the Chandrasekhar mass limit are usually assumed to be black holes.
In the present paper, we have tried to address the question whether it is possible to observationally test the weak cosmic censorship conjecture by measuring the radiation emitted by a collapsing body. In order to have a first understanding of the basic features of the problem, we decided to begin by studying the simplest theoretical collapse model, the LTB model, for which an analytical solution is known and easily calculated and to simplify as much as possible the assumptions related to the emitted radiation. In this scenario the final product of collapse can be either a black hole or a naked singularity, depending on the values chosen for the parameters that determine the density profile. The naked singularity that forms as the endstate of dust collapse is naked only for a “short time” in comoving coordinates. Nevertheless this time may have been quite large once the photons emitted from the high density region reach observers at infinity.
We have computed the radiation emitted by these collapsing objects and their light curves, which can potentially track the evolution of the collapse, in order to find observational signature capable of distinguishing the birth of a black hole from the one of a naked singularity. Our collapse model is very simple and assume that the object is optically thin to the emitted radiation, which should make much easier the possibility of distinguishing the two scenarios than a realistic case with a lot of astrophysical complications. The answer to our question is not intuitively accessible, as the final result depends on several relativistic effects, like the gravitational redshift and the time delay between the collapsing star and the distant observer. Within our simple model, we did not find any specific signature to identify the naked singularity scenario. As shown in Figs. \[f-delta\] and \[f-exp\], the light curves for black holes and naked singularities do not seem to be qualitatively different. Roughly speaking, the reason is that the high density region formed just before the formation of the singularity is too small to produce an observational signature in the flux reaching the distant observer. While our finding cannot definitively exclude the possibility of observationally probing the high density region where classically we would expect the formation of a spacetime singularity, observational tests of the weak cosmic censorship conjecture seem to be at least extremely challenging, even in principle and even in the simplest case where we neglect all the possible astrophysical complications.
This work was supported by the NSFC grant No. 11305038, the Innovation Program of Shanghai Municipal Education Commission grant No. 14ZZ001, the Thousand Young Talents Program, and Fudan University.
[99]{}
S. W. Hawking and R. Penrose, Proc. Roy. Soc. Lond. A [**314**]{}, 529 (1970).
S. W. Hawking and G. F. R. Ellis, [*The Large scale structure of space-time*]{} (Cambridge University Press, Cambridge, UK, 1973).
R. Penrose, Riv. Nuovo Cim. [**1**]{}, 252 (1969) \[Gen. Rel. Grav. [**34**]{}, 1141 (2002)\].
P. S. Joshi and D. Malafarina, Int. J. Mod. Phys. D [**20**]{}, 2641 (2011) \[arXiv:1201.3660 \[gr-qc\]\].
P. S. Joshi, [*Gravitational Collapse and Spacetime singularities*]{} (Cambridge University Press, Cambridge, UK, 2007).
Z. Li and C. Bambi, Phys. Rev. D [**87**]{}, 124022 (2013) \[arXiv:1304.6592 \[gr-qc\]\].
C. Bambi, D. Malafarina and L. Modesto, Phys. Rev. D [**88**]{}, 044009 (2013) \[arXiv:1305.4790 \[gr-qc\]\].
C. Bambi, D. Malafarina and L. Modesto, arXiv:1306.1668 \[gr-qc\].
J. G. Williams, S. G. Turyshev and D. H. Boggs, Phys. Rev. Lett. [**93**]{}, 261101 (2004) \[gr-qc/0411113\].
C. W. F. Everitt, D. B. DeBra, B. W. Parkinson, J. P. Turneaure, J. W. Conklin, M. I. Heifetz, G. M. Keiser and A. S. Silbergleit [*et al.*]{}, Phys. Rev. Lett. [**106**]{}, 221101 (2011) \[arXiv:1105.3456 \[gr-qc\]\].
B. Bertotti, L. Iess and P. Tortora, Nature [**425**]{}, 374 (2003).
J. M. Weisberg and J. H. Taylor, ASP Conf. Ser. [**328**]{}, 25 (2005) \[astro-ph/0407149\]
M. Kramer, I. H. Stairs, R. N. Manchester, M. A. McLaughlin, A. G. Lyne, R. D. Ferdman, M. Burgay and D. R. Lorimer [*et al.*]{}, Science [**314**]{}, 97 (2006) \[astro-ph/0609417\].
C. M. Will, Living Rev. Rel. [**9**]{}, 3 (2006) \[gr-qc/0510072\].
C. Bambi and N. Yoshida, Class. Quant. Grav. [**27**]{}, 205006 (2010) \[arXiv:1004.3149 \[gr-qc\]\].
C. Bambi, Phys. Rev. D [**85**]{}, 043002 (2012) \[arXiv:1201.1638 \[gr-qc\]\].
C. Bambi, Phys. Rev. D [**86**]{}, 123013 (2012) \[arXiv:1204.6395 \[gr-qc\]\].
C. Bambi, JCAP [**1209**]{}, 014 (2012) \[arXiv:1205.6348 \[gr-qc\]\].
C. Bambi, Astrophys. J. [**761**]{}, 174 (2012) \[arXiv:1210.5679 \[gr-qc\]\].
C. Bambi, Phys. Rev. D [**87**]{}, 023007 (2013) \[arXiv:1211.2513 \[gr-qc\]\].
C. Bambi, JCAP [**1308**]{}, 055 (2013) \[arXiv:1305.5409 \[gr-qc\]\].
D. F. Torres, Nucl. Phys. B [**626**]{}, 377 (2002) \[hep-ph/0201154\].
T. Johannsen and D. Psaltis, Astrophys. J. [**726**]{}, 11 (2011) \[arXiv:1010.1000 \[astro-ph.HE\]\].
T. Johannsen and D. Psaltis, Astrophys. J. [**773**]{}, 57 (2013) \[arXiv:1202.6069 \[astro-ph.HE\]\].
C. Bambi, Phys. Rev. D [**87**]{}, 084039 (2013) \[arXiv:1303.0624 \[gr-qc\]\].
C. Bambi, Mod. Phys. Lett. A [**26**]{}, 2453 (2011) \[arXiv:1109.4256 \[gr-qc\]\].
C. Bambi, Astron. Rev. [**8**]{}, 4 (2013) \[arXiv:1301.0361 \[gr-qc\]\].
S. D. Mathur, Fortsch. Phys. [**53**]{}, 793 (2005) \[hep-th/0502050\].
G. Dvali and C. Gomez, arXiv:1112.3359 \[hep-th\].
G. Dvali and C. Gomez, Phys. Lett. B [**719**]{}, 419 (2013) \[arXiv:1203.6575 \[hep-th\]\].
G. Dvali and C. Gomez, arXiv:1212.0765 \[hep-th\].
C. Bambi and K. Freese, Phys. Rev. D [**79**]{}, 043002 (2009) \[arXiv:0812.1328 \[astro-ph\]\].
C. Bambi, T. Harada, R. Takahashi and N. Yoshida, Phys. Rev. D [**81**]{}, 104004 (2010) \[arXiv:1003.4821 \[gr-qc\]\].
C. Bambi and N. Yoshida, Phys. Rev. D [**82**]{}, 064002 (2010) \[arXiv:1006.4296 \[gr-qc\]\].
P. S. Joshi, D. Malafarina and R. Narayan, arXiv:1304.7331 \[gr-qc\].
C. Bambi and D. Malafarina, Phys. Rev. D [**88**]{}, 064022 (2013) \[arXiv:1307.2106 \[gr-qc\]\].
K. ’i. Nakazato, K. Sumiyoshi and S. Yamada, Astrophys. J. [**721**]{}, 1284 (2010) \[arXiv:1001.5084 \[astro-ph.HE\]\].
K. ’i. Nakazato, K. Sumiyoshi, H. Suzuki and S. Yamada, Phys. Rev. D [**81**]{}, 083009 (2010) \[arXiv:1004.0291 \[astro-ph.HE\]\].
G. Lemaitre, Gen. Rel. Grav. [**29**]{}, 641 (1997) \[Annales Soc. Sci. Brux. Ser. I Sci. Math. Astron. Phys. A [**53**]{}, 51 (1933)\].
R. C. Tolman, Proc. Nat. Acad. Sci. [**20**]{}, 169 (1934) \[Gen. Rel. Grav. [**29**]{}, 935 (1997)\].
H. Bondi, Mon. Not. Roy. Astron. Soc. [**107**]{}, 410 (1947).
P. S. Joshi and I. H. Dwivedi, Class. Quantum Grav. **16**, 41 (1999). B. Waugh and K. Lake, Phys. Rev. D [**38**]{}, 1315 (1988).
R. P. A. C. Newman, Class. Quantum Grav. [**3**]{}, 527 (1986).
D. Christodoulou, Commun. Math. Phys. [**93**]{}, 171 (1984).
D. M. Eardley and L. Smarr, Phys. Rev. D [**19**]{}, 2239 (1979).
J. R. Oppenheimer and H. Snyder, Phys. Rev. [**56**]{}, 455 (1939).
R. C. Tolman, Phys. Rev. **55**, 364 (1939).
W. Israel, Nuovo Cim. B [**44**]{}, 1 (1966); Nuovo Cim. B [**48**]{}, 463 (1966).
P. Yodzis, H.-J. Seifert and H. Muller zum Hagen, Commun. Math. Phys. **34**, 135 (1973).
C. Hellaby and K. Lake, Astrophys. J. **290**, 381 (1985); Astrophys. J. **300**, 461 (1986).
P. S. Joshi and R. V. Saraykar, Int. J. Mod. Phys. D [**22**]{}, 1350027 (2013). \[arXiv:1205.3263 \[gr-qc\]\].
P. S. Joshi and I. H. Dwivedi, Phys. Rev. D [**47**]{}, 5357 (1993).
P. S. Joshi and I. H. Dwivedi, Commun. Math. Phys. [**146**]{}, 333 (1992).
P. S. Joshi and I. H. Dwivedi, Lett. Math. Phys. [**27**]{}, 235 (1993).
P. S. Joshi, N. Dadhich and R. Maartens, Phys.Rev. D [**65**]{}, 101501 (2002).
U. Miyamoto, S. Jhingan and T. Harada, Prog. Theor. Exp. Phys. 053E01 (2013).
P. S. Joshi, D. Malafarina and R. V. Saraykar, Int. J. Mod. Phys. D, [**21**]{}, 12500 (2012).
N. Ortiz and O. Sarbach, Class. Quantum Grav. [**28**]{}, 235001 (2011).
R. Penrose, in [*Gravitational Radiation and Gravitational Collapse*]{}, Proceedings of the IAU Symposium, edited by C. DeWitt-Morette, IAU Symposium n. 64 (Reidel, Dordrecht, 1974).
K.I. Nakao, N. Kobayashi and H. Ishihara, Phys.Rev. D [**67**]{}, 084002 (2003).
C. Bambi, Phys. Rev. D [**87**]{}, 107501 (2013) \[arXiv:1304.5691 \[gr-qc\]\].
F. C. Mena and B. C. Nolan. Class. Quantum Grav. [**18**]{}, 4531 (2001) \[arXiv:gr-qc/0108008 \[gr-qc\]\].
S. S. Deshingkar, P. S. Joshi, and I. H. Dwiwedi, Phys. Rev. D [**65**]{}, 084009 (2002).
E. Lund, L. Bugge, I. Gavrilenko, and A. Strandlie, J. INSTRUM, **4**, P04001 (2009).
L. Kong, D. Malafarina and C. Bambi, arXiv:1310.1320 \[gr-qc\].
[^1]: Though it has been suggested that matter might approach a dust-like behavior close to the formation of the singularity where very strong gravitational fields are present (essentially particles falling in close to the speed of light are not able to interact) [@Penrose2].
[^2]: If we restrict to the case of radial null geodesics ($L=0$) we get $\frac{dt}{dr}=\pm\rho'$ where the plus sign denotes outgoing geodesics, while the minus sign is for ingoing geodesics. Then the problem of studying the behavior of radial null geodesics emanating from the center of the dust cloud translates into the Cauchy problem given by Eq. with the initial value $t(0)=t_0$, where $t_0\in[t_i,t_s(0)]$.
[^3]: Another way of proceeding might have been to fix the same boundary for both models and thus have different values of $M_0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider constraints on gauge mediated supersymmetry breaking models imposed by the requirement of grand unification. In particular, we demonstrate several ways to reduce the number of parameters coming from the dynamical supersymmetry breaking sector. One of approaches exploits nonperturbative unification of gauge couplings in multi-messenger models.'
author:
- 'S.L.Dubovsky, D.S.Gorbunov, and S.V.Troitsky\'
title: ' Grand unification with gauge mediated supersymmetry breaking[^1] '
---
[**1.**]{} [*Introduction.*]{} Almost all existing experimental particle physics data fit well the Standard Model predicitons, so every reason to invent new physics beyond this theory is purely aesthetical. From the particle theory point of view, more beautiful models are obviously those which have more symmetries, and those which are predictive, i.e. have less free parameters. Two of the most popular extensions of the Standard Model deal with supersymmetry and with Grand Unification, respectively. The latter concept shares both pleasant features, namely, larger symmetries in gauge interactions and in matter content together with restrictions on the existing parameters of the Standard Model. On the other hand, supersymmetry requires much more symmetries restricting particle dynamics but often fails to be predictive in realistic models — it introduces quite a few new parameters instead of constraining the existing ones. Merging the two concepts, however, can lead to really aesthetically appealing models – Supersymmetric Grand Unified Theories (GUTs).
The main purpose of this talk is to demonstrate how the concept of Grand Unification helps to constrain the parameters in one particular class of supersymmetric models – models with gauge mediated supersymmetry breaking (GMSB). The latter models are known to be rather predictive by themselves; we show here that in some cases the number of free parameters may be even more reduced once a particular model is specified. We begin with a brief sketch of GMSB models, then discuss possible general constraints coming from the Grand Unification postulates. We turn then to particular examples and consider the ways the models of direct gauge mediation may be constrained.
[**2.**]{} [*Sketch of GMSB.*]{} In most realistic supersymmetric models, supersymmetry is broken in the sector of fields different from the Standard Model and its minimal supersymmetric extension (MSSM). The (rather weak) interaction between the two sectors transfers supersymmetry breaking to the visible fields, and so determines the parameters of the MSSM. In the most general case, there are of order $10^2$ of these parameters, and to start comparing the model with experiments, one first needs to constrain the parameter space. The most attractive way is to specify the interaction between the two sectors and to calculate the values of the parameters, but in practice several relations between them are often conjectured.
It is the kind of interaction between the dynamical supersymmetry breaking (DSB) sector and the observable world which distinguishes between different classes of supersymmetric models. This interaction should be suppressed at low energies, and two conventional choices are gravitational (which operates at energies of order Planck scale) and Yukawa or/and gauge (operating somewhere between the electroweak and Planck scales) interactions. The latter case corresponds to the GMSB theories, and makes it possible to calculate the supersymmetry breaking parameters of the MSSM by means of the field theoretical methods, without invoking less understood theories of quantum gravity. In the former case, some simple constraints on the MSSM parameters are often postulated (see, e.g., Ref.[@Nilles] for a review of gravity mediated supersymmetry breaking).
The key ingredient of the GMSB scenario is a set of messenger fields which fall in the vectorlike multiplets of the Standard Model gauge symmetry. They are the only fields from the visible sector which directly interact with the DSB sector. By means of this interaction (either pure Yukawa in the minimal models or both Yukawa and gauge in the so-called direct mediation models) the messengers obtain supersymmetry breaking masses of their component fields. The parameters of the MSSM – soft gaugino and scalar masses and trilinear couplings — are generated via loop effects by MSSM gauge interactions between messengers and ordinary particles. Thus, the masses of the superpartners are determined by their $SU(3)\times SU(2) \times U(1)$ quantum numbers and the parameters describing the spectrum of messengers (typically, there are two such parameters, the mass of the fermionic component, and the mass splitting between bosonic messengers). The gauge mediation mechanism naturally suppresses flavour violating processes and is highly predictive — all MSSM parameters (not counting 17 parameters of the non-supersymmetric Standard Model) are calculated in terms of four — two in the messenger sector and two in the Higgs sector. The GMSB phenomenology and model building are widely discussed, e.g., in the review Ref. [@Review]; here we will concentrate on ways to further reduce the number of free parameters by making use of Grand Unification constraints.
[**3.**]{} [*Simple GUT constraints.*]{}
Let us list the well known constraints which are characteristic to Grand Unified theories and enable one to restrict parameters of the low-energy theory. These constraints are
1. unification of gauge couplings;
2. unification of Yukawa couplings;
3. matter content in full GUT multiplets (or explanation of splitting, as in the case of Higgs doublets and triplets);
4. interactions which may be written in terms of GUT multiplets.
All these types of constraints may be used to rule out some models, and the first and the second constraints may provide quantitative bounds on parameters of a given model. Consider, as an example, the minimal gauge mediated model with one (5+$\bar 5$) set of messengers. Perturbative gauge unification in the visible sector is unchanged by construction (at least in one loop). One may, however, impose an additional constraint (which seems very plausible) that the Grand Unified Theory, if exists, should contain the secluded (DSB) sector as well. With this assumption, one can exploit the unification of gauge couplings of the DSB and visible sectors to gain information about the values of the coupling constant in the secluded sector and, notably, about the scale where it becomes strongly coupled. The latter scale $\Lambda_s$ determines the value of one of the supersymmetry breaking parameters for a given model. Phenomenological and cosmological constraints on this parameter allow us to extract “unifiable” models from a plethora of known DSB schemes. Only three of them (the “3-2” model [@ADS-85], its extension with extra matter, and the model with $SU(2)$ group and vector-like matter content [@Izawa]) were found to satisfy these criteria [@Izawa; @couplings] without introducing additional matter thresholds. Another example concerns the restrictions on the minimal $SU(5)$ GUT with GMSB that come from the analysis of gauge and $b$–$\tau$ Yukawa coupling unification [@GUT_fall]. These restrictions in fact have shown the inconsistency of the minimal model with unification.
One should note, however, that a unified theory containing both DSB and visible sectors is still missing. Nevertheless, as we will see in the next section, even the requirement of unification in the visible sector may be sometimes very restrictive.
[**4.**]{} [*Constraining multi-messenger direct gauge mediation scenarios.*]{} Let us turn to the models with direct gauge mediation where the messenger fields themselves are part of the DSB sector, namely, they carry charges under the secluded gauge group. From the visible sector viewpoint, the latter gauge group plays a role of a flavour group, so several copies of messenger fields should appear in the spectrum – the number of copies being equal to the dimensionality of the corresponding representation of the secluded gauge group. Since the messengers are charged under the Standard Model too, their contribution to the gauge beta functions of the Standard Model may lead to the loss of the asymptotic freedom. For example, if the number of $5+\bar 5$ multiplets exceeds four, the gauge coupling constants become large below $M_{\rm GUT}\sim 10^{16}$ GeV which contradicts the idea of perturbative unification. On the other hand, it is often that the DSB group is not very small, so it has no representation with dimension 4 or less[^2]. The most common approach to the question of saving gauge unification in this case is to make the messengers heavy [@direct; @Dimop-1]. For high enough thresholds of the messengers, Landau poles might be “pushed” away to energies higher than $M_{\rm GUT}$. Though particle phenomenology does not suffer from messengers being heavy, they often are problematic for cosmology [@Dimop-1; @new-nucl]. This fact suggests to look for other ways of solving the gauge unification problem in multi-messenger models. One of the ways is to invoke completely new physics at intermediate scales between $M_{\rm SUSY}$ and $M_{\rm
GUT}$. If the Standard model fields and/or messengers are composite states made of a few fundamental preons, the latter transforming as complete GUT multiplets, then at high energies only preons contribute to the beta functions of $SU(3)\times SU(2)\times U(1)$. Generally, their contribution is smaller than the contribution of the composite fields, and the coupling constants at $M_{\rm GUT}$ can remain small. A few examples of gauge mediated supersymmetry breaking models with direct gauge mediation and compositeness were constructed, but they are either toy models [@we:compose] or too complicated to be realistic [@compose]. All models require dynamical assumptions about uncalculable dynamics at strong coupling.
Let us turn now to the last, and the least explored possibility. Its main idea is to replace the perturbative unification of couplings by a controllable and phenomenologically acceptable unification at the strong coupling.
The possibility of gauge coupling unification in the strong coupling regime has been considered in the framework of both the Standard Model and its supersymmetric extensions [@Maiani]. Recently, this problem attracted some interest again [@nonpert-unif] after more precise measurements of the gauge coupling constants at $M_Z$ have been carried out. The latter results differ from the two-loop unification predictions by more than one standard deviation (see, e.g., Ref. [@kabak]).
Note that running gauge couplings of the MSSM $\alpha_1$ and $\alpha_2$ increase with energy, so $SU(2)\times U(1)$ is not asymptotically free. These couplings, however, run relatively slow, so Landau poles of these two groups appear at energies higher than the unification scale. Together with the asymptotic freedom of QCD this means that below $M_{GUT}$ all gauge couplings are small, and perturbative analysis is valid. This picture impies the existence of the “desert”, i.e. absence of particles huge region of masses between superparticle and unification scales. When new particles, like several multiplets of messengers, are introduced, the first coefficients of the $\beta$ functions increase, so gauge couplings may become large at the unification scale.
Despite the fact that unification in this case occurs at the strong coupling, it is unexpectedly controllable from the low energy point of view, especially in the supersymmetric case [@Ross]. Consider one-loop evolution of the coupling constants in an asymptotically non-free unified theory. If $M_G$ is the unification scale and $\alpha_G$ is the value of the unified gauge coupling at that scale, then the renormalisation group equations $${d\alpha_i\over dt}=b'_i\alpha_i^2
\label{RG}$$ have a solution $$\alpha_i^{-1}(Q)=\alpha_G^{-1}+b'_i t,$$ where $t={1\over 2\pi}\ln{Q\over M_G}$ and $b'_i$ are the first coefficients of the beta functions of the gauge couplings in the asymptotically non-free theory. Consider running of the [*ratios*]{} of pairs of the gauge couplings, $${d\over dt}\ln{\alpha_i\over\alpha_j}=b'_i\alpha_i-b'_j\alpha_j.
\label{r}$$ At one loop, these ratios have infrared fixed points, $${\alpha_i\over\alpha_j}={b'_j\over b'_i}.$$ These fixed points are reached at the energies which are model-dependent and may be read out from the solution to eq.[(\[r\])]{}, $${\alpha_i(Q)\over\alpha_j(Q)}={\alpha_G^{-1}+b'_j t \over
\alpha_G^{-1}+ b'_i t}.$$ The condition that the fixed point is almost reached is $|t|\gg\alpha_G^{-1}/b'_i$. In the case of MSSM without additional matter, one has $\alpha_G^{-1}\sim 24$, so that for $\alpha_2$ the fixed point occurs at $|t|\gg 24$, i.e., at $Q\ll
M_G\cdot\exp(-48\pi)\sim 10^{-66} M_G$ which certainly rules out the possibility of the fixed point analysis.However, with new matter added, the situation changes drastically — $b_i$ increase and $\alpha_G^{-1}$ decreases. Suppose that additional (messenger) multiplets fall in the complete vector-like representations of the $SU(5)$ unified gauge group, for example, ($5+\bar{5}$) or ($10+\bar{10}$). Then $b'_i=b_i+n$, where $b_i$ are $\beta$ function coefficients of the MSSM. Each ($5+\bar{5}$) set adds 1 to $n$ while each ($10+\bar{10}$) adds 3. For $n\ge 5$ the unification occurs at strong coupling. To estimate the energy scale where ratios of couplings get close to the fixed point value, let us take $\alpha_G=1$. Then even for $n=5$ the ratios are almost constant at $Q<0.04 M_G$.
For given $n$, the threshold corresponding to messenger mass is uniquely determined. Indeed, the low energy running of MSSM couplings is known, and at the threshold the couplings [*should*]{} have the ratios equal to $b'_i/b'_j$. The energy where ratios of MSSM running couplings, determined experimentally at $M_Z$, get to their fixed point values corresponds to the messenger threshold. Note that at $n\ge 5$ the corresponding thresholds are deep in the region of the attraction of the fixed points. For $n=5$, for example, the threshold is between 1 and 10 TeV, much lower than $0.04\cdot M_G$. This means that the fixed-point approach is self-consistent. The values of thresholds can be read out from Ref.[@Ross]; values of $6\le n \le 20$ are consistent with current bounds on the messenger mass [@to-appear].
So, from the low energy MSSM point of view we just have new boundary conditions for running of the gauge couplings. Instead of requiring the equality of couplings at $M_G$ (as in the case of perturbative unification), one should fix their ratios at the messenger scale. Details of evolution of the couplings near $M_G$, where they are large, are unknown; however, they do not affect significantly the low-energy predictions [@Ross; @NewRoss].
We conclude that having quite large number of messenger fields at scales between $M_{\rm SUSY}$ and $M_{\rm GUT}$ does not contradict the gauge coupling unification, the latter occuring in the nonperturbative regime.
The most interesting feature of this scenario is that the strong unification constrains significantly the parameter space of multi-messenger gauge mediation models. Namely, the mass scale of the messenger fields – one of the two parameters describing the superpartner masses – is determined for a given effective number of messengers $n$.
It is worth noting, however, that the superparticle spectrum does not depend significantly on the messenger mass $M$. Instead, the scale of superpartner masses is set by the product $\Lambda=Mx$, where $0<x<1$. In most supersymmetry breaking models, the scale $\Lambda$ depends on the scale $\Lambda_S$ where the gauge coupling constant of the secluded sector becomes strong. To determine $\Lambda_S$, one has to put some boundary conditions on that gauge coupling. This can be done either at $M_{\rm GUT}$, from the condition of “total” unification, like in Ref.[@couplings], or at some intermediate scale by means of the fixed point formalism. If all contributions to the soft $B_\mu$ term come from the gauge mediation, and the supersymmetric $\mu$ parameter is tuned so that electroweak symmetry breaking is radiative, then $\Lambda$ and $M$ determine the superparticle spectrum completely — the model thus has no free parameters. Whether this spectrum is realistic, depends on the choice of model [@to-appear].
[**5.**]{} [*Conclusion.*]{} We have demonstrated that simple requirements of the Grand Unification of both visible and secluded sectors, and/or (in the multi-messenger case) of non-perturbative Grand Unification in the visible sector can significantly reduce the number of parameters of theories with gauge mediated supersymmetry breaking, thus ruling out some of them or predicting superparticle spectrum for others.
We thank M.V.Libanov, Yu.F.Pirogov, and V.A.Rubakov for interesting discussions. This work is supported in part by the RFFI grant 96-02-17449a. Work of S.T. is supported in part by the U.S. Civilian Research and Development Foundation for Independent States of FSU (CRDF) Award No. RP1-187. \#1\#2\#3[[*Phys. Rev. Lett.* ]{}[**\#1 **]{} \#3 (19\#2)]{} \#1\#2\#3[[*Nucl. Phys.* ]{}[**B\#1 **]{} \#3 (19\#2)]{} \#1\#2\#3[[*Phys. Rev.* ]{}[**D\#1 **]{} \#3 (19\#2)]{} \#1\#2\#3[[*Phys. Lett.* ]{}[**B\#1**]{} \#3 (19\#2)]{}
[99]{} Nilles H P [*Phys. Rep.*]{} [**110**]{} 1 (1984) Giudice G F, Rattazzi R, hep-ph/9801271 Affleck I, Dine M, Seiberg N Izawa K, Nomura Y, Tobe K, Yanagida T , hep-ph/9705228 Dubovsky S L, Gorbunov D S, Troitsky S V , hep-ph/9707357;\
Eung Jin Chun, Haewon Lee, Won Sik l’Yi, hep-ph/9705457 Carone C, Murayama H , hep-ph/9510219;\
Blok B, Lu C, Zhang D , hep-ph/9602310;\
Hamidian H [*et al.*]{}, hep-ph/9803228 Agashe K, hep-ph/9804450 Poppitz E, Trivedi S P , hep-ph/9609529;\
Arkani-Hamed N, March-Russell J, Murayama H [*Nucl. Phys.*]{} [**B509**]{} 3 (1998), hep-ph/9701286;\
Murayama H , hep-ph/9705271;\
Dimopoulos S, Dvali G, Rattazzi R , hep-ph/9707537;\
Shirman Y , hep-ph/9709383 Dimopoulos S [*et al.*]{} , hep-ph/9705307 Gherghetta T, Giudice G F, Riotto A, hep-ph/9808401 Dubovsky S L, Gorbunov D S, Troitsky S V [*Phys. Lett.*]{} [**B423**]{} 301 (1998), hep-ph/9712397 Arkani-Hamed N, Luty M A, Terning J , hep-ph/9712389 Maiani L, Parisi G, Petronzio R [*Nucl. Phys.*]{} [**B136**]{} 115 (1978);\
Maiani L, Petronzio R [*Phys. Lett.*]{} [**176B**]{} 120 (1986); Erratum – ibid. [**178B**]{} 457 (1986) ;\
Cabibbo N, Farrar G R [*Phys. Lett.*]{} [**110B**]{} 107 (1982) Brahmachari B, Sarkar U, Sridhar K [*Mod. Phys. Lett.*]{} [**A8**]{} 3349 (1993);\
Hempfling R , hep-ph/9502201;\
Babu K S, Pati J C , hep-ph/9606215;\
Kolda C, March-Russell J , hep-ph/9609480 Langacker P, Polonsky N , hep-ph/9503214 Ghilencea D, Lanzagorta M, Ross G G , hep-ph/9707462 Gorbunov D S, Troitsky S V, to appear Amelino-Camelia G, Ghilencea D, Ross G G, hep-ph/9804437
[^1]: Talk presented by S.T. at the 10th International Seminar “Quarks-98”, Suzdal, Russia, May 18-24, 1998; to appear in the Proceedings.
[^2]: Recently, a model of direct mediation with the DSB group $SU(2)$ and two sets of $5+\bar
5$ messengers has been constructed [@Agashi]. We will not discuss this model here since analysis in this section is valid for multi-messenger models only.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This article uses relative symplectic cohomology, recently studied by the second author, to understand rigidity phenomena for compact subsets of symplectic manifolds. As an application, we consider a symplectic crossings divisor in a Calabi-Yau symplectic manifold $M$ whose complement is a Liouville manifold. We show that, for a carefully chosen Liouville structure, the skeleton as a subset of $M$ exhibits strong rigidity properties akin to super-heavy subsets of Entov-Polterovich.
Along the way, we expand the toolkit of relative symplectic cohomology by introducing products and units. We also develop what we call the contact Fukaya trick, concerning the behaviour of relative symplectic cohomology of subsets with contact type boundary under adding a Liouville collar.
address:
- '(DT) Harvard University'
- '(DT) University of California, Berkeley'
- '(UV) Stanford University'
author:
- Dmitry Tonkonog
- Umut Varolgunes
bibliography:
- 'Symp\_bib.bib'
title: 'Super-rigidity of certain skeleta using relative symplectic cohomology'
---
Introduction
============
Motivation
----------
Let $M$ be a symplectic manifold. Let us recall some standard definitions (see e.g. [@EnPo09]). Two subsets $K,S\subset M$ are called [*displaceable*]{} from each other if there exists a Hamiltonian diffeomorphism $\phi\co M\to M$ such that $\phi(K)\cap S=\emptyset$. Otherwise, $S$ is called non-displaceable from $K$ and vice versa. A subset $K\subset M$ is called (non-)displaceable if it is (non-)displaceable from itself. If $K,S\subset M$ are not displaceble from each other via a symplectomorphism, then we call them [*strongly non-displaceable*]{}.
Subsets $K,S\subset M$ are called [*stably displaceable*]{} if $K\times S^1$ is displaceable from $S\times S^1$ as subsets of the symplectic manifold $M\times T^*S^1$, with the product symplectic form, where $S^1$ represents the zero section. Otherwise, they are called [*stably non-displaceable*]{}, and $K\subset M$ is called stably non-displaceable if its stably non-displaceable from itself.
When the subset $K\subset M$ is a Lagrangian submanifold, Lagrangian Floer theory, see e.g. [@FO3Book], presents a powerful machine for detecting non-displaceability. Entov and Polterovich [@EnPo09] introduced the notions of heavy and superheavy sets (using their construction of a partial symplectic quasi-state via spectral invariants [@EP06]), giving access to non-displaceability questions for general compact subsets. These two techniques were combined in the seminal work [@FO3memoir].
In this paper, we approach these questions via another set of tools which uses recent advancements in Floer theory. Let us mention Corollory \[mainresult\] as our main Floer-theory-free result. Some of our readers might want to start reading from this statement and work their way backwards (and then hopefully forwards) in the introduction.
Symplectic-cohomological visibility
-----------------------------------
Let $(M,\omega)$ be a closed symplectic manifold. Recall the definitions of the Novikov ring $$\Lambda_{\ge 0}=\left\{\sum_{i=1}^\infty a_iT^{\omega_i}: a_i\in \mathbb{Q},\ \omega_i\in {{\mathbb R}}_{\ge 0},\ \omega_i\to+\infty \right\}$$ and the Novikov field $$\Lambda=\left\{\sum_{i=1}^\infty a_iT^{\omega_i}: a_i\in \mathbb{Q},\ \omega_i\in {{\mathbb R}},\ \omega_i\to+\infty \right\}.$$
We approach non-displaceability questions for general compact subsets using [*relative symplectic cohomology*]{}, recently studied by the second author [@VaThesis; @Va18]. The reader might benefit from taking a quick look at Section \[ssremindersh\] below, where we provided a summary of the construction. For a compact subset $K\subset M$, let $SH_M(K,\Lambda)$ be the $\Lambda$-vector space as defined in Equation (\[eqdefsh\]).
A compact subset $K$ is called $SH$**-invisible** if $SH_M(K,\Lambda)=0$, and $SH$**-visible** otherwise.
Below we list some properties of this notion. Recall that a [*compact domain*]{} in $M$ is a compact codimension-zero submanifold with boundary. For $K\subset M$, we say that a sequence of compact domains $D_1,D_2,\ldots $ [*approximate*]{} $K$, if $\bigcap D_i=K$ and $D_{i+1}\subset int(D_i)$ , for all $i\geq 1$.
\[thm:visib\_properties\] The following properties hold.
1. $M\subset M$ is $SH$-visible.
2. SH-visibility is invariant under symplectomorphisms.
3. If a compact subset $K$ is stably displaceable, then it is $SH$-invisible.
4. If a compact subset $K$ is $SH$-invisible, then any compact subset $K'\subset K$ is also $SH$-invisible.
5. If $K$ and $K'$ are $SH$-invisible compact subsets which can be approximated by compact domains $D_1,D_2,\ldots $ and $D_1',D_2',\ldots $ such that for evey $i\geq 1$, $\partial D_i\cap \partial D_i'=\emptyset$, then $K\cup K'$ is also $SH$-invisible.
6. Let $K$ be an $SH$-invisible compact domain. Then $K$ does not contain any tautologically unobstructed[^1] orientable Lagrangian submanifold (with a relative Pin structure) in its interior.
Items (1), (2), (3) and (5) were proved in [@VaThesis], where (5) is a special case of the Mayer-Vietoris property, for which a more polished reference is [@Va18] (specifically Section 4.3). The proof of (4) and (6) is one of the main contributions of the present paper.
\[conjecture\] Item (6) in Theorem \[thm:visib\_properties\] can be upgraded to Lagrangian submanifolds admitting bounding cochains (in the sense of [@FO3Book]) with nonzero self-Floer cohomology.
The main piece missing in proving this conjecture is a systematic treatment of the full Hamiltonian isotopy invariance package (including all higher homotopy coherences between continuation maps and the required bounds on the quantitative contributions of Floer solutions via topological energy) for the Floer theory of Lagrangians equipped with bounding cochains. Such a treatment is possible but is outside of the scope of this paper. We expect every statement we make about tautologically obstructed Lagrangians to be also true for Floer theoretically essential Lagrangians.
A compact subset $K$ is called $SH$**-full** if every compact set contained in its complement $M\setminus K$ is $SH$-invisible.
Below is a direct corollary of Theorem \[thm:visib\_properties\].
\[cor:diagram\_vis\_and\_full\] The following implications are true. $$\begin{array}{cccc}
(i)&
K \text{ is $SH$-visible }&\Rightarrow& K\text{ is stably non-displaceable from itself};
\vspace{0.4em}
\\
(ii)&
K \text{ is $SH$-full }&\Rightarrow& K\text{ is strongly non-displaceable~from any $SH$-visible subset};
\vspace{0.4em}
\\
(iii)&
K \text{ is $SH$-full }&\Rightarrow&
\begin{array}{c}
K\text{ is strongly non-displaceable~from any}\\
\text{tautologically unobstructed Lagrangian submanifold.}
\end{array}
\end{array}$$
We shall refer to the items from Theorem \[thm:visib\_properties\] simply by their numbers. The first implication is equivalent to (3).
Let us prove the second implication. Suppose $K,S\subset M$ are compact subsets, $K$ is $SH$-full and $S$ is $SH$-visible. Suppose $K,S$ are displaceable by a symplectomorphism; by (2), we may assume that they are actually disjoint. This is a contradiction.
Let us prove the third implication. Let $K$ be an $SH$-full subset, $L$ a tautologically unobstructed Lagrangian submanifold, and assume $L$ is disjoint from $K$. Then the closure of some tubular neighbourhood $\overline{U(L)}$ is still disjoint from $K$. By (ii), $\overline{U(L)}$ is $SH$-invisible. This contradicts (6).
The existence of such notions as $SH$-fullness, and $SH$-visibility is not surprising. Their behavior is similar to the notions of superheaviness, and heaviness (resp.) of Entov-Polterovich [@EnPo09]. We hope the precise relationship between the two frameworks will be explained in a collective effort.
Entov-Polterovich in fact define heaviness and superheaviness with respect to any idempotent in quantum cohomology. Since our main application is to Calabi-Yau manifolds[^2] in this paper, where the only possible idempotent is the unit, we only provide a brief remark about the analogous construction in our framework. A priori, we can define for any ideal $I$ of $SH_M(M,\Lambda)=QH(M,\Lambda)$, the notions of $SH$-invisibility, $SH$-visibility and $SH$-fullness with respect to $I$, e.g. $K$ is $SH$-invisible for $I$, if the submodule $I\cdot SH_M(K)=0$ and so on. Here we are using that $SH_M(K,\Lambda)$ is naturally a $SH_M(M,\Lambda)$-module of course.
Nevertheless, let us show that a principal ideal of $QH^{even}(M,\Lambda)$ generated by an idempotent does have a special role in our story. First, note that using the techniques of Sections \[sec:prod\_unit\], it follows that the connecting homomorphisms in the Mayer-Vietoris sequence of [@Va18] are in fact $QH(M,\Lambda)$-module homomorphisms.
Let us use only the module structures over the even part of the quantum cohomology from now on (by restricting scalars). Let $I$ be an ideal inside the even part of $QH(M,\Lambda)$. Then, we can multiply the Mayer-Vietoris sequence by this ideal (as a diagram of $QH^{even}(M,\Lambda)$-modules), and ask when the result is still an exact sequence. It follows from elementary homological algebra that this is the case if $QH^{even}(M,\Lambda)/I$ is flat over $QH^{even}(M,\Lambda)$, in other words if $I$ is pure (see [@stacks-project [Section 04PQ](https://stacks.math.columbia.edu/tag/04PQ)]). Since $I$ is clearly finitely generated, this is equivalent to $I$ being generated by an idempotent [@stacks-project [Lemma 05KK](https://stacks.math.columbia.edu/tag/05KK)].
In particular, for a given idempotent $a$, the vector spaces $a\cdot SH_M(K,\Lambda)$ satisfy the same Mayer-Vietoris property. We say $K$ is $a$-$SH$-invisible if $SH_M(K,\Lambda)$ is annihilated by $a$, and define $a$-$SH$-visible and $a$-$SH$-full as before. For example, Corollary \[cor:full\_is\_vis\] can be generalized to any $a$.
We expect that $SH$-fullness implies $SH$-visibility, but we can only show a slightly weaker statement.
\[cor:full\_is\_vis\] If $K$ is $SH$-full, then any compact domain that contains $K$ in its interior is $SH$-visible.
Consider a compact domain $D$ containing $K$. Then the compact domains $D$ and $\overline{M\setminus D}$ satisfy the conditions of Theorem \[thm:visib\_properties\], Item (5). Since the second set is $SH$-invisible, the first one is $SH$-visible by Item (1).
A simple closed curve in $S^2$ is $SH$-full if it divides $M$ into two disks of equal areas, using items (3) and (5) of Theorem \[thm:visib\_properties\], and $SH$-invisible otherwise. A tubular neighborhood of a non-contractible simple closed curve in the two torus is $SH$-visible, by item (6) of Theorem \[thm:visib\_properties\], but it is not $SH$-full (e.g. by item (iii) of Corollary \[cor:diagram\_vis\_and\_full\]).
Trivial Liouville cobordisms
----------------------------
A **neck** in $M$ is a symplectic embedding of a trivial (compact) Liouville cobordism $$\Sigma\times [1 -\alpha,s+\alpha]\subset M$$ for some $s>1$ and $\alpha>0$. Here $\Sigma$ is a closed manifold equipped with a contact form, and $[1 -\alpha,s+\alpha]\times \Sigma$ is a subset of its symplectization. The coordinate $$r\in [1-\alpha,s+\alpha]$$ is the exponential of the Liouville coordinate. That is, if $\rho$ is the Lioville coordinate such that ${\mathcal L}_{\partial/\partial \rho}\omega=\omega$, then $r=e^\rho$. We, in addition, assume that the hypersurface $\Sigma\times\{1\}$ is separating in $M$. One last requirement is that the periodic orbits of the Reeb vector field on $\Sigma$ should be all transversely non-degenerate.
Given a seperating contact hypersurface $\Sigma$ in $M$, we can talk about its convex and concave fillings which are both compact domains with boundary $\Sigma$.
If $\Sigma\times [1 -\alpha,s+\alpha]\subset M$ is a neck, we can make the convex filling $D$ of $\Sigma\times \{1\}$ larger by adding $\Sigma\times [1,s]\subset M$ to it and hence making it the convex filling $\tilde{D}$ of $\Sigma\times \{s\}$. There is a similar statement for the concave fillings. We are interested in the question: when is the restriction map $SH^*_M(\tilde{D},\Lambda)\to SH^*_M(D,\Lambda)$ an isomorphism?
Suppose $c_1(M)=0$. A contact hypersurface $(\Sigma,\theta)$ is called **index bounded** if all of its Reeb orbits are contractible inside $M$, and for any integer $k$, the periods of the Reeb orbits of Conley-Zehnder index $k$ are bounded above and below. We say that a neck $\Sigma\times [1 -\alpha,s+\alpha]\subset M$ is index bounded if $\Sigma\times \{1\}$ is index bounded. Similarly, we call a Liouville subdomain of $M$ index bounded if its boundary is index bounded.
\[prop:index\_bd\_collar\_invt\] Assume that $c_1(M)=0$, $\Sigma\times [1 -\alpha,s+\alpha]\subset M$ is an index bounded neck and $W\subset M$ is either the convex filling of $\Sigma\times \{1\}$ or the concave filling of $\Sigma\times \{s\}$. Then, there exists an isomorphism $$SH^*_M(W\cup \Sigma\times [1 ,s],\Lambda)\to SH^*_M(W,\Lambda).$$
The isomorphism is explicitly constructed using what we call the contact Fukaya trick. Using the argument in Lemma 4.1.1 of [@VaThesis] we can actually show that this is isomorphism is given by the restriction map as we had initially asked for. Proving this would make Section \[secfuk\] even more technical and this strengthening does not help us in the paper.
Using the well-known index computations for the Reeb orbits of contact boundaries of ellipsoids in $\mathbb{C}^n$ (e.g. Equation (2-6) from [@gutt]), we immediately obtain the following corollary.
Suppose $c_1(M)=0$. Let us take disjoint embeddings of symplectic ellipsoids in $\mathbb{C}^n$ into $M^{2n}$. Then we obtain that their union $D$ is $SH-$invisible and hence that the closure of the complement of $D$ is $SH-$full.
This is far from the best statement we could prove, but it makes the point, namely that here the images of the ellipsoids do not have to be displaceable, but they are $SH$-invisible. We could prove the statement of the corollary, with the same ease, for convex or concave toric domains $\mathbb{C}^n$ (see Sections 2.2 and 2.3 of [@gutt]). The statement is also true for convex domains in $\mathbb{C}^n$ but this requires a slightly modified argument using the fact that convex domains are dynamically convex, rather than index bounded (which might also be true but we do not know it in general). We do not spell this out because the symplectic consequences of $SH$-fullness in this case was already covered by Ishikawa’s superheaviness result from [@ishikawa].
Our next step is to discuss skeleta in symplectic manifolds. In the present paper they will be the main source of examples to which we will apply our non-displaceability results.
Giroux divisors
---------------
We refer the reader to [@tehrani; @mclean] for the notion of an [*SC divisor*]{} $D=\bigcup D_i$ in a symplectic manifold $(M,\omega)$. Briefly, this is a union of cleanly intersecting codimension two symplectic submanifolds $D_i$, such that the intersection orientations coincide with the symplectic orientations on all intersection strata.
As repopularized by McLean, a consequence of the $h$-principle for open symplectic embeddings of codimension two is the following proposition.
\[prop:D\_Stably\_disp\] Let $D$ be an SC divisor in $M$. Then, $D$ is stably displaceable.
To compare, it follows from Theorem \[thm:visib\_properties\] that $M\subset M$ is not stably displaceable. One may be interested in the intermediate behaviour: which neighborhoods of $D$ are stably displaceable? This is a hard question; for example, it is not known whether a disk inside a two-torus enclosing more than half of the total area is stably displaceable.
We introduce the following definition.
\[defgir\] A **Giroux divisor** $D=\bigcup D_i$ is an SC divisor with the property that there exist integers $w_i>0$ and a real number $c>0$ such that $$\sum w_i[D_i]=c\cdot PD[\omega]\ \in\ H_2(M).$$
Below is a structural result about the complements of Giroux divisors. The result follows from the work of McLean [@mclean], along with a construction we learned from [@Gi17]. A sketch proof is given in the Appendix.
\[prop:giroux\_skeleton\] Let $D\subset M$ be a Giroux divisor, then there exists a Liouville subdomain $W\subset M\setminus D$ such that:
- $\overline{M\setminus W}$ is stably displaceable,
- $\overline{M\setminus W}$ deformation retracts onto $D$.
Moreover, if $c_1(M)=0$, one can choose $W$ to be index bounded.
\[rem:giroux\_kahler\] Suppose $M$ is a complex projective variety and $D$ is an ample normal crossings divisor. Equip $M$ with the Kähler structure given by the canonical section of $\mathcal{O}(D)$ vanishing at $D$. Then, $D$ becomes an example of a Giroux divisor.
The complement of $D$ in $M$ is exhausted by Weinstein domains which are index bounded if $D$ is smooth and $c_1(M)=0$. This is also true for Donaldson type divisors in general symplectic manifolds [@Gi17 Theorems 1, 2]. By Proposition \[prop:D\_Stably\_disp\], for a sufficiently large such domain, $M\setminus W$ is stably displaceable. So $W$ satisfies all conditions in Proposition \[prop:giroux\_skeleton\]. Also, see Example \[examplediskbundle\].
If $D$ is not smooth, then Proposition \[prop:giroux\_skeleton\] modifies the standard Weinstein structures to a Liouville deformation equivalent“nice” Liouville structure in the sense of McLean.
\[defsket\] Assume $c_1(M)=0$, and let $D\subset M$ be a Giroux divisor. By a **skeleton** of $M\setminus D$ we mean the skeleton of any index bounded Liouville subdomain $W\subset M\setminus D$ as in Proposition \[prop:giroux\_skeleton\].
\[examplediskbundle\] Assume that $D$ is smooth, and $[D]=PD([\omega])$. Then $$M=K\sqcup U_1(D)$$ where $K$ is a skeleton of $M\setminus D$ and $U_1(D)$ is an open unit disk symplectic bundle over $D$ as in Figure \[fig:compl\_d\], see Corollary 8 of [@Gi17], cf. [@Bi01] in the Kähler case. Then for $0<r_2<r_1<1$, $\hat W=M\setminus U_{r_2}$ is the result of neck attachment to $W=M\setminus U_{r_1}$.
![Decomposition of a symplectic manifold into a disk bundle over a smooth symplectic divisor dual to $[\omega]$, and a skeleton.[]{data-label="fig:compl_d"}](Fig_Compl_D)
Putting the above results together, we obtain the following super-rigidity result for skeleta.
\[th:sk\_superrigid\] $M$ is a closed symplectic manifold. Assume that $c_1(M)=0$, $D\subset M$ is a Giroux divisor as in Definition \[defgir\] and $K\subset M$ is a skeleton of $M\setminus D$ as in Definition \[defsket\]. Then, $K$ is $SH$-full inside $M$.
Let $W\subset M\setminus D$ be as in Proposition \[prop:giroux\_skeleton\], and $K$ the skeleton of $W$. Let $W_i$ be the image of $W$ under its time-$i$ negative Liouville flow; this gives the nested sequence $$W=W_0\supset W_1\supset W_2\ldots \supset W_i\supset\ldots\quad \supset K.$$ Recall that $\overline{M\setminus W}$ is stably displaceable. Hence $W$ is $SH$-full. For any $i\geq 1$, $\overline{M\setminus W_i}$ is the result of neck attachment to $\overline{M\setminus W}$, so by Proposition \[prop:index\_bd\_collar\_invt\], $\overline{M\setminus W_i}$ is $SH$-invisible. Using Theorem \[thm:visib\_properties\], Item (3) and the definition of the skeleton we obtain the result.
\[mainresult\] Under the conditions of Theorem \[th:sk\_superrigid\]:
- $K$ is stably non-displaceable from itself inside $M$;
- $K$ is strongly non-displaceable from any tautologically unobstructed Lagrangian submanifold $L
\subset M$.
This immediately follows from Theorem \[th:sk\_superrigid\] and Corollary \[cor:diagram\_vis\_and\_full\].
Assume that we are in the situation of Remark \[rem:giroux\_kahler\]: $M$ is a complex projective Calabi-Yau and $D$ is an ample normal crossings divisor. Let $K$ be a skeleton of $M\setminus D$. Our theorem implies that $K$ is stably non-displaceable in $M$. In particular, it is stably non-displaceable in $W$. Given the expected results on $h$-principles for symplectic embeddings of flexible Weinstein domains, it follows that $W$ is not flexible. Here we also use the fact that the product of a flexible Weinstein manifold with $T^*S^1$ is still flexible [@Eli17]. This remark also applies to Donaldson divisors.
Recall that Casals and Murphy [@CM19] found examples of smooth affine varieties which are flexible (but are not subcritical). It follows that they do not admit smooth Calabi-Yau compactifications by normal crossings divisors. We do not know if this can be proved using purely algebro-geometric techniques (for example by computing their log Kodaira dimensions). A simpler proof of a more general version of this fact follows from uniruledness considerations by Zhou [@zzhou], see Theorem 5.10.
We point out that there are examples of Giroux divisors and $W$, which are not even deformation equivalent to a Weinstein domain, see [@Gi17 Proposition 9] for an example. The nature of the skeleta in these cases is rather mysterious and we are not aware of any rigidity results for them in the literature. Our results apply to these cases equally well.
It is possible to prove non-trivial special cases of the second bullet point of Corollary \[mainresult\] by more elementary means. Assume that $M$ is a complex projective Calabi-Yau of real dimension $2n$ and $D$ is a smooth ample divisor, let $U=M-D$. Then it can be shown using weak and hard Lefschetz theorems, and the Gysin sequence that $H^n_{prim}(M)\to H^n(U)$ is injective. Note that here we are using $[\omega]:=PD[D]$ as our Kahler class, and recall that a deRham cohomology class $A\in H^n(M)$ is called primitive if $A\wedge [\omega]=0$.
Take a smooth oriented Lagrangian $L$ in $M$. Assume that it does not intersect the skeleton. This implies that the integral over $L$ of any closed compactly supported form in $U$ (extended by $0$ to $M$) is zero, since we can always find a cohomologous form supported arbitrarily close to the skeleton with the help of the Liouville flow.
Now let $\alpha$ be a Thom form for $L$ inside $M$, meaning for any closed $n$-form $\beta$ on $M$: $\int_L \beta=\int_M \alpha\wedge\beta$. It is easy to see that $[\alpha]$ is a primitive class using that if it were not, there would have had to be a closed form $\gamma$ such that $\int_M\alpha\wedge\omega\wedge\gamma\neq 0$ by Poincare duality (i.e. existence of Hodge star). This is impossible because $\omega$ vanishes identically on $L$.
Moreover, the image of $[\alpha]$ under $H^n_{prim}(M)\to H^n(U)$ is zero by Poincare duality for $U$. Namely, if it were not, we could find a compactly supported closed $n$-form on $U$ which would pair non-trivially with it. This would be a contradiction to our previous findings. The upshot is that $[L]=0$.
Therefore, a smooth homologically essential Lagrangian can not displaced from the skeleton by a symplectomorphism. In fact being a little more careful we could replace symplectomorphism with a homeomorphism that preserves $[\omega]$. We feel it is worth exploring how far such a proof (with no reference to $J$-holomorphic curves) can pushed. For example is there an extension to the case when $M$ is Calabi-Yau, but $D$ is allowed to have normal crossing singularities?
We expect that the techniques of Ishikawa from [@ishikawa] also can be used to prove Corollary \[mainresult\]. For the experts we mention that this would hinge upon some index computations for the constant periodic orbits “on the divisor". There would also be some technical details for allowing a finite number of non-constant periodic orbits (of multiples of a distance to the divisor type Hamiltonian) which live near the “boundary" of $K$ and has the right CZ index to affect the spectral number of the unit. We hope that an expert will check the validity of our expectation.
\[ex:two\_torus\] Let $M$ be the two-torus and $D$ a point. In this case, one can take a skeleton $K$ of $M\setminus D$ to be the union of a meridian and a longtitude. This is an example falling under Corollary \[mainresult\]. For a non-example, one can take $M=S^2$ and $D$ a point.
Let us also state the following theorem, which is easy to prove.
\[thm:higher\_genus\] Assume that $M$ is a closed surface of genus at least $1$ and let $K$ be the complement of a finite disjoint union of open subsets of $M$ all diffeomorphic to disks. Then $K$ is $SH$-full as a subset of $M$.
That such $K$ is superheavy is also a special case of the main result of [@ishikawa].
As Theorem \[thm:higher\_genus\] suggests (by taking $D$ a finite set of points, and $K\subset M$ any skeleton of $M\setminus D$), Corollary \[mainresult\] should be more general, in particular, should hold in many non-Calabi-Yau cases. On the other hand, its failure is in some sense even more interesting. We plan to explore both directions in future work.
Algebraic structures for relative symplectic cohomology
-------------------------------------------------------
To prove Items (4) and (6) from Theorem \[thm:visib\_properties\], we need some general structural results regarding relative symplectic cohomology.
First, we show that $SH_M(K;\Lambda)$ is actually a unital $\Lambda$-algebra, and the units are functorial under restriction maps. While this is a totally expected structure, setting up the unit in the context of relative symplectic cohomology is not that straightforward and requires technical care. Our solution is to define the unit via [*raised symplectic cohomology*]{}, a modification of the original construction which allows us to use strictly positive Hamiltonians as perturbations.
Second, let $L\subset M$ be a tautologically unobstructed Lagrangian submanifold satisfying technical conditions related to gradings and signs, and $K\subset M$ be a compact subset. We introduce relative Lagrangian (self-)Floer cohomology, $HF^*_M(L,K,\Lambda)$ which is also a unital $\Lambda$-algebra. We also introduce the closed-open map $${\mathcal C}{\mathcal O}\co SH^*_M(K,\Lambda)\to HF^*(L,K,\Lambda).$$ We shall prove that it is a map of $\Lambda$-vector spaces respecting units, but it is easy to modify our arguments and show that it is a ring map.
In this paper, we do not discuss most of the expected properties of the ring structures, as we have no use for them. Namely, we do not prove that our ring structures do not depend on the choices, that they are associative, or in the closed string case that they are commutative.
Relative Lagrangian (self-)Floer cohomology also satisfies the Mayer-Vietoris property. The relevant statement for this paper is the last item in Proposition \[prop:lag\_analog\_thesis\] below, but one can construct a Mayer-Vietoris sequence with the same level of generality as in [@Va18]. With this in mind, let us mention another consequence of $SH$-fullness.
Let $L\subset M$ be a tautologically unobstructed oriented Lagrangian submanifold with a relative Pin structure, and $K$ be $SH$-full. Let $D$ be any compact domain containing $K$, then the restriction map $HF(L,\Lambda)\to HF(L,D,\Lambda)$ is an isomorphism.
The proof is similar to the proof of Theorem \[thm:visib\_properties\] Item (6) given at the end of Section \[sslag\].
Structure of the paper {#structure-of-the-paper .unnumbered}
----------------------
In Section 2, we review the construction of relative symplectic cohomology and introduce its open string version. We reduce the proofs of properties (4) and (6) of Theorem \[thm:visib\_properties\] to the existence of units with special properties. In Section 3, we discuss some chain level algebra which will be used to put algebra structures and define units on $H(\widehat{tel}(\mathcal{C}))$, where $\mathcal{C}$ is a $1$-ray. In Section 4, we apply the contact Fukaya trick to prove Proposition \[prop:index\_bd\_collar\_invt\]. In Section 5, we define the unit in relative symplectic cohomology and show that it is preserved under restriction maps. In Section 6, we define the unit in relative Lagrangian cohomology and construct closed-open maps. Then, we show that units are preserved under restriction maps and closed-open string maps.
In Appendix A, we give a proof of Proposition \[prop:giroux\_skeleton\].
Acknowledgements {#acknowledgements .unnumbered}
----------------
We have drawn inspiration from the work of Mark McLean. We thank him for his comprehensive papers. We also thank Paul Seidel for giving us comments on an earlier version of the Introduction to the paper. Finally, we thank KIAS for giving us a chance to collaborate at their institution for a week at the end of 2019 Summer.
DT was partially supported by the Simons Foundation under grant \#385573, Simons Collaboration on Homological Mirror Symmetry.
Relative Floer theoretic invariants {#sec:rel_sh}
===================================
In this section we go over the construction of relative symplectic cohomology from [@Va18], and the Lagrangian Floer cohomology version which will also be used in this paper. We prove the missing Items (4) and (6) from Theorem \[thm:visib\_properties\], modulo important technical propositions which are later proved in Sections \[sec:prod\_unit\] and \[sec:lag\].
Reminder on relative symplectic cohomology {#ssremindersh}
------------------------------------------
Let $M$ be a closed symplectic manifold. Let us assume that $c_1(M)=0$ for simplicity, and also fix a trivialization of the canonical bundle of $M$. Note that this choice does not play a role in the grading of contractible orbits. We refer the reader to [@Va18] for the construction without the Calabi-Yau assumption, which requires virtual techniques.
Let $K\subset M$ be a compact subset. We call the following datum an acceleration datum for $K$:
- $H_1\leq H_2\leq\ldots$ a monotone sequence of non-degenerate one-periodic Hamiltonians $H_i\co M\times S^1\to {{\mathbb R}}$ cofinal among functions satisfying $H\mid_{S^1\times K}<0$. In other words, $$H_i(x,t)\xrightarrow[i\to+\infty]{}\begin{cases}
0,& x\in K,\\
+\infty,& x\notin K.
\end{cases}$$
- A monotone homotopy of Hamiltonians $H_{i,i+1}:[i,i+1]\times M\times S^1\to\mathbb{R}$, for all $i$, which is equal to $H_i$ and $H_{i+1}$ in a neighborhood of the corresponding end points.
One can combine an acceleration datum into a single family of time-dependent Hamiltonians $H_s\co M\times S^1\to {{\mathbb R}}$, $s\in {{\mathbb R}}_{\geq 1}$. We also fix a Morse function on $[0,1]$ with critical values at the end points once and for all, which turns a $[0,1]$-family of Hamiltonians to a $(-\infty,\infty)$-family which is then used to write down the Floer equations.
Given an acceleration datum and a choice of a generic time dependent almost complex structure $J$, Hamiltonian Floer theory provides a $1$-ray of Floer chain complexes over $\Lambda_{\geq 0}$, called a [*Floer 1-ray*]{}: $$\begin{aligned}
\mathcal{C}(H_r,J):= CF^*(H_1)\to CF^*(H_2)\to\ldots \end{aligned}$$ Each $CF^*(H_i)$ is the Floer complex of $H_i$ over $\Lambda_{\ge 0}$, with the usual Floer differential. The horizontal arrows are Floer continuation maps defined using the monotone homotopies appearing in the acceleration datum. Recall that a holomorphic cylinder $u$ contributing to a Floer differential or a continuation map, does so with Novikov weight $$T^{E_{top}(u)}$$ where $$\label{eq:E_top}
E_{top}(u)=\int_{S^1}\gamma_{out}^* H_{out}\,dt-\int_{S^1}\gamma_{in}^* H_{in}\,dt+\omega(u),$$ $\gamma_{out}$, $\gamma_{in}$ are the asymptotic orbits of $u$, and $H_{out}$, $H_{in}$ are the Hamiltonians at the correspondind ends. (For Floer differentials, $H_{out}=H_{in}=H_i$ and for continuation maps, $H_{out}=H_{i+1}$, $H_{in}=H_i$ for some $i$.)
One defines the $\Lambda_{\geq 0}$-cochain complex $$tel(\mathcal{C}(H_s,J)) \text{ and } \widehat{tel}(\mathcal{C}(H_s,J)),$$ as in [@Va18]. see also Section \[sec:alg\] below. We stress that we always take the degreewise completion.
For two different choices of acceleration data for $K$ and almost complex structures, $H_s,J$ and $H_s',J'$, there is a canonical $\Lambda_{\geq 0}$-module map between $H^*(\widehat{tel}(\mathcal{C}(H_s,J)))$ and $H^*(\widehat{tel}(\mathcal{C}(H'_s,J')))$ defined using monotone continuation maps. Moreover, these maps are isomorphisms.
Hence, we define $$\begin{aligned}
SH_M^*(K):=H^*(\widehat{tel}(\mathcal{C}(H_s,J))).\end{aligned}$$
In what follows we will only be interested in the torsion-free part of $SH_M^*(K)$: $$\begin{aligned}
\label{eqdefsh}SH_M^*(K,\Lambda):=SH_M^*(K)\otimes\Lambda.\end{aligned}$$
Typically in Floer theory, one requires $J$ to be compatible with $\omega$. In fact, the theory works the same under the following weaker assumption: $J$ is tamed by $\omega$, and compatible in a neighbourhood of all 1-periodic orbits of all of the $H_i$. This greater flexibility will be later used in setting up the contact Fukaya trick.
When $M$ is Calabi-Yau, Hamiltonian Floer theory on $M$ can be set up using classical techniques, see Lecture 3 of [@sal], using generic choices of almost complex structures and Hamiltonians, which is what we are doing here. Note that in the Calabi-Yau case the genericity does not fail even if we use higher-dimensional parametric families of Floer equations. Indeed, the loci of Chern number zero $J$-holomorphic spheres are always codimension $4$ in $M$ times any parameter space. Therefore they generically do not interact with rigid or one-dimensional moduli spaces of Floer solutions. Note that here the key property is that there are no negative Chern number spheres, and therefore it is not possible to converge to configurations that involve Floer solutions that belong to higher dimensional families. The details of how transversality can be achieved by perturbing almost complex structures (for fixed Hamiltonian data) was worked out in Appendix B of [@mclean], which we will also be using here.
\[remanalytic\]Let us now explain a slightly different way of obtaining $SH_M^*(K)\otimes\Lambda$ using a more analytic language. Note that $\Lambda$ is a non-archimedean valued field. We define $CF^*(H_i,\Lambda)$ to be the (non-archimedean) normed/valued[^3] $\Lambda$-vector space freely generated by the $1$-periodic orbits which all have valuation $0$. Similarly, we define $tel(\mathcal{C}(H_s,J,\Lambda))$ as a valued $\Lambda$-vector space, which is notably infinite dimensional. Note that as $\Lambda$-vector spaces $tel(\mathcal{C}(H_s,J,\Lambda))$ and $tel(\mathcal{C}(H_s,J))\otimes \Lambda$ are canonically isomorphic. Now we take the completion of $tel(\mathcal{C}(H_s,J,\Lambda))$ and obtain a $\Lambda$-Banach space: $\widehat{tel}(\mathcal{C}(H_s,J,\Lambda))$. The homology of this complex is canonically isomorphic to $SH_M^*(K)\otimes\Lambda$.
Recall from [@Va18] that if $K\subset K'\subset M$, there is a canonical restriction map $$SH_M^*(K')\to SH_M^*(K).$$
Units and visibility
--------------------
We move on to new statements, no longer from [@Va18]. The next proposition will be proved in Section \[sec:prod\_unit\].
\[prop:unit\_sh\] For every compact set $K\subset M$, there is a distinguished element $1_K\in SH_M(K,\Lambda)$, called the unit, with the following properties.
- $SH_M(K,\Lambda)=0$ if and only if $1_K=0$.
- Restriction maps send units to units.
Although this proposition does not mention the existence of the product structure, we do actually prove its existence and unitality in Section \[sec:prod\_unit\].
We use Proposition \[prop:unit\_sh\]. Suppose $K'\subset K$, then the restriction map $SH^*_M(K,\Lambda)\to SH^*_M(K',\Lambda)$ sends $1_K$ to $1_{K'}$. Suppose that $SH^*_M(K,\Lambda)=0$, then $1_{K}=0$, so $1_{K'}=0$. Consequently, $SH^*_M(K',\Lambda)=0$.
Relative Lagrangian Floer theory {#sslag}
--------------------------------
Let $L\in M$ be an oriented Lagrangian with a relative Pin structure. Assume that there is a compatible (time-independent) almost complex structure $J_L$ such that $L$ does not bound any $J_L$-holomorphic discs. We want to define the relative Lagrangian Floer homology of $(L,J_L)$ for any compact subset $K\in M$: $$HF(L,J_L,K).$$ This will be a $\mathbb{Z}/2\mathbb{Z}$-graded $\Lambda$-vector space. The construction is very similar to the definition of relative symplectic cohomology, so we will be brief. We are using the results of [@Sei15] and [@mclean] here.
For a Hamiltonian $H:[0,1]\times M\to \mathbb{R}$ such that $\phi_H^1(L)$ is transverse to $L$, and a generic $[0,1]$-dependent almost complex structure $J:=\{J_t\}_{t\in [0,1]}$ with $J_t(x)=J_L(x)$ for every $x\in L$ and $t\in [0,1]$, we obtain $$CF^*(L,,J_L,H,J,\Lambda_{\geq 0}):=(\bigoplus_{\text{H-chords}}\Lambda_{\geq 0},d_{Fl}),$$ which is a chain complex over $\Lambda_{\geq 0}$ generated by the $1$-chords of $H$, and the differential counts Floer solutions $u: \mathbb{R}\times [0,1]\to M$ with boundary mapping to $L$ with weights $$\label{eq:E_top_lag}
E_{top}(u)=\int_{[0,1]}\gamma_{out}^* H_{out}\,dt-\int_{[0,1]}\gamma_{in}^* H_{in}\,dt+\omega(u).$$
For a monotone homotopy $H:[0,1]\times [0,1]\times M\to \mathbb{R}$ with $H|_0=H_0$ and $H|_1=H_1$, and a generic $[0,1]_t\times [0,1]_s$-dependent almost complex structure $J:\{J_{s,t}\}_{s,t\in [0,1]}$ with $J_{s,t}(x)=J_L(x)$ for every $x\in L$ and $s,t\in [0,1]$ and $J_{0,t}$ and $J_{1,t}$ generic as in the previous paragraph, we obtain a chain map $$CF^*(L,J_L,H_0,J_{0,t},\Lambda_{\geq 0})\to CF^*(L,J_L,H_1,J_{1,t},\Lambda_{\geq 0}).$$ A generic homotopy rel endpoints of such data gives rise to a chain homotopy as usual. We can then define the $1$-ray: $$\mathcal{C}:=CF^*(L,H_1,J,\Lambda_{\geq 0})\to CF^*(L,H_2,J,\Lambda_{\geq 0})\to \ldots\to ,$$ and define $$CF^*(L,H_s,J):=\widehat{tel}(\mathcal{C}).$$
Standard arguments show that $H(CF^*(L,H_s,J))$ is invariant under the choices of $H_s$ and $J$. Hence, we defined our $HF^*(L,J_L,K).$ As before, we want to throw away the torsion part of this, and define $$HF^*(L,J_L,K,\Lambda):=HF^*(L,J_L,K)\otimes\Lambda.$$
We will also drop the $J_L$ from the notation from now on, and declare that $L$ is tautologically unobstructed Lagrangian means that it has a specified $J_L$ implicit in writing $L$.
\[prop:lag\_analog\_thesis\]
1. There are canonical restriction maps for $K\subset K'$: $$HF^*(L,K')\to HF^*(L,K).$$
2. $HF^*(L,M,\Lambda)$ is isomorphic to Lagrangian Floer homology of $L$, which is isomorphic to $H^*(L,\Lambda)$ as a $\Lambda$-vector space.
3. If $L$ lies in the complement of $K$ in $M$, then $HF^*(L,K)=0$.
4. Let $K_1$ and $K_2$ be compact domains with equal boundaries such that $K_1\cup K_2=M$, then we have a Mayer-Vietoris sequence: $$\begin{tikzcd}
HF^*(L)\ar[r]&HF^*(L,K_1)\oplus HF^*(L,K_2)\ar[dl]\\HF^*(L,K_1\cap K_2)\ar[u]
\end{tikzcd}$$
The first two statements are straightforward. For the third statement, note that we can choose an acceleration data for $K$, so that all the $1$-chords of $L$ lie outside $K$, and moreover the value of the Hamiltonian $H_n$ at the chords of $H_n$ is approximately $n$. Now by the “adiabatic” argument in [@VaThesis], we obtain a uniform lower bound on the topological energies of all possible continuation maps for slowed down acceleration data, which proves that none of the generators survive in the completion. For the last one, note that the proof from [@Va18] applies verbatim here as one simply replace the $1$-periodic orbits in Lemma 4.1.1 from [@Va18] with $1$-chords on $L$, and the acceleration data that is constructed there would also satisfy this modified requirement.
\[prop:unit\_hf\] There is a distinguished element $1_{K,L}\in HF^*(L,K,\Lambda)$, called the unit, with the following properties.
- $HF^*(L,K,\Lambda)=0$ if and only if $1_{K,L}=0$.
- Restriction maps send units to units.
We need one more piece of information, which uses the existence relative closed-open string maps with good properties.
\[prop:zero\_unit\_hf\_sh\] For any $K,L$ as above, if $1_K=0\in SH^*_M(K,\Lambda)$, then $1_{K,L}=0\in HF^*(L,K,\Lambda)$.
The proofs of Propositions \[prop:unit\_hf\] and \[prop:zero\_unit\_hf\_sh\] are given in Section \[sec:lag\].
Suppose $L\subset M$ be such a Lagrangian submanifold, in particular $L\subset int(K)$.
By definition, we have that $SH^*_M(K,\Lambda)=0$. First, we claim that $HF^*(L,K,\Lambda)=0$. This follows by the unitality trick: we have that $1_K=0$ by Proposition \[prop:unit\_sh\], hence $1_{K,L}=0$ by Proposition \[prop:zero\_unit\_hf\_sh\], so $HF^*(L,K,\Lambda)=0$ by Proposition \[prop:unit\_hf\].
Let $N$ be the compact domain that is the closure of $M\setminus K$. By Proposition \[prop:lag\_analog\_thesis\], $HF^*(L,N,\Lambda)=0$.
Note that since units are preserved under restriction maps, we also have that $HF^*(L,K\cap N,\Lambda)=0$. Now the Lagrangian Mayer-Vietoris sequence from Proposition \[prop:lag\_analog\_thesis\] implies that $HF^*(L)=0$, which is a contradiction.
Chain-level algebra {#sec:alg}
===================
This section sets up some algebraic preliminaries used later. The first subsection reminds the basic notions from [@Va18]; the rest contains more technical material which will be necessary in Sections \[sec:prod\_unit\] and \[sec:lag\].
Rays, completion, telescope
---------------------------
This subsection reminds the algebraic setup from [@Va18]; we assume the reader is familiar with this reference.
Let $Ch_{\Lambda_{\geq 0}}$ be the category of $\mathbb{Z}$- or $\mathbb{Z}/2\mathbb{Z}$- graded chain complexes over the Novikov ring $\Lambda_{\geq 0}$. A 1-ray ${\mathcal C}$ is the following diagram in $Ch_{\Lambda_{\geq 0}}$, infinite to the right: $${\mathcal C}=C_1\xrightarrow{c_1} C_2\xrightarrow{c_2} C_3\xrightarrow{c_3} \ldots$$ Here each $C_i$ is a chain complex over $\Lambda_{\geq 0}$, and each $c_i$ is a chain map. Let $1-ray$ be the category of $1$-rays of with underlying modules assumed to be free. Morphisms in this category are given by maps of $1$-rays, and composition of morphisms is defined by composing the squares in the finite direction.
In $1-ray$ we have a notion of two morphisms being [*equivalent*]{}, defined by the existence of a homotopy of maps of $1$-rays.
In $1-ray$, there special morphisms called [*weak compression morphisms*]{}. We call a morphism between $f\co {\mathcal C}\to {\mathcal C}'$ between $1$-rays a *weak equivalence* if there exist morphisms $g\co {\mathcal C}'\to{\mathcal C}$ and $g'\co {\mathcal C}\to{\mathcal C}'$ such that $f\circ g'$ and $g\circ f$ are weak compression morphisms.
The telescope construction provides a functor $$tel: 1-ray\to Ch_{\Lambda_{\geq 0}}.$$ We also have the degree-wise completion functor $$\widehat{\cdot}:Ch_{\Lambda_{\geq 0}}\to Ch_{\Lambda_{\geq 0}}.$$ Composing the telescope and the completion functor, we obtain the completed telescope functor $$\widehat{tel}: 1-ray\to Ch_{\Lambda_{\geq 0}}.$$
Note that $tel$ sends equivalent morphisms to homotopy equivalent chain maps. Moreover, weak equivalences are sent to quasi-isomorphisms. The same statements hold true after completion as well, noting that quasi-isomorphisms stay quasi-isomorphisms after completion whenever the underlying modules are free.
Filtered direct limits and strictification
------------------------------------------
Another way to express the data of a $1$-ray is the following. Let $N$ be the category with objects positive integers, and precisely $1$ morphism from $n$ to $m$, whenever $n\leq m$, and no other morphisms. Then a functor $N\to Ch_{\Lambda_{\geq 0}}$ is precisely the same data as a $1$-ray. This also gives rise to a functor $N\to Mod_{\Lambda_{\geq 0}}$, and a self natural transformation of this functor given by the differentials.
With this in mind, given a $1$-ray $\mathcal{C}=C_1\to C_2\to C_3\to\ldots$, we can also define a chain complex $$\lim{(\mathcal{C})},$$ which is obtained by applying the standard construction of filtered direct limits of modules to the corresponding functor $N\to Mod_{\Lambda_{\geq 0}}$ and its self natural transformation.
Let us call a map of $1$-rays $\mathcal{C}\to \mathcal{C}'$ **strict** if for all $i\geq 1$, $C_i\to C'_{i+1}[1]$ (i.e. the homotopies) are identically zero. Let us denote the corresponding subcategory of $1-ray$ by $st-1-ray$. Note that $\lim$ defines a functor $$\lim: st-1-ray\to Ch_{\Lambda_{\geq 0}}.$$
We define the $1$-ray $$F(\mathcal{C})=F^1(tel(\mathcal{C}))\to F^2(tel(\mathcal{C}))\to F^3(tel(\mathcal{C}))\to\ldots,$$ where $$F^n(tel(\mathcal{C}))=\left(\bigoplus_{i\in [1,n-1]}C_i[1]\oplus C_i\right)\oplus C_n$$ with the differential depicted below (for $n=3$, the general form is clear) $$\begin{aligned}
\xymatrix{
C_1\ar@{>}@(ul,ur)^{d } &C_2\ar@{>}@(ul,ur)^{d} &C_3\ar@{>}@(ul,ur)^{d}\\
C_1[1]\ar@{>}@(dl,dr)_{-d} \ar[u]^{\text{id}}\ar[ur]^{f_1} &C_2[1]\ar@{>}@(dl,dr)_{-d} \ar[u]^{\text{id}}\ar[ur]^{f_2}&}\end{aligned}$$ and the maps are the canonical inclusion maps. Note that by definition $\lim{(F(\mathcal{C}))}$ is equal to $tel(\mathcal{C})$.
Then by the discussion of functioriality of cones and telescopes as in [@Va18], it is clear that we can extend $F$ to a functor $$F:1-ray\to st-1-ray.$$ We call this the **strictification** functor.
Let us also note that there exists canonical commutative diagrams $$\begin{aligned}
\label{c2estrictification}
\xymatrix{
F^n(tel(\mathcal{C})) \ar[d]\ar[r] &F^{n+1}(tel(\mathcal{C}))\ar[d] \\ C_n \ar[r] &C_{n+1}},\end{aligned}$$ where the vertical arrows are quasi-isomorphisms given by the direct sum of maps $C_i\to C_n$, $i\in [1,n]$ and the zero maps $C_i[1]\to C_n$, $i\in [1,n-1]$.
These diagrams give a morphism in $st-1-ray$, which induce quasi-isomorphisms after applying the $\lim$ functor.
Tensor product {#subsec:tensor}
--------------
The remainder of this section is no longer taken from [@Va18]. It will be used to set up the product and units on relative symplectic cohomology.
Let us define the tensor product $\mathcal{C}\otimes \mathcal{C}'$ of two one-rays $\mathcal{C}$ and $\mathcal{C}'$ as their slice-wise tensor product.That is, if $
{\mathcal C}=C_1\to C_2\to\ldots
$ and ${\mathcal C}'=C_1'\to C_2'\to\ldots$, then
$${\mathcal C}\otimes{\mathcal C}'=C_1\otimes C_1'\to C_2\otimes C_2'\to\ldots$$ with the obvious structure maps. Note that the differential of the tensor product of two cochain complexes involves the Koszul sign as usual. We shall use the following.
\[lem:tel\_tensor\] Let $\mathcal{D}:=\lim{(F^1(tel(\mathcal{C}))\otimes F^1(tel(\mathcal{C}')\ldots \to F^n(tel(\mathcal{C}))\otimes F^n(tel(\mathcal{C}')\to \ldots)}$. There exists a canonical zig-zag of quasi-isomorphisms $$tel(\mathcal{C}\otimes \mathcal{C}')\to \lim{({\mathcal C}\otimes{\mathcal C}')}\leftarrow \mathcal{D}\to tel(\mathcal{C})\otimes tel(\mathcal{C}').$$ Moreover, the maps $tel(\mathcal{C}\otimes \mathcal{C}')\to \lim{({\mathcal C}\otimes{\mathcal C}')}$ and $\mathcal{D}\to tel(\mathcal{C})\otimes tel(\mathcal{C}')$ stay quasi-isomorphisms after completion.
Note that we have canonical quasi-isomorphisms $$F^n(tel(\mathcal{C}\otimes \mathcal{C}'))\to C_n\otimes C_n'\leftarrow F^n(tel(\mathcal{C}))\otimes F^n(tel(\mathcal{C}')),$$ which are compatible for different $n$’s. Taking the direct limits of these maps we obtain a zig-zag of quasi-isomorphisms: $$tel(\mathcal{C}\otimes \mathcal{C}')\to \lim{({\mathcal C}\otimes{\mathcal C}')}\leftarrow \mathcal{D}.$$
Clearly, $$F^1(tel(\mathcal{C}))\otimes F^1(tel(\mathcal{C}')\to\ldots \to F^n(tel(\mathcal{C}))\otimes F^n(tel(\mathcal{C}')\to \ldots$$ is cofinal in the diagram $$\begin{aligned}
\xymatrix{
F^1(tel(\mathcal{C}))\otimes F^1(tel(\mathcal{C}') \ar[d]\ar[r] &F^1(tel(\mathcal{C}))\otimes F^2(tel(\mathcal{C}')\ar[d]\ar[r]& F^1(tel(\mathcal{C}))\otimes F^3(tel(\mathcal{C}')\ar[r]\ar[d]& \\ F^2(tel(\mathcal{C}))\otimes F^1(tel(\mathcal{C}')\ar[r]\ar[d] &F^2(tel(\mathcal{C}))\otimes F^2(tel(\mathcal{C}')\ar[r]\ar[d]&& \\ F^3(tel(\mathcal{C}))\otimes F^1(tel(\mathcal{C}')\ar[r]\ar[d]&&\ldots &\\&&&&}\end{aligned}$$
Moreover, the filtered direct limit of the diagram above is precisely $tel(\mathcal{C})\otimes tel(\mathcal{C}')$. Therefore we obtain a quasi-isomorphism $$\mathcal{D}\to tel(\mathcal{C})\otimes tel(\mathcal{C}'),$$which finishes the proof of the first statement.
That $tel(\mathcal{C}\otimes \mathcal{C}')\to \lim{({\mathcal C}\otimes{\mathcal C}')}$ stays a quasi-isomorphism after completion follows from Lemma 2.3.7 of [@Va18]. More straightforwardly, the same is true for $\mathcal{D}\to tel(\mathcal{C})\otimes tel(\mathcal{C}')$ because the involved modules are free.
Assume that we are given a morphism $\mathcal{C}\otimes \mathcal{C}'\to \mathcal{D}$ in $1-ray$. We can turn this into a map $$\begin{aligned}
H^*(\widehat{tel}(\mathcal{C}))\otimes H^*(\widehat{tel}(\mathcal{C}'))\to H^*(\widehat{tel}(\mathcal{D}))\end{aligned}$$ by composing the following natural maps.
1. $H^*(\widehat{tel}(\mathcal{C}))\otimes H^*(\widehat{tel}(\mathcal{C}'))\to H^*(\widehat{tel}(\mathcal{C})\otimes \widehat{tel}(\mathcal{C}'))$;
2. $H^*(\widehat{tel}(\mathcal{C})\otimes \widehat{tel}(\mathcal{C}'))\to H^*(\widehat{tel(\mathcal{C})\otimes tel(\mathcal{C}')})$;
3. $H^*(\widehat{tel(\mathcal{C})\otimes tel(\mathcal{C}')})\to H^*(\widehat{tel(\mathcal{C}\otimes \mathcal{C}')})$;
4. $H^*(\widehat{tel(\mathcal{C}\otimes \mathcal{C}')})\to H^*(\widehat{tel(\mathcal{D}}))$.
The third map comes from Lemma \[lem:tel\_tensor\], and the other maps are obvious.
\[lemmafunctorial\] Assume that we have a morphism $\mathcal{C}\to\tilde{\mathcal{C}}$ and an object $\mathcal{C}'$ in $1-ray$, then we get a morphism $\mathcal{C}\otimes \mathcal{C}'\to\tilde{\mathcal{C}}\otimes \mathcal{C}'$. In turn, if we have a morphism $\tilde{\mathcal{C}}\otimes \mathcal{C}'\to\mathcal{D}$, by composition we can obtain $\mathcal{C}\otimes \mathcal{C}'\to\mathcal{D}$.
Further assume that $\mathcal{C}\to\tilde{\mathcal{C}}$ is strict. Then, the diagram below commutes: $$\begin{tikzcd}
H^*(\widehat{tel}(\mathcal{C}))\otimes H^*(\widehat{tel}(\mathcal{C}'))
\ar[d]
\ar[r]
&
H^*(\widehat{tel}(\mathcal{D}))
\\
H^*(\widehat{tel}(\tilde{\mathcal{C})})\otimes H^*(\widehat{tel}(\mathcal{C}'))
\ar[ur]
\end{tikzcd}$$
A direct computation shows that the diagram below is a $1$-cube. $$\begin{tikzcd}
C_i\otimes C_i'
\ar[d, " g_i\otimes id"']
\ar[r, "f_i\otimes f_i'"]
\ar[dr,"h_i\otimes f_i'"]
&
C_{i+1}\otimes C_{i+1}'
\ar[d, " g_{i+1}\otimes id"]
\\
\tilde{C}_i\otimes C_i'
\ar[r, "\tilde{f}_i\otimes f_i'"']
&
\tilde{C}_{i+1}\otimes C_{i+1}'
\end{tikzcd}$$
This finishes the first part. The second part follows from showing the naturality of maps (1)-(3). The strictness assumption helps with (3) as $\lim$ is only functorial for strict morphisms of $1$-rays. Here we also use that the diagrams (\[c2estrictification\]) are compatible with strict morphisms.
Units {#subsec:algebra_units}
-----
Let $U$ be the $1$-ray $$\Lambda_{\geq 0}\to\Lambda_{\geq 0}\to\ldots ,$$ where all complexes $\Lambda_{\ge 0}$ are in degree $0$ and have zero differential, and the morphism maps are all the identity. Note that, as suggested by the notation, we are remembering the unit element of $\Lambda_{\ge 0}$ here. Both $H^*(tel(U))$ and $H^*(\widehat{tel}(U))$ are canonically isomorphic to $\Lambda_{\geq 0}$ in degree $0$, and zero otherwise.
\[defrealization\]Let $$f:\mathcal{C}\to \mathcal{D},\ U_{\mathcal{C}'}:U\to \mathcal{C}',\text{ and }p:\mathcal{C}'\otimes \mathcal{C}\to \mathcal{D}$$ be morphisms. If the composition $$\mathcal{C}= U \otimes \mathcal{C}\xrightarrow{} \mathcal{C}'\otimes \mathcal{C}\xrightarrow{p} \mathcal{D}$$ is equivalent to $f$, we call $U_{\mathcal{C}'}$ a realization of $f$ via $p$.
Let us analyze the situation more concretely. A map of $1$-rays $U_{\mathcal{C}'}:U\to \mathcal{C}'$ is equivalent to the following data.
- a closed element $u_i\in C_i^{0'}$ for each $i=1,2,\ldots$
- an element $p_{i+1}\in C_{i+1}^{-1'}$ such that the image of $u_i$ in $C_{i+1}'$ is equal to $u_{i+1}+dp_{i+1}$.
Let $u_{\mathcal{C}'}$ be the image of $1$ under the map $\Lambda_{\geq 0}=H^*(\widehat{tel}(U))\to H(\widehat{tel}(\mathcal{C}'))$. The following lemma is easy.
\[lemstrictunit\] Let $x\in C_1'$ be any element cohomologous to $u_1$, then the homology class of its image under the canonical chain map $C_1'\to \widehat{tel}(\mathcal{C}')$ is equal to $u_{\mathcal{C}'}$.
Now let us take a map of $1$-rays $U_{\mathcal{C}'}^*:U\to \mathcal{C}'$ equivalent to $U_{\mathcal{C}'}$. The data of $U_{\mathcal{C}'}^*$ is equivalent to elements $u_i^*,p_{i+1}^*$ as above. The homotopy between $U_{\mathcal{C}'}$ and $U_{\mathcal{C}'}^*$ is equivalent to the data of
- an element $h_i\in C_i^{-1'}$ such that $u_i-u_i^*=dh_i$ for each $i=1,2,\ldots$
- an element $q_{i+1}\in C_{i+1}^{-2'}$ such that the image of $h_i$ in $C_{i+1}'$ is equal to $h_{i+1}+dq_{i+1}$.
Finally we want to analyze the maps $$\mathcal{C}= U\otimes \mathcal{C}\xrightarrow{} \mathcal{C}'\otimes \mathcal{C}$$ for $U_{\mathcal{C}'}$ and $U_{\mathcal{C}'}^*$. More precisely, we want to show that these two maps are equivalent.
$$\begin{tikzcd}
C_i
\ar[d, " u_i\otimes", bend left]
\ar[d, " u_i^*\otimes"', bend right]
\ar[r, "f_i"'']
&
C_{i+1}
\ar[d, "u_{i+1}\otimes ", bend left]
\ar[d, " u_{i+1}^*\otimes"', bend right]
\\
C_i'\otimes C_i
\ar[r, "f_i'\otimes f_i"']
&
C_{i+1}'\otimes C_{i+1}
\end{tikzcd}$$
We define $C_i\to C_i\otimes C_i'[1] $ by $c\mapsto h_i\otimes c$. Let us check that this indeed is a homotopy: $$u_i\otimes c- u_i^*\otimes c=dh_i \otimes c = d(h_i\otimes c)+h_i\otimes dc$$.
We then recall that the two maps $C_i\to C_{i+1}\otimes C_{i'+1}[1] $ (corresponding to the two squares on the sides of the $2$-slit) are by definition $c\mapsto p_i\otimes f_i(c) $ and $c\mapsto p_i^*\otimes f_i(c)$.
Finally we define $C_i\to C_i\otimes C_i'[2] $ by $c\mapsto q_{i+1}\otimes f_i(c)$. Again a similar computation shows that this completes the diagram to a $3$-slit. Note that there are two sign changes in the computation, which result in the correct equation.
Hence, we proved the following.
Assume that $U_{\mathcal{C}'}$ is a realization of $f$ via $p$ as above as in Definition \[defrealization\]. Then any morphism that is equivalent to $U_{\mathcal{C}'}$ is also a realization of $f$ via $p$. Moreover, it gives rise to the same element $u_{{\mathcal C}'}\in H(\widehat{tel}(\mathcal{C}'))$.
The next lemma is the main result of this section.
\[lem:realis\_h\] Assume that $U_{\mathcal{C}'}$ is a realization of $f$ via $p$. Let $u_{\mathcal{C}'}$ be the image of $1$ under the map $\Lambda=H^*(\widehat{tel}(U))\to H(\widehat{tel}(\mathcal{C}'))$ as before. Then the map obtained by inputting $u_{\mathcal{C}'}$ in $$H(\widehat{tel}(\mathcal{C}'))\otimes H(\widehat{tel}(\mathcal{C}))\to H(\widehat{tel}(\mathcal{D}))$$ is equal to the map induced by $f$.
It is easy to see that $U_{\mathcal{C}'}$ is equivalent to a strict morphism. Therefore, using Lemma \[lemstrictunit\], we can assume that $U_{\mathcal{C}'}$ is strict.
Then using functoriality as in Lemma \[lemmafunctorial\], we can reduce to the case $C'=U$. We finish the proof if we can prove that the composition of all the maps below
- $H^*(\widehat{tel}(\mathcal{C}))\to H^*(\widehat{tel}(\mathcal{U}))\otimes H^*(\widehat{tel}(\mathcal{C}))$
- $H^*(\widehat{tel}(\mathcal{U}))\otimes H^*(\widehat{tel}(\mathcal{C}))\to H^*(\widehat{tel}(\mathcal{U})\otimes \widehat{tel}(\mathcal{C}))$;
- $H^*(\widehat{tel}(\mathcal{U})\otimes \widehat{tel}(\mathcal{C}))\to H^*(\widehat{tel(\mathcal{U})\otimes tel(\mathcal{C})})$;
- $H^*(\widehat{tel(\mathcal{U})\otimes tel(\mathcal{C})})\to H^*(\widehat{tel(\mathcal{U}\otimes \mathcal{C})})$;
is the same as the map $H^*(\widehat{tel}(\mathcal{C}))\to H^*(\widehat{tel(\mathcal{U}\otimes \mathcal{C})})$ induced by $\mathcal{C}= \mathcal{U}\otimes \mathcal{C}$. This is straightforward.
\[cor:real\_iso\] In the seting of Lemma \[lem:realis\_h\], let ${\mathcal C}={\mathcal C}'$. Assume that the map $$H(\widehat{tel}(\mathcal{C}))\otimes\Lambda\to H(\widehat{tel}({\mathcal{D})})\otimes \Lambda$$ induced by $f$ is an isomorphism. Then, $H(\widehat{tel}({\mathcal C}))\otimes\Lambda=0$ if and only if $u_{\mathcal{C}}$ is torsion, i.e. $u_{{\mathcal C}}=0$ in $H(\widehat{tel}(\mathcal{C}))\otimes\Lambda$.
Contact Fukaya trick {#secfuk}
====================
Necks and admissible functions
------------------------------
Let $(M,\omega)$ be a closed symplectic manifold and $\Sigma\times [1 -\alpha,s+\alpha]\subset M$ be a neck.
Let $U_1$ be the convex filling of $\Sigma=\Sigma\times \{1\}$ and define $$U_s=U_1\cup (\Sigma\times [1,s]).$$
Here is the plan of what follows. The idea is to fix a specially chosen diffeomorphism $g\co M\to M$ which is identity outside of the neck, and acts only on the coordinate $r$ strictly monotonically inside the neck, taking $U_1$ to $U_s$. The next step is to construct cofinal families of Hamiltonians $f_i$ for $U_1$ and $F_i$ for $U_s$, whose Hamiltonian vector fields are related by $g$. Also note that under $g$ the push forward of a compatible almost complex structure, which is cylindrical in the neck, is still tame for the original symplectic structure. We will also make sure that it is compatible near the $1$-periodic orbits. This construction requires care because $g$ is not a symplectomorphism.
If we use the above cofinal families and almost complex structures, then $g$ gives a bijection between the Floer solutions contributing to the Floer 1-rays for $U_1$ and $U_s$. However, the topological energies of a Floer cylinder and its $g$-image are different. This is the content of what we call the [*contact Fukaya trick*]{}, see Proposition \[prop:fukaya\_trick\]. In particular, $g$ does not necessarily give rise to a quasi-isomorphism of the corresponding completed telescopes. However, we show that it does give a quasi-isomorphism in the $c_1(M)=0$, index bounded case as claimed in Proposition \[prop:index\_bd\_collar\_invt\]. The “concave filling" case is analogous, so we focus on the convex filling case we just explained.
We begin with the diffeomorphism $g$. Consider a compactly supported diffeomorphism $$g(r)\co [1-\alpha, s+\alpha]\to [1-\alpha, s+\alpha],\quad g(1)=s,$$ with the following properties: $$\begin{cases}
g(r)\text{ is monotone onto } [1-\alpha,s-\epsilon],& r\in[1-\alpha,1-\epsilon]\\
g(r)=sr,& r\in[1-\epsilon,1]\\
g(r)\text{ is monotone onto }[s,
\tilde s(1+\epsilon)],&r\in [1,1+\epsilon]\\
g(r)=\tilde s r,& r\in[1+\epsilon,1+2\epsilon]\\
g(r)\text{ is monotone onto }[\tilde s(1+2\epsilon),s+\alpha],&r\in [1+2\epsilon, s+\alpha].
\end{cases}$$ Recall that $s$ was fixed above; $\tilde s$ and $\epsilon$ are free parameters, $\tilde s>s$. We fix $\epsilon<\min(\alpha,s)$ to be sufficiently small, and define $\tilde s$ by the equation $$\label{eq:tilde_s}
\epsilon=\tilde s (1+\epsilon)-s.$$ In other words, $\tilde s=(s+\epsilon)/(1+\epsilon)$. The reason for this choice will be seen later.
Let $\phi$ be the diffeomorphism of $M$ induced by $g$ on the neck, extended by the identity outside of the neck.
Let ${\mathcal R}$ be pullback of the Reeb vector field on $\Sigma$ under the projection $[1,s+\alpha]\times \Sigma\to \{1\}\times\Sigma$. Then for an arbitrary function $f(r)$ of the collar coordinate, its Hamiltonian vector field is $$\label{eq:Ham_f_neck}
X_{f(r)}=f'(r)\cdot {\mathcal R}.$$
Consider a $C^\infty$ function $f(r)$, $r\in[1-\alpha,s+\alpha]$, with the following properties: $$\begin{cases}
f(r)=cr+b,& r\in[1-\alpha,1-\epsilon]\\
f(r)\text { is monotone and has total increase $\delta$,}& r\in[1-\epsilon,1],\\
f(1)=0,\\
f(r)=Kr +c_K,&r\in [1,1+\epsilon]\\
f(r)\text { is monotone and has total increase $\delta$,}& r\in[1+\epsilon,1+2\epsilon]\\
f(r)=cr+B,&r\in [1+2\epsilon, s+\alpha].
\end{cases}$$ We think of $\delta$ and the slopes $c,K$ as free parameters, whereas the constants $b,c_K,B$ are determined by continuity. Total increase $\delta$ means, for instance in the first case, that $f(1)-f(1-\epsilon)=\delta$. We call such functions of the coordinate $r$ [*profile functions*]{}.
Note that for any given $\delta>0$, for sufficiently small $c$ one has: $$\label{eq:osc_collar}
f(1)-f(1-\alpha)<2\delta\quad\text{and}\quad
f(s+\alpha)-f(1+\epsilon)<2\delta.$$
Now fix some $\delta>0$. We call a time-dependent Hamiltonian $H\co M\to{{\mathbb R}}$ [*admissible with profile $f$*]{} if it:
- restricts to $f(r)$ on the neck $\Sigma\times [1 -\alpha,s+\alpha]$,
- has non-degenerate periodic orbits outside of the neck, for convenience all assumed to be constant.
Moreover, we demand that the profile $f(r)$ of an admissible Hamiltonian satisfies:
- $K$ is not a period of ${\mathcal R}$;
- $c$ is smaller than the periods of the periodic orbits of ${\mathcal R}$;
- if $f'(r)$ is equal to a period of ${\mathcal R}$, then $f''(r)\neq 0$;
- $(s-1+\alpha)c<c(1-\alpha)+b$ and $c(1-\alpha)+b+\epsilon>0$.
Matching profiles
------------------
Recall the diffeomorphism $\phi\co M\to M$ defined above, determined by the function $g(r)$. For a profile function $f(r)$, consider the pushforward of its Hamiltonian vector field under $\phi$: $$\label{eq:phi_X}
\phi_*X_{f(r)}=f'(g^{-1}(r))\cdot {\mathcal R}.$$
We claim that $\phi_*X_{f(r)}$ is a Hamiltonian vector field of the smooth function $F(r)$: $$\begin{cases}
F(r)=cr+\tilde b,& r\in[1-\alpha,s(1-\epsilon)]\\
F(r)=sf(r/s),& r\in[s(1-\epsilon),s]\\
F(r)=Kr+\tilde{c}_K,&r\in [s,\tilde s(1+\epsilon)]\\
F(r)=\tilde s f(r/\tilde s)+\tilde d,& r\in[\tilde s(1+\epsilon),\tilde s(1+2\epsilon)]\\
F(r)=cr+\tilde{B},&r\in [\tilde s(1+2\epsilon),s+\alpha].
\end{cases}$$ The constants $\tilde b,\tilde{c}_K, \tilde d,\tilde B$ are determined by continuity. See Figure \[fig:neck\_push\].
Clearly, $F(r)$ is determined up to an additive constant, and we have choosen the normalisation $F(s)=0$. This is seen from the fact that $f(1)=0$.
![A profile function $f$ (solid plot), and a matching function $F$ (dotted plot).[]{data-label="fig:neck_push"}](Fig_Neck_Push)
We claim that $$\label{eq:X_F}
\phi_*X_{f(r)}=X_{F(r)}.$$ Indeed, in view of (\[eq:phi\_X\]) and (\[eq:Ham\_f\_neck\]) this is equivalent to: $$F'(r)=f'(g^{-1}(r)),$$ which is easy to check. We say that this function $F(r)$ [*matches*]{} $f(r)$.
\[lem:match\_dif\_1\] Fix a constant $C>0$. Then for all $K>0$, and all sufficiently small $c,\delta>0$, the following holds. Suppose $f(r)$ is a profile with sufficiently small parameters $c,\delta$ and arbitrary $K$. Let $F(r)$ match $f$. It holds that: $$\label{eq:F_f_dif}
|F(1-\alpha)-f(1-\alpha)|<C,\quad |F(s+\alpha)-f(s+\alpha)|<C,$$ and $$\label{eq:F_f_curvy_region}
|F(g(r))-f(r)|<C\quad\text{for all}\quad\text{for}\quad r\in [1-\epsilon,1]\cup[1+\epsilon,1+2\epsilon].$$
The important point is that the above bounds do not depend on $K$ which can be arbitrarily large. Note that as $K\to+\infty$, $f(r)\to+\infty$ on $(1,s+\alpha]$ and $F(r)\to+\infty$ on $(s,s+\alpha]$.
Recall that $f(1)=F(s)=0$. By construction, $$\begin{array}{l}
f(1-\alpha)=-\delta-c(\alpha-\epsilon),\\ F(1-\alpha)=-\delta s-c(s(1-\epsilon)-(1-\alpha)).
\end{array}$$ Both values do not depend on $K$, and are small when $c,\delta$ are small.
Recall that according to (\[eq:tilde\_s\]), $\epsilon=\tilde s (1+\epsilon)-s$. This choice has the following crucial significance. By construction, $f$ has slope $K$ on the segment $[1,1+\epsilon]$ of length $\epsilon$, so it has total increase $K\epsilon$ on this segment. Again by construction, $F$ has slope $K$ on the segment $[s,\tilde s(1+\epsilon)]$, which is of length $\epsilon$ by (\[eq:tilde\_s\]). So
$$\begin{array}{l}
f(1+\epsilon)=K\epsilon,\\
F(\tilde s(1+\epsilon))=K\epsilon.
\end{array}$$ Continuing the computation, $$\begin{array}{l}
f(s+\alpha)=K\epsilon+\delta+c(s+\alpha-(1+\epsilon)),\\
F(s+\alpha)=K\epsilon +\delta\tilde s+c(s+\alpha-\tilde s(1+2\epsilon)).
\end{array}$$ We see that the difference is small, if $\delta,c$ are small. We have proved (\[eq:F\_f\_dif\]).
To prove (\[eq:F\_f\_curvy\_region\]), one easily checks: $$\begin{array}{l}
f([1-\epsilon,1])=[-\delta,0],\\
F(g([1-\epsilon,1]))= [-\delta s,0],
\end{array}$$ and $$\begin{array}{l}
f([1+\epsilon,1+2\epsilon])=[K\epsilon,K\epsilon+\delta],\\
F(g([1+\epsilon,1+2\epsilon]))\subset [K\epsilon, K\epsilon+\delta\tilde s].
\end{array}$$ Now (\[eq:F\_f\_curvy\_region\]) easily follows.
Let $H$ be an admissible Hamiltonian with profile $f$, and let $F$ match $f$. We say that a Hamiltonian $\tilde H$ on $M$ [*matches $H$*]{} if:
- $\tilde H$ restricts to $F(r)$ on the neck $\Sigma\times [1 -\alpha,s+\alpha]$, and
- on each of the two connected components of the complement of the neck, $H-\tilde H$ is constant. Explicitly, $$\label{eq:H_h_dif_col_1}
\tilde H=H+F(1-\alpha)-f(1-\alpha) \quad \text{on}\quad U_{1-\alpha},$$ and $$\label{eq:H_h_dif_col_2}
\tilde H=H+F(s+\alpha)-f(s+\alpha) \quad \text{on}\quad X\setminus U_{s+\alpha}.$$
\[lem:phi\_H\] If $\tilde H$ matches $H$, then $ \phi_*X_{H}=X_{\tilde H}$.
On the neck, the claim follows from the analogous fact about $F(r)$ and $f(r)$, see Equation (\[eq:phi\_X\]). Away from the neck, $\phi$ is the identity and $\tilde H-H$ is locally constant, so the claim holds there too.
Cofinal families
----------------
Consider a neck $\Sigma\times [1 -\alpha,s+\alpha]\subset M$. As above, fix $\epsilon>0$, a diffeomorphism $g$ of $[1-\alpha,s+\alpha]$ and the induced diffeomorphism $\phi$ of $M$. Fix monotonically decreasing sequences of positive numbers $\delta_i\to 0$, $c_i\to 0$, $\beta_i\to 0$, and a monotonically increasing sequence $K_i\to +\infty$. For each $i$, choose a profile function $f_i(r)$ with the given parameters $\epsilon ,\delta_i,c_i,K_i$. We require that $f_i(r)\le f_{i+1}(r)$ for all $r\in[1-\alpha,s+\alpha]$. Note that $$f_i(r)-\beta_i \to
\begin{cases}
0,& r\le 1\\
+\infty,&r>1
\end{cases}$$ The same is true about $f_i(r)$ without the summand $
\beta_i$. We have subtracted $\beta_i$ to make sure that $f_i(r)-\beta_i<0$ for $r\le 1$ as required in the definition of a cofinal family (recall that $f_i(1)=0$). Let $F_i(r)$ match $f_i(r)$, then $$F_i(r)-\beta_i \to
\begin{cases}
0,& r\le s\\
+\infty,&r>s
\end{cases}.$$ Let $H_i$ be an admissible Hamiltonian with profile $f_i$, and $\tilde H_i$ be an admissible Hamiltonian with profile $F_i$. Then $\{H_i\}$ are a cofinal family for $U_1\subset M$, and $\{\tilde H_i\}$ are a cofinal family for $U_s\subset M$.
Let $J$ be a tame almost complex structures which is cylindrical on $$\label{eq:sigma_times_int_1}
\Sigma\times \left([1-\alpha,s+\alpha]\right),$$ and compatible with $\omega$ on
Then $\phi_*J$ is tame, and moreover, it is compatible on $$\label{eq:sigma_times_int_2}
\Sigma\times \left([s(1-\epsilon)s,s]\cup[\tilde{s}(1+\epsilon),\tilde{s}(1+2\epsilon)]\right).$$
Since $\phi$ is identity outside of the neck, it suffices to check this on the neck. First, recall that $J$ is cylindrical and $\phi$ only depends on the radial coordinate on the region (\[eq:sigma\_times\_int\_1\]), so $\phi_*J$ is tame on the image of that region.
Second, recall that $\phi$ scales the neck coordinate $r$ by a fixed factor on each of the two components of the region (\[eq:sigma\_times\_int\_2\]). Because the neck coordinate $r$ is the exponential of the Liouville coordinate, $\phi$ acts by a shift of the Liouville coordinate on (\[eq:sigma\_times\_int\_2\]). Hence, $\phi$ is conformally symplectic onto the image of (\[eq:sigma\_times\_int\_2\]). So $\phi_*J$ is still cylindrical on that region, in particular it is compatible.
Let $f_i$ and $F_i$ be as above, choose Hamiltonians $\{H_i\}$ with profile $f_i$, and let $\tilde H_i$ match $H_i$ (so that $\tilde H_i$ has profile $F_i$). We can arrange that $H_i$ is a cofinal family for $U_1$, and $\tilde H_i$ is a cofinal family for $U_s$.
Next, choose monotone homotopies between the $f_i$ within the space of profile functions. Extend them to monotone homotopies between the $H_i$. Consider the matching monotone homotopies between the $F_i$, and extend them to monotone homotopies between the $\tilde H_i$. This produces Floer 1-rays ${\mathcal C}$ (for $U_1$ using $\{H_i\}$) and ${\mathcal C}'$ (for $U_s$ using $\{\tilde H_i\}$).
\[lem:phi\_bij\] The diffeomorphism $\phi$ takes solutions of the Floer equation contributing to $CF^*(H_i,J_i)$ bijectively to solutions for $CF^*(\tilde H_i,\phi_*J_i)$. The analogous holds for continuation maps.
In other words, $\phi$ takes all solutions contributing to the structure maps of ${\mathcal C}$ bijectively onto Floer solutions contributing to the structure maps of ${\mathcal C}'$.
This is a reformulation of Lemma \[lem:phi\_H\]. The analogous property is true for matching monotone homotopies constructed using matching profile functions.
Contact Fukaya trick {#contact-fukaya-trick}
--------------------
Consider cofinal families $H_i$ for $U_1$ and $\tilde H_i$ for $U_s$, with Floer 1-rays ${\mathcal C}$ and ${\mathcal C}'$ as above. We will translate Lemma \[lem:phi\_bij\] to a comparison between ${\mathcal C}$ and ${\mathcal C}'$.
A generator of ${\mathcal C}$ is a 1-periodic orbit of $H_i$ for some $i$. These generators can be of two types:
- constant orbits in $U_{1-\alpha}$ and $X\setminus U_{s+\alpha}$,
- non-constant orbits in $\Sigma\times ([1-\epsilon,1]\cup [1+\epsilon,1+2\epsilon])$.
Generators of ${\mathcal C}'$ are 1-periodic orbits of $\tilde H_i$ for some $i$; they are of two types as above, with the difference that non-constant orbits belong to $\Sigma\times ([s(1-\epsilon),s]\cup [\tilde s(1+\epsilon),\tilde s(1+2\epsilon)])$.
By Lemma \[lem:phi\_H\], $\phi$ induces a bijection between the generators of ${\mathcal C}$ and ${\mathcal C}'$, $\gamma\mapsto \phi(\gamma)$. Lemma \[lem:phi\_bij\] states that there is a bijection at the level of structure maps; however, note that the topological energies of $u$ and $\phi(u)$ are different. The proposition below expresses this difference.
\[prop:fukaya\_trick\] Consider the natural bijection between the generators of ${\mathcal C}$ and ${\mathcal C}'$ discussed above: $\gamma\mapsto \tilde\gamma\coloneqq \phi(\gamma)$. It upgrades to a strict isomorphism between the chain complexes $$\Phi\co tel({\mathcal C})\otimes \Lambda \to tel({\mathcal C}')\otimes \Lambda,$$ which has the following form: $$\Phi\co \gamma\mapsto T^{\Delta(\gamma)}\cdot \tilde\gamma,$$ for both generators of $ tel({\mathcal C})$ corresponding to $\gamma$. The numbers $\Delta(\gamma)\in {{\mathbb R}}$ are defined in (\[eq:delta\_gamma\]) below, and satisfy: $$\label{eq:delta_O1}
\Delta(\gamma)=\omega(C(\gamma,\tilde \gamma))+O(1).$$ Here the first summand for $\gamma$ a non-constant orbit is the area of the trivial cylinder $C(\gamma,\tilde \gamma)\subset X$ spanning $\gamma$ and $\tilde \gamma$, and contained in the neck; for $\gamma$ a constant orbit it is zero. The second summand $O(1)$ is a quantity depending on $\gamma$ which is uniformly bounded in absolute value across all generators.
Consider a Floer solution $u$ contributing to ${\mathcal C}$ with asymptotic orbits $\gamma_{in},\gamma_{out}$. By Lemma \[lem:phi\_bij\], $\phi(u)$ is a Floer solution contributing to ${\mathcal C}'$. Recall the definition of topological energy (\[eq:E\_top\]): $$E_{top}(u)=\int_{S^1}\gamma_{out}^* H_{out}\,dt-\int_{S^1}\gamma_{in}^* H_{in}\,dt+\omega(u).$$ Define, for a $1$-periodic orbit $\gamma$ of $H_i$: $$\label{eq:delta_gamma}
\Delta(\gamma)=\left(\int_{S^1} \tilde \gamma^* \tilde H_i-\gamma^*H_i\right)+\omega(C(\gamma,\tilde\gamma)).$$ Then for a Floer solution $u$ as above one has: $$E_{top}(\phi(u))-E_{top}(u)=\Delta(\gamma_{out})-\Delta(\gamma_{in}),$$ because of the obvious diffeotopy from $\phi$ to identity and Stokes theorem. Recall that a Floer solution $u$ contributes to ${\mathcal C}$ as follows: $$\gamma_{in}\mapsto T^{E_{top}(u)}\gamma_{out},$$ while $\phi(u)$ contributes to ${\mathcal C}'$ as follows: $$\tilde \gamma_{in}\mapsto T^{E_{top}(u)+\Delta(\gamma_{out})-\Delta(\gamma_{in})}\tilde \gamma_{out},$$ or equivalently $$T^{\Delta(\gamma_{in})}\tilde \gamma_{in}\mapsto T^{\Delta(\gamma_{out})}T^{E_{top}(u)}\tilde \gamma_{out}.$$ This can be rewritten as $$\Phi(\tilde \gamma_{in})\mapsto T^{E_{top(u)}}\Phi(\tilde \gamma_{out}).$$ So $\Phi$ is a chain isomorphism. To show (\[eq:delta\_O1\]), in view of (\[eq:delta\_gamma\]) it remains to show that there exists a constant $C$ such that: $$\label{eq:diff_H}
\left| \int_{S^1} \tilde \gamma^* \tilde H_i-\gamma^*H_i\right|<C$$ for all $1$-periodic orbits $\gamma$ of $H_i$, for all $i$. We claim that this follows from Lemma \[lem:match\_dif\_1\].
To see why, assume first that $\gamma$ is a constant orbit, then it belongs to $U_\alpha$ or $E_{s+\alpha}$, and $\tilde \gamma=\gamma$. The above difference is simply $\tilde H(\gamma)-H(\gamma)$, where $\gamma$ is seen as a point in $X$. Then (\[eq:diff\_H\]) follows from (\[eq:F\_f\_dif\]), (\[eq:H\_h\_dif\_col\_1\]) and (\[eq:H\_h\_dif\_col\_2\]).
Now assume $\gamma$ is a non-constant orbit, then $\gamma$ belongs to the neck where $H_i=f_i$, $\tilde H_i=F_i$. By construction of the $f_i$, $\gamma$ belongs to the region of the neck where $r\in [1-\epsilon,1]\cup[1+\epsilon,1+2\epsilon]$. Here (\[eq:F\_f\_curvy\_region\]) applies to guarantee (\[eq:diff\_H\]).
Index boundedness
-----------------
Let us recall the setup of this section. We started with a neck $\Sigma\times [1-\alpha,s+\alpha]\subset M$ in a symplectic manifold $M$. We denoted by $U_1$ the domain bounded by $\Sigma\times\{1\}$, and took $U_s=U_1\cup (\Sigma\times [1,s])$. We constructed Floer 1-rays ${\mathcal C},{\mathcal C}'$ which compute $SH^*_M(U_1)$ and $SH^*_M(U_s)$, in particular: $$H^*(\widehat{tel}({\mathcal C}))\otimes\Lambda=SH^*_M(U_1, \Lambda),\quad H^*(\widehat{tel}({\mathcal C}'))\otimes\Lambda=SH^*_M(U_s,\Lambda).$$ We explained how to carefully choose these Floer 1-rays so that they satisfy Proposition \[prop:fukaya\_trick\].
At this point we would like to remind the reader Remark \[remanalytic\]. With that in mind, in general, Proposition \[prop:fukaya\_trick\] does not give any straightforward relationship between $SH^*_M(U_1,\Lambda)$ and $SH^*_M(U_s,\Lambda)$. This is because the map $\gamma\mapsto T^{\Delta(\gamma)}\cdot \tilde\gamma$ and/or its inverse may not extend to the completions, as they might not be continuous, or equivalently bounded.
We now assume that the neck is index bounded, i.e. for every integer $i$, the periods of the periodic orbits of $\Sigma$ of Conley-Zehnder index $i$ have a uniform upper and lower bound.
In our current notation, the statement is equivalent to $$SH^*_M(U_1,\Lambda)\cong SH^*_M(U_s,\Lambda).$$
Note that there exists a constant $a>0$ such that for any non-constant $1$-periodic orbit $\gamma$, the quantity $\omega(C(\gamma,\tilde\gamma))$ is bounded by $a$ times the period of the Reeb orbit corresponding to $\gamma$ in $\Sigma$.
Therefore, the index bounded condition implies that the map from Proposition \[prop:fukaya\_trick\] and its inverse are in fact both continuous, and therefore they can be extended to the completions. Since this extension is a functorial operation, they are still chain maps and strict inverses of each other, which finishes the proof.
Product and unit via raised cohomology {#sec:prod_unit}
======================================
Raised cohomology and perturbation spaces
-----------------------------------------
\[dfn:raised\_sh\] Let $K\subset M$ be a compact subset of a symplectic manifold and $\epsilon>0$. Define the raised symplectic cohomology $$SH^*_M(K,\epsilon)$$ analogously to $SH^*_M(K)$ using increasing families $H_i$ which are cofinal for $\{H:H|_{S^1\times K}<\epsilon\}$, equivalently, $$H_i(x)\xrightarrow[i\to+\infty]{}\begin{cases}
\epsilon,& x\in K,\\
+\infty,& x\notin K.
\end{cases}$$ Moreover, we will always assume $H_i>0$ everywhere on $X$ for such cofinal families; note that this would not be possible with the usual definition of $SH^*_M(K)$.
Let ${\mathcal M}_{k;1}$ be the moduli space of genus zero Riemann surfaces with $k$ input punctures and one output puncture, modulo automorphisms. We assume that they are equipped with compatible choices of cylindrical ends. We will only be interested in the cases $k=0,1,2$, when ${\mathcal M}_{k;1}$ consists of a unique element, a curve $C\in {\mathcal M}_{k;1}$.
Fix weight parameters $${\mathbf w}=(w^i>0)_{i=1,\ldots, k}\cup (w^0>0),$$ one for each input and output (so that in the case $k=0$ we only have $w^0$). The superscripts are used as indices here. Also fix $${\mathbf H}=(H^i)_{i=1,\ldots, k}\cup (H),$$ a collection of time-dependent everywhere positive Hamiltonians on $X$, one for each input and output.
Define $$\label{eq:S}
{\mathcal S}_{k;1}({\mathbf w},{\mathbf H})=\{
(\alpha, H)\}
\subset \Omega^1(C)\times C^\infty(C\times X)$$ to be the subset consisting of all $\alpha,H$ with the following properties: $$\left\{
\begin{array}{l}
d\alpha\ge 0,\ H\ge 0,\\
d_{C\times X}(H\alpha)\mid_{C\times\{x\}}\ge 0, \text{ for every }x\in X\\
\alpha\equiv w^idt\text { on cylindrical ends,}\\
H\equiv H^i(t)\text { on cylindrical ends.}\\
\end{array}
\right.$$
The first two requirements ensure that the topological energies of the solutions of the Floer equation $$(du-X_H\otimes \alpha)^{0,1}=0,$$ with respect to any domain dependent compatible almost complex structure, are non-negative. Let us demonstrate this.
Let $(\Sigma, j) \in {\mathcal M}_{k;1}$, and $(\alpha, H)\in {\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$ for some weights ${\mathbf w}$ and time-dependent Hamiltonians ${\mathbf H}$. Consider the trivial bundle $\pi:M\times \Sigma\to \Sigma$ with the connection $2$-form $$\Omega:=pr_M^*\omega+d_{C\times X}(H\alpha).$$ Recall the one to one correspondance between $\Sigma$-dependent almost complex structures $J_M$ on $M$ and almost complex structures $J^{tot}$ on $M\times\Sigma$ which are split for the horizontal subbundles defined by $\Omega$, and which make $\pi$ a $(J^{tot},j)$-holomorphic map. Under this correspondence, solutions $u$ of the Floer equation $(du-X_H\times \alpha)^{0,1}$ for $J_M$ are in turn in one to one correspondance with $(j,J^{tot})$-holomorphic sections $\tilde{u}=(u,id)$ of $\pi$. Moreover, we have the equality: $$topE(u)=topE(\tilde{u}),$$ by definition.
We now claim that our assumption $d_{C\times X}(H\alpha)\mid_{C\times\{x\}}\geq 0$ implies that $\Omega$ and $J^{tot}$ form a semi-tame pair for any semi-tame $J_M$, namely that $\Omega(\cdot, J^{tot}\cdot)\geq 0$. One can easily reduce to showing this for horizontal vectors. A simple computation shows that for any $v\in T_p\Sigma$, its unique horizontal lift to $(p,x)$ is given by $$v+\theta(v)X_H(p,x)$$ via the splitting $T_{(p,x)}(M\times \Sigma) =T_xM\oplus T_p\Sigma.$ The desired result follows by plugging in all of these in the semi-tameness equation and using $d_{C\times X}(H\alpha)\mid_{C\times\{x\}}\geq 0$.
This implies that any $J^{tot}$-holomorphic curve inside $M\times\Sigma$ has non-negative geometric energy, which in turn can be shown to equal topological energy by the standard argument. The upshot is that is $u$ is a solution of the Floer equation, then $topE(u)\geq 0$, as desired.\
It will be convenient to also refer to the underlying space of 1-forms: $$\label{eq:Omega_space}
\Omega_{k;1}({\mathbf w})=\{\alpha\in \Omega^1(C):\ d\alpha\ge 0 \text{ on }C,\ \alpha\equiv w^idt\text{ on cylindrical ends}\}.$$ Clearly, there is a projection $$\pi\co {\mathcal S}_{k;1}({\mathbf w},{\mathbf H})\to \Omega_{k;1}({\mathbf w}).$$
We shall use the spaces ${\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$ to define (raised) symplectic cohomology, products, and units. Table \[tab:weights\] may be helpful to keep track of the definitions.
[|c|c|c|]{}
\
\
$k=1$ & $w^1=w^0=1$ & ${\mathcal S}_c({\mathbf H})$\
[|c|c|c|]{}\
$k=1$ & $w^1=1$, $w^0=2$ & ${\mathcal S}_{\tilde c}({\mathbf H})$\
[|c|c|c|]{}
\
$k=0$ & $w^0=1$ & ${\mathcal S}_u({\mathbf H})$\
[|c|c|c|]{}
\
$k=2$ & $w^1=w^2=1$, $w^0=2$ & ${\mathcal S}_p({\mathbf H})$\
There are natural gluing operations on these perturbations spaces. Their existence is important to ensure the properties of the operations we are going to define using these perturbations. For example, there are maps $$\#_\rho\co {\mathcal S}_c\sqcup {\mathcal S}_c\to {\mathcal S}_c$$ which glue perturbation data on the strips using gluing parameter $\rho\gg 0$ in two possible orders, whenever the Hamiltonians at the glued cylindrical ends match. As another example, there is a map $$\#_\rho\co {\mathcal S}_u\sqcup {\mathcal S}_p\to {\mathcal S}_c,$$ gluing the unit datum to the product datum at, say, the second input, which will be used to show that $-*_\epsilon 1_\epsilon$ is chain homotopic to $c_{\epsilon,2\epsilon}$. There are other natural gluing maps like this. Rather than writing them all down, we summarize the main point and do not mention it further: all gluing operations needed below can be performed staying within the spaces (\[eq:S\]) of perturbation data, unless explicitly mentioned otherwise. A full treatment would have introduced families of choices parametrized by certain manifolds with corners where near the boundaries existence of “gluing coordinates" is assumed. We would then frame contractibility as families defined on boundaries of these manifolds with corners admitting extensions to the interior. This is standard especially for our rudimentary needs with very explicit diagrams containing only homotopies of homotopies. We will omit further mention of this and only prove the main point, which is that our space of unbroken choices is contractible.
The first thing we need to understand is the conditions under which the spaces (\[eq:S\]) are non-empty and contractible. For two time-dependent Hamiltonians $H_1,H_2\co S^1\times X\to {{\mathbb R}}$, let us write $$\begin{array}{c}
H_1\succeq H_2 \textit{ if }
\exists\textit{ a time-independent } H^\textit{ref}\co X\to{{\mathbb R}}\textit{ s.~t.~}
\\
H_1(t,x)\ge H^\textit{ref}(x)\ge H_2(t,x) \textit{ for every }(t,x)\in S^1\times X.
\end{array}$$
\[lem:S\_nonempty\] Assume that $$w^0\ge \sum_{i=1}^k w^i\quad\textit{and}\quad H^0\succeq H^i\textit{ for each }i=1,\ldots, k.$$ (When $k=0$, the conditions say $w^0\ge 0$, $H^0\ge 0$.) Then $\Omega_{k;1}({\mathbf w})$ is non-empty. Furthermore, for each $\alpha\in \Omega^1_{k;1}({\mathbf w})$, the space $\pi^{-1}(\alpha)\subset {\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$ is non-empty.
![The Hamiltonian part of the Floer data from Lemma \[lem:S\_nonempty\] on the pair-of-pants.[]{data-label="fig:S_Nonempty"}](Fig_S_Nonempty)
It is well known that $\Omega_{k;1}({\mathbf w})$ is non-empty: there exists a 1-form $\alpha$ satisfying $d\alpha\ge 0$ and having the desired behaviour at the cylindrical ends. It remains to construct an $H$ such that $(\alpha,H)\in {\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$; see Fig. \[fig:S\_Nonempty\]. For this, consider a monotone homotopy from $H^i$ to $H^\textit{ref}$ on a cylindrical end near each input, and a monotone homotopy from $H^\textit{ref}$ to $H^0$ near the output. This defines $H$ on the union of three cylindrical ends. We assume that $\alpha$ is proportional to $dt$ on these cylindrical ends, so $d(H\otimes \alpha)\ge 0$ on them. Outside of these regions, set $H\equiv H^\textit{ref}$.
\[lem:S\_contr\] Whenever ${\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$ is non-empty, it is contractible.
Consider the projection $\pi\co {\mathcal S}_{k;1}({\mathbf w},{\mathbf H})\to \Omega_{k;1}({\mathbf w})$. The base space $\Omega_{k;1}({\mathbf w})$ is convex (whenever it is non-empty), hence contractible. Assuming the non-emptiness, each fibre of the projection is non-empty by Lemma \[lem:S\_nonempty\]. Next, each fibre is obviously convex, hence contractible. It follows that ${\mathcal S}_{k;1}({\mathbf w},{\mathbf H})$ is contractible.
\[lem:s\_c\_std\] The space ${\mathcal S}_c$ contains the usual continuation map data coming from monotone homotopies: namely, $\alpha=dt$ and $H\ge 0$ is non-decreasing in $s$.
Extrinsic continuation
----------------------
We assume that the reader is familiar with the algebra from [@VaThesis]; also see Section \[sec:alg\].
\[dfn:c\_e1e2\] Let $\epsilon_1<\epsilon_2$. Let $\{H_i^1\}$, $\{H_i^0\}$ be cofinal families computing $SH^*_M(K,\epsilon_1)$ resp. $SH^*_M(K,\epsilon_2)$, chosen so that $H_i^0\succeq H_i^1$ for each $i$. Consider two 1-rays from the diagram below. $$\begin{tikzcd}
\ldots\ar[r]&CF^*(H_{i-1}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i+1}^1)\ar[r]\ar[d]&\ldots\\
\ldots\ar[r]&CF^*(H_{i-1}^0)\ar[r]&CF^*(H_{i}^0)\ar[r]&CF^*(H_{i+1}^0)\ar[r]&\ldots
\end{tikzcd}$$ The top ray is a Floer 1-ray for $SH^*_M(K,\epsilon_1)$, and the bottom ray is a Floer 1-ray for $SH^*_M(K,\epsilon_2)$.
Construct a morphism between these 1-rays (i.e. the vertical and diagonal arrows in the above diagram) by counting Floer solutions on the cylinder with perturbation data chosen consistently from the spaces ${\mathcal S}_c({\mathbf H})$. We use ${\mathbf H}=(H^1_i,H^0_i)$ for the vertical arrows, and ${\mathbf H}=(H^1_i,H^0_{i+1})$ for the diagonal arrows. (Recall that we use 1-dimensional families of data for the latter, as usual [@VaThesis].)
The induced $\Lambda_{\ge 0}$ module map $$c_{\epsilon_1,\epsilon_2}\co SH^*_M(K,\epsilon_1)\to SH^*_M(K,\epsilon_2),$$ is called an extrinsic continuation map.
\[dfn:c\_e\] Let $\epsilon>0$. Let $\{H_i^1\}$, $\{H_i^0\}$ be cofinal families computing $SH^*_M(K)$ resp. $SH^*_M(K,\epsilon)$, chosen so that $H_i^0\geq H_i^1$ (the standard order). Note that the $H_i^1$ are not positive in this case. The $\Lambda_{\ge 0}$ module map $$c_{0,\epsilon_1}\co SH^*_M(K)\to SH^*_M(K,\epsilon)$$ is defined analogously to [@Va18], using continuation maps from monotone homotopies. (The perturbation data no longer live in the spaces ${\mathcal S}_c({\mathbf H})$, because the Hamiltonians are not everywhere positive.)
\[lem:c\_commut\] The maps from Definition \[dfn:c\_e1e2\] and \[dfn:c\_e\] are well-defined. For $0\le\epsilon_0<\epsilon_1<\epsilon_2$ there is a commutative diagram $$\begin{tikzcd}
SH^*_M(K,\epsilon_0)\arrow[r, "c_{\epsilon_0,\epsilon_1}"]\arrow[rd, "c_{\epsilon_0,\epsilon_2}"']
&SH^*_M(K,\epsilon_1)\ar[d, "c_{\epsilon_1,\epsilon_2}"]
\\
&SH^*_M(K,\epsilon_2),
\end{tikzcd}$$ where, if $\epsilon_0=0$, we formally put $SH^*_M(K,0)\coloneqq SH^*_M(K)$.
\[lem:c\_iso\] One has $$c_{0,\epsilon}=T^\epsilon\cdot f,$$ where $f$ is an isomorphism of $\Lambda_{\ge 0}$-modules.
Let $\{H^1_i\}$ be a cofinal family for $SH^*_M(K)$, then $\{H^0_i\coloneqq H^1_i+\epsilon\}$ is a cofinal family for $SH^*_M(K,\epsilon)$. For each $i$, choose a monotone homotopy from $H^1_i$ to $H^0_i$ which has the following form: $$H_i^c(s,t)=H^1_i(t)+\rho(s)$$ where $\rho(s)\co (-\infty,+\infty)\to [0,\epsilon]$ is a monotone function such that $$\rho(s)\equiv 0,\ s\ll 0,\quad \rho(s)\equiv \epsilon,\ s\gg 0.$$ Observe that the Hamiltonian vector field of $\rho(s)$ vanishes. All rigid solutions to this continuation map are constant, but contribute with topological energy $T^\epsilon$.
Let $H_{i,i+1}^c(s,t)$ be a monotone homotopy from $H_i^1$ to $H_{i+1}^1$, and $c_{i,i+1}$ the induced continuation map between the Floer complexes. We choose the homotopy from $H_i^0$ to $H_{i+1}^0$ (for the Floer 1-ray) to be $$H_{i,i+1}^c(s,t)+\epsilon$$ and the homotopy from $H_i^1$ to $H_{i+1}^0$ (for the extrinsic continuation) to be $$H_{i,i+1}^c(s,t)+\rho(s).$$ Floer solutions for the two above homotopies are precisely the same as for the unmodified homotopy $H_{i,i+1}^c(s,t)$. So, using the obvious identifications of the Floer complexes, the first continuation map is again $c_{i,i+1}$, and the second one is $T^\epsilon c_{i,i+1}.$ The diagram from Definition \[dfn:c\_e1e2\] becomes the following. $$\begin{tikzcd}
\ldots \ar[r]&CF^*(H_{i}^1)\ar[r, "c_{i,i+1}"]\ar[d, "T^\epsilon\cdot \mathrm{Id}"']\ar[dr, "h_i"]&CF^*(H_{i+1}^1)\ar[r]\ar[d, "T^\epsilon\cdot \mathrm{Id}"]&\ldots\\
\ldots
\ar[r]&CF^*(H_{i}^0)\ar[r, "c_{i,i+1}"]&CF^*(H_{i+1}^0)\ar[r]&\ldots
\end{tikzcd}$$ We can arrange for the homotopies $h_i$ to be identically zero here. We again use that changing Hamiltonians involved in the Floer equations by functions only depending on $s$ and $t$ do not change the equations. Hence we can construct a one parameter family of continuation map equations between the two broken homotopies from $H_i^1$ to $H_{i+1}^0$ by $s$-dependent Hamiltonians with the homotopy giving $c_{i,i+1}$ implanted at different $s$-levels. The ends of this family is what one would obtain by gluing the two broken data. Using that $c_{i,i+1}$’s were defined by regular Floer data, we obtain that this parametrized family is also regular (transversality is clearly satisfied). But, notice that solutions of different members of this family of equations differ by translations in the $s$-direction. Hence, there is no rigid solution in this parametrized problem, proving the claim.
The upshot is that this map of 1-rays induces $T^\epsilon\cdot \mathrm{Id}$ at the level of relative symplectic cohomology, using the obvious generator-wise identification of all Floer complexes and hence the telescopes.
It also holds that $c_{\epsilon_1,\epsilon_2}$ equals $T^{\epsilon_2-\epsilon_1}$ times an isomorphism. This follows from Lemmas \[lem:c\_commut\] and \[lem:c\_iso\]; alternatively, the proof of Lemma \[lem:c\_iso\] adapts.
Variant extrinsic continuation
------------------------------
We will now define an auxuliary map, $\tilde c_{\epsilon,2\epsilon}$, and show that it is chain homotopic to $c_{\epsilon,2\epsilon}$. We will later use $\tilde c_{\epsilon,2\epsilon}$ to prove the unitality property.
\[dfn:c\_tilde\] Let $\epsilon>0$. Let $\{H_i^1\}$, $\{H_i^0\}$ be two different cofinal families computing $SH^*_M(K,\epsilon)$, chosen so that $H_i^0\succeq H_i^1$. Consider the 1-rays from the diagram below. $$\begin{tikzcd}
\ldots\ar[r]&CF^*(H_{i-1}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i+1}^1)\ar[r]\ar[d]&\ldots\\
\ldots\ar[r]&CF^*(2H_{i-1}^0)\ar[r]&CF^*(2H_{i}^0)\ar[r]&CF^*(2H_{i+1}^0)\ar[r]&\ldots
\end{tikzcd}$$ The top ray is a Floer 1-ray for $SH^*(K,\epsilon_1)$, and the bottom is one for $SH^*(K,2\epsilon_2)$. For the curves defining the bottom ray, we can assume that they solve Floer’s equation using $\alpha=2dt$ as the 1-form and $H_i^0$ as the Hamiltonians, rather than $\alpha=dt$ and $2H_i^0$ as the Hamiltonians. This is tautological since $(2dt)\otimes H=dt\otimes (2H)$.
Construct a morphism between these 1-rays (i.e. the vertical and diagonal arrows) by counting Floer solutions on the cylinder with perturbation data chosen consistently from the spaces ${\mathcal S}_{\tilde c}({\mathbf H})$. That is, we use ${\mathbf H}=(H^1_i,H^0_i)$ for the vertical arrows, and ${\mathbf H}=(H^1_i,H^0_{i+1})$ for the diagonal arrows.
The induced $\Lambda_{\ge 0}$ module map $$\tilde c_{\epsilon,2\epsilon}\co SH^*_M(K,\epsilon)\to SH^*_M(K,2\epsilon),$$ is called variant continuation map.
\[lem:c\_versions\] The maps $c_{\epsilon,2\epsilon}$ and $\tilde c_{\epsilon,2\epsilon}$ agree at the cohomology level.
Consider the spaces $${\mathcal S}_{\tilde c}= {\mathcal S}_{1;1}({\mathbf w},{\mathbf H}),\quad {\mathbf w}=(1,2),\quad {\mathbf H}=(H^1,H^0)$$ used in the definition of $\tilde c_{\epsilon,2\epsilon}$ and $${\mathcal S}_c\coloneqq {\mathcal S}_{1;1}({\mathbf w},{\mathbf H}),\quad {\mathbf w}=(1,1),\quad {\mathbf H}=(H^1,2H^0)$$ used in the definition of $c_{\epsilon,2\epsilon}$. Here $H^1$ would be an element of the cofinal family $\{H^1_i\}$ and $H^0\succeq H^1$ an element of the cofinal family $\{H^0_i\}$. But we can consider $H^1,H^0$ as two arbitrary Hamiltonians, $H^0\succeq H^1$.
It is enough to show that we can interpolate between perturbation data from ${\mathcal S}_{\tilde c}$ and ${\mathcal S}_{c}$ within the space $$\{\alpha\otimes H:\ d(\alpha\otimes H)\ge 0,\ \alpha\otimes H\textit{ fixed near the punctures }\},$$ to define a homotopy between $c_{\epsilon,2\epsilon}$ and $\tilde c_{\epsilon,2\epsilon}$ defined over the Novikov ring. We will now present a way of constructing such interpolations. (The rest of the argument is formal, and similar to the proof of Lemma \[lem:c\_commut\].)
Consider the space $${\mathcal C}(H^1,H^0)=
\{\rho(s)\cdot dt\otimes H(s)\}\subset\Omega^1(C)\otimes C^{\infty}(C,\mathbb{R}).$$ where: $$\begin{array}{l}
H(s)\co X\times S^1\times {{\mathbb R}}\to {{\mathbb R}},\quad H\ge 0,\quad \partial_sH\ge 0,\\
H(s)\equiv H^1 \textit{ for }s\ll 0,\quad H(s)\equiv H^0 \textit{ for }s\gg 0,\\
\rho(s)\co {{\mathbb R}}\to[1,2],\quad \partial_s\rho\ge 0,\\
\rho(s)\equiv 1 \textit{ for }s\ll 0,\quad \rho(s)\equiv 2 \textit{ for }s\gg 0.
\end{array}$$
We claim that this space is non-empty and contractible for every $H^0\succeq H^1$. Indeed, it is the quotient under the multiplicative group ${{\mathbb R}}_{>0}$ of the space of pairs $(\rho(s),H(s))$ where $\rho(s)$, $H(s)$ satisfying the above properties. This space of pairs is obviously non-empty and contractible when $H^0\succeq H^1$, so its quotient by ${{\mathbb R}}_{>0}$ is also contractible.
Consider the quotient map $$q:(\alpha,H)\mapsto \alpha\otimes H,$$ which has already appeared above implicitly. Clearly, $${\mathcal C}\subset q({\mathcal S}_1)\cap q({\mathcal S}_2),$$ since one can write $(\rho(s)\cdot dt)\otimes H(s)=dt \otimes (\rho(s)\cdot H(s))$. Moreover, the spaces ${\mathcal C}(H^1,H^0)$ are compatible with gluing Floer solutions and continuation maps. So they are suitable interpolation spaces.
Product and unit on raised cohomology
-------------------------------------
\[dfn:prod\_e\] Let $\epsilon>0$. Let $\{H_i^1\}$, $\{H_i^2\}$, $\{H_i^0\}$ be three different cofinal families computing $SH^*_M(K,\epsilon)$, chosen so that $H_i^0\succeq H_i^1,H^2_i$ for all $i$. They give rise to three Floer 1-rays. Consider two 1-rays from the diagram below. $$\begin{tikzcd}
\ldots \ar[r]\ar[dr]&CF^*(H_{i-1}^1)\otimes CF^*(H_{i-1}^2)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i}^1)\otimes CF^*(H_{i}^2)\ar[r]\ar[d]\ar[dr]& \ldots\\
\ldots \ar[r]& CF^*(2H_{i-1}^0)\ar[r]&CF^*(2H_{i}^0)\ar[r]& \ldots
\end{tikzcd}$$ The bottom ray is the Floer 1-ray corresponding to $\{2H_i^0\}$. The top ray is the slice-wise tensor product of the Floer 1-rays corresponding to $\{H_i^1\}$, $\{H_i^2\}$.
Construct a morphism between these rays by counting Floer solutions on pairs-of-pants where the perturbation data are consistently chosen from ${\mathcal S}_{p}({\mathbf H})$. As in Definition \[dfn:c\_tilde\], we use the fact that $(2dt)\otimes H=dt\otimes (2H)$ to treat the output orbits of the pairs-of-pants as elements of $CF^*(2H_i^0)$.
Using this morphism, induce a $\Lambda_{\ge 0}$ module map called the product: $$*_{\epsilon}\co SH^*_M(K,\epsilon)\otimes SH^*_M(K,\epsilon) \to SH^*_M(K,2 \epsilon)$$ as explained in Subsection \[subsec:tensor\].
In this paper, we do not bother with proving that this product is independent of the choices, as we have no use for it.
\[dfn:unit\_e\] Let $\epsilon>0$, and $\{H_i^0\}$, $H_i^0\ge 0$, be a cofinal family computing $SH^*_M(K,\epsilon)$.
Consider the 1-ray $U$ from Subsection \[subsec:algebra\_units\] and the Floer 1-ray for $\{H_i^0\}$. They are the rows of the diagram below. $$\begin{tikzcd}
\ldots\ar[r]&\Lambda_{\ge 0}\ar[r]\ar[d]\ar[dr]&\Lambda_{\ge 0}\ar[r]\ar[d]\ar[dr]&\Lambda_{\ge 0}\ar[r]\ar[d]&\ldots\\
\ldots\ar[r]&CF^*(H_{i-1}^0)\ar[r]&CF^*(H_{i}^0)\ar[r]&CF^*(H_{i+1}^0)\ar[r]&\ldots
\end{tikzcd}$$ Construct a morphism between these rays by counting Floer solutions over ${{\mathbb C}}P^1\setminus\{z_0\}$ where the perturbation data are consistently chosen from ${\mathcal S}_{u}({\mathbf H})$, and ${\mathbf H}$ runs across the $H_i^0$.
Using this morphism, define the elements $$1_{K,\epsilon}\in SH^*_M(K,\epsilon)$$ as in Subsection \[subsec:algebra\_units\].
We do need this element to not be well-defined.
\[lem:prod\_and\_cont\] The elements $1_{K,\epsilon}\in SH^*_M(K,\epsilon)$ do not depend on the choices. Moreover, They are compatible with the maps $c_{\epsilon_1,\epsilon_2}$ for $\epsilon_1<\epsilon_2$, i.e. $$1_{K,\epsilon_2}=c_{\epsilon_1,\epsilon_2}(1_{K,\epsilon_1}).$$
This is again a standard application of Lemmas \[lem:S\_nonempty\] and \[lem:S\_contr\], along with Lemma \[lemstrictunit\].
\[lem:unit\_real\_cont\] Via the chain-level product from Definition \[dfn:prod\_e\], the chain-level unit from Definition \[dfn:unit\_e\] is a realisation, in the sense of Subsection \[subsec:algebra\_units\], of the chain-level variant continuation map $\tilde c_{\epsilon,2\epsilon}$ from Definition \[dfn:c\_tilde\].
There are natural gluing maps of spaces of perturbation data $$\#_\rho\co {\mathcal S}_u\sqcup {\mathcal S}_p\to {\mathcal S}_{\tilde c},\quad \rho\gg 0.$$ Using the gluing maps and the contractibility of these spaces, one can fill in the arrows of the diagram for the desired homotopy.
\[cor:1\_e\_c\] At the cohomology level, product with the unit $-*_\epsilon 1_\epsilon$ equals the extrinsic continuation map $c_{\epsilon,2\epsilon}$.
This follows from Lemma \[lem:unit\_real\_cont\] and Lemma \[lem:c\_versions\].
Product and unit on relative cohomology
---------------------------------------
From now on, let us consider the symplectic cohomologies over $\Lambda$: $$SH^*_M(K,\epsilon,\Lambda)\coloneqq SH^*_M(K,\epsilon)\otimes \Lambda.$$
Define the product $$*\co SH^*_M(K,\Lambda)\otimes SH^*_M(K,\Lambda) \to SH^*_M(K,\Lambda)$$ as the unique map making the diagram below commutative: $$\begin{tikzcd}
SH^*_M(K,\epsilon,\Lambda)\otimes SH^*_M(K,\epsilon,\Lambda)
\ar[r, "*_\epsilon"]
&
SH^*_M(K,2\epsilon,\Lambda)
\\
SH^*_M(K,\Lambda)\otimes SH^*_M(K,\Lambda)
\ar[r, "*"]
\ar[u, "c_{0,\epsilon}\otimes c_{0,\epsilon}"]
&
SH^*_M(K,\Lambda)
\ar[u, "c_{0,2\epsilon}"']
\end{tikzcd}$$ Recall that the vertical arrows are isomorphisms over $\Lambda$ by Lemma \[lem:c\_iso\].
Define the unit $1_K\in SH^*_M(K,\Lambda)$ as $$1= c_{0,\epsilon}^{-1}(1_{K,\epsilon}).$$ Again, recall that $c_{0,\epsilon}$ is an isomorphism over $\Lambda$ by Lemma \[lem:c\_iso\].
The element $1_K$ are well-defined. Moreover, $1_K$ is a 2-sided unit for $*$.
Let us check that $1_K$ is a unit. Consider the following diagram: $$\begin{tikzcd}
SH^*_M(K,\epsilon,\Lambda)\otimes \Lambda\langle 1_{K,\epsilon}\rangle
\ar[r, "-*_\epsilon 1_{K,\epsilon}", bend left]
\ar[r, "c_{\epsilon,2\epsilon}", bend right]
&
SH^*_M(K,2\epsilon,\Lambda)
\\
SH^*_M(K,\Lambda)\otimes \Lambda\langle 1_K \rangle
\ar[r, "-*1_K"]
\ar[u, "c_{0,\epsilon}\otimes c_{0,\epsilon}"]
&
SH^*_M(K,\Lambda)
\ar[u, "c_{0,2\epsilon}"']
\end{tikzcd}$$ The outer square is commutative by the definition of units and products. The slit is commutative by Corollary \[cor:1\_e\_c\]. Looking at the inner square and Lemma \[lem:c\_commut\] implies that $-*1$ is the identity.
\[lem:1\_zero\_sh\_zero\] If $1_K=0\in SH^*_M(K)$, then $SH^*_M(K)=0$.
This follows from Corollary \[cor:real\_iso\].
Notice that this statement does not make any reference to the product, but its proof relied on the existence of a product structure.
Restriction maps respect units $1_K\in SH^*_M(K)$.
It is straightforward to define restriction maps $r_\epsilon\co SH^*_M(K,\epsilon)\to SH^*_M(K',\epsilon)$ for $K\subset K'$ and prove that the following diagram is commutative: $$\begin{tikzcd}
SH^*_M(K',\epsilon)
\ar[r, "r_\epsilon"]
&
SH^*_M(K,\epsilon)
\\
SH^*_M(K')
\ar[r, "r"]
\ar[u, "c_{0,\epsilon}"]
&
SH^*_M(K)
\ar[u, "c_{0,\epsilon}"]
\end{tikzcd}$$ The bottom row is the usual restriction map. It is also straightforward to show that $r_\epsilon$ respects the elements $1_{K,\epsilon}$ using Lemma \[lemstrictunit\]. The lemma follows.
Relative Lagrangian Floer theory and closed-open maps {#sec:lag}
=====================================================
Lagrangian Floer theory
-----------------------
Let $L\subset M$ be a tautologically unobstructed oriented Lagrangian submanifold with a relative Pin structure.
As explained in Section \[sec:rel\_sh\], for every compact set $K\subset M$ there exists a Lagrangian version of relative symplectic cohomology, $HF^*(L,K)$.
Completely analogously to the previous section, one defines the raised version $HF^*_M(L,K,\epsilon)$ as the homology of the completed telescope of $$\begin{tikzcd}
\ldots\ar[r]&CF^*(L,H_{i-1})\ar[r]&CF^*(L,H_{i})\ar[r]&CF^*(L,H_{i+1})\ar[r]&\ldots
\end{tikzcd}$$ Here $CF^*(L,H)$ is the usual Floer complex of $L$ with itself using the Hamiltonian $H$, and the $H_i$ are an increasing cofinal family Hamiltonians as in the beginning of Section \[sec:prod\_unit\]. Also analogously to Section \[sec:prod\_unit\], one defines extrinsic continuation maps, product, and the unit $$1_{K,L,\epsilon}\in HF^*(L,K,\epsilon).$$ Finally, one defines the product and unit on $HF^*(L,K)$ by repeating the formal trick from Section \[sec:prod\_unit\]. They satisfy obvious versions of the properties from that section.
The only detail which is slightly different is the definition of the spaces of perturbation data. Let $C$ be the disk with $k\in\{0,1,2\}$ input punctures and one output puncture. One uses the spaces $$\label{eq:S_L}
{\mathcal S}^{op}_{k;1}({\mathbf w},{\mathbf H})=\{
(\alpha, H)\}
\subset \Omega^1(C)\times C^\infty(C\times X)$$ where $\alpha,H$ with the following properties: $$\left\{
\begin{array}{l}
d\alpha\ge 0,\ H\ge 0,\\
d(\alpha\otimes H)\ge 0,\\
\alpha\equiv w^idt\text { on strip-like ends,}\\
H\equiv H^i(t)\text { on strip-like ends,}\\
T(\partial C)\subset \ker \alpha.
\end{array}
\right.$$
The last property has no analogue in the closed-string case. It is required to guarantee that the topological energy of a Floer strip is indeed topological. Analogues of Lemmas \[lem:S\_nonempty\] and \[lem:S\_contr\] on non-emptiness and contractibility of these spaces hold under similar conditions.
![The Riemann surfaces used to define the product structure for the open string version (left) and the closed-open maps (right).[]{data-label="fig:slit"}](Fig_Slit)
One point worth a clarification is why, given $w^0\ge \sum_{i=1}^k w_i$, there exists at least one $\alpha$ with $d\alpha\ge 0$ and $T(\partial C)\subset \ker\alpha$, compare with the beginning of the proof of Lemma \[lem:S\_nonempty\]. Assume, for instance, $k=2$ and $w^0= w_1+w_2$ (it is easy to upgrade the example to the case of the inequality). Consider the domain $C$, the disk with three boundary punctures, conformally represented in Figure \[fig:slit\] by a strip of width $w^0$ with an extra slit dividing the widths into $w^1$ and $w^2$. The form $\alpha=dt$ in this representation has the desired properties: it restricts to $w^idt$ on strip-like ends (recall that a strip-like end, by definition, conformally reparametrises a neighboirhood of a puncture to a strip of width one), and $\alpha$ vanishes on $T(\partial C)$, including the tangent directions to the slit.
Closed-open maps
----------------
The last ingredient is to relate the closed- and open-string theories by the closed-open map, with the aim of establishing Proposition \[prop:zero\_unit\_hf\_sh\].
For this, let $C$ be the unit disk with one boundary puncture (considered as output) and one interior puncture (considered as input) modulo automorphisms. Equip $C$ with a strip-like end at the boundary puncture, and cylindrical end at the interior puncture. One can assume that the punctures are at points 1 and 0, respectively. Define $$\label{eq:S_co}
{\mathcal S}_{1;0;1}({\mathbf w},{\mathbf H})=\{
(\alpha, H)\}
\subset \Omega^1(C)\times C^\infty(C\times X)
\textit{ where }
{\mathbf w}=(w^1,w^0),\ {\mathbf H}=(H^1,H^0).$$ to be the subset consisting of all $\alpha,H$ with the following properties: $$\left\{
\begin{array}{l}
d\alpha\ge 0,\ H\ge 0,\\
d(\alpha\otimes H)\ge 0,\\
\alpha\equiv w^1dt\text { on the cylindrical end,}\\
H\equiv H^1(t)\text { on the cylindrical end,}\\
\alpha\equiv w^0dt\text { on the strip-like end,}\\
H\equiv H^0(t)\text { on the strip-like end,}\\
T(\partial C)\subset \ker \alpha.
\end{array}
\right.$$
Assume that $w^0\ge w^1$ and $H^0\succeq H^1$, then again ${\mathcal S}_{1;0;1}({\mathbf w},{\mathbf H})$ is non-empty and contractible. To set up closed-open maps between raised cohomologies, we use perturbation spaces of the form $${\mathcal S}_{co}({\mathbf H})={\mathcal S}_{1;0;1}({\mathbf w},{\mathbf H})\quad\textit{with}\quad w^1=w^0=1.$$
Let $\epsilon>0$, and $\{H_i^1\}$, $\{H_i^0\}$ be two cofinal families for $SH^*_M(K,\epsilon)$ chosen so that $H_i^0\succeq H_i^0$ for each $i$. Consider two Floer 1-rays from the diagram below, and construct the morphism between them using curves on the disk $C$ with one interior input and one boundary output as above, with perturbation data chosen consistently from ${\mathcal S}_{co}({\mathbf H})$ using ${\mathbf H}=(H_i^1,H_i^0)$ and $(H_i^1,H_{i+1}^0)$. $$\begin{tikzcd}
\ldots\ar[r]&CF^*(H_{i-1}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i}^1)\ar[r]\ar[d]\ar[dr]&CF^*(H_{i+1}^1)\ar[r]\ar[d]&\ldots\\
\ldots\ar[r]&CF^*(L,H_{i-1}^0)\ar[r]&CF^*(L,H_{i}^0)\ar[r]&CF^*(L,H_{i+1}^0)\ar[r]&\ldots
\end{tikzcd}$$ This gives rise to $\Lambda_{\ge 0}$ module map $${\mathcal C}{\mathcal O}_\epsilon\co SH^*_M(K,\epsilon)\to SH^*(L,K,\epsilon).$$ One formally defines $${\mathcal C}{\mathcal O}\co SH^*_M(K,\Lambda)\to SH^*(L,K,\Lambda)$$ by requiring the following diagram to commute (recall that the vertical arrows are isomorphisms over $\Lambda$:
$$\begin{tikzcd}
SH^*_M(K,\epsilon)\otimes \Lambda \arrow[r, "{\mathcal C}{\mathcal O}_\epsilon"]& SH^*(L,K,\epsilon)\otimes \Lambda \\
SH^*_M(K)\otimes \Lambda \arrow[r, "{\mathcal C}{\mathcal O}"]\arrow[u,"c_{0,\epsilon}"]& SH^*(L,K)\otimes \Lambda\arrow[u,"c_{0,\epsilon}"]
\end{tikzcd}$$
\[lem:co\_unit\] The closed-open map satisfy $${\mathcal C}{\mathcal O}(1_{K,\epsilon})=1_{K,L,\epsilon},\quad {\mathcal C}{\mathcal O}(1_K)=1_{K,L}.$$
Let ${\mathcal C}$ be a Floer 1-ray for $SH^*_M(K,\epsilon)$, ${\mathcal C}'$ be one for $SH^*(L,K,\epsilon)$, and $U$ the 1-ray from Subsection \[subsec:algebra\_units\]. One constructs a homotopy $$\begin{tikzcd}
U\arrow[r, "1_{K,\epsilon}"]\arrow[rr, bend right, "1_{K,L,\epsilon}"'] &{\mathcal C}\arrow[r, "{\mathcal C}{\mathcal O}"] &{\mathcal C}'
\end{tikzcd}$$ by counting Floer solutions of appropriate continuation maps. This uses the fact that there are natural gluing maps $$\#_\rho\co {\mathcal S}_{u}\sqcup {\mathcal S}_{co}\to {\mathcal S}^{op}_{u},\quad \rho\gg 0.$$ The perturbation spaces ${\mathcal S}_{u}$ and ${\mathcal S}_{co}$ were introduced above (they are used to define the unit and the closed-open map), and ${\mathcal S}^{op}_{u}({\mathbf H})\coloneqq {\mathcal S}^{op}_{0;1}({\mathbf w},{\mathbf H})$ with ${\mathbf w}=w^0=1$ are the perturbation spaces from the definition of unit in Lagrangian Floer cohomology. Hence, the first equality follows from Lemma \[lemstrictunit\].
The second equality from the lemma follows from the first one, and the definitions of $1_K$, $1_{K,L}$.
Liouville complements
=====================
First of all, note that McLean’s Proposition 6.17 reduces the problem to showing that there is a primitive $\theta$ defined on $M-D$ such that the relative cohomology class in $H^2(M,M-D)$ defined by $(\omega,\theta)$ is equal to $\sum \frac{w_i}{c}[D_i]$. It suffices to prove this for $c=1$, so let us assume that.
Let $(L,h,\nabla)$ be the pre-quantization complex line bundle for $(M,\omega)$. More precisely, $h$ is a Hermitian metric on $L$ and $\nabla$ is a compatible connection such that the curvature $2$-form of $\nabla$ is equal to $\omega$.
Let $O_M(D)$ be the complex line bundle associated to the SC symplectic divisor $D$ as explained in the discussion near Equations (6) and (7) of [@tehrani]. Then $O_M(D)$ and $L$ are isomorphic as complex line bundles because they have the same first Chern class. This in particular shows that $L$ has a section which vanishes precisely along $D$ with multiplicity $w_i$ on $D_i$ by taking the tensor product of the “defining" sections of $O(D_i)$. Let us call this section $s$.
Now let $P\to M$ be the $U(1)$-bundle associated to $L$. Then we have a connection one form $\theta'$ on $P$ which is a primitive of the pullback of $\omega$ to $P$. We can construct a section $s'$ of $P\to M$ over $M-D$ using $s$, by $s':=s/{|s|}$. Pulling back $\theta'$ by $s'$, we obtain a primitive $\theta$ of $\omega$ on $M-D$, which satisfies the desired conditions.
[^1]: A Lagrangian submanifold $L$ is called tautologically unobstructed if there exists a compatible almost complex structure $J$ such that $L$ bounds no non-constant $J$-holomorphic disks.
[^2]: This only means $c_1(M)=0$ throughout the paper.
[^3]: It is more convenient to talk about valuations for us. For translation to the more familiar language, as it is used in functional analysis for the archimeden case: note that norm is given by $e^{-val}.$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'While the number of asteroids with known shapes has drastically increased over the past few years, little is known on the the time-evolution of shapes and the underlying physical processes. Here we propose an averaged abrasion model based on micro-collisons, accounting for asteroids not necessarily evolving toward regular spheroids, rather (depending on the fall-back rate of ejecta) following an alternative path, thus confirming photometry-derived features, e.g. existence of large, relatively flat areas separated by edges. We show that our model is realistic, since the bulk of the collisions falls into this category.'
author:
- 'G. Domokos, A. Á. Sipos, Gy. M. Szabó, and P. L. Várkonyi'
title: Formation of sharp edges and planar areas of asteroids by polyhedral abrasion
---
Introduction
============
The number of asteroids with known shapes has increased significantly in the past years. 8 asteroids were visited by fly-by missions, providing high-resolution images (e.g. Veverka et al. 1994, Chapman et al. 1995, Robinson et al. 1995, Veverka et al. 1999, Oberst et al. 2001, Duxbury et al. 2004, Saito et al. 2006, Schulz et al. 2008). Radar observations can also reveal the envelope of the shape (e.g. Ostro et al. 1988). A very promising tool of shape identification is the inversion of photometric light curves (Kaasalainen and Troppa, 2001, Kaasalainen et al 2002, KTP2002 hereafter), which resulted in photometric models for $\approx$100 asteroids, so far. These new efforts and techniques lead to the rapid increase of the number of known asteroid shapes and we can expect this number to keep growing in the future.
These shapes have not been globally classified in the literature, but one can easily recognize some typical shape properties. On one hand, there are some asteroids with highly nonconvex shape (e.g. 3908 Nyx), there are bifurcated objects (Kleopatra), earthnut-like forms (Eros), all of which may be results of typical formation processes. On the other hand, several asteroids have a strikingly simple shape envelope with large flat surfaces bordered by relatively sharp edges. A significant portion of photometric shapes apparently share these features (e.g. 6053 [1993 BW]{}, Ďurech et al. 2002, 10115 1992 KS, Busch et al. 2005, 1580 Betulia, 1980 Tezcatlipoca, 2100 Rha-Shalom, Kaasalainen et al. 2004). These observations are often attributed to local concavities (e.g. Ďurech 2002) or albedo effects (KTP2002), i.e. they have been claimed to be artefacts of the photometric method. However, close-up images from spacecrafts also reveal that the presence of prominent flat areas and sharp edges is typical for at least half of the small asteroids (e.g. Ida, Gaspra, Annefrank, Steins). Recently, Troppa et al (2008) confirmed the presence of flat areas on photometric [shapes]{} and found that the number of flat areas/facets decreases with the elongation of the shape (or its convex hull). While photometric imaging certainly does not provide sufficient information to quantify the uncertainities associated with the mentioned qualitative geometric features, these images suggest that at least some asteroids may not evolve towards spheroids. Our goal is to explore this avenue by applying a simple averaged abrasion model.
Recently, several studies have been dedicated to explain the shapes and surface properties. The conclusions point to different evolution scenarios; they also help to better understand the relation between collisions and the global shape. The following basic scenarios were proposed:
- Formation of primordial shapes in [disrupting,]{} energetic collisions (Michel et al 2003), followed by continuous formation and erosion of craters (Housen and Holsapple, 2003, O’Brien et al. 2006).
- Rubble pile structure asteroids can get very elongated forms as a result of tidal effects (Bottke et al. 1999).
- Evolution of rubble piles with elongated shapes toward a spherical form by seismic shaking of collisions with moderate energy ( Korycansky and Asphaug, 2003, KA2003 hereafter, Szabó and Kiss, 2008)
- Continuous alteration of the surface by low-energy collisions and space weathering that form small craters (Lazzarin et al. 2006).
The widespread concept is that small impactors always smoothen the global shape by eroding sharp edges (A. Conrad, personal communication). The root of this concept is the obvious $O(3)$-symmetry of the problem: the impacts have uniform radial intensity and one would expect that abrasion in this environment converges to an $O(3)$-symmetric object (sphere).
Low-energy impactors remove small pieces from the surface of the asteroid, which either fall back to the surface (captured by the gravitational field), or they escape. A model based exclusively on the first scenario (KA2003) predicted convergence to spheres (or oblate ellipsoids for rotating asteroids). In this paper, we adress the second one. In contrast to natural expectations, abrasion by small, escaping fragments breaks the $O(3)$-symmetry, and leads to the formation of large planar areas and sharp edges. The global outcome for most initial shapes are low-order convex polyhedra (i.e. convex shapes bordered by a few planes), similar to at least a considerable fraction of asteroids. We also present preliminary results for a more realistic model, combining the effects of escaping material and the redistribution of ejecta on the asteroid’s surface. The reported features appear to preserved in a certain parameter-range.
Description of the model
========================
Abrasion due to micro-impacts is a complex, stochastic process, in which minor, discrete pieces are removed from the asteroid’s surface in small collisions. Our model is an averaged continuum approach where the effect of large numbers of small, discrete impactors is modeled by continuous abrasion of the surface. The first step towards the construction of the partial diffeential equation (PDE) is determining the average rates of abrasion on an asteroid’s surface under the simplest assumptions for the discrete impactors (e.g. uniform distribution for the direction).
Let the asteroid surface at time $t$ be given by the endpoints of a two-parameter family of vectors $\textbf{x}(p,q,t)$. We assume that a large number of small particles (relative to the asteroid) move along random straight lines (in coordinates fixed to the asteroid) and collide with the asteroid; subsequently they escape together with small, broken portions of the asteroid (Ronca and Furlong 1979). In case of impact trajectories we ignore gravitational effects as well as spinning; both are negligible at high-speed impacts. (The role of gravity, however, can not be neglected when considering material falling back to the surface; the model we first discuss assumes that all material escapes, later in the paper we develop a general model including non complete ejecta escape.)
In particular, we assume constant and uniform (i.e. time- space- and direction-independent) radial intensity $\Phi$ of particle collisions (average number of collisions per unit time, unit area and unit solid angle), resulting, in case of convex surfaces, in constant and uniform surface intensity $m$ of particle collisions (average number of collisions per unit time, and unit surface area). We assume that the volume of material abraded from unit surface per unit time, or, equivalently, the abrasion speed $\partial
\textbf{x}(p,q,t)/ \partial t$ in the direction of the inward directed surface normal $\textbf{n}(p,q,t)$ is determined by the surface intensity $m$ and the abrasion “efficiency” of the impactors. The latter could be expressed as a function $f(\beta)$ of the impact angle $\beta$ ($0<\beta <
\pi/2$) between $\textbf{n}$ and the direction of the incoming particle. Abrasion speed can be obtained by integrating $\Phi f(\beta)$ over the $HS$ half-sphere and considering that the number of the incoming particles depend on $\beta$, $$\begin{aligned}
\frac{\partial \textbf{x}(p,q,t)}{\partial t} = \int_{HS}\Phi
\cos\beta
f(\beta)dS\cdot \textbf{n}(p,q,t)\nonumber\\
=\Phi
\int_0^{\pi/2}f(\beta)\cos\beta\sin\beta d\beta \cdot\textbf{n}(p,q,t)
\end{aligned}$$ In case of $f(\beta)=1$, Eq. 1 yields $|\partial \textbf{x}/ \partial t| =
\Phi = m$, however, for any other $f(\beta)$ we still obtain the simple PDE called Eikonal equation for the abrasion of convex surfaces due to uniform micro-impacts: $$\frac{\partial \textbf{x}(p,q,t)}{\partial t} = constant\cdot
\textbf{n}(p,q,t),$$
Of course, primordial asteroid shapes are often locally concave. In these regions equation (1) should be integrated over its zenith angle, which is only a part of the half-sphere. Thus, the surface intensity $m$ (and the abrasion speed) is lower for concave surface points than for typical (generic) points on the convex hull, and therefore concave parts tend to disappear. In the rest of the paper, we focus on the convex case.
After establishing the appropriate averaged continuum model, our next goal is to show the resulting stable polyhedral geometric patterns.
Solutions of (2) are known as wave-fronts (Arnold, 1986), and singularity theory describes their evolution, Figure 1 illustrates the propagation of a planar wave-front starting from a convex, smooth curve. One of the most relevant features is that initially smooth surfaces evolve structurally stable singularities (cf. Figure 1b) and self-intersections along edges(Figure 1c). In a physical abrasion process, the solution is limited to the domain free of self-intersections (which is a weak solution of the PDE, (Sethian, 1985)): the aforementioned self-intersecting edges and vertices appear on the surface of the physically relevant domain (Figure 1d). In three dimensions, the situation is analogous: edges and vertices appear in the solution for all generic (typical) smooth initial shapes (Figure 2a,b). Structural stability means in this context that the appearance of these geometric features can be expected for almost all initial surfaces. Another feature of this process is that polyhedra evolve to polyhedra, with the number of faces monotonically decreasing in time (Figure 2c). These features imply that this process produces shapes with flat areas separated by marked edges and vertices, almost regardless of the starting configuration. One can also provide a simple and complete list of final limit shapes (depending on the initial shape): tetrahedron, elongated cigar-shape or a flat disc (Várkonyi et al., 2008, Domokos et al, 2009.). We demonstrate two of these limits numerically (Figure 2). Needless to say, more sophisticated discrete models would predict the process more accurately in the quantitative sense, however, our compact PDE model captures the most essential qualitative features.
{width="\columnwidth"}
Previously, the evolution of asteroid shapes was modelled in two dimensions by Ronca and Furlong (1979) under similar assumptions, however, involving two misleading concepts: (i) rate of volume loss was associated with *radial* abrasion speed (toward the barycenter) rather than speed normal to the surface, and (ii) a false interpretation of zenith angle lead to nonconstant abrasion speeds on convex surfaces and also to inconsistent results. Their model also predicts surface singulaities, however, otherwise radically different morphological characters then ours.
Simulations
===========
Eq. (2) is a Hamilton-Jacobi type PDE, which can be efficiently simulated by level-set methods (Osher and Fedkiw, 2002) and fast-marching methods (Sethian, 1999). Both methods handle well singularities (edges and vertices). In this paper, a MATLAB-based implementation of the level-set method (Mitchell, 2008) was used to compute the evolution of various initial shapes.
As reported from experiments e.g. by Cappacioni et al. (1984), Catullo et al (1984) and Ryan et al (2000), the average axis ratio of asteroids resulting from fragmentation of larger objects is 1/1.4/2 with large scatter in both ratios. We used initial axis ratios even closer to the sphere to demonstrate that the emergence of elongated shapes and edges is independent of the initial shapes.
Figure 2 shows the evolution of three different random initial shapes of which the first two have been picked from relatively smooth surfaces. Observe emerging sharp edges and plane areas, analogous to the planar wave front in Figure 1. Also, observe the tetrahedral limit in the first case and the increasing prolateness leading to a ’cigar-shape’ in the second case. The third series starts with a multi-faceted polyhedral shape (which could be interpreted as the convex hull of a more general shape). Here the number of faces is reduced in abrasion, implying the emergence of low-order polyhedra (and finally, a flat tetrahedron). The emerging prolateness for many initial shapes confirms the observation of Troppa et al. (2008). The figure also contains photometric shapes (Troppa et al 2002, KTP2002, Kaasalainen et al. 2004) showing visual resemblance to the intermediate stage of the simulated examples. We emphasize, however, that it was not our goal to reconstruct the exact history of any particular asteroid shape, we merely intended to illustrate the strong qualitative agreement between photometric shapes and the simulated geometries.
All asteroid images shown in Figure 2 are the result of photometric modeling providing only approximate information. We also tried to reconstruct the evolution of one asteroid for which the shape is known in detail. Asteroid 5535 Annefrank, visited by Stardust (Duxbury et al. 2004) apparently has two components, the major part is similar to a tetrahedron. Figure 3a shows a hypothetical time-series chosen by trial-and error from simulations of randomly perturbed spheres. One stage of the evolution is a rounded polyhedron, strongly resembling 5535 Annefrank (Fig 3c).
{width="8cm"}
{width="7.5cm"}
So far, all simulations were based on our extreme assumption that all abraded material escapes the gravitational field. To check the robustness of the described qualitative features we included in our PDE model the effect of ejecta falling back to the surface. We integrated the gravitational potential $U$ for the evolving shape and assumed that some fraction $p$ of ejecta resettles to the surface with intensity proportional to $e^{U/\overline{U}}$ which is a maximum entropy approximation of the fallback rate. Here, $\overline{U}$ denotes the average value of $U$ on the surface. All our previous computations correspond to the case $p=0$, whereas the case $p=1$ represents the scenario described in more detail by (KA2003). Detailed evaluation of this unified model is beyond the scope of the current paper, however we briefly comment on the emerging behavior: for $p=1$ the shapes converge to a sphere (which would be perturbed if we considered rotational effects); for smaller $p$, the final shape is a round disk with two smooth sides surrounded by a sharp edge; the flatness of the disk is decreasing function of $p$. Figure 3b shows an example with $p=0.25$, illustrating another remarkable feature of the process: the initial behavior resembles the case $p=0$ (Figure 3a), and the emerging polyhedral shape remains visible for a long time before being transformed to a disk.
Conclusions
===========
We showed in this paper that a simple continuum model, based solely on the assumptions that asteroids have
- numerous, high-speed collisions with
- relatively small impactors
- arriving uniformly randomly from all directions
accounts for the appearance of flat areas, separated by edges and vertices, which is observable in the available data. Photometric imaging certainly does not provide sufficient information to quantify the uncertainities associated with the mentioned qualitative geometric features. Based on our present model we can conclude that asteroids do not necessarily evolve preferentially toward regular spheroids, but may (depending on the fall-back rate of ejecta) follow an alternative path, thus confirming photometry-derived features. As far as we know, ours is the only scenario reproducing these features.
Our simple model does not include several effects which also influence the abrasion. (For example, collisions with larger objects have totally different geometric effects: vertices and edges will be worn and appear less pronounced, craters are created, etc.). The global effects listed in Sec. 1 also play an important role in forming the final geometry. Quantitative predictions of shape evolution could only be achieved by more sophisticated, discrete models. After including the effect of gravity, as long as the majority of particles escape, edges remain a visible geometric feature.
The relevance of our model relies on the assumption that high-speed collisions with small impactors dominate the abrasion process. This is quite plausible, since the size distribution of asteroids (e.g. Tedesco et al. 2005, Parker et al. 2008) in the $>500$ m range was found to be a quickly decreasing function following a power law of exponent $-2.5\ldots -3$ The size distribution of smaller particles has not been examined directly, however numerical models for impact evolution predict even steeper (up to the $-7$ power) size distribution functions especially for small sizes (Michel et al. 2004). While it is beyond the scope of the current paper, it is possible to estimate an order of magnitude for the number of collisions needed to model the body (KA2003).
Abrasion processes governed by equation (2) occur also on much smaller scale: pebbles shaped by wind-blown sand , also known as ventifacts, evolve in a similar manner (Várkonyi et al., 2008, Domokos et al., 2009), the details of which process we are currently investigating.
Supports from OTKA grant T72146 (GD, AS, PV), OTKA K 76816 (GyMSz) and the “Bolyai János” Research Fellowship of the Hungarian Academy of Sciences (GyMSz) are gratefully acknowledged. The comments by the anonymous Referee greatly improved this paper.
Arnold, V.I., 1986, Catastrophe Theory, Springer, Berlin (2nd edition)
Botke, W. F. Jr., Richardson, D. C., Michel, P. & Love, S. G., 1999, AJ, 117, 1921
Busch, M. W. et al., 2005, Icarus, 181, 145
Cappacioni, F et al., 1984, Nature, 308 832
Catullo, V., Zappala, V., Farinella, P. & Paolicchi, P., 1984, A&A, 138, 4464
Chapman, C. et al., 1995, Nature, 374, 783
Domokos, G. et al., 2009, Periodica Polytechnica, Arch. 40, 1.
Ďurech, J, 2002, Icarus, 159, 192
Duxbury, T. C. et al., 2004, J. Geophysical Research, 109, 2
Housen, K. R. & Holsapple, K. A., 2003, Icarus, 163, 102
Kaasalainen, M., Troppa, J. & Piironen, J. 2002, Icarus, 159, 369 (KTP2002)
Kaasalainen et al., 2004, Icarus, 167, 178
**** Korycansky and Asphaug 2003, Icarus 163, 374
Lazzarin, M et al., 2006, ApJ, 647, 179
Michel, P., Benz, W. & Richardson, D. C., 2003, Nature, 421, 608
Michel, P. et al., P&SS, 52
Mitchell, I. M., 2008, The Flexible, Extensible and Efficient Toolbox of Level Set Methods, J. Scientific Computing, 35, 300-329.
Oberst, J. et al., 2001, Icarus, 153, 160
O’Brien, D. P., Greenberg, R. & Richardson, J. E., 2006, Icarus, 183, 79
Osher, S., Fedkiw, R., 2002, Level Set Methods and Dynamic Implicit Surfaces. Springer-Verlag.
Ostro, S. J., Connelly, R. & Belkora, L., 1988, Icarus, 73, 150
Robinson, M. et al., 1995, Meteoritics, 30, 566
Ronca, L.B. & Furlong, R.B, 1979, EMP, 21, 409
Ryan, E.V., Ann. Rev. P&SS, 28, 367
Saito, J. et al., 2006, Science, 312, 1341
Schulz, R. et al., 2008, DPS, 40, 2829
Sethian, J.A. 1999, Level Set Methods and Fast Marching Methods, Cambridge University Press.
Sethian, J.A., 1985, Comm. Math. Phys., 101, 487.
Szabó, Gy.M. & Kiss, L. L., 2008, Icarus, 196, 135
Troppa, J., et al., 2008, Icarus, 198, 91
Troppa, J. et al., 2003, Icarus, 164, 346
Veverka J., Belton, M., Klassen, K. & Chapman, C., 1994, Icarus, 107, 2
Veverka J. et al, 1999, Icarus, 140, 3
Várkonyi P. L., Domokos G. & Sipos A. Á. 2008, Proc. of the $1^{st}$ Workshop on Mathematical Geosciences, June 11-12 2008, Freiberg, Germany. Available at: http://www.iamg.tu-freiberg.de/assets/docs/pdf/ProceedingsMatGeoS08/Peter Varkonyi.pdf
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the last years it has been shown that some properties of strongly coupled superconductors can be potentially described by classical general relativity living in one higher dimension, which is known as holographic superconductors. This paper gives a quick and introductory overview of some holographic superconductor models with s-wave, p-wave and d-wave orders in the literature from point of view of bottom-up, and summarizes some basic properties of these holographic models in various regimes. The competition and coexistence of these superconductivity orders are also studied in these superconductor models.'
author:
- |
Rong-Gen Cai$^1$[^1] , Li Li$^2$[^2] , Li-Fang Li$^3$[^3] , Run-Qiu Yang$^1$[^4]\
\
$^1$State Key Laboratory of Theoretical Physics,\
Institute of Theoretical Physics, Chinese Academy of Sciences,\
Beijing 100190, China.\
$^2$ Crete Center for Theoretical Physics, Department of Physics,\
University of Crete, 71003 Heraklion, Greece.\
$^3$State Key Laboratory of Space Weather,\
Center for Space Science and Applied Research, Chinese Academy of Sciences,\
Beijing 100190, China.
title: '**Introduction to Holographic Superconductor Models**'
---
Introduction
=============
The phenomenon of superconductivity was discovered in the early part of the last century that the electrical resistivity of a material suddenly drops to zero below a critical temperature $T_c$. More importantly, the so-called Meissner effect tells us that the magnetic field is expelled in the superconducting phase, which is distinguished from the perfect conductivity. In the latter case a pre-existing magnetic field will be trapped inside the sample. Conventional superconductors are well described by BCS theory [@Bardeen:1957mv], where the condensate is a Cooper pair of electrons bounded together by phonons. According to the symmetry of the spatial part of wave function of the Cooper pair, superconductors can be classified as the s-wave, p-wave, d-wave, f-wave superconductors, etc. However, some materials of significant theoretical and practical interest, such as high temperature cuprates and heavy fermion compounds, are beyond BCS theory. There are indications that the involving physics is in strongly coupled regime, so one needs a departure from the quasi-particle paradigm of Fermi liquid theory [@Carlson:2002]. Condensed matter theories have very few tools to do this.
On the other hand, although some of the deeper questions arising from the Anti-de Sitter/Conformal Field Theory(AdS/CFT) correspondence [@Maldacena:1997re; @Gubser:1998bc; @Witten:1998qj] remain to be understood from first principles, this duality creating an interface between gravitational theory and dynamics of quantum field theory provides an invaluable source of physical intuition as well as computational power. In particular, in a “large $N$ and large $\lambda$" limit, the gravity side can be well described by classical general relativity, while the dual field theory involves the dynamics with strong interaction.[^5] It is often referred to as “holography" since a higher dimensional gravity system is described by a lower dimensional field theory without gravity, which is very reminiscent of an optical hologram. There are indeed many physical motivations that lead to this amazing holographic duality in the literature, such as renomalization group flow and black hole membrane paradigm. A very new perspective was proposed, which is called the exact holographic mapping [@Qi:2013caa]. By constructing a unitary mapping from the Hilbert space of a lattice system in flat space (boundary) to that of another lattice system in one higher dimension (bulk), it provides a more explicit and complete understanding of the bulk theory for a given boundary theory and can be compared with AdS/CFT correspondence.
It has been shown that the AdS/CFT correspondence can indeed provide solvable models of strong coupling superconductivity, see refs. [@Horowitz:2010gk; @Herzog:2009xv; @Iqbal:2011ae; @Musso:2014efa] for reviews. The physical picture is that some gravity background would become unstable as one tunes some parameter, such as temperature for black hole and chemical potential for AdS soliton, to developing some kind of hair. The emergency of the hair in the bulk corresponds to the condensation of a composite charged operator in the dual field theory. More precisely, the dual operator acquires a non-vanishing vacuum expectation value breaking the U(1) symmetry spontaneously. It has been uncovered that this simple holographic setup shows similar properties with real superconductors.
The holographic s-wave superconductor model known as Abelian-Higgs model was first realized in refs. [@Hartnoll:2008vx; @Hartnoll:2008kx]. According to the AdS/CFT correspondence, in the gravity side, a Maxwell field and a charged scalar field are introduced to describe the U(1) symmetry and the scalar operator in the dual field theory, respectively. This holographic model undergoes a phase transition from black hole with no hair (normal phase/conductor phase) to the case with scalar hair at low temperatures (superconducting phase). The holographic model for the insulator/superconductor phase transition has been realized in ref. [@Taka] at zero temperature by taking the AdS soliton as the gravitational background in the Abelian-Higgs model. Holographic d-wave model was constructed by introducing a charged massive spin two field propagating in the bulk [@Chen:2010mk; @Benini:2010pr]. The superconductivity in the high temperature cuprates is well known to be of d-wave type.
In recent years, evidence from several materials suggests that we now have examples of p-wave superconductivity, providing us new insights into the understanding of unconventional superconductivity in strongly correlated electron systems [@Mackenzie]. To realize a holographic p-wave model, one needs to introduce a charged vector field in the bulk as a vector order parameter. Ref. [@Gubser:2008wv] presented a holographic p-wave model by introducing an SU(2) Yang-Mills field into the bulk, where a gauge boson generated by one SU(2) generator is dual to the vector order parameter. The authors of refs. [@Cai:2013pda; @Cai:2013aca] constructed a holographic p-wave model by adopting a complex vector field charged under a U(1) gauge field, which is dual to a strongly coupled system involving a charged vector operator with a global U(1) symmetry. An alternative holographic realization of p-wave superconductivity emerges from the condensation of a two-form field in the bulk [@Aprile:2010ge; @Donos:2011ff; @Donos:2012gg].
The philosophy for holographic setups is that even though the underlying microscopic description of the theory with a gravity dual is quite likely to be different form that arising in materials of experimental interest, it may uncover some universal aspects of the strongly coupled dynamics and kinematics, thus would help the development of new theories of superconductivity. By mapping the quantum physics of strongly correlated many body systems to the classical dynamics of black hole physics in one higher dimension, the holographic approach provides explicit examples of theories without a quasi-particle picture in which computations are nevertheless feasible.
The models studied in the literature can be roughly divided into two classes, i.e., the bottom-up and top-down models. In the former approach the holographic model is constructed phenomenologically by picking relevant bulk fields corresponding to the most important operators on the dual field theory and then writing down a natural bulk action considering general symmetries and other features of dual system. Thus the holographic description is necessarily effective and can be used to describe a wide class of dual theories instead of a definite single theory. In the top-down approach, the construction of a model is uniquely determined by a consistent truncation from string theory or supergravity. One can usually have a much better control over the dual field theory, nonetheless, the resulted models are much more complicated.
This paper aims at providing a quick and introductory overview of those three kinds of holographic superconductor models from the point of view of bottom-up.[^6] The organization of the paper is as follows. In the next section, we review basic elements of Ginzburg-Landau theory of superconductivity and holographic duality, as a warm up. A brief introduction to the Abelian-Higgs model is presented in section \[sect:higsmodel\]. In section \[sect:p-wave\], we first introduce the SU(2) Yang-Mill model, focusing on its condensate and conductivity, then study the Maxwell-vector model, paying more attention to the vector condensate induced by magnetic fields and its complete phase diagram in terms of temperature and chemical potential, and finally discuss the third p-wave model by introducing a two-form in the bulk. This model can exhibit a novel helical superconducting phase. Section \[sect:dwave\] is devoted to holographic d-wave models. In the next two sections, we pay attention to the competition and coexistence among different orders, including different superconducting orders in section \[sect:competition\] as well as superconducting order and magnetic order in section \[sect:M&S\]. The conclusion and some discussions are included in section \[sect:conclusion\].
Preliminary {#sect:prelimi}
===========
Ginzburg-Landau theory
----------------------
The microscopic origin of traditional superconductivity is well understood by BCS theory, which explained the superconducting current as a superfluid of pairs of electrons interacting through the exchange of phonons. However, at phenomenological level, the Ginzburg-Landau theory of superconductivity [@Ginzburg:1950] had great success in explaining the macroscopic properties of superconductors. [^7]
In this phenomenological theory, the free energy of a superconductor can be expressed in terms of a complex order parameter field, $\Psi$, which is directly related to the density of the superconducting component. By assuming smallness of $|\Psi|$ and smallness of its gradients, the free energy near the superconducting critical temperature $T_c$ has the following form,
$$\label{GL}
F=F_n+\alpha|\Psi|^2+\frac{\beta}{2}|\Psi|^4+\frac{1}{2m^*}|(-i\nabla-e^*\vec{A})\Psi|^2+\frac{|\vec{B}|^2}{2\mu_0},$$
where $F_n$ is the free energy in the normal phase, $\vec{A}$ is the vector potential and $B=\nabla\times\vec{A}$ is the magnetic field. $m^*$ and $e^*$ are effective mass and charge of condensate. If one considers the BCS theory, $m^*=2m$ and $e^*=2e$ with $m$ and $e$ the mass and charge of electrons forming Copper pairs. $\alpha$ and $\beta$ are two phenomenological parameters, which behave as $\alpha=\alpha_0(T_c-T)$ with $\alpha_0$ and $\beta$ two positive constants. Note that we work in the unites $\hbar=c=1$. It is obvious that the free energy is invariant under the following transformation, $$\label{uone}
\Psi\rightarrow\Psi e^{i\theta(x)},\quad \vec{A}\rightarrow\vec{A}+\frac{1}{e^*}\nabla\theta(x),$$ which is known as the U(1) gauge symmetry.
Minimising the free energy with respect to the order parameter and the vector potential, one obtains the Ginzburg-Landau equations, $$\begin{aligned}
&&\alpha\Psi+\beta|\Psi|^2\Psi+\frac{1}{2m^*}(-i\nabla-e^*\vec{A})^2\Psi= 0, \\
&& \vec{J_s}=-\frac{ie^*}{2m^*}(\Psi^*\nabla\Psi-\Psi\nabla\Psi^*)-\frac{e^{*2}}{m^*}\Psi^*\Psi\vec{A}.\label{london}\end{aligned}$$ The first equation determines the order parameter $\Psi$, and the second one provides the superconducting current $\vec{J_s}$ which is the dissipation-less electrical current. The Ginzburg-Landau theory actually can be derived from the BCS microscopic theory. Thus, the electrons that contribute to superconductivity would form a superfluid and $|\Psi|^2$ indicates the fraction of electrons condensed into a superfluid.
This phenomenological theory can give many useful information even in homogeneous case. Let us consider a homogeneous superconductor with no superconducting current, so the equation for $|\Psi|$ simplifies to $$\alpha\Psi+\beta|\Psi|^2\Psi=0.$$ Above the superconducting transition temperature, $T>T_c$, one only gets a trivial solution $\Psi=0$, which corresponds to the normal state of the superconductor. Below the critical temperature, $T<T_c$, apart from the trivial solution, there are a series of non-trivial solutions which read $$\label{break}
|\Psi_0|=\sqrt{-\frac{\alpha}{\beta}}=\sqrt{\frac{\alpha_0(T_c-T)}{\beta}}.$$ Furthermore, compared with $\Psi=0$, those solutions have lower potential energy, thus are dominant. Note that there are infinite solutions giving the ground state of superconducting phase. However, the true ground state can only choose one solution from them. Therefore, the ground state will change under the U(1) transformation . In such case, we call that the U(1) symmetry is spontaneously broken. From one sees that $\Psi$ approaches zero as $T$ gets closer to $T_c$ from below, which is a typical behaviour of a second order phase transition. One can find from that in the homogeneous case one can neglect the contribution from the first term and thus the superconducting current is proportional to the vector potential, i.e., $\vec{J_s}\propto\vec{A}$. If one takes a time derivative on both sides, one will obtain $\vec{E}=-\partial_t\vec{A}\propto\partial_t\vec{J_s}$. This means that the electric fields accelerate superconducting electrons resulting in the infinite DC conductivity. If one takes the curl and combines with Maxwell’s equations, one will find $\nabla^2\vec{B}\propto\vec{B}$ indicating the decay of magnetic fields inside a superconductor, i.e., the Meissner effect.
The Ginzburg-Landau equations predict two characteristic lengths in a superconductor. The first one is the coherence length $\xi$ which is given by $$\label{glcoherence}
\xi=\sqrt{\frac{1}{2m^*|\alpha|}}.$$ It is the characteristic exponent of the variations of the density of superconducting component. In the BCS theory $\xi$ denotes the characteristic Cooper pair size. The other one is the penetration length $\lambda$ which reads $$\label{glpentration}
\lambda=\sqrt{\frac{m^*}{\mu_0 e^{*2}|\Psi_0|^2}}=\sqrt{\frac{m^* \beta}{\mu_0 e^{*2}|\alpha|}},$$ where $\Psi_0$ is the equilibrium value of the order parameter in the absence of electromagnetic fields. This length characterises the speed of exponential decay of the magnetic field at the surface of a superconductor.
Note that from definitions and the temperature dependences near $T_c$ behave as $$\begin{aligned}
\xi & \propto& (T_c-T) ^{-1/2},\\
\lambda &\propto& (T_c-T) ^{-1/2}.\end{aligned}$$ Both diverge as $T\rightarrow T_c$ from below with the critical exponent $1/2$. Nevertheless, the ratio $\kappa=\lambda/\xi$ known as the Ginzburg-Landau parameter is temperature independent. Type-I superconductors correspond to cases with $0<\kappa<1/\sqrt{2}$, and type-II superconductors correspond to cases with $\kappa>1/\sqrt{2}$. One of the most important findings from the Ginzburg-Landau theory was that in a type-II superconductor, strong enough magnetic fields can penetrate the superconductor by forming the hexagonal lattice of quantised tubes of flux, called the Abrikosov vortex lattice.
Finally one point we would like to emphasize is that in the Ginzburg-Landau theory the U(1) symmetry is broken spontaneously in the superconducting phase transition. Actually, only the spontaneous symmetry breaking feature itself can lead to many fundamental phenomenological properties of superconductivity, without any precise detail of the breaking mechanism specified [@Weinberg:1986]. In this review, we will show how a similar effective approach constitutes the basis of superconductivity in terms of holographic description.
Holographic duality
-------------------
The original conjecture proposed by Maldacena [@Maldacena:1997re] was that type-IIB string theory on the product spacetime $AdS_5 \times S^5$ should be equivalent to $\mathcal{N}= 4$ SU(N) supersymmetric Yang-Mills theory on the 3+1 dimensional boundary. This super-Yang-Mills theory is a conformal field theory, so this duality is named AdS/CFT correspondence. Later, this conjecture has been generalized to more general gravitational backgrounds and cases without supersymmetry and conformal symmetry.[^8] From a modern perspective, the correspondence is an equality between a quantum field theory (QFT) in d dimensional spacetime and a (quantum) gravity theory in d+1 spacetime dimensions. This correspondence is also sometimes called gauge/gravity duality, gauge/string duality or holographic correspondence (or duality).
A remarkable usefulness of the correspondence comes from the fact that it is a strong-weak duality: when the quantum field theory is strongly coupled, the dual gravitational theory is in a weakly interacting regime and thus more mathematically tractable, and vice versa. So the holographic duality provides us a powerful toolkit for studying strongly interacting systems. The backbone of the correspondence was elaborated by the authors of refs. [@Gubser:1998bc; @Witten:1998qj]. For every gauge invariant operator $\mathcal{O}$ in the QFT, there is a corresponding dynamical field $\Phi$ in the bulk gravitational theory. The partition function in gravity side is equal to the generating functional of the dual boundary field theory. More specifically, adding a source $J$ for $\mathcal{O}$ in the QFT is equivalent to impose a boundary condition for the dual field $\Phi$ at the boundary of the gravity manifold (say at $z\rightarrow 0$), i.e., the field $\Phi$ tends towards the value $\Phi\rightarrow\phi_0= J$ at the boundary up to an overall power of $z$. The formula reads $$\label{gkpw}
Z_{{\rm bulk}}[\Phi\rightarrow \phi_0=J]=\left<{\rm exp}\left(i\int\sqrt{-g_0}\;d^dx\; J\mathcal{O}\right)\right>_{\rm QFT},$$ where $g_0$ is the determinant of the background metric of dual field theory. If we want to study a strongly coupled field theory, we can translate it into a weakly coupled gravity system. In the semiclassical limit, the partition function is equal to the on-shell action of the bulk theory and thus one only needs to solve particular differential equations of motion. Therefore we can compute expectation values and correlation functions of the operator $\mathcal{O}$ in the (strongly coupled) QFT by differentiating the left side with respect to $J=\phi_0$.
In order to get familiar with the calculation by holography, let us consider $\mathcal{O}$ as a scalar operator which is dual to the bulk scalar field also denoted as $\Phi$. The minimal bulk action for $\Phi$ is given by $$\label{acscalar}
S_0=\int d^{d+1}x\sqrt{-g}\left[-\frac{1}{2}(\partial\Phi)^2-\frac{1}{2}m^2\Phi^2\right].$$ For illustration the gravity background is fixed as the pure $AdS_{d+1}$ in Poincáre coordinates $$ds^2=\frac{L^2}{z^2}(dz^2-dt^2+d\vec{x}\cdot d\vec{x}),$$ together with a profile for the scalar field $\Phi=\Phi(z,t,\vec{x})$. Here $\vec{x}$ are $d-1$ spatial coordinates, $t$ is a timelike coordinate and $z$ is the radial spatial coordinate. $L$ is known as AdS radius and the conformal boundary of AdS is located at $z\rightarrow 0$. Note that the geometry is invariant under the scaling transformation $(z, t, \vec{x})\rightarrow(\lambda z,\lambda t, \lambda \vec{x})$. Actually, the full isometry group of $AdS_{d+1}$ is identical to the conformal group in d dimensional boundary spacetime.
To calculate the on-shell action for $\Phi$, we need to solve the equation of motion derived from action . Working in Fourier space $\Phi(z, t, \vec{x})\rightarrow\Phi(z, \omega, \vec{k})=\Phi(z) e^{-i\omega t+i\vec{k}\cdot\vec{x}}$, we obtain $$\partial^2_z\Phi(z)-\frac{d-1}{z}\partial_z\Phi(z)-(k^2+\frac{m^2L^2}{z^2})\Phi(z)=0,\quad k^2=-\omega^2+\vec{k}^2.$$ Near the boundary $z\rightarrow 0$, the above equation admits the general asymptotic solution, $$\Phi(z, \omega, \vec{k})\sim A(k)\; z^{d-\Delta}+B(k)\; z^\Delta, \quad z\rightarrow 0,$$ with $\Delta=d/2+\sqrt{m^2L^2+d^2/4}$.[^9] The relation between A and B is determined by the interior of AdS. Making Fourier transformation back into real space, we then obtain $$\label{asypsa}
\Phi(z, t, \vec{x})\sim A(t,\vec{x})\; z^{d-\Delta}+B(t,\vec{x})\; z^\Delta, \quad z\rightarrow 0,$$
We now try to identify which term can be considered as the source $J$ of the dual operator $\mathcal{O}$. It turns out that, as long as $m^2L^2>-d^2/4+1$, the mode $A$ is non-normalizable with respect to the inner product $$\label{innerp}
(\Phi_1,\Phi_2)=-i\int_{\Sigma_t}dz d\vec{x}\sqrt{-g}g^{tt}(\Phi^*_1\partial_t\Phi_2-\Phi_2\partial_t\Phi^*_1),$$ with $\Sigma_t$ a constant-$t$ slice. The $B$ mode in this case is normalizable. We identify the coefficient $A$ as the source term, i.e., $$J(t, \vec{x})=\phi_0(t, \vec{x})=A(t, \vec{x})=\lim_{z\rightarrow 0} z^{\Delta-d}\;\Phi (z, t, \vec{x}).$$ This means that the on-shell action, and thus the partition function in AdS is a functional of $J(t, \vec{x})$. We can now argue that the scaling dimension of $\mathcal{O}$ is $\Delta$ without further calculation. Consider the scale transformation $(z, t, \vec{x})\rightarrow\lambda(z, t, \vec{x})$, the scalar $\Phi$ under such operation transforms as $\tilde{\Phi}(\lambda z,\lambda t,\lambda\vec{x})=\Phi(z,t,\vec{x})$. So the source term $J=A$ must transform as $\tilde{J}(\lambda t,\lambda\vec{x})=\lambda^{\Delta-d}J(t,\vec{x})$, and thus according to $\mathcal{O}$ transforms as $\mathcal{\tilde{O}}=\lambda^{-\Delta} \mathcal{O}$ which suggests the dimension of $\mathcal{O}$ should be $\Delta$.
Calculating the on-shell bulk action in terms of the solution with asymptotic expansion , one will find that the on-shell action will diverge near the boundary $z\rightarrow 0$. This divergence is interpreted as dual to UV divergences of the boundary field theory. Actually, the infrared (IR) physics of the bulk near the boundary corresponds to the ultraviolet (UV) physics of dual QFT, and vice versa. This is called UV/IR relation [@Peet:1998wn] and the radial direction $z$ plays the role of energy scale in the dual boundary theory. Physical processes in the bulk occurring at different radial positions correspond to different field theory processed with energies which scale as $E\sim 1/z$.
The divergence can be cured by adding local counter terms at the boundary, known as holographic renormalization [@Balasubramanian:1999re; @Bianchi:2001de; @Bianchi:2001kw]. For the present case the counter term to be introduced is $$S_{ct}=\frac{\Delta-d}{2L}\int_{z\rightarrow 0} d^dx\sqrt{-\gamma}\; \Phi^2,$$ where $\gamma$ is the determinant of the induced metric at the boundary. So the renormalized on-shell action should be $S^{ren}=S_0+S_{ct}$. We can then compute the expectation value by using the basic formula . That relation implies $$\left<\mathcal{O}\right>=-i\frac{\delta Z_{bulk}[J]}{\delta J}\sim\frac{\delta S^{ren}[J]}{\delta J},$$ where we have taken the semiclassical limit $Z_{bulk}=e^{i S^{ren}}$. Straightforward calculation shows that $$\left<\mathcal{O}\right>(t,\vec{x})=\frac{2\Delta-d}{L} B(t,\vec{x}).$$ This is often summarised as saying that the “non-normalizable" mode $A$ gives the source in the dual field theory, whereas the “normalizable" mode $B$ encodes the response.
In the real world, many important experimental processes such as transport and spectroscopy involve small time dependent perturbations about equilibrium. Those phenomena can be described by linear response theory, in which the basic quantity is the retarded Green’s function. The retarded Green’s function is defined to linearly relate sources and corresponding expectation values. In frequency space, it can be written as $$\delta \left<\mathcal{O}\right>(\omega,\vec{k})=G^{R}(\omega,\vec{k})\; \delta J(\omega,\vec{k}).$$ Using above formula one can continue to compute $G^R(\omega,\vec{k})$ which is given by $$G^R(\omega,\vec{k})=\frac{2\Delta-d}{L} \frac{B(\omega,\vec{k})}{J(\omega,\vec{k})}.$$ However, there is something subtle we shall discuss here. Different from the case in Euclidean signature where the bulk solution can be uniquely determined by additional requirement of regularity in the IR, while in the real-time Lorentzian signature, we must choose an appropriate boundary condition in far IR region of the geometry. This ambiguity reflects multitude of real-time Green’s functions (Feynman, retarded, advanced) in the QFT. Since the retarded Green’s function describes causal response of the system to a perturbation, we involve an in-going condition describing stuff falling into the IR, i.e., moving towards larger $z$ as time passes. The advanced Green’s function corresponds to the choice of out-going condition enforced in the IR region. [^10]
Let us briefly consider the case with $-d^2/4< m^2L^2<-d^2/4+1$ where the second restriction comes from the unitary bound. One can easily check that both terms in are normalizable with the inner product . So either one can be considered as a source, and the other one as a response. These two ways to quantize a scalar field in the bulk by imposing Dirichlet or Neumann like boundary conditions correspond to two different dual field theories [@Klebanov:1999tb], respectively. In the standard quantization, the corresponding operator has dimension $\Delta$, while in the alternative quantization, the corresponding operator has dimension $d -\Delta$. [^11]
The above discussion only uses the near boundary expansion and thus applies to generic asymptotically AdS geometries. It can also be applied to other fields such as components of the metric and Maxwell fields. To sum up, we first obtain a solution which satisfies appropriate boundary conditions, especially the condition in the deep IR. Then we compute the properly renormalized on-shell action, identify the source and response from the asymptotic behaviour of the solution near the boundary, and compute the Green’s function through linear response. In the next section we will use this procedure to compute the optical conductivity.
Another essential entry in the holographic dictionary is that the thermodynamic data of the QFT is entirely encoded in the thermodynamics of the black hole in the dual geometry. QFT states with finite temperature are dual to black hole geometries, where the Hawking temperature of the black hole is identified with the temperature in the QFT. Turning on a chemical potential in this QFT corresponds to gravity with a conserved charge. The thermal entropy of QFT is identified as the area of black hole horizon and the free energy is related to the Euclidian on-shell bulk action. As space is limited, we only introduce essential issues which will be needed to discuss holographic superconductors.
Before the end of this subsection, let us point out that the holographic duality can be used to understand some hard nuts in quantum gravity from dual field theory side. A typical example is the black hole information paradox. It was first suggested by Hawking [@Hawking:1976] that black holes destroy information which seemed to conflict with the unitarity postulate of quantum mechanics. The black hole information paradox can be resolved, at least to some extent, by holography, because it shows how a black hole can evolve in a manner consistent with quantum mechanics in some contexts, i.e., evolves in a unitary fashion [@Hawking:2005kf; @Hawking:2014tga]. There are some excellent papers talking about aspects of holographic duality, see, for example, refs. [@Aharony:1999ti; @Hartnoll:2009sz; @McGreevy:2009xe; @CasalderreySolana:2011us; @Adams:2012th; @Sachdev:2011wg] for more details.
Holographic S-wave Models {#sect:higsmodel}
=========================
The Abelian-Higgs model
-----------------------
In this subsection, we begin with the Abelian-Higgs model [@Hartnoll:2008vx] by introducing a complex scalar field $\Psi$, with mass $m$ and charge $q$, into the $(3+1)$ dimensional Einstein-Maxwell theory with a negative cosmological constant. The complete action can be written down as [^12] $$\label{higsaction}
S =\frac{1}{2\kappa^2} \int d^4 x
\sqrt{-g} \left ( \mathcal{R}+\frac{6}{L^2}-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-|\nabla\Psi-i qA\Psi|^2-m^2 |\Psi|^2 \right),$$ where $ 2 \kappa^2=16\pi G $ with $G$ being the Newtonian gravitational constant, $\mathcal{R}$ is the scalar curvature of spacetime, $L$ is the AdS radius and Maxwell field strength $F_{\mu\nu}=\nabla_\mu A_\nu-\nabla_\nu A_\mu$. If one rescales $A_\mu\rightarrow A_\mu/q$ and $\Psi\rightarrow \Psi/q$, then the matter part has an overall factor $1/q^2$ in front of its Lagrangian, thus the back reaction of the matter fields on the metric becomes negligible when $q$ is large. The limit $q\rightarrow\infty$ with $q A_\mu$ and $q\Psi$ fixed is called the probe limit. Here we will review the results obtained in this probe approximation, which can simplify the problem while retains most of the interesting physics. The study including the back reaction of matter fields can be found in ref. [@Hartnoll:2008kx].
The background metric is the AdS-Schwarzschild black hole with planar horizon $$\label{AdSswtz}
ds^2=-f(r)dt^2+\frac{dr^2}{f(r)}+r^2(dx^2+dy^2),~~~f(r)=r^2(1-r_h^3/r^3),$$ where we have set the AdS radius $L$ to be unity. The conformal boundary is located at $r\rightarrow\infty$. The Hawking temperature of the black hole is determined by the horizon radius $r_h$: $T=3r_h/4\pi$. The solution describes a thermal state of dual field theory in (2+1)-dimensions with temperature $T$. In addition, it is clear that the AdS-Schwarzschild black hole is an exact solution of the action (\[higsaction\]) when the matter sector is negligible. It will be seen shortly that when the temperature is lowered enough, the black hole solution will become unstable and a new stable black hole solution appears with nontrivial scalar field.
To see the formation of scalar hair, we are interested in static, translationally invariant solutions, thus we consider the ansatz [@Hartnoll:2008vx] $$\begin{aligned}
\Psi=\psi(r)\;, \ \ \ A=\phi(r)\,dt\;.\end{aligned}$$ The $r$ component of Maxwell equations implies that the phase of $\psi$ must be constant. Therefore, for convenience, one can take $\psi$ to be real. This leads to the equations of motion [^13] $$\label{higsEOMs}
\begin{split}
\psi''+(\frac{f'}{f}+\frac{2}{r})\psi'+\frac{\phi^2}{f^2}\psi-\frac{m^2}{f}\psi=0,\\
\phi''+\frac{2}{r}\phi'-\frac{2\psi^2}{f}\phi=0.
\end{split}$$ As pointed out in ref. [@Gubser:2008px], the coupling of the scalar to the Maxwell field produces a negative effective mass for $\psi$ (see the third term in the first equation). Since this term becomes more important at low temperatures, we expect an instability towards forming nontrivial scalar hair.
The asymptotic behaviours of scalar field and gauge field near the AdS boundary are $$\begin{aligned}
\psi=\frac{\psi_-}{r^{\Delta_-}}+\frac{\psi_+}{r^{\Delta_+}}+\cdots,\ \ \ \phi=\mu-\frac{\rho}{r}+\cdots,\end{aligned}$$ where $\Delta_{\pm}=(3\pm\sqrt{3^2+4m^2})/2$, $\mu$ is the chemical potential and $\rho$ is the charge density in the dual field theory. According to the AdS/CFT dictionary, the leading coefficient $\psi_-$ is regarded as the source of the dual scalar operator $\mathcal{O}$ with scaling dimension $\Delta_+$. Since we want the U(1) symmetry to be broken spontaneously, we should turn off the source, i.e., $\psi_-=0$. Therefore the subleading term $\psi_+$ provides the vacuum expectation value $\langle\mathcal{O}\rangle$ in the absence of any source. [^14]
![\[sconden\] The condensate as a function of temperature. The critical temperature $T_c$ is proportional to $\sqrt{\rho}$. We choose $m^2=-2$.](scondensate.pdf)
Figure \[sconden\] shows how the condensate $\langle\mathcal{O}\rangle$ behaves as a function of temperature in a canonical ensemble with $\rho$ fixed to be one. As one can see that there is a critical temperature $T_c$ below which the condensate appears, then rises quickly as the system is cooled and finally goes to a constant for sufficiently low temperatures. This behaviour is qualitatively similar to that obtained in BCS theory and observed in many materials. Near the critical temperature $T_c$, $\langle\mathcal{O}\rangle\sim (T_c-T)^{1/2}$, which is the typical result predicated by Ginzburg-Landau theory, see equation . By comparing the free energy of these hairy configurations to the solution $\psi=0, \phi=\rho(1/r_h-1/r)$ with no scalar hair, one finds that the hairy phase is thermodynamically favoured and the difference of free energies behaves like $(T_c-T)^2$ near the critical point, indicating a second order phase transition.
We now compute the optical conductivity, i.e., the conductivity as a function of frequency $\omega$, which is related to the retarded current-current two-point function for the U(1) symmetry, $\sigma(\omega)=\frac{1}{i\omega}G^{R}(\omega,\vec{k}=0)$. According to the holographic duality, this can be obtained by calculating electromagnetic fluctuations in the bulk. By symmetry, it is sufficient to turn on the perturbation $\delta A=A_x(r)e^{-i\omega t}dx$, then the linearized equation of motion for $A_x$ is $$A_x''+\frac{f'}{f}A_x'+(\frac{\omega^2}{f^2}-\frac{2\psi^2}{f})A_x=0\;.$$ To obtain the real time correlation functions for the dual boundary theory, the holographic description associates in-going and out-going boundary conditions at the black hole horizon to retarded and advanced boundary correlators respectively [@Son:2002sd]. To consider causal behaviour, one should impose the in-going wave condition at the horizon: $A_x\sim f^{-i\omega/3r_h}$. Near the AdS boundary, the asymptotic behaviour of $A_x$ is given by $$A_x=A^{(0)}+\frac{A^{(1)}}{r}+\cdots.$$ According to the AdS/CFT correspondence, $A^{(0)}$ is the source, while $A^{(1)}$ is dual to the current. Thus one can obtain $$\sigma(\omega)=\frac{1}{i\omega}G^R(\omega)=\frac{1}{i\omega}\frac{A^{(1)}}{ A^{(0)}}\;.$$
![\[sconduc\] The optical conductivity as a function of frequence. The solid lines in the left plot are the real part of the conductivity, while the dashed lines in the right plot are the imaginary part of the conductivity. We choose $m^2=-2$. The horizontal lines correspond to temperatures above $T_c$. Other curves from the left to right correspond to $T/T_c\simeq0.888$ (blue), $T/T_c\simeq0.222$ (green) and $T/T_c\simeq0.105$ (red), respectively. There is a delta function at the origin for the real part of the conductivity in the condensed phase.](scondre.pdf "fig:") ![\[sconduc\] The optical conductivity as a function of frequence. The solid lines in the left plot are the real part of the conductivity, while the dashed lines in the right plot are the imaginary part of the conductivity. We choose $m^2=-2$. The horizontal lines correspond to temperatures above $T_c$. Other curves from the left to right correspond to $T/T_c\simeq0.888$ (blue), $T/T_c\simeq0.222$ (green) and $T/T_c\simeq0.105$ (red), respectively. There is a delta function at the origin for the real part of the conductivity in the condensed phase.](scondim.pdf "fig:")
The AC conductivity as a function of frequency is presented in figure \[sconduc\]. Above the critical temperature, the conductivity is a constant. As the temperature is lowered below $T_c$, the optical conductivity develops a gap at some special frequency $\omega_g$ known as gap frequency. As suggested in ref. [@Horowitz:2008bn], it can be identified with the one at the minimum of the imaginary part of the AC conductivity. Re$[\sigma(\omega)]$ is very small in the infrared and rises quickly at $\omega_g$.[^15] There also exists a small “bump" slightly above $\omega_g$, which is reminiscent of the behaviour due to fermionic pairing [@Gubser:2008wv]. For different choice of parameters, one can obtain a robust feature $\omega_g\simeq8T_c$ with deviations of less than $10\%$. Compared to the corresponding BCS value $\omega_g\simeq3.5T_c$, the result shown here is consistent with the fact that the holographic model describes a system at strong coupling. There is also a delta function at $\omega=0$ appearing as soon as $T<T_c$. This can be seen from the imaginary part of the conductivity. According to the Kramers-Kronig relation $$\label{kkrelation}
\text{Im}[\sigma(\omega)] = - \frac{1}{\pi} {\mathcal{P}}\int_{-\infty}^{\infty} \frac{\text{Re}[\sigma(\omega')] d\omega'}{\omega'-\omega},$$ one can conclude that the real part of the conductivity contains a Dirac delta function at $\omega=0$ if and only if the imaginary part has a pole, i.e., Im$(\sigma)\sim1/\omega$.
From above discussion, we see that this simple model can provide a holographically dual description of a superconductor. It predicts that a charged condensate emerges below a critical temperature via a second order transition, that the DC conductivity becomes infinite, and that the optical conductivity develops a gap at low frequency. The temperature dependences of the coherence length $\xi$ as well as the penetration length $\lambda$ in the holographic model are both proportional to $(T_c-T)^{-1/2}$ near the critical temperature [@Horowitz:2008bn; @Maeda:2008ir]. It has been shown that this holographic superconductor is type-II [@Hartnoll:2008kx]. The condensate can form a lattice of vortices and the minimum of the free energy at long wavelength corresponds to a triangular array [@Maeda:2009vf]. The effects of a superconducting condensate on holographic Fermi surfaces have been studied [@Faulkner:2009am; @Bagrov:2014mqa]. All these features are very reminiscent of real superconductors. Although the holographic model is very simple, it indeed captures some significant characteristics for superconductivity, thus helping us to understand real, strongly coupled superconductors.
Holographic insulator/superconductor phase transition
-----------------------------------------------------
In this subsection, let us consider a five-dimensional Einstein-Abelian-Higgs theory with following action $$\label{5dhigsaction}
S= \frac{1}{2\kappa^2} \int d^5x \sqrt{-g} \left( {\cal R} +\frac{12}{L^2} -\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-|\nabla_{\mu}\Psi -i q A_{\mu} \Psi |^2 -m^2 |\Psi|^2\right).$$ When one does not include the matter sector, the theory has a five-dimensional AdS-Schwarzschild black hole solution. It is interesting to note that there also exists another exact solution, so-called AdS soliton, in the theory (\[5dhigsaction\]). The AdS soliton solution can be obtained by double Wick rotation from the AdS-Schwarzschild black hole as $$\begin{aligned}
\label{soliton}
ds^2 = f(r)d\chi^2+\frac{dr^2}{f(r)}+r^2(-dt^2+dx^2+ dy^2), \quad f(r) = r^2(1-r_0^4/r^4).\end{aligned}$$ To remove the potential conical singularity, the spatial coordinate $\chi$ has to be periodic with a period $\pi/r_0$. If one considers the coordinates $(\chi, r)$, the geometry looks like a cigar and the tip is given by $r=r_0$. The AdS soliton has no horizon, and therefore no entropy is associated with this solution. Due to the existence of an IR cutoff at $r=r_0$ for the soliton solution, the field theory dual to this gravity background turns out to be in confined phase at zero temperature. Furthermore, this solution can be explained as a gravity dual to an insulator in condensed matter theory. If one increases the chemical potential to a critical value, the AdS soliton solution becomes unstable to developing a scalar hair with nontrivial scalar profile. It is shown that the new solution can describe a superconducting phase [@Taka]. In this way, the holographic insultor/superconductor phase transition at zero temperature can be realized in the Abelian-Higgs model (\[5dhigsaction\]).
More precisely, let us also consider the following ansatz in the probe limit $$\Psi = \psi (r), \ \ \ A_{\mu}= \phi (r) dt.$$ In the AdS soliton (\[soliton\]) background, the equations of motions turn out to be $$\begin{aligned}
\begin{split}
& \psi'' +\left( \frac{f'}{f}+\frac{3}{r}\right) \psi' -\left( \frac{m^2}{f}-\frac{q^2 \phi^2}{r^2 f}\right) \psi =0, \\
& \phi'' +\left( \frac{f'}{f}+\frac{1}{r}\right) \phi' -\frac{2 q^2 \Psi^2}{f}\phi =0.
\end{split}\end{aligned}$$ In the five-dimensional case, the BF bound is $m^2_{BF}=-4$. For simplicity, let us consider the case with $m^2=-15/4$. To solve the equations of motion, we have to specify the boundary conditions both at the tip and the AdS boundary. Near the AdS boundary, we have the following asymptotical form $$\psi = \frac{\psi_-}{r^{3/2}} +\frac{\psi_+}{r^{5/2}}+ \cdots, \quad \phi = \mu -\frac{\rho}{r^2}+ \cdots.$$ Note that in this case, both terms proportional to $\psi_-$ and $\psi_+$ are normalizable, so the corresponding operators ${\cal O}_1$ and ${\cal O}_2$ have dimensions $\triangle =3/2$ and $\triangle =5/2$, respectively. On the other hand, near the tip of the soliton, these fields behave like $$\begin{aligned}
\label{Newmann}
\begin{split}
&\psi= a +b \log (r-r_0)+ c (r-r_0) +\cdots, \\
&\phi= A +B \log (r-r_0) +C (r-r_0)+ \cdots,
\end{split}\end{aligned}$$ where $a, b, c$ and $A, B, C$ are all constants. The field regularity at the tip requires us to take $b=B=0$. As in the previous subsection we can set $q=1$ and further set $r_0=1$ without loss of generality.
![\[Condensate\] The behaviour of condensation for the operator $\langle {\cal O}_1 \rangle$ (left) and $\langle {\cal O}_2 \rangle$ (right) with respect to chemical potential. Used with permission from ref. [@Taka].](CondenseO1.pdf "fig:") ![\[Condensate\] The behaviour of condensation for the operator $\langle {\cal O}_1 \rangle$ (left) and $\langle {\cal O}_2 \rangle$ (right) with respect to chemical potential. Used with permission from ref. [@Taka].](CondenseO2.pdf "fig:")
With the boundary conditions, solving the equations of motion, one can find that when the chemical potential $\mu$ is beyond some critical value, the condensation happens. Concretely, for the operator ${\cal O}_1$, the critical chemical potential is $\mu_1=0.84$, while the critical chemical potential $\mu_2= 1.88$ for the operator ${\cal O}_2$. The behaviour of condensation is plotted in figure \[Condensate\] with respect to chemical potential. In figure \[rhomu\] the charge density $\rho$ with respect to chemical potential is plotted. We can see that at the phase transition point, its derivative is discontinuous, which verifies that the phase transition is indeed second order, since one has $\rho =\partial \Omega/\partial \mu$, where $\Omega$ is the Gibbs free energy density.
![ \[rhomu\] The charge density $\rho$ as a function of $\mu$ when $\langle {\cal O}_1\rangle \neq 0$ (left) and $\langle {\cal O}_2\rangle \neq 0$ (right). Its derivative jumps at the phase transition point. Used with permission from ref. [@Taka].](2ndOrderO1.pdf "fig:") ![ \[rhomu\] The charge density $\rho$ as a function of $\mu$ when $\langle {\cal O}_1\rangle \neq 0$ (left) and $\langle {\cal O}_2\rangle \neq 0$ (right). Its derivative jumps at the phase transition point. Used with permission from ref. [@Taka].](2ndOrderO2.pdf "fig:")
To calculate conductivity $\sigma (\omega)$ we can consider the perturbation of the component $A_x$ in the soliton background. Assuming it has the form $A_x \sim e^{-i\omega t}$, its equation then turns out to be $$A''_x +\left ( \frac{f'}{f}+\frac{1}{r}\right) A'_x +\left (\frac{\omega^2}{r^2f}- \frac{2 q^2 \Psi^2}{f}\right) A_x=0.$$ At the tip one takes the Newmann boundary condition as in (\[Newmann\]). Near the AdS boundary, one has the asymptotical form as $$A_x = A_x^{(0)}+ \frac{A_x^{(1)}}{r^2}+\frac{A_x^{(0)}\omega^2}{2}\frac{ \log \Lambda r}{r^2}+ \cdots,$$ where $\Lambda$ is a cutoff. The holographic conductivity can be obtained as $$\sigma(\omega)= -\frac{2 i A_x^{(1)}}{\omega A_x^{(0)}}+ \frac{i \omega}{2}.$$ Since the background has no horizon, the real part of the conductivity always vanishes. This means that there is no dissipation. The imaginary part is plotted in figure \[BubbleSC\]: The left plot corresponds to the case of pure AdS solution without scalar hair, while the right one to the case with nontrivial scalar hair. There exist poles periodically at the points where $A_x^{(0)}$ vanishes. These correspond to normalized modes dual to vector operators. One can see that when $\omega$ is large, both case are similar, while when $\omega \to 0$, they are quite different. In the case without condensation, the imaginary part goes to zero when $\omega \to 0$, while it diverges in the case with condensation. According to the Kramers-Kronig relation , it shows that there is a delta functional support for the real part of conductivity at $\omega=0$. Therefore the AdS soliton background with nontrivial scalar hair should be identified with the superconductivity.
![\[BubbleSC\] The imaginary part of the conductivity for the AdS soliton without condensation $\langle {\cal O}_{1,2} \rangle =0$ (left) and with condensation $\langle {\cal O}_1\rangle \neq 0$ (right). $\rho=0.0094$ and $\mu=0.84$ are taken in the right plot. Used with permission from ref. [@Taka].](BubbleInsulator.pdf "fig:") ![\[BubbleSC\] The imaginary part of the conductivity for the AdS soliton without condensation $\langle {\cal O}_{1,2} \rangle =0$ (left) and with condensation $\langle {\cal O}_1\rangle \neq 0$ (right). $\rho=0.0094$ and $\mu=0.84$ are taken in the right plot. Used with permission from ref. [@Taka].](BubbleSC2.pdf "fig:")
In the Einstein-Abelian-Higgs theory (\[5dhigsaction\]), besides the two phases described above, as in the four dimensional case, there exist another two solutions: AdS Reissner Nordström (AdS RN) black hole[^16] without scalar hair and AdS RN black hole with scalar hair, the latter can be identified with a superconductivity phase, while the former is dual to a conductor phase. Combining the four phases together, one could have the phase digram of the theory, which is schematically plotted in figure \[schematicPhase\]. The green line in the figure denotes the first order phase transition, while two red lines represent second order phase transition. Considering back reaction of matter sector, the complete phase diagrams in terms of temperature and chemical potential for the Abelian-Higgs model have been constructed in ref. [@Horowitz:2010jq]. It is interesting to note that the behaviour of the entanglement entropy with respect to chemical potential is non-monotonic and seems to be universal in this kind of insulator/superconductor models [@Cai:2012sk; @Cai:2012es; @Yao:2014fua].
![\[schematicPhase\] The phase diagram of AdS soliton and AdS black hole with a charged scalar field obtained in the large $q$ limit. Used with permission from ref. [@Taka].](PhaseGR.pdf)
Holographic P-wave Models {#sect:p-wave}
=========================
The SU(2) Yang-Mills P-wave model {#sect:yangmills}
---------------------------------
The first holographic p-wave model is constructed by introducing a SU(2) Yang-Mills field in asymptotically AdS spacetime. One of three U(1) subgroups is regarded as the gauge group of electromagnetism and the off-diagonal gauge bosons which are charged under this U(1) gauge field are supposed to condense outside the horizon. The full action is given by [@Gubser:2008wv] $$\label{su2action}
S =\int d^4 x
\sqrt{-g}[\frac{1}{2\kappa^2}(\mathcal{R}+\frac{6}{L^2})-\frac{1}{4\hat{g}^2}
F^a_{\mu\nu} F^{a\mu \nu}],$$ where $\kappa$ is the four dimensional gravitational constant, $\hat{g}$ is the Yang-Mills coupling constant and $L$ is the AdS radius. The field strength for the SU(2) gauge field $A=A^a_{\mu}\tau^a dx^{\mu}$ is $$\label{supstrength}
F^a_{\mu\nu}=\partial_\mu A^a_\nu-\partial_\nu A^a_\mu + \epsilon^{abc}A^b_\mu
A^c_\nu,$$ where $\mu,\nu=(t,r,x,y)$ denote the indices of spacetime and $a,b,c=(1,2,3)$ are the indices of the SU(2) group generators $\tau^a=\sigma^a/2i$ ($\sigma^a$ are Pauli matrices). $\epsilon^{abc}$ is the totally antisymmetric tensor with $\epsilon^{123}=+1$.
Note that the ratio $\kappa/\hat{g}$ measures the influence of Yang-Mills field on the background geometry. For the case $\kappa/\hat{g}\ll 1$ with $A^a_\mu$ fixed, the back reaction of the matter field can be ignored, thus the metric is simply $AdS_4$ Schwarzschild black hole $$ds^2=\frac{r^2}{L^2}\left[-\left(1-\frac{r_h^3}{r^3}\right)dt^2+dx^2+dy^2\right]+\frac{L^2}{r^2}\frac{dr^2}{1-r_h^3/r^3},$$ with the temperature given by $T=\frac{3r_h}{4\pi L^2}$. Without loss of generality, we shall choose $L=1$, and we also fix a scale by setting $r_h=1$.
### Vector condensate
To realize the p-wave condensate, one takes the ansatz $$A=\phi(r)\tau^3 dt+\varpi(r)\tau^1 dx\;.$$ It is clear that the non-trivial profile of $\varpi(r)$ picks out the $x$ direction as special, thus the condensed phase breaks the gauge group $\text{U}(1)_3$ generated by $\tau^3$ and $SO(2)$ rotational symmetry in $x-y$ plane. The relevant equations are [@Gubser:2008wv] $$\label{su2eoms}
\begin{split}
\phi'' + {2 \over r} \phi' -{1 \over r(r^3-1)} \varpi^2 \phi &= 0,\\
\varpi'' + {1+2r^3 \over r(r^3-1)} \varpi'+{r^2 \over (r^3-1)^2} \phi^2 \varpi &= 0,
\end{split}$$ with primes representing the derivative with respect to $r$.
The regularity at the horizon $r=1$ demands the behaviour like $$\phi=\phi_1(r-1)+\cdots,\quad \varpi=w_0+w_2(r-1)^2+\cdots,$$ while the asymptotical expansion near the boundary $r\rightarrow\infty$ takes the form $$\phi=\mu-{\rho\over r}+\cdots,\quad \varpi=W_0+{W_1\over r}+\cdots.$$ According to the holographic dictionary, $\mu$ is regarded as chemical potential and $\rho$ is the total charged density, and $W_0$ is the source of the dual operator $J^x$. To spontaneously break the U(1) symmetry, we should impose $W_0=0$, then the coefficient $W_1$ gives the vacuum expectation value of $J^x$. According to the two-fluid model, the total charge density $\rho$ can be divided into two components $\rho=\rho_n+\rho_s$, where $\rho_n$ is the normal component, while $\rho_s$ is the superconducting component. In the holographic setup, the normal charge density $\rho_n$ is proportional to the $\tau^3$ part of the electric field at the horizon, i.e., $\rho_n=\phi_1$. Therefore the superconducting charge density is $\rho_s=\rho-\rho_n$.
![The fraction $\rho_s/\rho$ (right) and the condensate (left) as a function of temperature. Adapted with permission from ref. [@Gubser:2008wv].[]{data-label="sucondensate"}](sucondens.pdf)
By numerically solving the equations , one finds that the condensate is non-vanishing only when the rescaled temperature $T/\sqrt{\rho}$ is small enough, i.e., lower than $T_c$ at which the condensate first turns on. As one can see in the right plot of figure \[sucondensate\], as the temperature is lowered, $\langle J^x\rangle$ increases continuously. Near $T_c$, $\langle J^x\rangle$ vanishes as $\sqrt{T_c-T}$, which is the typical behaviour predicted by Ginzburg-Landau theory. The fraction $\rho_s/\rho$ of the charge carried by the superconducting condensate goes to zero linearly near $T_c$. [^17]
We have interpreted $\text{U}(1)_3$ generated by $\tau^3$ as the gauge group of electromagnetism. The condensate of $\langle J^x\rangle$ spontaneously breaks this U(1) symmetry as well as the rotational symmetry, thus resulting in an anisotropic superconducting phase. To see this much more clearly, we shall calculate the optical conductivity, which can be deduced by the retarded Green’s function of the $\text{U}(1)_3$ current. Similar to the previous section, in gravity side the linear response to electromagnetic probes is turned out to study how linear perturbations of the $\tau^3$ component of the gauge field propagate.
### Conductivity {#sect:suconducticity}
In the presence of the condensate $\varpi\tau^1 dx$, the $x$ direction is special, so the conductivity $\sigma_{xx}$ along the $x$ direction is expected to be different from $\sigma_{yy}$ along the $y$ direction. To obtain consistent linearized equations, we can turn on the perturbation [@Gubser:2008wv] $$\label{supertubation}
\delta A=e^{-i\omega t} \left[(a_t^1 \tau^1 + a_t^2 \tau^2) dt + a_x^3 \tau^3 dx + a_y^3 \tau^3 dy \right],$$ where all the functions depend on $r$ only. By plugging the perturbation into the linearized Yang-Mills equation, one finally obtains four second order equations $$\label{supertubationay}
{a_y^3}''+\frac{2r^3+1}{r(r^3-1)}{a_y^3}'+\left[\frac{\omega^2r^2}{(r^3-1)^2}-\frac{\varpi^2}{r(r^3-1)}\right]a_y^3=0,$$
\[supertubationax\] $$\begin{aligned}
{a_x^3}''+\frac{2r^3+1}{r(r^3-1)}{a_x^3}'+\frac{r^2}{(r^3-1)^2}\left(\omega^2a_x^3-\phi\varpi{a_t^1}-i\omega{a_t^2} \right)=0,\\
{a_t^1}''+\frac{2}{r}{a_t^1}'+\frac{\phi\varpi}{r(r^3-1)}{a_x^3}=0,\\
{a_t^2}''+\frac{2}{r}{a_t^2}'-\frac{\varpi}{r(r^3-1)}\left(\varpi {a_t^2}+i\omega {a_x^3}\right)=0,\end{aligned}$$
and two first order constraint equations $$\label{supertubationax1}
\begin{split}
i\omega{a_t^1}'+\phi {a_t^2}'-\phi'{a_t^2}=0,\\
i\omega{a_t^2}'-\phi{a_t^1}'-(1-\frac{1}{r^3})\varpi{a_x^3}'+\phi'{a_t^1}+(1-\frac{1}{r^3})\varpi'{a_x^3}=0.
\end{split}$$ It is clear that the equation of motion of the ${a_y^3}$ mode decouples from the others, and the conductivity $\sigma_{yy}$ exhibits similar “soft gap" behaviour to the s-wave model [@Hartnoll:2008vx].[^18] What we are interested in is the conductivity $\sigma_{xx}$ in the $x$ direction. The conductivity $\sigma_{xx}$ can be determined by solving the coupled equations with the constraints given by . More precisely, we impose the ingoing wave condition at the horizon, which corresponds to a retarded Green’s function, $$\label{suingoing}
\begin{split}
a_x^3 &= (r-1)^{-i\omega/4\pi T} \left[ 1 +a_x^{3(1)} (r-1) + a_x^{3(2)} (r-1)^2 + \cdots \right],\\
a_t^1 &= (r-1)^{-i\omega/4\pi T} \left[ a_t^{1(2)} (r-1)^2 + a_t^{1(3)} (r-1)^3 + \cdots \right],\\
a_t^2 &= (r-1)^{-i\omega/4\pi T} \left[ a_t^{2(1)} (r-1) + a_t^{2(2)} (r-1)^2 + \cdots \right],
\end{split}$$ where all the coefficients can be fixed once $w_0$, $\phi_1$ and $\omega$ are specified. Near the conformal boundary $r\rightarrow\infty$, one has a generic solution to the equations of motion $$\label{suboundaty}
\begin{split}
a_x^3 &= A_x^{3(0)} + {A_x^{3(1)} \over r} + \cdots,\\
a_t^1 &= A_t^{1(0)} + {A_t^{1(1)} \over r} + \cdots,\quad a_t^2= A_t^{2(0)} + {A_t^{2(1)} \over r} + \cdots.
\end{split}$$ As pointed out in ref. [@Gubser:2008wv], there exists a residual gauge invariance. After fixing this residual gauge freedom, one can finally obtain the gauge invariant conductivity along $x$ direction $$\sigma_{xx} = -{i\over\omega A_x^{3(0)}}\left(A_x^{3(1)}+W_1 {i\omega A_t^{2(0)}+\mu A_t^{1(0)} \over \mu^2 - \omega^2}\right)\;.$$ Numerical calculation can only display the continuous part of $\sigma_{xx}(\omega)$. One can reveal the non-analytic behaviour by virtue of the Kramers-Krong relations, which tells us that a simple pole in $\text{Im}[\sigma_{xx}(\omega)]$ at $\omega_0$ implies a delta function $\delta(\omega-\omega_0)$ to $\text{Re}[\sigma_{xx}(\omega)]$. Further more, the positivity constraint on the real part of conductivities requires any pole of $\text{Im}[\sigma_{xx}(\omega)]$ on the real axis to have a positive residue.
The behaviour of conductivities as a function of frequency $\omega$ is shown in figure \[suconductivity\], from which one can see the following features [@Gubser:2008wv]. First, both $\sigma_{xx}$ and $\sigma_{yy}$ approach constant for sufficiently large $\omega$. This is because the condensate involves dynamics with a characteristic energy scale set by $\sqrt{\rho}$. If $\omega\gg\sqrt{\rho}$, the propagation of the gauge boson should become insensitive to the condensate and can be approximated by the case in pure $AdS_4$, thus is a constant. Second, $\sigma_{yy}$ exhibits gapped dependence similar to the Abelian-Higgs model in figure \[sconduc\]. $\text{Re}(\sigma)$ is very small in the infrared, then rises quickly at $\omega =\omega_g \simeq \sqrt{\rho}$. There is a slight “bump” a little above $\omega_g$ which is reminiscent of the behaviour expected for fermionic pairing. Third, there is a pole in $\text{Im}[\sigma_{xx}]$ at $\omega=\omega_0\simeq1.8\sqrt{\rho}$. Therefore, there is a delta function contribution to $\text{Re}[\sigma_{xx}]$ at $\omega=\omega_0$. Finally, in the small $\omega$ region, $\text{Re}[\sigma_{xx}]$ can be well parameterized in terms of the Drude model $$\text{Re}[\sigma_{\text{Drude}}] = {\sigma_0 \over 1 + \omega^2 \tau^2}\;,$$ where $\sigma_0$ gives the DC conductivity and $\tau$ is the scattering time. The best fit gives a narrow Drude peak in $\sigma_{xx}$ and suggests conductivity due to quasi-particles with scattering time to diverge as $T\rightarrow0$.
![Conductivities $\sigma_{xx}$ and $\sigma_{yy}$ with respect to frequency at $T/\sqrt{\rho}\simeq0.0779$. The dotted curves are the best fits of the Drude model prediction to $\text{Re}~\sigma_{xx}(\omega)$. Adapted with permission from ref. [@Gubser:2008wv].[]{data-label="suconductivity"}](suconduc.pdf)
We do not have a microscopic description of the condensate in the language of the dual theory without gravity. However, we know clearly that there is an SU(2) current algebra, and the component $J^x$ develops an expectation for sufficiently large chemical potential. Yet we only turn on $\tau^1 dx$ mode corresponding to the p-wave background. The $(p+ip)$-wave case can be realized by involving the combination $\tau^1dx+\tau^2dy$. This mode results in an isotropic superconducting phase which exhibits a pseudogap[^19] at low temperatures and a nonzero Hall conductivity with no external magnetic field [@Roberts:2008ns]. However, it should be pointed out that $(p+ip)$ configurations are unstable against turning into pure p-wave background. The insulator/superconductor phase transition for the SU(2) p-wave model has been studied in ref. [@Akhavan:2010bf].
A new ground state can be found when a magnetic component of the gauge field is larger than a critical value, which forms a triangular Abrikosov lattice in the spatial directions perpendicular to the magnetic field [@Ammon:2011je; @Bu:2012mq]. In the same spirit, a p-wave superconductor for which the dual field is explicitly known has been constructed in refs. [@Ammon:2008fc; @Basu:2008bh; @Ammon:2009fe] by embedding a probe of two coincident D7-branes in the AdS black hole background. From this top-down approach one can try to identify the SU(2) chemical potential as an isospin chemical potential and the condensate as a $\rho$ meson. The back reaction of the gauge field on the metric in the SU(2) Yang-Mills model has been considered in refs. [@Ammon:2009xh]. It is interesting to note that when the back reaction is strong enough, the phase transition will be a first order one. The holographic SU(2) p-wave superconductor model has been extended to include, for example, the Gauss-Bonnet term [@Cai:2010zm; @CNZ] and Chern-Simons coupling [@Zayas:2011dw]. In addition, based on the backreacted metric, the behaviour of entanglement entropy in the holographic superconducting phase transitions has been studied in refs. [@Cai:2012nm; @Arias:2012py; @Cai:2013oma].
The Maxwell-Vector P-wave model {#sect:vector}
-------------------------------
Let us introduce a charged vector field into the $(d+1)$ dimensional Einstein-Maxwell theory with a negative cosmological constant. The full action reads [@Cai:2013pda; @Cai:2013aca] $$\label{vecaction}
\begin{split}
S=\frac{1}{2\kappa^2}\int d^{d+1} x
\sqrt{-g}(\mathcal{R}+\frac{d(d-1)}{L^2}+\mathcal{L}_m),\\
\mathcal{L}_m=-\frac{1}{4}F_{\mu\nu} F^{\mu \nu}-\frac{1}{2}\rho_{\mu\nu}^\dagger\rho^{\mu\nu}-m^2\rho_\mu^\dagger\rho^\mu+iq\gamma \rho_\mu\rho_\nu^\dagger F^{\mu\nu},
\end{split}$$ where a dagger denotes complex conjugation and $\rho_\mu$ is a complex vector field with mass $m$ and charge $q$. We define $\rho_{\mu\nu}=D_\mu\rho_\nu-D_\nu\rho_\mu$ with the covariant derivative $D_\mu=\nabla_\mu-iqA_\mu$. The last non-minimal coupling term characterizes the magnetic moment of the vector field $\rho_\mu$.
Since $\rho_\mu$ is charged under the U(1) gauge field, according to AdS/CFT correspondence, its dual operator $\hat{J^\mu}$ will carry the same charge under this symmetry and a vacuum expectation value of this operator will then trigger the U(1) symmetry breaking spontaneously. Thus, the condensate of the dual vector operator will break the U(1) symmetry as well as the spatial rotational symmetry since the condensate will pick out one direction as special. Therefore, viewing this vector field as an order parameter, the holographic model can be used to mimic a p-wave superconductor (superfluid) phase transition. The gravity background without vector hair ($\rho_\mu=0$)/with vector hair ($\rho_\mu\neq0$) is used to mimic the normal phase/superconducting phase in the dual system.
Indeed, it was shown in ref. [@Cai:2013pda] that working on the probe limit, as one lowers the temperature, the normal phase becomes unstable to developing nontrivial configuration of the vector field. The calculation of the optical conductivity reveals that there is a delta function at the origin for the real part of the conductivity, which means the condensed phase is indeed superconducting. In this subsection, we shall review the effect of a background magnetic field on the model and its complete phase diagram in terms of temperature and chemical potential.
### Condensate induced by magnetic field
Generally speaking, to consider the case with a magnetic field, one needs to solve coupled partial differential equations which is much more involved in practice. However, if one is interested in the instability induced by the magnetic field, one can overcome this difficulty by only focusing the dynamics near the critical point at which the condensate is very small. More precisely, one can introduce a deviation parameter $\epsilon$ from the critical point at which the condensate begins to appear. The coupled equations of motion can then be solved order by order in terms of the power of $\epsilon$.
Following the above procedure, we now turn on a magnetic field to study how the applied magnetic field influences the system. The background is taken to be a (3+1) dimensional AdS-Schwarzschild black hole . A consistent ansatz is as follows [@Cai:2013pda] $$\label{vecmatterB}
\begin{split}
\rho_\nu dx^\nu=[\epsilon\rho_x(r,x)e^{ipy}+\mathcal{O}(\epsilon^3)]dx+[\epsilon\rho_y(r,x)e^{ipy}e^{i\theta}+\mathcal{O}(\epsilon^3)]dy,\\
A_\nu dx^\nu=[\phi(r)+\mathcal{O}(\epsilon^2)]dt+[Bx+\mathcal{O}(\epsilon^2)]dy,
\end{split}$$ where $\rho_x(r,x)$, $\rho_y(r,x)$ are all real functions, $p$ is a real constant and the constant $\theta$ is the phase difference between the $x$ and $y$ components of the vector field $\rho_\mu$. The constant magnetic field $B$ is perpendicular to the $x-y$ plane.
The profile of $\phi$ can be uniquely determined at the zeroth order of $\epsilon$, which takes the form $$\label{vecat}
\phi(r)=\mu(1-r_h/r)\;,$$ with $\mu$ interpreted as the chemical potential. The equations of motion for $\rho_x$ and $\rho_y$ can be deduced from at order $\mathcal{O}(\epsilon)$. We further separate the variables as $\rho_x(r,x)=\varphi_x(r)X(x)$ and $\rho_y(r,x)=\varphi_y(r)Y(x)$. Then one can get the following equations [^20] $$\label{veceomabc}
\begin{split}
\varphi_x''+\frac{f'}{f}\varphi_x'+\frac{q^2\phi^2}{f^2}\varphi_x-\frac{m^2}{f}\varphi_x-\frac{E}{r^2f}\varphi_x=0,\\
-\ddot{X}\mp (1+\gamma)qBY+(qBx-p)^2X=EX,\\
-\ddot{Y}\mp (1+\gamma)qB X+(qBx-p)^2Y=EY,
\end{split}$$ where the prime denotes the derivative with respect to $r$ and the dot denotes the derivative with respect to $x$. We have also made a consistent assumption $\varphi_x=\varphi_y$ and $E$ is a constant coming from variables separation. The last two equations for $X(x)$ and $Y(x)$ can be solved analytically and the eigenvalue is given by $E=(2n+1)|qB|\pm (1+\gamma)qB$ where $n$ can be chosen as a non-negative integer.
We are interested in how the applied magnetic field influences on the transition temperature from the normal phase to the condensed phase. The effective mass of the charged vector field in the lowest energy state, i.e., in the lowest Landau level $n=0$ depends on the magnetic field $B$ and the non-minimal coupling parameter $\gamma$ as $$\label{veclowestmass}
m_{\rm eff}^{2}=m^2-\frac{|\gamma qB|}{r^2}-\frac{q^2\phi^2}{f}\;.$$ It is clear that the increase of the magnetic field $B$ decreases the effective mass and thus tends to raise the transition temperature, even in the case that the electric field is turned off. Only the magnetic field itself can trigger the phase transition. This result has an analogy to the QCD vacuum instability induced by a strong magnetic field to spontaneously developing the $\rho$-meson condensate. It is clear that the last term in describing a non-minimal coupling of the vector field $\rho^\mu$ to the gauge field $A_\mu$ plays a crucial role in the instability. Note that similar coupling can be found in many formalisms used to describe the coupling of magnetic moment to the background magnetic field for charged vector particles [@Djukanovic:2005ag; @Young:1963zza].
![\[vecmagneticn\] The transition temperature from the normal phase to the condensed phase as a function of magnetic field. $T_c$ is the critical temperature in the case without magnetic field. The magnetic field raises the transition temperature. One has chosen $m^2=3/4$. This figure was taken from ref. [@Cai:2013pda].](magneticn.pdf)
The $(T, B)$ phase diagram for the lowest Landau level is depicted in figure \[vecmagneticn\] in the case with fixed charge density $\rho=\mu r_h$. To determine which side of the phase transition line is the condensed phase, we can consider the equation . It suggests that the magnetic field decreases the effective mass. So if we increase the magnetic field at a fixed temperature, the normal state will become unstable for sufficiently large magnetic field.
It is clear that the transition temperature increases with the applied magnetic field. In ordinary superconductors an external magnetic field suppresses superconductivity via diamagnetic and Pauli pair breaking effects. However, it has also been proposed that the magnetic field induced superconductivity can also be realized in type-II superconductors [@PhysRevLett.58.1482; @Rasolt:1992zz], in which the Abrikosov flux lattice may enter a quantum limit of the low Landau level dominance with a spin-triplet pairing. And possible experimental evidence for the strong magnetic induced superconductivity can be found, for example, in refs. [@levy2005; @uji2010].
Due to the degeneracy in $p$, a linear superposition of the solutions with different $p$ is also a solution of the model at $\mathcal{O}(\epsilon)$. We can take this advantage to construct a class of vortex lattice solutions. As a typical example, the triangular lattice is shown in figure \[veclattice\]. It should be stressed that it is the special combinations $J_{\pm}=\langle\hat{J^x}\pm i\hat{J^y}\rangle$ which exhibit the vortex lattice structure. Strictly speaking, to obtain the true ground state, one should calculate the free energy of the solutions with different lattice structures from the action to find which configuration minimizes the free energy. It turns out that the linear analysis presented here is not sufficient to determine the most stable solution, thus should include higher order contributions. Furthermore, it is worthwhile to mention that in the AdS soliton background, the external magnetic field triggered phase transition and vortex lattice structure also happen for the vector field p-wave model [@Cai:2013kaa].
![\[veclattice\] The vortex lattice structure for the triangular lattice in $x-y$ plane. The contour plot is also drawn in the bottom. In particular, the condensate vanishes in the core of each vortex. The figure was taken from ref. [@Cai:2013pda].](lattices.pdf)
The response of this system to the magnetic field is quite different from the behaviour of ordinary superconductor where the magnetic field makes the transition more difficult. But the result here is quite similar to the case of QCD vacuum instability induced by strong magnetic field to spontaneously developing the $\rho$-meson condensate [@Chernodub:2010qx; @Chernodub:2011mc]. Although so, it was shown that in model the condensate of the vector operator forms a vortex lattice structure in the spatial directions perpendicular to the magnetic field. Of course, the non-minimal coupling term in the action plays a crucial role in both cases. Therefore in some sense, this model is a holographic setup of the study of $\rho$-meson condensate.
### The complete phase diagram
The probe approximation neglecting the back reaction of the matter fields can indeed uncover many key properties. Nevertheless, it still loses some important information, such as the phase structure of the system. In the following paragraphs, we will discuss both the black hole background and soliton background in full back reaction case. Then a complete phase diagram in terms of temperature and chemical potential will be shown. We shall consider a $(4+1)$-dimensional bulk theory [@Cai:2014ija].
We would like to study a dual theory with finite chemical potential or charge density accompanied by a U(1) symmetry, so we turn on $A_t$ in the bulk. We want to allow for states with a non-trivial current $\langle\hat{J_x}\rangle$, for which we further introduce $\rho_x$ in the bulk. Because a non-vanishing $\langle\hat{J_x}\rangle$ picks out $x$ direction as special, which obviously breaks the rotational symmetry in spatial plane, thus we should introduce an additional function in the $xx$ component of the metric in order to describe the anisotropy. Therefore, for the matter part, we consider the ansatz $$\label{vecmatter}
\rho_\nu dx^\nu=\rho_x(r)\,dx\;,\quad A_\nu dx^\nu=\phi(r)\,dt\;.$$ We will consider black hole and soliton backgrounds separately.\
\
[ *(1) AdS black hole with vector hair.*]{}\
For the black hole background, we adopt the following metric ansatz $$\label{vecBHans}
ds^2=-a(r)e^{-b(r)}dt^2+\frac{dr^2}{a(r)}+r^2\,(c(r)dx^2+dy^2+dz^2)\;.$$ The position of horizon is denoted as $r_h$ at which $a(r_h)=0$ and the conformal boundary is located at $r\rightarrow\infty$. One finds that the $r$ component of Maxwell equations implies that the phase of $\rho_x$ must be constant. Without loss of generality, we can take $\rho_x$ to be real. Then, the independent equations of motion in terms of above ansatz are deduced as follows $$\label{vecBHeom}
\begin{split}
\phi''+(\frac{c'}{2c}+\frac{b'}{2}+\frac{3}{r})\phi'-\frac{2q^2\rho_x^2}{r^2ac}\phi=0,\\
\rho_x''+(\frac{a'}{a}-\frac{c'}{2c}-\frac{b'}{2}+\frac1r)\rho_x'+\frac{e^{b}q^2\phi^2}{a^2}\rho_x-\frac{m^2}{a}\rho_x=0,\\
b'-\frac{2a'}{a}-\frac{c'}{c}+\frac{2\rho_x'^2}{3rc}-\frac{re^b\phi'^2}{3a}-\frac{2e^b q^2\rho_x^2\phi^2}{3ra^2c}+\frac{8r}{a}-\frac{4}{r}=0,\\
c''+(\frac{a'}{a}-\frac{c'}{2c}-\frac{b'}{2}+\frac{3}{r})c'+\frac{2{\rho_x'}^2}{r^2}-\frac{2e^b q^2\rho_x^2\phi^2}{r^2a^2}+\frac{2m^2\rho_x^2}{r^2a}=0,\\
(\frac{3}{r}-\frac{c'}{2c})\frac{a'}{a}+(\frac{1}{r}+\frac{b'}{2})\frac{c'}{c}-\frac{\rho_x'^2}{r^2c}+\frac{e^b\phi'^2}{2a}+\frac{3e^b q^2\rho_x^2\phi^2}{r^2a^2c}-\frac{m^2\rho_x^2}{r^2ac}-\frac{12}{a}+\frac{6}{r^2}=0,
\end{split}$$ where the prime denotes the derivative with respect to $r$.
When $\rho_x=0$, there exists an exactly analytical black hole solution, namely, AdS Reissner-Nordström black hole which reads $$\label{RNmetric}
\begin{split}
ds^2=-f(r)dt^2+\frac{dr^2}{f(r)}+r^2(dx^2+dy^2+dz^2),\\
f(r)=r^2\left[1-\left(1+\frac{\mu^2}{3r_h^2}\right)\left(\frac{r_h}{r}\right)^4+\frac{\mu^2}{3r_h^2}\left(\frac{r_h}{r}\right)^6\right], \quad \phi(r)=\mu\left[1-\left(\frac{r_h}{r}\right)^2\right].
\end{split}$$ This solution is dual to a conductor phase in the dual field theory. However, the full coupled equations of motion do not admit an analytical solution with non-trivial $\rho_x$. Therefore, we have to solve them numerically. We will use shooting method to solve equations . In order to find the solutions for all the five functions, i.e., $\rho_x(r),\phi(r), a(r), b(r)$ and $c(r)$ one must impose suitable boundary conditions both at conformal boundary $r\rightarrow\infty$ and at the horizon $r=r_h$.
In order to match the asymptotical AdS boundary, the general falloff near the AdS boundary behaves as $$\label{vecboundarybh}
\begin{split}
\phi=\mu-\frac{\rho}{r^2}+\cdots,\quad \rho_x=\frac{{\rho_x}_-}{r^{{\Delta}_-}}+\frac{{\rho_x}_{+}}{r^{{\Delta}_+}}+\cdots,\\
a=r^2(1+\frac{a_4}{r^4})+\cdots,\quad c=1+\frac{c_4}{r^4}+\cdots,\quad b=0+\frac{b_4}{r^4}+\cdots,
\end{split}$$ where the dots stand for the higher order terms in the expansion of $1/r$ and ${\Delta}_\pm=1\pm\sqrt{1+m^2}$.[^21] In general, in the above expansion we must impose ${\rho_x}_-=0$, which meets the requirement that the condensate appears spontaneously. According to the AdS/CFT dictionary, up to a normalization, the coefficients $\mu$, $\rho$, and ${\rho_x}_{+}$ are regarded as chemical potential, charge density and the $x$ component of the vacuum expectation value of the vector operator $\hat{J^\mu}$ in the dual field theory, respectively.
We focus on black hole configurations that have a regular event horizon located at $r_h$ and require the regularity conditions at the horizon $r=r_h$, which means that all five functions $\{\rho_x,\phi, a, b, c\}$ would have finite values at $r_h$ and admit a series expansion in terms of $(r-r_h)$. After substituting such series expansion into equations , one finds there are only six independent parameters at the horizon, i.e., $\{r_h,\rho_x(r_h),\phi'_x(r_h),c(r_h),b(r_h)\}$ and other coefficients can be expressed in terms of those parameters.
Two free parameters $b(r_h)$ and $c(r_h)$ can be fixed by AdS boundary conditions that $b(r\rightarrow\infty)=0$ and $c(r\rightarrow\infty)=1$. Without loss of generality, the location of $r_h$ can be fixed to be one in our numerical calculation. We are then left with two independent parameters $\{\rho_x(r_h),\phi'(r_h)\}$. By choosing $\phi'(r_h)$ as the shooting parameter to match the source free condition, i.e., ${\rho_x}_-=0$, we finally have a one-parameter family of solutions labeled by the value of $\rho_x$ at the horizon. After solving the set of equations, we can read off the condensate $\langle \hat{J^x}\rangle$, chemical potential $\mu$ and charge density $\rho$ directly from the asymptotical expansion .\
\
[*(2) AdS soliton with vector hair.*]{} To construct homogeneous charged solutions with vector hair in the soliton background, we take the metric as $$\label{ansatz}
ds^2=\frac{dr^2}{r^2g(r)}+r^2(-f(r)dt^2+h(r)dx^2+dy^2+g(r)e^{-\chi(r)}d\eta^2)\;,$$ where $g(r)$ vanishes at the tip $r=r_0$ of the soliton. The asymptotical AdS boundary is located at $r\rightarrow\infty$. Further, in order to obtain a smooth geometry at the tip $r_0$, $\eta$ should be made with an identification $$\label{Gamma}
\eta\sim\eta+\Gamma\;,\qquad \Gamma=\frac{4\pi e^{\frac{\chi(r_0)}{2}}}{r_0^2 g'(r_0)}\;.$$ This gives a dual picture of the boundary theory with a mass gap, which is reminiscent of an insulating phase.
The independent equations of motion are deduced as follows $$\label{vecsleoms}
\begin{split}
\phi''-(\frac{f'}{2f}-\frac{g'}{g}-\frac{h'}{2h}+\frac{\chi'}{2}-\frac{3}{r})\phi'-\frac{2q^2\rho_x^2}{r^4gh}\phi=0,\\
\rho_x''+(\frac{f'}{2f}+\frac{g'}{g}-\frac{h'}{2h}-\frac{\chi'}{2}+\frac{3}{r})\rho_x'+\frac{q^2\phi^2}{r^4fg}\rho_x-\frac{m^2}{r^2g}\rho_x=0, \\
f''-(\frac{f'}{2f}-\frac{g'}{g}-\frac{h'}{2h}+\frac{\chi'}{2}-\frac{5}{r})f'-\frac{\phi'^2}{r^2}-\frac{2q^2\rho_x^2\phi^2}{r^6gh}=0,\\
\chi'-\frac{f'}{f}-\frac{2g'}{g}-\frac{h'}{h}+\frac{2{\rho_x'}^2}{3rh}-\frac{\phi'^2}{3rf}-\frac{2q^2\rho_x^2\phi^2}{3r^5fgh}+\frac{8}{rg}-\frac{8}{r}=0,\\
h''+(\frac{f'}{2f}+\frac{g'}{g}-\frac{h'}{2h}-\frac{\chi'}{2}+\frac{5}{r})h'+\frac{2{\rho_x'}^2}{r^2}-\frac{2q^2\rho_x^2\phi^2}{r^6fgh}+\frac{2m^2\rho_x^2}{r^4gh}=0,\\
(\frac{6}{r}-\frac{f'}{f}-\frac{h'}{h})\frac{g'}{g}+(\frac{f'}{f}+\frac{h'}{h})\chi'-\frac{f'h'}{fh}-\frac{2\rho_x'^2}{r^2h}+\frac{\phi'^2}{r^2f}+\frac{6q^2
\rho_x^2\phi^2}{r^6fgh}-\frac{2m^2\rho_x^2}{r^4gh}\\-\frac{24}{r^2g}+\frac{24}{r^2}=0,
\end{split}$$ where the prime denotes the derivative with respect to $r$. Similar to the black hole case, we will solve those coupled equations of motion numerically by use of shooting method. In order to find the solutions for all the six functions $\mathcal{F}=\{\rho_x,\phi,f,g,h,\chi\}$ one must impose suitable boundary conditions at both conformal boundary $r\rightarrow\infty$ and the tip $r=r_0$.
The asymptotical expansion for metric fields and matter fields near the boundary $r\rightarrow\infty$ is as follows $$\label{vecboundarysl}
\begin{split}
&\phi=\mu-\frac{\rho}{r^2}+\cdots,\quad \rho_x=\frac{{\rho_x}_-}{r^{{\Delta}_-}}+\frac{{\rho_x}_{+}}{r^{{\Delta}_+}}+\cdots,\quad f=1+\frac{f_4}{r^4}+\cdots\\
&g=1+\frac{g_4}{r^4}+\cdots,\quad h=1+\frac{h_4}{r^4}+\cdots,\quad \chi=0+\frac{\chi_4}{r^4}+\cdots,
\end{split}$$ where the dots stand for the higher order terms of $1/r$. We choose the source free condition ${\rho_x}_-=0$ as before. The coefficients $\mu$, $\rho$, and ${\rho_x}_{+}$ are directly related to the chemical potential, charge density and $x$ component of the vacuum expectation value of the vector operator $\hat{J^\mu}$ in the dual system, respectively.
We impose the regularity conditions at the tip $r=r_0$, which means that all functions have finite values and admit a series expansion in terms of $(r-r_0)$ as $$\label{vecseries}
\mathcal{F}=\mathcal{F}(r_0)+\mathcal{F}'(r_0)(r-r_0)+\cdots.$$ By plugging the expansion into , one can find that there are six independent parameters at the tip $\{r_0,\rho_x(r_0),\phi(r_0),f(r_0),h(r_0),\chi(r_0)\}$. However, there exist four useful scaling symmetries in the equations of motion, which read $$\label{vecscaling1}
\chi\rightarrow \chi+\lambda,\quad \eta\rightarrow e^{\lambda/2}\eta\;,$$ $$\label{vecscaling2}
\phi\rightarrow\lambda \phi,\quad t\rightarrow\lambda^{-1} t,\quad f\rightarrow\lambda^2 f\;,$$ $$\label{vecscaling3}
\rho_x\rightarrow\lambda \rho_x,\quad x\rightarrow\lambda^{-1} x,\quad h\rightarrow\lambda^2 h\;,$$ and $$\label{vecscaling4}
r\rightarrow\lambda r,\quad (t,x,y,\eta)\rightarrow{\lambda^{-1}}(t,x,y,\eta),\quad(\phi,\rho_x)\rightarrow\lambda(\phi,\rho_x)\;,$$ where in each case $\lambda$ is a real positive constant.
By using above four scaling symmetries, we can first set $\{r_0=1,f(r_0)=1,h(r_0)=1,\chi(r_0)=0\}$ for performing numerics. After solving the coupled differential equations, one should use the first three symmetries again to satisfy the asymptotic conditions $f(\infty)=1$, $h(\infty)=1$ and $\chi(\infty)=0$. We choose $\phi(r_0)$ as the shooting parameter to match the source free condition, i.e., ${\rho_x}_-=0$. Finally, for fixed $m^2$ and $q$, we have a one-parameter family of solutions labeled by $\rho_x(r_0)$. After solving the set of equations, we can read off the condensate $\langle \hat{J^x}\rangle$, chemical potential $\mu$ and charge density $\rho$ from the corresponding coefficients in . It should be noticed that different solutions obtained in this way will have different periods $\Gamma$ for $\eta$ direction. We should use the last scaling symmetry to set all of the periods $\Gamma$ equal in order to obtain same boundary geometry. We shall fix $\Gamma$ to be $\pi$ in this section.
![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD1.pdf "fig:"){width="50.00000%"} ![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD2.pdf "fig:"){width="50.00000%"}\
![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD3.pdf "fig:"){width="50.00000%"} ![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD4.pdf "fig:"){width="50.00000%"}\
![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD5.pdf "fig:"){width="50.00000%"} ![\[completephase\] The complete phase diagrams of the Maxwell-vector model with S=pure AdS soliton, BH=AdS Reissner-Nordström black hole, SC=hairy soliton, and BC=hairy black hole. In each region the thermodynamically stable phase is labeled. As $m^2$ and $q$ are changed, the shape of each region gets modified. This figure is described in ref. [@Cai:2014ija] in more detail.](combinePD6.pdf "fig:"){width="50.00000%"}\
These two kinds of situations have been well studied in ref. [@Cai:2014ija]. There are four different bulk solutions given by the pure AdS soliton, AdS Reissner-Nordström black hole and their vector hairy counterparts. According to the AdS/CFT dictionary, the hairy solution is dual to a system with a non-zero vacuum expectation value of the charged vector operator which breaks the U(1) symmetry and the spatial rotation symmetry spontaneously. The above four solutions in the bulk correspond to an insulating phase, a conducting phase, a soliton superconducting phase and a black hole superconducting phase, respectively. Since we do not turn on magnetic field, the model is left with two independent parameters, i.e., the mass $m$ of the vector field giving the scaling dimension of the dual vector operator and its charge $q$ controlling the strength of the back reaction on the background geometry. The phase structure of the model heavily depends on those two parameters. There exist second order, first order and zeroth order[^22] phase transitions as well as the “retrograde condensation" in which the hairy solutions exist only above a critical temperature or below a critical chemical potential with the free energy much larger than the solutions without hair.
With four kinds of phases at hand, the complete phase diagrams can be constructed in terms of temperature and chemical potential. At each point in $T$-$\mu$ plane, one should find the phase which has the lowest free energy. Since there are many types of phase transitions in both soliton and black hole backgrounds, the $T$-$\mu$ phase diagrams are expected to be much more complicated than the holographic s-wave model [@Horowitz:2010jq] and the Yang-Mills p-wave model [@Akhavan:2010bf]. Some typical examples are shown in figure \[completephase\]. We can see from the complete phase diagrams that in some cases, more than one superconducting phase appears in a phase diagram in the model. The phase diagrams for some realistic superconducting materials are usually complicated, and indeed, more than one superconducting phase can occur, for example, see refs. [@Kordyuk; @Chubukov; @Yuan]. Definitely, it is of great interest to see whether this model is relevant to those superconducting materials.
The Helical P-wave model {#sect:helical}
------------------------
The gravity solutions above mainly describe spatially homogeneous superconducting states. However, it has long been known that it is possible to have superconducting states that are spatially inhomogeneous. A well known example is the Fulde-Ferrell-Larkin-Ovchinnikov (FFLO) phase, for which a Cooper pair consisting of two fermions with different Fermi momenta condenses leading to an order parameter with non-vanishing total momentum [@Fulde:1964zz; @larkin:1964zz]. In this section, we shall introduce a holographic model which can realize p-wave superconducting phase with a helical order. That is to say, the order parameter points in a given direction in a plane which then rotates as one moves along the direction orthogonal to the plane.
We consider a (4+1) dimensional model with a gauge field $A_\mu$ and a charged two-form $C_{\mu\nu}$ [@Donos:2011ff; @Donos:2012gg] $$\label{eomhelical}
\begin{split}
S=\int d^5 x\sqrt{-g}\big[\mathcal{R}+12&-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-\frac{1}{4}C^{\mu\nu}C^\dagger_{\mu\nu}+\frac{{i}}{24m}\epsilon^{\mu\nu\rho\sigma\lambda}C_{\mu\nu}
H^\dagger_{\rho\sigma\lambda}\big]\,,
\end{split}$$ where we have chosen units where the AdS radius is unity, a dagger denotes complex conjugation and the field strengths read $$\label{ders}
F=dA\;,\qquad H=dC+{i} e\, A\wedge C\;.$$ The gauge field $A_\mu$ is dual to a current in the dual theory and the two-form $C_{\mu\nu}$ corresponds to a self-dual rank two tensor operator with scaling dimension $\Delta=2+|m|$. In particular, this charged operator has angular momentum $l=1$ and thus can serve as an order parameter for $p$-wave superconductors. Since what we are interested in is a system at finite temperature and chemical potential with respect to the global U(1) symmetry, we will construct electrically charged asymptotically AdS black holes in gravity side. The normal phase with no condensate is described by the electrically charged Reissner-Nordström AdS black hole, which is spatially homogeneous and isotropic. This model is specified by two parameters $m$ and $e$. It was shown in ref. [@Donos:2011ff] that when $e^2>m^2/2$ this black hole is unstable to developing non-trivial two-form hair that is dual to p-wave superconductors with helical order.
### Boundary conditions
The helical black hole solution was constructed in ref. [@Donos:2012gg] in which the authors adopted the ansatz $$\label{helicalansatz}
\begin{split}
ds^{2}&=-g\,f^{2}\,dt^{2}+g^{-1}{dr^{2}}+h^{2}\,\omega_{1}^{2}+r^{2}\,\left(e^{2\alpha}\,\omega_{2}^{2}+e^{-2\alpha}\,\omega_{3}^{2}\right),\\
C&=(i\, c_{1}\,dt+c_2 dr)\wedge\omega_{2}+c_{3}\,\omega_{1}\wedge\omega_{3}\;, \qquad A=a\, dt\;,
\end{split}$$ where the one-forms $\omega_i$ are given by $$\label{helicalforms}
\begin{split}
&\omega_{1}=dx_{1}\;,\\
&\omega_{2}=\cos\left(kx_{1}\right)\,dx_{2}-\sin\left(kx_{1}\right)\,dx_{3}\;,\\
&\omega_{3}=\sin\left(kx_{1}\right)\,dx_{2}+\cos\left(kx_{1}\right)\,dx_{3}\;.
\end{split}$$ Note that the constant $t$ and $r$ slices in the above metric are spatially homogeneous of Bianchi type $\text{VII}_0$. All eight functions in the ansatz depend on the radial coordinate $r$ only and $k$ is a constant. After substituting the ansatz into the action , one finds that $c_1$ and $c_2$ can be determined by other functions, thus we are left with six independent functions including $f$, $g$, $h$, $\alpha$, $c_{3}$ and $a$. More precisely, $f$ and $g$ satisfy first order differential equations and other functions satisfy second order equations.
To solve the coupled equations of motion for above six functions, one needs to specify suitable boundary conditions in the horizon $r_h$ and the conformal boundary $r\rightarrow\infty$. Regularity at the horizon demands that $g(r_h)=a(r_h)=0$ and all of them have analytic expansion in terms of $(r-r_h)$. We then find that the full expansion at the horizon is fixed by six parameters, i.e., $r_h, f(r_h), h(r_h), \alpha(r_h), a'(r_h)$ and $c_3(r_h)$. Near the boundary $r\rightarrow\infty$, one demands asymptotically AdS geometry with the fall-off $$\begin{aligned}
\label{uvexpan}
\begin{split}
&g=r^{2}\,\left(1-{M}{r^{-4}}+\cdots \right),\quad f=f_{0}\left(1-{c_{h}}{r^{-4}}+\cdots\right),\\
&h=r\,\left(1+{c_{h}}{r^{-4}}+\cdots \right),\quad \alpha={c_{\alpha}}{r^{-4}}+\cdots,\\
&a=f_{0}\,\left(\mu-{\rho}{r^{-2}}/2+\cdots\right),\quad c_{3}={c_{v}}{r^{-\left|m\right|}}+\cdots,
\end{split}\end{aligned}$$ which is determined by eight parameters $M, f_0, c_h, c_\alpha, \mu, \rho, c_v$ and $k$. One should note that the expansion of $c_3$ is chosen so that the charged operator dual to the two-form $C$ has no source, thus can spontaneously acquire an expectation value proportional to $c_v$ which is spatially modulated in the $x_1$ direction with period $2\pi/k$. $\mu$ and $\rho$ are regarded as the chemical potential and charge density in the dual system respectively. The holographic interpretation of the other UV parameters will be given below. Observe that when $k\neq0$ the order parameter rotates in the $(x_2, x_3)$ plane as one moves along the $x_1$ direction thus there is a reduced helical symmetry.
There are two scaling symmetries of the coupled equations which can be used to set $\mu=f_0=1$. To solve the six differential equations, we need to specify ten integration constants. However, we have fourteen parameters in two boundaries minus two for the scaling symmetries. Therefore, we expect to leave with a two parameter family of black hole solutions which can be selected as temperature $T$ and wave number $k$.
### Thermodynamics {#sect:thermo}
We shall work in grand canonical ensemble with the chemical potential $\mu$ fixed. The thermodynamic potential of the boundary thermal state is identified with temperature $T$ times the on-shell bulk action in Euclidean signature. We denote $w$ as the density of thermodynamic potential per spatial volume in dual field theory. Then one can obtain the following expression for the free energy density [^23] $$w=-M=3M+8c_h-\mu\rho-s\,T\;,$$ where the entropy density $s=4\pi r_h^2 h(r_h)$ and $f_0$ is set to be one. From above equation one can immediately obtain the Smarr-type formula $$4M+8c_h-\mu\rho-s\,T=0\;.$$ An on-shell variation of the total action for fixed $k$ gives us the first law $$\delta w=-s\delta T-\rho\delta\mu,$$ and hence $w=w(T, \mu)$.
The expectation value of the dual stress-energy tensor is given, after setting $f_0=1$, by $$\begin{split}
&T_{tt}=3M+8c_{h},\quad
T_{x_{1}x_{1}}=M+8c_{h},\\
&T_{x_{2}x_{2}}=M +8c_{\alpha}\,\cos\left(2kx_{1} \right) ,\\
&T_{x_{3}x_{3}}=M -8c_{\alpha}\,\cos\left(2kx_{1} \right) ,\\
&T_{x_{2}x_{3}}=-\,8\,c_{\alpha}\,\sin\left(2kx_{1}\right).
\end{split}$$ Obviously the stress-energy tensor is traceless as a consequence of the underlying conformal symmetry. We further extract the energy density $\varepsilon=T_{tt}=3M+8c_{h}$ from which we can rewrite $w=\varepsilon-s T-\mu\rho$ and the first law takes the more familiar form $\delta \varepsilon=T \delta s+\mu\delta\rho$. The average hydrostatic pressure $\bar p$ is defined as minus the average of the trace of the spatial components. We get $\bar p=M+8c_h/3$, and hence the system satisfies the thermodynamical relation $\varepsilon+\bar p=Ts+\mu\rho$.
### Helical p-wave solutions {#sect:helicalsolu}
We focus on the specific case with $m=1.7$ and $e=1.88$ [^24] and set $\mu=f_0=1$. Starting from the AdS Reissner-Nordström black hole solution, as the temperature is lowered, the first instability appears at $T_c\simeq0.0265$ and $k=k_c\simeq0.550$. Below $T_c$, there is a continuum of hairy black hole solutions appearing with different values of $k$.
Figure \[helicalfig1\] summaries the free energy density $w$ as a function of temperature $T$ and wave number $k$. One can see that all hairy solutions have smaller free energy than the normal solutions at the same temperature and the transition to the p-wave preferred branch is second order. For a given temperature $T<T_c$, there is a one parameter family of solutions specified by $k$, and the most thermodynamically preferred solution is denoted by the red line. One can prove that while the general hairy solutions in figure \[helicalfig1\] have $c_h\neq0$, the solutions on the red line do have vanishing $c_h$ [@Donos:2013woa].
![The free energy density as a function of $T$ and $k$ for the case of $m=1.7$ and $e=1.88$. The red curve denotes the thermodynamically favored p-wave superconducting phase minimizing the free energy with respect to $k$ at fixed $T$. The black curves correspond to curves with constant $T$. The blue line is for the case of some domain wall solutions. Reprinted with kind permission from ref. [@Donos:2012gg]. \[helicalfig1\]](helicalfree.pdf)
The helical superconducting order can be fixed by the condensate $c_v$ and wave number $k$, which are shown in figure \[helicalfig2\] with respect to $T$ for the red line in figure \[helicalfig1\]. Near $T_c$ one can find the critical phenomenon $c_{v}\simeq1.7\times 10^5 T_c^{3.7}\left(1-T/T_{c} \right)^{1/2}$, which is the famous mean field behaviour. As the temperature is lowered, the red line moves smoothly down to sufficiently low temperature at which $k\equiv k_0\simeq0.256$. In particular, the ground state at $T=0$ is also spatially modulated.
The $T=0$ limit of hairy solutions approach a smooth domain wall solution which interpolates between $AdS_5$ in the UV and a new IR fixed point with an anisotropic scaling. This fixed point in the IR reads $$\begin{split}
&g=K\,r^{2},\quad f=\bar f_0r^{z-1},\quad h=k h_{0}\,,\quad \alpha=\alpha_{0}\,,\\
&a=a_{0}r^{z},\quad
c_{3}=k c_{0}\,r\,,
\end{split}$$ with $K, h_0, \alpha_0, a_0, c_0$ and $z$ all constants. This fixed point solution is invariant under the anisotropic scaling $r\rightarrow\lambda^{-1}r$, $t\rightarrow \lambda^{z}t$, $x_{2,3}\rightarrow\lambda x_{2,3}$ and $x_1\to x_1$. All those constants can be determined by the equations of motion.[^25] As a typical example, choosing $m=1.7$ and $e=1.88$, one can obtain $$\begin{split}
&z\simeq1.65 ,\qquad K\simeq 0.995,\qquad h_{0}\simeq0.993,\\
&\alpha_{0}\simeq-0.380,\quad a_{0}\simeq0.265, \quad ~c_{0}\simeq 3.69.
\end{split}$$
![The condensate $c_v$ and wave number $k$ as a function of $T$ for the thermodynamically preferred hairy solutions. The blue dots represent the quantities for the domain wall solutions. Used with permission from ref. [@Donos:2012gg]. \[helicalfig2\]](helicalcondensate.pdf "fig:") ![The condensate $c_v$ and wave number $k$ as a function of $T$ for the thermodynamically preferred hairy solutions. The blue dots represent the quantities for the domain wall solutions. Used with permission from ref. [@Donos:2012gg]. \[helicalfig2\]](helicalwave.pdf "fig:")
The domain wall solutions interpolating between the UV fixed point and the IR fixed point can be specified by the wave number $k$ [@Donos:2012gg]. One can see in figure \[helicalfig1\] that the $T\rightarrow0$ limit of the hairy solutions approach these domain wall solutions (the blue line). Similarly, in figure \[helicalfig2\], the condensate $c_v$ and wave number $k= k_0\simeq0.256$ for the domain wall denoted by blue dots smoothly connect with the corresponding black hole solution.
To summarize, a holographic p-wave model with helical superconducting order is introduced in this subsection. As the temperature is lowered, a helical superconducting state emerges spontaneously breaking both the abelian symmetry and the three-dimensional spatial Euclidean symmetry down to Bianchi $\text{VII}_0$ symmetry. These homogeneous, but anisotropic ground states at $T=0$ are holographically described by smooth domain wall solutions, which exhibit zero entropy density and an emergent scaling symmetry in the far IR.
Further nature of the model has been well studied in ref. [@Donos:2013woa]. For example, some of the p-wave solutions can exhibit the phenomenon of pitch inversion[^26] and the symmetry of the black hole solutions is enhanced at the pitch inversion temperature. The superconducting phase can also be $(p+ip)$ order. Depending on the mass and charge of the two-form, both the p-wave and the $(p+ip)$-wave can be thermodynamically favored. The two kinds of orders will compete with each other and there can be first order transition between them.
Holographic D-wave Models {#sect:dwave}
=========================
It is remarkable to see that rather simple and generic gravity models can capture many features of the phase structure of superconducting systems. Nevertheless, in order to construct more sophisticated and more realistic models one clearly needs to include additional ingredients. The focus of this part is on realising an important missing phase, i.e. d-wave superconductivity (superfluidity). The importance is self-evident since many unconventional superconductors admit either d-wave or mixed symmetry. A natural candidate for modelling the d-wave condensate is to use a charged spin two field in the bulk, instead of a charged scalar field or a vector field. Based on this approach, there are two acceptable holographic models describing the d-wave condensate in the literature.
The authors of ref. [@Chen:2010mk] first constructed a minimal gravitational model by introducing a symmetric, traceless rank-two tensor field minimally coupled to a U(1) gauge field in the background of an AdS black hole. The d-wave condensate appears below a critical temperature via a second order phase transition, resulting in an isotropic superconducting phase but no hard gap for its optical conductivity. Let us call it CKMWY d-wave model in terms of the initials of the five authors. The other effective holographic d-wave model was proposed soon after the first one with the same matter fields but with much more complex interactions [@Benini:2010pr]. The phase diagram, optical conductivity, as well as fermionic spectral function were investigated in detail. With a fixed metric, this model has advantages such as being ghost-free and having the right propagating degrees of freedom. This model will be named as BHRY d-wave model for short in what follows.
The CKMWY d-wave model {#subsect:CKMWY model}
----------------------
To construct a holographic d-wave model, the minimal effective bulk action including gravity, U(1) gauge field and tensor field reads [@Chen:2010mk] $$\begin{aligned}
\label{wwdwave}
S =\frac{1}{2\kappa ^{2}}\int d^{4}x\sqrt{-g}\left[\mathcal{R}+\frac{6}{L^{2}}-(D_{\mu }B_{\nu \gamma
})^{^\dagger }D^{\mu }B^{\nu \gamma }-m^{2}B_{\mu \nu }^{^\dagger}B^{\mu \nu }-
\frac{1}{4}F_{\mu \nu }F^{\mu \nu }\right],\end{aligned}$$where $D_\mu=\nabla_\mu-iq A_\mu$ is the covariant derivative in the black hole background, $L$ is the AdS radius that will be set to unity, and $q$ and $m^{2}$ are the charge and mass squared of $B_{\mu \nu }$, respectively. Working in the probe limit, i.e. $q\rightarrow\infty$ with $q A_\mu$ and $qB_{\mu \nu }$ fixed, the matter part can be treated as perturbations in the 3+1 dimensional AdS black hole background .
We would like to realize a d-wave superconductor on the boundary such that a condensate emerges on the $x-y$ plane with translation invariance and the rotational symmetry is broken down to Z(2) with the condensate changing its sign under a $\pi/2$ rotation on the $x-y$ plane. Therefore, we use an ansatz for $B_{\mu \nu }$ and $A_{\mu }$ as $$\begin{aligned}
B_{xx}=-B_{yy}=\psi(r)\;,\quad\quad A=\phi(r)\,dt\;,\end{aligned}$$with all other field components being turned off and $\psi(r)$ and $\phi(r)$ being real functions. The background geometry is fixed as AdS-Schwarzschild black hole given in . Then the final equations of motion read $$\begin{split}
\psi''+(\frac{f'}{f}-\frac{2}{r})\psi'-\left(\frac{2f'}{rf}+\frac{m^2}{f}-\frac{q^2\phi^2}{f^2}\right)\psi=0\;,\\
\phi''+\frac{2}{r}\phi'-\frac{4 q^2 \psi^2}{r^4f}\phi=0\;.
\end{split}$$ These two equations are very similar as the case for the Abelian-Higgs model (see equations ). Therefore, it is natural to expect $\psi$ to condense spontaneously below a critical temperature. More precisely, we demand the following asymptotic form near the AdS boundary $r\rightarrow\infty$ $$\begin{aligned}
\phi=\mu-\rho/r+\cdots,\quad \psi=f_1 r^{\Delta_-}+\cdots,\end{aligned}$$with $\Delta_-=\frac{1- \sqrt{17+4m^{2}}}{2}$. Note that the expansion of $\psi$ is chosen such that the charged operator dual to $B_{\mu\nu}$ has no source, thus can acquire an expectation value proportional to $f_1$ spontaneously. According to holographic dictionary, $\mu$ is interpreted as the chemical potential, and $\rho$ as the charge density in the dual theory. The order parameter of the boundary theory can be obtained by reading the asymptotic behaviour of $B$, i.e. $$\langle \mathcal{O}_{ij}\rangle =\left(
\begin{array}{cc}
f_{1} & 0 \\
0 & -f_{1}\end{array}\right),$$ where $(i,j)$ are the indexes in the boundary coordinates $(x,y)$. In what follows, we shall keep the chemical potential $\mu$ fixed and choose $q$ to be minus one, which is the setup adopted by ref. [@Chen:2010mk].
![\[dcondenw\] The d-wave condensate as a function of temperature. The condensate goes to zero at the critical temperature $T_c$. We choose $m^2=-1/4$. Used with permission from ref. [@Chen:2010mk].](dcondensatew.pdf)
The d-wave condensate as a function of temperature can be obtained numerically, which is shown in figure \[dcondenw\]. One can see clearly that below $T_c$, the tensor field is Higgsed to break the U(1) symmetry spontaneously in the boundary theory. Numerical calculation further ensures that the phase transition characterized by the d-wave condensate is second order with the mean field critical behaviour $f_1\sim (T_c-T)^{1/2}$. The conductivity has also been studied, which uncovered that the AC conductivity is isotropic and below $T_c$, the DC conductivity becomes infinite but has no hard gap.
The BHRY d-wave model {#subsect:BHRY model}
---------------------
The approach of the CKMWY d-wave model just writes down a minimal action for the spin two field without looking in detail at the constraint equations required to get the correct number of propagating degrees of freedom. Soon, the authors of ref. [@Benini:2010pr] analyzed in more detail the effective action for the spin two field and how the constraint equations could be satisfied. The desired theory for a charged, massive spin two field in a fixed Einstein background takes the following form $$\label{dwBHRY}
\begin{split}
S=\frac{1}{2\kappa^2}\int d^{d+1}x\sqrt{-g}(-\frac{1}{4} F_{\mu\nu} F^{\mu\nu}+ \mathcal{L}_d),\\
\mathcal{L}_d=-|D_\rho\varphi_{\mu\nu}|^2+2|D_\mu\varphi^{\mu\nu}|^2+|D_\mu\varphi|^2
-\big[D_\mu\varphi^{\dagger\mu\nu}D_\nu \varphi+\text{h.c.}\big]-i q F_{\mu\nu} \varphi^{\dagger\mu\lambda} \varphi^\nu_\lambda\\
-m^2\big(|\varphi_{\mu\nu}|^2-|\varphi|^2\big)+2\mathcal{R}_{\mu\nu\rho\lambda} \varphi^{\dagger\mu\rho}\varphi^{\nu\lambda}-\frac{1}{d+1} \mathcal{R} |\varphi|^2\;,
\end{split}$$ where $D_{_\mu}=\nabla_\mu-i q A_\mu$, $\varphi\equiv{\varphi^\mu}_\mu$, $\varphi_\rho\equiv D^{\mu}\varphi_{\mu\rho}$ and ${\mathcal{R}^\mu}_{\nu\rho\lambda}$ is the Riemann tensor of the background metric. The above theory is ghost-free and describes the correct number of propagating degrees of freedom. The disadvantage is that one has to be restricted to work in a fixed background spacetime that satisfies the Einstein condition $\mathcal{R}_{\mu\nu}=\frac{2\Lambda}{d-1}g_{\mu\nu}$. In the context of holographic superconductors, this restriction forces us to work in the probe approximation where the spin two field and gauge field do not influence on the metric. One such a geometry is given by the AdS-Schwarzschild black hole with a planar horizon $$ds^2 = \frac{L^2}{z^2} \Big(-f(z)\,dt^2 + d \vec x_{d-1}^2 + \frac{dz^2}{f(z)} \Big), \quad f(z)=1-\Big(\frac{z}{z_h} \Big)^d \;.$$ The black hole horizon is located at $z=z_h$, while the conformal boundary of the spacetime is located at $z=0$. The temperature of this black hole is $$T = \frac{d}{4 \pi z_h} \;.$$
### The d-wave condensate
We consider an ansatz where $\varphi_{\mu\nu}$ and $A_\mu$ depend only on the radial coordinate $z$ and only the space components of $\varphi_{\mu\nu}$ are turned on. According to ref. [@Benini:2010pr], it is consistent to turn on a single component of $\varphi_{\mu\nu}$ and to set other components of the gauge field except for $A_t$ to be zero. Then our ansatz is $$\label{dhansatz}
A= \phi(z) \, dt \;, \qquad \varphi_{xy}(z)=\frac{L^2}{2z^2} \, \psi(z) \;,$$ with all other components of $\varphi_{\mu\nu}$ set to zero, and $\phi$ and $\psi$ real.
With the above ansatz , the equations of motion for $\phi$ and $\psi$ are given by $$\begin{split}
\psi'' + \left( \frac{f'}{f} - \frac{d-1}{z} \right) \psi' + \left( \frac{ q^2 \phi^2}{f^2} - \frac{m^2 L^2}{z^2 f} \right) \psi=0\;,\\
\phi'' + \frac{3-d}z \, \phi' - \frac{q^2L^2}{z^2 f} \, \psi^2 \, \phi=0\;.
\end{split}$$ Here the prime denotes the derivative with respect to the radial coordinate $z$. To solve the above coupled equations, one demands that two fields near the boundary $z=0$ should behave as $$\begin{aligned}
\phi=\mu-\rho z^{d-2}+\cdots, \quad \psi=\psi_{+} z^{\Delta}+\cdots,\end{aligned}$$ where $\Delta=d/2+\sqrt{d^2+4m^2L^2}/2$. The unitary bound implies that $\Delta\geq d$ for spin two operators. Therefore, the mass of $\varphi_{\mu\nu}$ has a lower bound, i.e. $m^2\geq 0$.[^27] Note that the fall-off of $\psi$ is chosen so that the dual charged operators have no deformation but can acquire expectation value spontaneously. Up to a normalization, the coefficients $\mu$, $\rho$ and $\psi_{+}$ are interpreted as chemical potential, charge density and the expectation value of the $xy$ component for the spin two operator $\mathcal{O}_{xy}$, respectively. At the horizon, one should require $\phi(z_h)=0$ in order to keep $g^{\mu\nu}A_{\mu}A_{\nu}$ being finite at the horizon.
![\[dhcondens\] The condensate $\langle \mathcal{O}_{xy} \rangle$ in $d=3$ dimensional spacetime as a function of the temperature for various values of $\Delta$. The curves form top to down correspond to $\Delta=6, 4$ and $3.5$, respectively. Used with kind permission from ref. [@Benini:2010pr].](dhcondensate)
In what follows we will focus on $d=3$. The resulting boundary value problem can be solved directly, for example, by shooting method. A typical dependence of $\langle \mathcal{O}_{xy}\rangle$ on the temperature is shown in figure \[dhcondens\]. As we lower the temperature, the normal phase becomes unstable to developing tensor hair at a certain critical temperature $T_c$. This is a typical second order phase transition.
### Conductivity {#conductivity}
We are interested in the electromagnetic response of the condensed phase. To incorporate this feature, we extract the optical conductivity of this d-wave model by linear response theory. The conductivity tensor $\sigma _{ij}$ can be defined through $$J_{i}=\sigma _{ij}\;E_{j}\;,$$ where $i, j=x, y$. $J$ and $E$ are the electric current and electric field, respectively. To compute the conductivity in a holographic framework we turn on a source for the current $J_i$ dual to the gauge field in the bulk. Following the standard approach discussed in previous sections, we perturb the gauge field by $\delta A=e^{-i\omega t}A_{x}(r)dx$. The bulk equations of motion couple linear fluctuations of the gauge field $A_x$ to some spin two components. To obtain a consistent set of equations, one should also turn on time dependent fluctuations of the complexified fields $\varphi_{ty}$, $\varphi_{ty}^\dagger$, $\varphi_{zy}$ and $\varphi_{zy}^\dagger$. The linearized equations of motion for the $e^{-i\omega t}$ component of these fluctuations are [@Benini:2010pr]
\[dhEOMlinear\] $$\begin{aligned}
\label{dhEOMAx}
0 &= A_x'' + \frac{f'}f \, A_x' + \frac{\omega^2}{f^2} \, A_x + \frac{q \psi}{2f^2} \, \big[ (\omega - 2q \phi) \varphi_{ty}^\dagger - (\omega + 2q\phi) \varphi_{ty} \big] \nonumber \\
&\quad - \frac{iq\psi}2 \, \big( {\varphi^\dagger_{zy}}' - \varphi_{zy}' \big) + \frac{iq}{2f} \, (\psi' f - \psi f') \big( \varphi_{zy}^\dagger - \varphi_{zy} \big) \ , \\
\label{dhEOMty}
0 &= \varphi_{ty}'' + \frac2z \, \varphi_{ty}' - \frac{2f + m^2L^2}{z^2 f} \, \varphi_{ty} + L^2 \, \frac{q\omega + 2q^2 \phi}{4z^2 f} \, \psi A_x + \frac i2\, \big[ 2(\omega + q \phi) \varphi_{zy}' + q \phi' \varphi_{zy} \big] \ , \\
\label{dhEOMRzy}
0 &= \big[ (\omega + q\phi)^2 z^2 - m^2L^2 f \big] \, \varphi_{zy} + \frac i4\, L^2 q f \psi A_x' + \frac i2 L^2 q f \psi' A_x \nonumber \\
&\quad - i(\omega + q\phi) z^2 \varphi_{ty}' - \frac i2 \, \big[ 4(\omega + q \phi)z + q \phi'z^2 \big] \, \varphi_{ty} \ .\end{aligned}$$
The equations for $\varphi_{ty}^\dagger$ and $\varphi_{zy}^\dagger$ can be obtained by complex conjugation and an additional transformation $\omega$ to $-\omega$ from the last two equations. The functions $\varphi_{zy}$ and $\varphi_{zy}^\dagger$ can be eliminated from the first two equations by virtue of , leaving three coupled differential equations for $A_x$, $\varphi_{ty}$ and $\varphi_{ty}^\dagger$. Since the fluctuation $A_y$ decouples from above set of fluctuations, we can conclude that the Hall conductivity $\sigma_{xy}(\omega)$ is vanishing.
The conductivity is related to the retarded Green’s function for the charge current. To calculate the retarded function, one should impose causal boundary conditions on the equations of motion. As a consequence, the near-horizon modes of the gauge field and spin two field are falling into the horizon, i.e., $A_x$, $\varphi_{ty}$ and $\varphi_{ty}^\dagger$ have the behaviour as $$(z_h-z)^{- i\omega z_h/ 3} \;.
\label{nearhorizon}$$ Near the boundary $z=0$, the asymptotical behaviour for the perturbation fields $A_x$, $\varphi_{ty}$ and $\varphi_{ty}^{\dagger}$ is given by $$\begin{aligned}
\nonumber
A_x&=&A_x^{(0)}+A_x^{(1)}z+\cdots,\\
\varphi_{ty}&=&\varphi_{ty-}z^{\Delta_-}+\varphi_{ty+}z^{\Delta_+}+\cdots,\\ \nonumber
\varphi_{ty}^{\dagger}&=&\varphi^\dagger_{ty-}z^{\Delta_-}+\varphi^\dagger_{ty+}z^{\Delta_+}+\cdots,\end{aligned}$$ with $\Delta_{\pm}= \frac{-1\pm \sqrt{9+4m^2L^2}}{2}$. Here $\varphi_{ty-}$ and $\varphi^\dagger_{ty-}$ are identified as the source terms, while $\varphi_{ty+}$ and $\varphi^\dagger_{ty+}$ are the normalizable fluctuations. Since the presence of source terms $\varphi_{ty-}$ and $\varphi^\dagger_{ty-}$ will also source the U(1) current, one should look for solutions where the source in the series expansion of $\varphi_{ty}$ and $\varphi_{ty}^{\dagger}$ should vanish. Finally, one can obtain the conductivity in the $x$ direction as $$\sigma_{xx}=\frac{A_x^{(1)}}{i \omega A_x^{(0)}}\;.$$
![\[dhconductivity\] The real part of the conductivity as a function of frequency for a $\Delta=4 $ condensate. Used with permission from ref. [@Benini:2010pr].](dhconduc)
In order to obtain the conductivity in the $y$ direction, we look at the effect of a $\pi/2$ rotation of the condensate in equations . This rotation operation flips the sign of $\psi$, which is equivalent to flipping the sign of $A_x$ by viewing of the equations of motion. Changing the sign of $A_x$ is equivalent to flipping the sign of both the electric field (the $A_x^{(0)}$ term) and the current (the $A_x^{(1)}$ term), thus conductivity will be unaffected under such a change of sign. This implies that the conductivity is proportional to the identity matrix. In fact, the isotropy of conductivity is a consequence of the symmetries that $\sigma_{ij}$ has in the d-wave case. An isotropic conductivity for a d-wave superconductor can be also produced in an explicit microscopic model [@isosigma].
The numerical results for the conductivity are shown in figure \[dhconductivity\] for a d-wave condensate of conformal dimension $\Delta=4$. As the temperature is lowered one can observe a spike in the conductivity, which is a signal of a bound state. This spike is localized at smaller values of $\omega$ as the temperature is lowered. A second spike in the conductivity appears to disappear as the temperature is decreased. One can also see that $\text{Re}[\sigma(\omega)]$ does not vanish even for arbitrary small frequency $\omega$, so there is no hard gap in the dual boundary superconducting phase.
The fermionic spectral function in this holographic superconductor with a d-wave condensate has been well studied in ref. [@Benini:2010qc]. It was showed that, with a suitable bulk Majorana coupling, the Fermi surface is anisotropically gapped. At low temperatures the gap shrinks to four nodal points, while at high temperatures the Fermi surface is partially gapped generating four Fermi arcs. The $(d+id)$ condensate for the BHRY model was investigated in ref. [@Chen:2011ny], in which the existence of fermi arcs is confirmed and a non-vanishing Hall conductivity is obtained in the absence of a magnetic field.
Although both d-wave models we reviewed above can be used to study the properties of a superconducting phase transition with a d-wave order parameter in a dual strongly interacting field theory, the construction is not ideal. For example, the BHRY d-wave model can only work in the probe limit. However, it is well known that including the back reaction of matter fields would lead to a much richer phase structure. It will be desirable to study a consistent holographic d-wave model with back reaction. To write down an action for a charged spin two field propagating in a curved spacetime is challenging, because it usually suffers from non-hyperbolic and non-causal behaviour of the spin two field. Apart from those two effective models, the authors of ref. [@Kim:2013oba] discussed top-down models for holographic d-wave superconductors in which the order parameter is a charged spin two field in the bulk.
Competition and Coexistence of Superconducting Order Parameters {#sect:competition}
===============================================================
The holographic models of s-wave, p-wave and d-wave superconductors, which have scalar, vector and spin-2 order parameters respectively, have been discussed in the previous sections. These models were based on a specific setup where the dynamics in the bulk involves only a single order parameter. It is desirable to generalize the single order parameter case to multi order parameters case because the real high $T_c$ superconducting systems indeed involve various orders, such as magnetic ordering and superconductivity, see figure \[hightphase\] and, e.g., refs. [@Berg2009; @Zaanen1012]. The holographic correspondence provides us a convenient way to investigate the interaction for these orders by simply introducing dual fields in the bulk as well as appropriate couplings among them. Following this strategy, several works on the competition of multi order parameters in the holographic superconductor models have already been made [@Cai:2013wma; @Nie:2013sda; @Amado:2013lia; @Li:2014wca; @Nishida:2014lta; @Basu:2010fa; @Musso:2013ija; @Liu:2013yaa; @Amoretti:2013oia; @Wen:2013ufa; @Donos:2012yu]. In the following, we will review the competition between two s-wave orders [@Cai:2013wma], the competition between s-wave order and p-wave order [@Nie:2013sda; @Amado:2013lia] and the competition between s-wave order and d-wave order [@Li:2014wca] one by one. The first case concerns the competition between two orders with the same symmetry and the last two cases are to study the competition of orders with different symmetry. The phase diagrams are also drawn for the corresponding models.
![\[hightphase\] Schematic phase diagrams of the cuprates (left) and the pnictide superconductor $Ba(Fe_{1-x}Co_x)As_2$ (right). In the right plot, the antiferromagnetic phase is labeled by AFM, the normal (tetragonal) phase is denoted by Tet, and the superconducting phase by SC. Note the similarities of the phase diagrams. Adapted with permission from ref. [@Norman:2011] .](cuprates.pdf "fig:") ![\[hightphase\] Schematic phase diagrams of the cuprates (left) and the pnictide superconductor $Ba(Fe_{1-x}Co_x)As_2$ (right). In the right plot, the antiferromagnetic phase is labeled by AFM, the normal (tetragonal) phase is denoted by Tet, and the superconducting phase by SC. Note the similarities of the phase diagrams. Adapted with permission from ref. [@Norman:2011] .](fesupc.pdf "fig:")
Competition and coexistence of two s-wave orders {#ss}
-------------------------------------------------
Historically, Ginzburg-Landau theory has proved to be an extraordinarily valuable phenomenological tool in understanding single-component superconductors. Its generalization to the two-component Ginzburg-Landau model (TCGL) was constructed, and its applicability to the two-band systems studied in refs. [@Silaev:2012; @Shanenko:2011; @Vagov:2012]. Upon switching on the interband coupling between the two components, this model can describe the phenomenon of the two gaps in materials such as MgB$_{2}$ ($s_{++}$) [@Carlstrom:2010; @Buzea:2001] and iron pnictides ($s_{+-}$) [@iron-based1; @iron-based2; @iron-based3]. Applying the multi-band Ginzburg-Landau theories to the gravity side, the holographic multi-band superconductor model can be realized involving some competing scalar fields coupled to a single gauge field. Such system exhibits rich phase structure. Next, we will describe the holographic model concretely and show results explicitly.
### The holographic model
Let us start a holographic superconductor model with $N$ scalar hairs in $(3+1)$ dimensional anti-de Sitter spacetime. The action reads [@Cai:2013wma] $$\label{action0}
S =\frac{1}{2\kappa^2}\int d^4 x
\sqrt{-g}[\mathcal{R}+\frac{6}{L^2}-\frac{1}{4}F_{\mu\nu} F^{\mu \nu}+\sum_{k=1}^{N}(-|\nabla\psi_k-ie_kA\psi_k|^2-m_k^2|\psi_k|^2)-\mathcal{V}_{intact}],$$ where $e_k$ and $m_k$ ($k=1,2,...,N$) are the charge and mass of the scalar field $\psi_k$, respectively. The term $\mathcal{V}_{intact}$ denotes the possible interaction among bulk matter fields. Here one can perform a rescaling of the type $A_\mu\rightarrow\frac{1}{e_2}A_\mu, \psi_k\rightarrow\frac{1}{e_2}\psi_k$ to set the charge of the scalar field $\psi_2$ to unity. We are interested in the dynamics and mutual interaction among different orders. Here we limit ourselves to the case with $N=2$. The concrete model we will study is described by the following action $$\label{action}
\begin{split}
S =\frac{1}{2\kappa^2}\int d^4 x
\sqrt{-g}[\mathcal{R}+\frac{6}{L^2}+\frac{1}{e_2^2}\mathcal{L}_m],\\
\mathcal{L}_m=-\frac{1}{4}F_{\mu\nu} F^{\mu \nu}-|D_1\psi_1|^2-m_1^2|\psi_1|^2-|D_2\psi_2|^2-m_2^2|\psi_2|^2,
\end{split}$$ where we have defined $D_{1\mu}=\nabla_\mu-i\frac{e_1}{e_2}A_\mu$ and $D_{2\mu}=\nabla_\mu-iA_\mu$. The parameter $e_2$ controls the strength of the back reaction and $e_1/e_2$ is the effective charge of $\psi_1$ or the ratio of two scalar charges.
The hairy black hole solution is assumed to take the following metric form $$\label{metric}
ds^2=-f(r)e^{-\chi(r)}dt^2+\frac{dr^2}{f(r)}+r^2(dx^2+dy^2),$$ together with homogeneous matter fields $$\label{matter}
\psi_1=\psi_1(r),\quad \psi_2=\psi_2(r),\quad A=\phi(r)dt.$$ The horizon $r_h$ is determined by $f(r_h)=0$ and the temperature of the black hole is given by $$\label{temp}
T=\frac{f'(r_h)e^{-\chi(r_h)/2}}{4\pi}.$$ One can use the U(1) gauge symmetry to set $\psi_1$ to be real. After using the $r$ component of Maxwell’s equations we can also safely choose $\psi_2$ to be real. We will work in the unites where $L=1$. Then, the independent equations of motion in terms of the above ansatz are deduced as follows $$\label{eomsss}
\begin{split}
\psi_1''+(\frac{f'}{f}-\frac{\chi'}{2}+\frac{2}{r})\psi_1'+(\frac{e_1^2}{e_2^2}\frac{\phi^2e^{\chi}}{f^2}-\frac{m_1^2}{f})\psi_1=0, \\
\psi_2''+(\frac{f'}{f}-\frac{\chi'}{2}+\frac{2}{r})\psi_2'+(\frac{\phi^2e^{\chi}}{f^2}-\frac{m_2^2}{f})\psi_2=0, \\
\phi''+(\frac{\chi'}{2}+\frac{2}{r})\phi'-\frac{2}{f}(\frac{e_1^2}{e_2^2}\psi_1^2+\psi_2^2)\phi=0,\\
\frac{f'}{f}+\frac{r}{2e_2^2}(\psi_1'^2+\psi_2'^2)+\frac{re^{\chi}\phi'^2}{4e_2^2f}+\frac{r}{2e_2^2f}(m_1^2\psi_1^2+m_2^2\psi_2^2)+\frac{re^{\chi}\phi^2}{2e_2^2f^2}
(\frac{e_1^2}{e_2^2}\psi_1^2+\psi_2^2)\\-\frac{3r}{f}+\frac{1}{r}=0 ,\\
\chi'+\frac{r}{e_2^2}(\psi_1'^2+\psi_2'^2)+\frac{re^{\chi}\phi^2}{e_2^2f^2}(\frac{e_1^2}{e_2^2}\psi_1^2+\psi_2^2)=0,
\end{split}$$ where a prime denotes the derivative with respect to $r$.
The gravity background describing the normal phase is just the AdS Reissner-Nordström black hole with a planar horizon $$\label{normalss}
\phi(r)=\mu(1-\frac{r_h}{r}),\quad \psi_1(r)=\psi_2(r)=0,\quad f(r)=r^2(1-\frac{r_h^3}{r^3})+\frac{r_h^2}{4r^2}\frac{\mu^2}{e_2^2}(1-\frac{r}{r_h}),$$ where $r_h$ is the black hole horizon and $\mu$ is the chemical potential of the black hole.
### Phase transition
The two band model is controlled by four model parameters, i.e., $m_1^2$, $m_2^2$, $e_2$, and $e_1/e_2$. Here we will choose $m_1^2=0$ and $m_2^2=-2$. One may expect that the model admits three different superconducting phases. The first superconducting phase corresponds to the case with $\psi_1\neq0$ and $\psi_2=0$ (Phase-). The second superconducting phase corresponds to the case with $\psi_2\neq0$ and $\psi_1=0$ (Phase-). The third superconducting phase admits the region where both scalars condense simultaneously.
![\[condensateABC\] The condensate as a function of temperature for three kinds of the coexisting phase. The red curve is for $\psi_1$, while the blue one is for $\psi_2$. The condensate for Phase-C with $e_1/e_2=1.95$ and $e_2=4$ is shown in the left plot. The condensate for Phase-A with $e_1/e_2=1.95$ and $e_2=2$ in the middle plot and for Phase-B with $e_1/e_2=1.9$ and $e_2=1.5$ in the right plot. Three plots were taken from ref. [@Cai:2013wma].](condensateC.pdf "fig:") ![\[condensateABC\] The condensate as a function of temperature for three kinds of the coexisting phase. The red curve is for $\psi_1$, while the blue one is for $\psi_2$. The condensate for Phase-C with $e_1/e_2=1.95$ and $e_2=4$ is shown in the left plot. The condensate for Phase-A with $e_1/e_2=1.95$ and $e_2=2$ in the middle plot and for Phase-B with $e_1/e_2=1.9$ and $e_2=1.5$ in the right plot. Three plots were taken from ref. [@Cai:2013wma].](condensateA.pdf "fig:") ![\[condensateABC\] The condensate as a function of temperature for three kinds of the coexisting phase. The red curve is for $\psi_1$, while the blue one is for $\psi_2$. The condensate for Phase-C with $e_1/e_2=1.95$ and $e_2=4$ is shown in the left plot. The condensate for Phase-A with $e_1/e_2=1.95$ and $e_2=2$ in the middle plot and for Phase-B with $e_1/e_2=1.9$ and $e_2=1.5$ in the right plot. Three plots were taken from ref. [@Cai:2013wma].](condensateB.pdf "fig:")
The numerical results are shown in figure \[condensateABC\]. The model admits three kinds of the coexisting phase. The first kind is shown in the left panel of figure \[condensateABC\]. As we lower temperature, the scalar $\psi_1$ first condenses at $T_c$ where the superconducting phase transition happens; when we continue lowering temperature to a certain value, say $T_2$, the scalar $\psi_2$ begins to condense, while the condensate of $\psi_1$ decreases, resulting in the phase with both orders; if one further lowers temperature, the first condensate quickly goes to zero at a temperature $T_3$; when temperature is lower than $T_3$, there exists only the condensate of $\psi_2$. This superconducting phase is denoted by Phase-C and it is the case uncovered in the probe limit in ref. [@Basu:2010fa]. The second kind of the coexisting phase is presented in the middle plot. It is different from the first one in that the coexisting phase survives even down to a low temperature. We denote this case by phase-A. Depending on the back reaction, the inverse of phase-A is also true: the condensate of $\psi_1$ emerges following the condensate of $\psi_2$, and then both orders are always present. This case is labeled as Phase-B drawn in the right plot.
To determine whether those above three coexistence phases are thermodynamically favored in their own parameter spaces, one should calculate the free energy of the system for each phase. Working in grand canonical ensemble, the chemical potential is fixed. In gauge/gravity duality the grand potential $\Omega$ of the boundary thermal state is identified with temperature times the on-shell bulk action with Euclidean signature. The Euclidean action must include the Gibbons-Hawking boundary term for a well-defined Dirichlet variational principle and further a surface counter term for removing divergence. Note that we are considering a stationary problem, the Euclidean action is related to the Minkowski one by a minus sign as $$\begin{split}
-2\kappa^2 S_{Euclidean}=\int d^4 x
\sqrt{-g}[&\mathcal{R}+\frac{6}{L^2}+\frac{1}{e_2^2}\mathcal{L}_m]+\int_{r\rightarrow\infty} d^3x
\sqrt{-h}(2\mathcal{K}-\frac{4}{L})\\
+&\frac{1}{e_2^2}\int_{r\rightarrow\infty} d^3x\sqrt{-h}(\frac{{\triangle_1}_+-3}{L}\psi_1^2+\frac{{\triangle_2}_+-3}{L}\psi_2^2),
\end{split}$$ where $h$ is the determinant of the induced metric on the boundary $r\rightarrow\infty$, and $\mathcal{K}$ is the trace of the extrinsic curvature. By using of the equations of motion and the asymptotical expansion of matter and metric functions near the AdS boundary, the grand potential $\Omega$ can be expressed as $$\label{grand1}
\frac{2\kappa^2\Omega}{V_2}=\varepsilon,$$ where $V_2=\int dx dy$ and the constant $\varepsilon$ is from the asymptotical expansion of $f=r^2+\varepsilon/r+\cdots$. For the normal phase given in , one has $\varepsilon=-r_h^3-\frac{r_h}{4}\frac{\mu^2}{e_2^2}$.
The free energy corresponding to phase-A(C) is drawn in figure \[freeAss\]. From each plot, phase-A(C) does have the lowest free energy, indicating that once phase-A(C) appears, it is thermodynamically favored. But for phase-C, there is only a narrow window admitting the two orders to coexist. Comparing phase-B with phase-A, the only difference is that $\psi_2$ condenses before $\psi_1$. From figure \[freeAss\], one can also see there exist two transition points in both cases. One is the critical superconducting phase transition and the other within the superconducting phase, which indicates the fact that our system is multi-band. In order to see this clearly, one can define the total condensate as $\langle\mathcal{O}\rangle=\langle O_{1+}\rangle^{1/\Delta_{1+}}+\langle O_{2+}\rangle^{1/\Delta_{2+}}$, and draw $\langle\mathcal{O}\rangle$ as a function of temperature in figure \[condensateAA\]. As one lowers temperature, $\langle\mathcal{O}\rangle$ emerges at the critical superconducting phase transition point, then at a certain temperature inside the superconducting phase it has a sudden increase, where the condensate of the other $\psi$ appears. Such a behaviour is very reminiscent of the one in the real multi-band superconductor.
![\[freeAss\] The free energy as a function of temperature for Phase-A(C) labeled as the solid red curve. The equations of motion also admit three other types of solutions, i.e., the normal phase (dotted black curve), Phase- (dashed green curve) and Phase- (solid blue curve). The curve in the insert of the left plot is the difference of free energy between Phase-A and Phase-. One can see phase-A(C) indeed has the lowest free energy. The plots were taken from ref. [@Cai:2013wma].](freeA.pdf "fig:") ![\[freeAss\] The free energy as a function of temperature for Phase-A(C) labeled as the solid red curve. The equations of motion also admit three other types of solutions, i.e., the normal phase (dotted black curve), Phase- (dashed green curve) and Phase- (solid blue curve). The curve in the insert of the left plot is the difference of free energy between Phase-A and Phase-. One can see phase-A(C) indeed has the lowest free energy. The plots were taken from ref. [@Cai:2013wma].](freeC.pdf "fig:")
![\[condensateAA\] The total condensate as a function of temperature for Phase-A. We set $e_1/e_2=2$ and $e_2=2$. The two special points at $T_c\simeq0.0488\mu$ and $T\simeq0.0298\mu$ correspond to the superconducting critical point and the point at which $\psi_2$ begins to emerge in Phase-A, respectively. The figure was taken from ref. [@Cai:2013wma].](condensateAAA.pdf)
### Conductivity {#sect:conduc}
In order to ensure the system is indeed in a superconducting state, one should calculate the conductivity $\sigma$. Since now the back reaction is included, one has to consider the fluctuations of $A_x$ and $g_{tx}$. Assuming both perturbations have a time dependence of the form $e^{-i\omega t}$, the final equation of motion to calculate the conductivity can be obtained as $$\label{Axeom}
A_x''+(\frac{f'}{f}-\frac{\chi'}{2})A_x'+[(\frac{\omega^2}{f^2}-\frac{\phi'^2}{e_2^2f})e^{\chi}-\frac{2}{f}(\frac{e_1^2}{e_2^2}\psi_1^2+\psi_2^2)]A_x=0.$$ Since the conductivity is related to the retarded two-point function of the U(1) current, i.e, $\sigma=\frac{1}{i\omega}G^R(\omega,k=0)$, one imposes the ingoing boundary condition near the horizon $$\label{ingoing}
A_x=(r-r_h)^{-\frac{i\omega}{4\pi T}}[a_0+a_1(r-r_h)+a_2(r-r_h)^2+\cdots],$$ with $a_0, a_1, a_2$ being constants. The gauge field $A_x$ near the boundary $r\rightarrow\infty$ falls off as $$\label{axbound}
A_x=A^{(0)}+\frac{A^{(1)}}{r}+\cdots.$$ According to the AdS/CFT dictionary, the retarded Green function can be read as $G^R=\frac{1}{2\kappa^2 e_2^2}\frac{A^{(1)}}{A^{(0)}}$, from which one can obtain the conductivity $$\label{conduc}
\sigma(\omega)=\frac{1}{i\omega}G^R(\omega,k=0)=\frac{1}{2\kappa^2 e_2^2}\frac{A^{(1)}}{i\omega A^{(0)}}.$$ The optical conductivity as a function of frequency in the region with two order parameters is presented in figure \[conductivityss\]. One can see clearly that the optical conductivity in two band model behaves qualitatively similar to the model with only one scalar order discussed in figure \[sconduc\]. In addition, from the Kramers-Kronig relations, one can conclude that the real part of the conductivity has a Dirac delta function at $\omega=0$ since the imaginary part has a pole, i.e., Im$[\sigma(\omega)]\sim\frac{1}{\omega}$.
![\[conductivityss\] The optical conductivity as a function of frequency at temperature $T=0.0273\mu$ for Phase-A. The red solid line is the real part of the conductivity, while the blue dashed line is the imaginary part of the conductivity. Here the parameter $e_1/e_2=1.95$ and $e_2=2$ are taken. There is a delta function at the origin for the real part of the conductivity. Figure taken from ref. [@Cai:2013wma].](conductivity_cox.pdf "fig:")
### Phase diagram
Constructing the parameter space is helpful to learn in which region the superconducting orders can coexist. One can complete this task by just turning the problem as an eigenvalue problem. Focus on the concrete model discussed in this paper, i.e., $m_1^2=0$ and $m_2^2=-2$, a good starting point is to find the critical valve of the ratio $e_1/e_2$ such that $T$ is a critical temperature at which $\psi_1$ begins to vanish or emerge. At such a temperature, $\psi_1$ is very small and can be treated as a perturbation on the background where only $\psi_2$ condenses $$\label{conduc}
-\psi_1''-(\frac{f'}{f}-\frac{\chi'}{2}+\frac{2}{r})\psi_1'+\frac{m_1^2}{f}\psi_1=\frac{e_1^2}{e_2^2}\frac{\phi^2e^{\chi}}{f^2}\psi_1,$$ where $\{\phi, f, \chi\}$ are functions describing the hairy AdS black hole with only $\psi_2$ non-vanishing. Imposing the appropriate boundary conditions, this equation can be considered as an eigenvalue problem with positive eigenvalue $e_1^2/e_2^2$. The full phase diagram for the five superconducting phases is shown in figure \[diagramss\].
![\[diagramss\] The full phase diagram for the five superconducting phases. Depending on $e_1/e_2$ and $e_2$, the phase diagram is divided into five parts. The most thermodynamically favored phase in each part is labeled. This figure was taken from ref. [@Cai:2013wma].](diagram.pdf)
From figure \[diagramss\], one has as many as five superconducting phases in the model apart from the normal phase. Depending on the model parameters $e_1/e_2$ and $e_2$, each phase can be most thermodynamically stable in some region of parameter space. As one increases the strength of the back reaction, the region for Phase-C with the coexisting behaviour of two order parameters only in a narrow window is gradually forced to shrink and finally vanishes at $e_2^{critical}$, while the regions for Phase-A and Phase-B where both order parameters always present enlarge. In this sense, one can conclude that the gravity which provides an equivalent attractive interaction between the holographic order parameters tends to make the coexistence of two orders much more easy rather than more difficult.
In this subsection \[ss\], a holographic superconductor model with more than one order parameter in four dimensions has been studied, where each complex scalar field in the bulk is minimally coupled to a same U(1) gauge field. This can be interpreted as a holographic multi-band superconductor model. Concretely, we have discussed the two-band case with mass squares $m_1^2=0$ and $m_2^2=-2$ for two bulk scalar fields $\psi_1$ and $\psi_2$, respectively. Depending on the strength of the back reaction $1/e_2^2$ and the relative charge ratio $e_1/e_2$ of the two scalar fields, the model admits as many as five different superconducting phases. Three of them, denoted by Phase-A, Phase-B and Phase-C, exhibit the coexistence region of two order parameters. More specifically, for Phase-C, as one lowers the temperature, the second scalar $\psi_2$ condenses following $\psi_1$ will completely suppress the condensate of the first order, i.e., $\psi_1$ will go to zero finally. The condensate behaviours in Phase-A and Phase-B are similar. One of the two orders condenses first, and once the other begins to condense, both always coexist. However, this model is limited to the competition of the order parameters with the same asymmetry. Therefore it is quite interesting to study the holographic models with superconducting order parameters with different spins. This will be done in the following subsections.
Competition between s-wave and p-wave orders {#sp}
--------------------------------------------
In this subsection, we will study two holographic superconductor models with both s-wave and p-wave condensed. One is proposed in ref. [@Nie:2013sda], where the authors built a holographic superconductor model with a scalar triplet charged under an SU(2) gauge field in the bulk. The other holographic s+p model in ref. [@Amado:2013lia] consists of a scalar doublet charged under an U(2) gauge field living in a planar Schwarzschild black hole geometry. The discussions for both models are limited to the probe limit case.
### The holographic s+p superconductor with a scalar triplet charged under an SU(2) gauge field
To realize the s-wave and p-wave superconductivity in one model, we first consider a real scalar triplet charged in an SU(2) gauge field in the gravity side. The full action is [@Nie:2013sda] $$\begin{aligned}
\begin{split}
&S =\frac{1}{2 \kappa_g ^2}\int d^{d+1}x \sqrt{-g} (R-2\Lambda)+S_M,\\
&S_M=\frac{1}{g_c^2}\int d^{d+1}x \sqrt{-g}(-D_\mu \Psi^{a} D^\mu \Psi^a-\frac{1}{4}F^a_{\mu\nu}F^{a\mu\nu}-m^2 \Psi^a\Psi^a),
\end{split}\end{aligned}$$ where $\Psi^a$ is an SU(2) charged scalar triplet in the vector representation of the SU(2) gauge group, and $$D_\mu\Psi^a=\partial_\mu \Psi^a+\epsilon^{abc}A^b_\mu\Psi^c.$$ $F^a_{\mu\nu}$ is the gauge field strength which is the same as and reads $$F^a_{\mu\nu}=\partial_\mu A^a_\nu-\partial_\nu A^a_\mu +\epsilon^{abc} A^b_\mu A^c_\nu.$$ $g_c$ is the Yang-Mills coupling constant as well as the SU(2) charge of $\Psi^a$. One can redefine the fields $A^a_{\mu}$ and $\Psi^a$ to get the standard expression where the coupling $g_c$ appears in the derivative operator $D_\mu$. Here we limit ourselves to the case of probe limit. This limit can be realized consistently by taking the limit $g_c\rightarrow\infty$.
In the probe limit, we consider the $d+1$ dimensional AdS black brane as the background with metric $$\label{metric_sp}
ds^2=-f(r)dt^2+\frac{1}{f(r)}dr^2+r^2dx_i dx^i.$$ $x^i$s are the coordinates of a $d-1$ dimensional Euclidean space. The function $f(r)$ is $$\label{fEinstein}
f(r)=r^2\left(1-\frac{r_h^d}{r^d}\right),$$ with $r_h$ the horizon radius. Here the AdS radius $L$ has been set to be unity. The temperature of the black brane is related to $r_h$ as $$\label{TemperatureE}
T=\frac{d }{4\pi}r_h.$$ This is just the temperature of dual field theory in the AdS boundary.
Let us consider the following ansatz for the matter fields $$\begin{aligned}
\Psi^3=\Psi_3(r),\quad A^1_t=\phi(r),\quad A^3_x=\Psi_x(r),\end{aligned}$$ with all other field components being turned off. In this ansatz, we take $A^1_\mu$ as the electromagnetic U(1) field. With this ansatz, the equations of motion of matter fields in the AdS black brane background read $$\begin{aligned}
\label{eoms}
\nonumber
\phi''+\frac{d-1}{r}\phi' -\Big(\frac{2 \Psi_3^2}{f}+\frac{\Psi_x^2}{r^2 f}\Big)\phi&=&0, \\
\Psi_x''+\Big(\frac{d-3}{r}+\frac{f'}{f}\Big)\Psi_x'+\frac{\phi^2}{f^2}\Psi_x&=&0, \\ \nonumber
\Psi_3''+\Big(\frac{d-1}{r}+\frac{f'}{f}\Big)\Psi_3'-\Big(\frac{m^2}{f}-\frac{\phi^2}{f^2}\Big)\Psi_3&=&0.\end{aligned}$$ One can see $\Psi^3$ and $\Psi_x$ are not coupled in their equations of motion, but they are both coupled to the same U(1) electromagnetic field. In this model, thus one can easily realize the s-wave and p-wave superconductivity consistently.
![\[SPcondensation\] Condensate of the operators in the s+p coexisting phase. The blue curve is for the condensate of the p-wave operator, while the red curve is for the condensate of the s-wave operator. The figure was taken from ref. [@Nie:2013sda].](SPcondensation.pdf){width="9.0cm"}
We take the case with $\Delta=\Delta_{eg}=(6+\sqrt{3})/4$ as an example. The condensate behaviour for the coexisting phase is drawn in the left plot of figure \[SPcondensation\]. We can see that the $s+p$ coexisting phase starts from the p-wave phase and ends with the pure s-wave condensate phase. Based on the calculation of free energy shown in figure \[FreeEnergysp\], we confirm that the s+p coexisting phase indeed has the lowest free energy and is thus thermodynamically favored in the temperature region. Thus the potential first order phase transition from the pure p-wave phase to the pure s-wave phase is replaced by the phase transitions from the p-wave phase to the s-wave phase through an s+p coexisting phase. And all the three phase transitions are continuous ones, and are of characteristic of second order phase transition within the numerical accuracy.
From figure \[FreeEnergysp\], we see that the Gibbs free energy curves of the s-wave and p-wave phases have an intersection when $\Delta_{cI}<\Delta<\Delta_{cII}$. The s+p coexisting phase just exists in this interval. By computing the values of $T_c^{sp1}$ and $T_c^{sp2}$ and getting the relations $T_c^{sp1}(\Delta)$ and $T_c^{sp2}(\Delta)$ in the region $\Delta_{cI}<\Delta<\Delta_{cII}$ , a phase diagram of the holographic model on the $\Delta$–$T$ plane can be shown in figure \[phasediagram\]. We can see from the figure that the system contains four kinds of phases known as the normal phase, the s-wave phase, the p-wave phase and the s+p coexisting phase. The s+p coexisting phase is favored in the area between the blue line and the red line. The region for the s+p coexisting phase is very narrow in the phase diagram. This is similar to the situation of the coexisting phase with two s-wave orders in the probe limit studied in ref. [@Basu:2010fa]. However, for the latter, the region with the coexisting phase is enlarged with the full back reaction [@Cai:2013wma]. This would be due to the additional interaction between the two scalar fields in the bulk through gravity and this interaction reduces the repellency between the two condensates.
![\[FreeEnergysp\] (Left) The Gibbs free energy versus temperature for various phases. The black solid curve is for the normal phase, the blue solid curve is for the p-wave phase, and the dashed lines from bottom to top are for the s-wave phase with operator dimension $\Delta=\Delta_{cI}$, $(6+\sqrt{3})/4$, and $\Delta_{cII}$, respectively. (Right) The Gibbs free energy in the region near the intersection point of the p-wave curve and the s-wave curve with $\Delta=(6+\sqrt{3})/4$. The plots were taken from ref. [@Nie:2013sda].](FreeEnergy.pdf "fig:"){width="7.5cm"} ![\[FreeEnergysp\] (Left) The Gibbs free energy versus temperature for various phases. The black solid curve is for the normal phase, the blue solid curve is for the p-wave phase, and the dashed lines from bottom to top are for the s-wave phase with operator dimension $\Delta=\Delta_{cI}$, $(6+\sqrt{3})/4$, and $\Delta_{cII}$, respectively. (Right) The Gibbs free energy in the region near the intersection point of the p-wave curve and the s-wave curve with $\Delta=(6+\sqrt{3})/4$. The plots were taken from ref. [@Nie:2013sda].](FreeEnergyL.pdf "fig:"){width="8.1cm"}
![\[phasediagram\]The $\Delta-T$ phase diagram. The normal phase, the s-wave phase, d-wave phase and the coexisting phase are colored differently. The right plot is an enlarged version of the coexisting region. The figures were taken from ref. [@Nie:2013sda].](phasediagramsp.pdf){width="16cm"}
Very recently, the back reaction effect was included in this model [@Nie:2014qma], which showed a rich phase structure and various condensate behaviours such as the “n-type" and “u-type" ones. The phase transitions to the p-wave phase or s+p coexisting phase become first order in strongly back reacted cases. The phase diagrams similar as figure \[phasediagram\] in different strength of back reaction were constructed, indicting that the region for the s+p coexisting phase is enlarged with a small or medium back reaction parameter, while is reduced in the strongly back reacted case.
### The holographic s+p superconductor with a scalar doublet charged under a U(2) gauge field
In this sector, we consider a holographic s+p model consisting of a scalar doublet charged under a U(2) gauge field living in a $(3+1)$-dimensional Schwarzschild-AdS black brane geometry. The action for the matter sector reads [@Amado:2013lia] $$S=\int d^4 x \sqrt{-g}\left(-\frac{1}{4}\tilde{F}^{\mu\nu}_{c}\tilde{F}_{\mu\nu}^c-m^2\Psi^\dagger \Psi-(D^\mu\Psi)^\dagger D_\mu\Psi\right)\,,
\label{accionsp}$$ with $$\begin{aligned}
&&\Psi=\sqrt{2} \begin{pmatrix}
\lambda \\
\psi
\end{pmatrix}\;,\quad
D_{\mu}=\partial_{\mu}-iA_{\mu}\;,\quad
A_{\mu}=A_{\mu}^{c}T_c\,, \\ \nonumber
&&
T_0=\frac{1}{2}\mathbb{I}\;,\quad
T_i=\frac{1}{2}\sigma_i\,.\end{aligned}$$ The system lives in the Schwarzschild-AdS background . Considering the following consistent ansatz for the fields $$A^{(0)}_0= \Phi(r)\,,\quad A^{(3)}_0 = \Theta(r)\,,\quad A^{(1)}_1 = w(r)\,,\quad \psi=\psi(r)\,,$$ with all functions being real-valued, the resulting equations of motion read $$\label{pu2eom}
\begin{split}
\psi''+\left(\frac{f'}{f}+\frac{2}{r}\right)\psi'+\left(\frac{(\Phi-\Theta)^2}{4f^2}-\frac{m^2}{f}-\frac{w^2}{4r^2f}\right)\psi=0\,,\\
\Phi''+\frac{2}{r} \Phi'-\frac{\psi^2}{f}(\Phi-\Theta)=0\,,\\
\Theta''+\frac{2}{r}\Theta'+\frac{\psi^2}{f}(\Phi-\Theta)-\frac{w^2}{r^2f}\Theta=0\,,\\
w''+\frac{f'}{f}w'+\frac{\Theta^2}{f^2}w-\frac{\psi^2}{f}w=0\,.
\end{split}$$ In what follows we choose the scalar to have $m^2 = -2$ and thus the corresponding dual operator has mass dimension 2.
The UV asymptotic behaviour of the fields, corresponding to the solution of equations in the limit $r\to\infty$, is given by $$\begin{split}
&\Phi=\mu-\rho/r+\cdots\,,\\
&\Theta=\mu_3-\rho_3/r+\cdots\,,\\
&w = w^{(0)}+w^{(1)}/r+\cdots\,,\\
&\psi = \psi^{(1)}/r+\psi^{(2)}/r^2+\cdots\,,
\end{split}$$ where in the dual field theory side, $\mu$ and $\rho$ are respectively the chemical potential and charge density corresponding to the overall $U(1)\subset U(2)$ generated by $T_0$, whereas $\mu_3$ and $\rho_3$ are the chemical potential and charge density corresponding to the $U(1)\subset SU(2)$ generated by $T_3$. $\psi^{(1)}$ is the source of a scalar operator of dimension 2, while $\psi^{(2)}$ is its expectation value. Finally $w^{(0)}$ and $w^{(1)}$ are the source and vacuum expectation value of the current operator $J_x^{(1)}$. Notice that in a background where $w(r)$ condenses the $SU(2)\subset U(2)$ is spontaneously broken, and moreover spatial rotational symmetry is spontaneously broken too.
We are looking for solutions of the equations where $\psi$, $w$, or both acquire non-trivial profiles. First we will switch on a chemical potential $\mu$ along the overall U(1), while requiring that the other chemical potential $\mu_3$ remains null. Therefore the UV boundary conditions are $$\psi^{(1)}=0\,,\quad w^{(0)}=0\,,\quad \mu_3 =0\,.
\label{uvconds}$$ In the IR regularity requires $A_t$ to vanish at the black hole horizon. So far, the holographic multi-component superfluid model has been realized.
![\[condensatespp\] Left: Condensates $ \psi^{(2)}$ (solid) and $w^{(1)}$ (dashed) as a function of $1/\mu$ in the s-wave (blue) and s+p-wave (red) phases. The p condensate appears at $\mu_{sp}$ such that $\mu_s/\mu_{sp}\simeq0.395$. The inset zooms in on the plot of $ \psi^{(2)}$ to show the difference in the scalar condensate between the s (blue) and the s+p (red) solutions. Right: Free energy of the different solutions versus $1/\mu$: normal phase in black, s-wave phase in blue, and s+p-wave phase in red. Reprinted with permission from ref. [@Amado:2013lia].](condz.pdf "fig:"){width="7.8cm"} ![\[condensatespp\] Left: Condensates $ \psi^{(2)}$ (solid) and $w^{(1)}$ (dashed) as a function of $1/\mu$ in the s-wave (blue) and s+p-wave (red) phases. The p condensate appears at $\mu_{sp}$ such that $\mu_s/\mu_{sp}\simeq0.395$. The inset zooms in on the plot of $ \psi^{(2)}$ to show the difference in the scalar condensate between the s (blue) and the s+p (red) solutions. Right: Free energy of the different solutions versus $1/\mu$: normal phase in black, s-wave phase in blue, and s+p-wave phase in red. Reprinted with permission from ref. [@Amado:2013lia].](FE2.pdf "fig:"){width="7.9cm"}
In the left plot of figure \[condensatespp\] the condensates $\langle \mathcal{O}_2 \rangle \sim \psi^{(2)}$ and $\langle J_x^{(1)}\rangle\sim w^{(1)}$ are plotted as a function of the chemical potential. Notice that the solution where both condensates coexist extends down to as low $1/\mu$ as where the decoupling limit is trustable. And the free energy for the different solutions is shown in the right plot of figure \[condensatespp\]. At small chemical potential only the normal phase solution exists. At $\mu=\mu_s \simeq 8.127$ there is a second order phase transition to the s-wave solution. If one keeps increasing $\mu$, at $\mu_{sp}\simeq 20.56$ there is a second order phase transition from the s-wave phase to the s+p-wave phase. The system stays in the s+p-wave phase for $\mu>\mu_{sp}$.
Next, we relax the condition $\mu_3 =0$ and study the phase diagram of the system as a function of $\mu$ and $\mu_3/\mu$. Notice that turning on a second chemical potential means to explicitly break $U(2)\rightarrow U(1)\times U(1)$. The system can now be interpreted as a holographic dual to an unbalanced mixture. The UV boundary conditions now read $$\psi^{(1)}=0\,,\quad w^{(0)}=0\,.\label{uvconds2}$$
![\[phasediagram\_2\] Phase diagram of the unbalanced system as a function of $1/\mu$ and $\mu_3/\mu$. Second order phase transitions are denoted by blue lines, whereas the red line corresponds to a first order phase transition. Reprinted with permission from ref. [@Amado:2013lia].](phasediagram_2.pdf)
In this case, by computing the free energy of the different solutions, the phase diagram of the system as a function of $1/\mu$ and $\mu_3/\mu$ is plotted in figure \[phasediagram\_2\]. For small values of $\mu_3/\mu$, the solution where both condensates coexisting extends down to as lower $1/\mu$ as where we can trust the decoupling limit. As $|\mu_3|/\mu$ gets larger, the transition to the s+p-wave phase happens at a higher value of $\mu$. For $|\mu_3|/\mu$ is large enough, the p-wave phase is preferred at intermediate values of $\mu$. Therefore, as $\mu$ is increased above a critical value $\mu_p$ the system goes from the normal to the p-wave phase through a second order phase transition. If $\mu$ is increased even further a first order phase transition takes the system from the p-wave to the s-wave phase. The tricritical point where the normal, s-wave and p-wave phases meet happens at $1/\mu\simeq 0.223$ and $|\mu_3|/\mu\simeq 0.815$, the p-wave solution is never energetically preferred for $|\mu_3|/\mu< 0.815$.
In summary, in this subsection we have reviewed the competition between s-wave and p-wave order through two holographic superconductor models. The first model is realized the s+p superconductor with a scalar triplet charged under an SU(2) gauge field and the other is constructed with a scalar doublet charged under a U(2) gauge field. The s+p coexisting phase exists in both models. In the first model, the s+p coexisting phase is narrow and one condensation tends to kill the other. This competing behaviour is similar to the case shown in the condensed matter system [@Pallab:2014]. However, in the second case, the condensates feed on different charge densities and the coexisting phase survives down to a low temperature. Therefore, it should be noted that the competing scenario is model dependent. In next subsection, we will study the competition between the scalar field and the tensor field, i.e., the competition between the s-wave and d-wave orders.
Competition between s-wave order and d-wave order {#sd}
-------------------------------------------------
In section \[sect:dwave\], we have mentioned two acceptable holographic models describing the d-wave condensation. The CKMWY d-wave model is reviewed in subsection \[subsect:CKMWY model\] and the BHRY d-wave model in subsection \[subsect:BHRY model\]. In order to realize the condensation of s-wave order and d-wave order in one holographic model, one can simply combine the Abelian-Higgs model with a d-wave model. Thus, one could have two holographic models with s-wave order and d-wave order. Here we will discuss the competition between s-wave order and d-wave order for both d-wave models in the probe limit where one neglects the back reaction of matter fields to the background geometry [@Li:2014wca]. The phase structures are given and the behaviours of the thermodynamic quantities for the s+d coexisting phase are also studied. The coexisting phase does appear in both models and is thermodynamically favored.
### The s-wave + BHRY d-wave model {#sect:superfluid}
To study the competition between s-wave and d-wave orders, let us first start with the holographic model by combining the Abelian-Higgs s-wave model and BHRY d-wave model. The holographic model with a scalar field $\psi_1$, a symmetric tensor field $\varphi_{\mu\nu}$ and a U(1) gauge field $A_\mu$ is described by the following action [@Li:2014wca]: [^28]
$$\label{BHRY}
\begin{split}
S=\frac{1}{2\kappa^2}\int d^{4}x\sqrt{-g}(- \frac{1}{4} F_{\mu\nu} F^{\mu\nu} -|D\psi_1|^2-m_1^2|\psi_1|^2+ \mathcal{L}_d),\\
\mathcal{L}_d=-|\tilde{D}_\rho\varphi_{\mu\nu}|^2+2|\tilde{D}_\mu\varphi^{\mu\nu}|^2+
|\tilde{D}_\mu\varphi|^2
-\big[\tilde{D}_\mu\varphi^{\dagger\mu\nu}\tilde{D}_\nu \varphi+\text{h.c.}\big]-i q_2 F_{\mu\nu} \varphi^{\dagger\mu\lambda} \varphi^\nu_\lambda\\
-m_2^2\big(|\varphi_{\mu\nu}|^2-|\varphi|^2\big)+2R_{\mu\nu\rho\lambda} \varphi^{\dagger\mu\rho}\varphi^{\nu\lambda}-\frac{1}{4} R |\varphi|^2,
\end{split}$$
where $D_{\mu} = \nabla_\mu - i q_1 A_\mu$ and $\tilde{D}_\mu = \nabla_\mu - i q_2 A_\mu$, $\varphi\equiv{\varphi^\mu}_\mu$, $\varphi_\rho\equiv g^{\mu\lambda}\tilde{D}_{\lambda}\varphi_{\mu\rho}$ and ${R^\mu}_{\nu\rho\lambda}$ is the Riemann tensor of the background metric. $\psi_1$ is the scalar order and $\psi_{\mu\nu}$ is the tensor order. The parameters $q_1$ and $q_2$ are the charges of the scalar and the tensor fields, respectively. One can perform a rescaling to set the charge $q_1$ of the scalar to be unity. Then the phase structure of this theory is determined by the ratio $q_2/q_1$ by fixing the mass square of the scalar field $m_1^2$ and the mass square of the tensor field $m_2^2$. We shall set $q_1=1$ without loss of generality in the following discussion.
Working in the probe limit, we choose the background metric to be the 3+1 dimensional AdS-Schwarzschild black hole with planar horizon . And we consider the following ansatz $$\label{ansatz_1sd}
A_\mu \, dx^\mu = \phi(r) \, dt \;, \quad \psi_1=\psi_1(r)\quad
\varphi_{xy}=\varphi_{yx} = \frac{r^2}{2} \, \psi_2(r) \;,$$ with $\phi(r)$, $\psi_1(r)$ and $\psi_2(r)$ all real functions.
With the above ansatz , the equations of motion for $\phi$, $\psi_1$ and $\psi_2$ are given by $$\begin{aligned}
\label{EOMs}
\begin{split}
\phi'' + \frac{2 \phi'}{r} - \frac{2}{f}\phi \psi_1^2- \frac{q_2^2} {f} \phi \psi_2^2=&0, \\
\psi_1'' + \frac{f'}{f} \psi_1'+ \frac{2}{r} \psi_1' + \frac{\phi^2}{f^2} \psi_1- \frac{m_1^2}{f} \psi_1=&0, \\
\psi_2'' + \frac{f'}{f} \psi_2' + \frac{2}{r} \psi_2' + \frac{ q_2^2 \phi^2}{f^2} \psi_2- \frac{m_2^2}{f} \psi_2=&0.
\end{split}\end{aligned}$$ Here the prime denotes the derivative with respect to $r$. With this model at hand, we can study the competition mechanism between the s-wave order and d-wave order. It is easy to see that equations have a symmetry $$\label{symmetry}
m_1^2\leftrightarrow m_2^2,\ q_2\rightarrow 1/q_2, \ \phi\rightarrow q_2 \phi,\ \psi_1\rightarrow q_2\psi_2 /\sqrt{2},\ \ \psi_2\rightarrow\sqrt{2} q_2\psi_1.$$ Under this symmetry transformation, the role of s-wave and d-wave would interchange each other. Without loss of generality, here we focus on the case $m_1^2<m_2^2$.
Before solving the set of coupled equations numerically, we make a briefly qualitative analysis on the possible phases for such a model. Following ref. [@Basu:2010fa], we rephrase the equations for the s-wave and d-wave as a potential problem. The evolution equations for s-wave and d-wave in equations can be rewritten as follows $$\begin{aligned}
\begin{split}
\frac{d^2}{d y^2}\tilde{\psi_1}-\tilde{V}_{1eff}(y)\tilde{\psi_1}&=&0, \\
\frac{d^2}{d y^2}\tilde{\psi_2}-\tilde{V}_{2eff}(y)\tilde{\psi_2}&=&0,
\end{split}\end{aligned}$$ where $dy=-\frac{dz}{z^2 f}$ with $z=1/r$ and $V_{1eff}(z)=-f^2(\frac{\phi^2}{f^2}-\frac{m_1^2}{f}+\frac{f_{,z}}{f}z^3)$ and $V_{2eff}(z)=-f^2(\frac{q_2^2 \phi^2}{f^2}-\frac{m_2^2}{f}+\frac{f_{,z}}{f}z^3)$. Now in terms of the new variable $y$, the equations of motion for s-wave and d-wave are rephrased as a potential problem on a semi infinite line, i.e., $y\in [0, \infty)$. Our qualitative discussion is based upon the lemma proven in ref. [@Basu:2010fa]. For the case $q_2^2<1$, no matter which gauge field configuration we choose, we always have $V_{1eff}<V_{2eff}$. Therefore the phase structure of the system is the same as that of s-wave holographic superconductor with a single scalar. While, for the case $q_2^2\geq1$, one may expect that the d-wave field with large charge $q_2$ will always dominate. However, the potential $V_{1eff}$ diverges like $\frac{1}{y^2}$ near the boundary $y=0$ when we lower the temperature. Therefore, lowering the temperature possibly makes the mass dependent potential more important and hence the s-wave tends to dominate. We will confirm this with the following numerical calculation.
Here we set the mass square $m_1^2=-2$ and $m_2^2=7/4$ and we take $q_2=2.66$ as a typical example. Our numerical results confirm that the model does admit the coexistence region of two orders with different symmetry, which is drawn in figure \[co\]. We find that the s+d coexisting phase starts from the d-wave phase and ends with the pure s-wave condensate phase. The calculation of the free energy confirms that the coexisting s+d phase is thermodynamically favored as shown in figure \[free\].
![\[co\] Condensate of the operators in the s+d coexisting phase. The blue curve is for the condensate of the d-wave operator, while the red curve is for the s-wave operator. The figure was taken from ref. [@Li:2014wca].](s_and_d.pdf "fig:")
![\[free\] The left plot shows the difference of Gibbs free energy between the superconducting phase and the normal phase. The blue curve is for the d-wave phase, the green line is for the s-wave phase, while the red curve is for the s+d coexisting phase. The right plot is an enlarged version of the left one to show the s+d phase more clearly. Figures taken from ref. [@Li:2014wca].](free.pdf "fig:") ![\[free\] The left plot shows the difference of Gibbs free energy between the superconducting phase and the normal phase. The blue curve is for the d-wave phase, the green line is for the s-wave phase, while the red curve is for the s+d coexisting phase. The right plot is an enlarged version of the left one to show the s+d phase more clearly. Figures taken from ref. [@Li:2014wca].](free_blowup.pdf "fig:")
Based on the above discussions, it can be seen clearly that there exist three particular points at which the derivative of the charge density with respect to temperature is discontinuous, indicating a second order phase transition. The one with the highest temperature is the critical point for the superconducting phase transition, while the remaining two points are inside the superconducting phase, indicating the appearance and disappearance of coexisting phase. We can also see the signal of phase transition from the behaviour of the total charge density as a function of temperature and the ratio $\rho_s/\rho$ versus temperature shown in figure \[ratio\_1\], where $\rho_s$ is the superconducting charge density $\rho_s=\rho-\rho_n$ and $\rho_n$ is the normal charge density carried by the black hole.
![\[ratio\_1\] Left: The total charge density as a function of temperature. Right: The ratio of the superconducting charge density over the total charge density $\rho_s/\rho$ versus temperature. The plots were taken from ref. [@Li:2014wca].](rho.pdf "fig:") ![\[ratio\_1\] Left: The total charge density as a function of temperature. Right: The ratio of the superconducting charge density over the total charge density $\rho_s/\rho$ versus temperature. The plots were taken from ref. [@Li:2014wca].](rhos.pdf "fig:")
In order to ensure the system is indeed in a superconducting state, we calculate the optical conductivity $\sigma(\omega)$, which corresponds to the red line shown in figure \[aa\]. We see that much more interesting phenomena happen in the low frequency region. Unlike the s-wave case which only has a bump at $\omega/T \simeq 400$ in figure \[aa\], for pure d-wave condensate, apart from a much more obvious bump at $\omega/T \simeq 500$, Re($\sigma_{xx}$) has an additional spike at a lower frequency. This spike may indicate the existence of a bound state [@Benini:2010pr]. One can see clearly that such peak becomes much more sharper in the s+d coexisting state, thus the bound state is enhanced due to the additional condensate of s-wave order.
![\[aa\] The real part (left) and imaginary part (right) of the conductivity as a function of frequency at temperature $T=0.018\mu$. The red curve is for the s+d coexisting phase, the green line is for the pure s-wave phase and the blue curve for the pure d-wave phase. Plots taken from ref. [@Li:2014wca]. ](real.pdf "fig:") ![\[aa\] The real part (left) and imaginary part (right) of the conductivity as a function of frequency at temperature $T=0.018\mu$. The red curve is for the s+d coexisting phase, the green line is for the pure s-wave phase and the blue curve for the pure d-wave phase. Plots taken from ref. [@Li:2014wca]. ](imagine.pdf "fig:")
![\[phase\_1\] The $q_2$-$T$ phase diagram. The four phases are colored differently and we label the most thermodynamically favoured phase in each region. The figure was taken from ref. [@Li:2014wca].](phase_1.pdf "fig:")
To be complete, we give the phase diagram with $m_1^2=-2$ and $m_2^2=7/4$ shown in figure \[phase\_1\], which can tell us in which region the coexisting phase appears. From figure \[phase\_1\], we see that the coexisting phase exists only in a narrow region in the phase diagram. We denote the critical temperature for a single s-wave or d-wave starting to condense as $T_{cs}$ and $T_{cd}$. If we set the charges of the s-wave and d-wave fields to unity, then $T_{cs}/\mu \simeq 0.0588$ and $T_{cd}/\mu\simeq0.0253$. We see that
- In the regime $q_2<T_{cs}/T_{cd}\simeq 2.323$, the s-wave dominates the system and there is no condensation of the d-wave order.
- As $q_2$ increases beyond $2.323$, the s+d phase appears, which emerges from the d-wave phase and ends with a pure s-wave.
- If we continue increasing $q_2$ to the case $q_2>1.155 T_{cs}/T_{cd}\simeq 2.683$, the s-wave order never condenses and the resulting phase diagram is the same as that of model with only d-wave order.
Finally, we try to give a qualitative explanation on the mechanism through which the condensation of one order affects the dynamics of the other order. Note that here the back reaction is not taken into account. Thus the two fields interact only through their effect on the gauge field once one or both has (have) condensed. Through looking at the gauge field we may give some insight into the competing mechanics between two orders.
![\[eff\_mass\] The blue curve is the effective mass square of s-wave without the condensation of d-wave. The red curve is the effective mass square of the s-wave under the condensation of d-wave. It can be seen clearly that the effective mass of s-wave increases after the condensation of d-wave. This figure was taken from ref. [@Li:2014wca].](effective_mass.pdf "fig:")
- First, after the d-wave order condenses, if one keeps lowering the temperature and reaches the critical temperature at which the pure s-wave would condense, this condensation does not happen. This is due to the fact that the condensation of the d-wave increases the effective mass of the s-wave, thus prevents the instability of the s-wave to happen, which can be seen from figure \[eff\_mass\]. This reflects the competition between s-wave and d-wave.
- However, further decreasing the temperature, the condensation of s-wave does happen. This is due to the fact that the effective mass of the s-wave is lowered and ultimately even if the condensation of the d-wave depleted the gauge potential, the background with only d-wave order becomes unstable.
- Finally, the condensate of the s-wave order kills the first one. This should be due to the effective mass square of the s-wave being lower.
It should be noted that this phenomenon is model dependent. This narrow coexistence region of two superconducting orders and the fact that one condensate can eventually kill the other also happen for two s-wave orders in ref. [@Basu:2010fa] and the s+p case in ref. [@Nie:2013sda]. The competition diagram is similar to the competition between the conventional s-wave and the triplet Balian-Werthamer or the B-phase pairings in the doped three dimensional narrow gap semiconductors, such as $\mathrm{Cu}_x\mathrm{Bi}_2\mathrm{Se}_3$ and $\mathrm{Sn}_{1-x}\mathrm{In}_x\mathrm{Te}$ in the condensed matter system [@Pallab:2014]. Although in ref. [@Pallab:2014] the competition is apparently between a s-wave order and a p-wave order, d-wave and p-wave are similar in some circumstances, for example, their excitations of the normal component can be probed using low frequency photons.
### The s-wave + CKMWY d-wave model
With the same strategy, in this subsection we study the competition between s-wave order and d-wave order in the model combining the Abelian-Higgs s-wave model [@Hartnoll:2008vx] with the CKMWY d-wave model [@Chen:2010mk]. The full action including a U(1) gauge field $A_\mu$, a complex scalar field $\psi_1$ and a symmetric, traceless tensor field $B_{\mu\nu}$ takes the following form [@Li:2014wca] $$\label{CKMWY}
S=\frac{1}{2\kappa^2}\int d^4
x\sqrt{-g}(-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-|D\psi_1|^2-m_1^2|\psi_1|^2+\tilde{\mathcal{L}}_d),$$ with $$\tilde{\mathcal{L}}_d=-g^{\mu\lambda}(\tilde{D}_{\mu}B_{\nu\gamma})^{\dagger}\tilde{D}_{\lambda}B^{\nu\gamma}-m_2^2B_{\mu\nu}^{\dagger}B^{\mu\nu}.$$ Here $D_\mu = \nabla_\mu - i q_1 A_\mu$ and $\tilde{D}_\mu = \nabla_\mu - i q_2 A_\mu$. In the probe limit, matter fields can be treated as perturbations in the 3+1 dimensional AdS black hole background . Let us consider the following ansatz $$\psi_1=\psi_1(r),\quad B_{xx}=-B_{yy}=\psi_2(r),\quad A_t=\phi(r)dt,$$ with all other field components being turned off and $\psi_1(r)$, $\psi_2(r)$ and $\phi(r)$ being real functions. Then the explicit equations of motion read $$\label{EOM_2}
\begin{split}
\phi''+\frac{2}{r}\phi'-\frac{4 q_2^2 \psi_2^2}{r^4f}\phi-\frac{2q_1^2 \psi_1^2}{f}\phi=0,\\
\psi_1''+(\frac{f'}{f}+\frac{2}{r})\psi_1'+\frac{q_1^2\phi^2}{f^2}\psi_1-\frac{m_1^2}{f}\psi_1=0,\\
\psi_2''+(\frac{f'}{f}-\frac{2}{r})\psi_2'+\frac{q_2^2\phi^2}{f^2}\psi_2-\frac{2f'}{rf}\psi_2-\frac{m_2^2}{f}\psi_2=0.
\end{split}$$
![\[condensate\_2\] The let plot shows the condensation in the s+d coexisting phase. The right plot shows the differences of Gibbs free energy between superconducting phases and the normal phase. Here the blue line stands for the d-wave phase, the green one for the s-wave phase and the red one for the s+d coexisting phase. The figure is reproduced from ref. [@Li:2014wca].](s_and_d_2.pdf "fig:") ![\[condensate\_2\] The let plot shows the condensation in the s+d coexisting phase. The right plot shows the differences of Gibbs free energy between superconducting phases and the normal phase. Here the blue line stands for the d-wave phase, the green one for the s-wave phase and the red one for the s+d coexisting phase. The figure is reproduced from ref. [@Li:2014wca].](free_blowup_2.pdf "fig:")
The numerical results are shown in figure \[condensate\_2\] for the case with $q_1=1$, $q_2=1.34$, $m_1^2=-2$ and $m_2^2=-13/4$. As seen, a new phase with both s-wave order and d-wave order coexistence can appear near $T^{cross}$ and this s+d coexisting phase has the lowest free energy and is thus thermodynamically preferred to the s-wave phase and d-wave phase. In more detail, as we lower the temperature of the system, it first undergoes a phase transition from the normal phase to the pure d-wave phase at $T_c^d$. Then at $T_c^{sd1}$, a new phase transition occurs, and the system goes into an s+d coexisting phase. Finally the system undergoes the third phase transition from the s+d coexisting phase to a pure s-wave phase at $T_c^{sd2}$. Note that all the three phase transitions are second order.
The feature of the phase transitions can also be seen clearly from the charge density as the function of temperature in figure \[charge\_2\]. One can see that the charge density with respect to temperature is continuous, but its derivative is discontinuous at three special points, indicating three second order phase transitions. These features are the same as those for the model in the previous subsection. But there is a little difference in the behaviour of the total charge density for the d-wave phase. In the s-wave + BHRY d-wave model, the total charge density changes monotonously with the temperature, while it behaves non-monotonous in the present case.
![\[charge\_2\] The total charge density as a function of the temperature. The red curve is for the normal phase, while the blue one corresponds to the superconducting phase. There are three special temperatures at which the derivatives of charge density with the temperature are discontinuous. Figure taken from ref. [@Li:2014wca].](rho_2.pdf "fig:")
![\[ratio\_2\] The ratio of the superconducting charge density over the total charge density, $\rho_s/\rho$, with respect to the temperature. The red curve describes the ratio $\rho_s/\rho$ when the system transfers from the d-wave phase to s-wave phase through the s+d coexisting phase. The green dashed blue curve is for the ratio $\rho_s/\rho$ of the pure s-wave phase and the blue dashed curve is the ratio for the pure d-wave phase. The figure was taken from ref. [@Li:2014wca].](rhos_2.pdf "fig:")
The information of the phase transitions can also be revealed via the behaviour of the ratio $\rho_s/\rho$ with respect to the temperature. From figure \[ratio\_2\], one can see that the ratio $\rho_s/\rho$ also has a small kink in the region of the coexisting phase. Comparing figure \[ratio\_1\] with figure \[ratio\_2\], we see that in the former case, the green dashed curve for the pure s-wave phase intersects with the blue dashed curve for the pure d-wave phase. In contrast, the green dashed curve in figure \[ratio\_2\] is always lower than the blue dashed curve. Therefore, as one lowers the temperature, the ratio $\rho_s/\rho$ in the s+d coexisting phase increases for the former , while it decreases for the latter . The authors of ref. [@Nie:2013sda] investigated an s+p coexisting phase and found the decrease of the ratio $\rho_s/\rho$ in the coexisting phase, similar to figure \[ratio\_2\]. They suggested that it might be an experimental signal of the phase transition from a single condensate phase to a coexisting phase. Nevertheless, the results here uncover that the ratio $\rho_s/\rho$ versus temperature is model dependent.
The phase diagram for the model with $m_1^2=-2$ and $m_2^2=-13/4$ in the $q_2-T$ plane is shown in figure \[phase\_2\]. As the s-wave + BHRY d-wave model, the system also contains four kinds of phases known as the normal phase, s-wave phase, d-wave phase and s+d coexisting phase. The normal phase dominates in the high temperature region, the s-wave phase dominates in the lower temperature region with small $q_2$ below the red curve, and the d-wave phase dominates in the higher temperature zone with large $q_2$ above the blue curve. The s+d coexisting phase is favoured in the area between the red and blue curves. The region for the s+d coexisting phase is very narrow in the phase diagram, which indicates that the s-wave and d-wave phases generally repel each other, but they can coexist in a very small range of temperature.
![\[phase\_2\] The $q_2$-$T$ phase diagram with $m_1^2=-2$ and $m_2^2=-\frac{13}{4}$. The most thermodynamically favored phase in each part is labeled. The s+d coexisting phase exists only in a narrow region. The right plot is an enlarged version for the coexisting region in order to see this more clearly. Plots taken from ref. [@Li:2014wca].](phase_2.pdf "fig:") ![\[phase\_2\] The $q_2$-$T$ phase diagram with $m_1^2=-2$ and $m_2^2=-\frac{13}{4}$. The most thermodynamically favored phase in each part is labeled. The s+d coexisting phase exists only in a narrow region. The right plot is an enlarged version for the coexisting region in order to see this more clearly. Plots taken from ref. [@Li:2014wca].](phase2_blow.pdf "fig:")
Comparing the two holographic setups, i.e., the model and the model , one can see some common features as follows.
- The s+d coexisting phase does exist in a region of the model parameter $q_2/q_1$. Once the coexisting phase appears, it is always thermodynamically favoured, compared to the pure s-wave and pure d-wave superconducting phases, which can be seen from the free energy in figure \[free\] and figure \[condensate\_2\].
- All phase transitions are second order in these two holographic models.
- One can see from figure \[phase\_1\] and figure \[phase\_2\] that the phase structure is very similar for both models. The region for the s+d coexisting phase is very narrow in the phase diagram, indicating that the s-wave and d-wave phases generally repel each other.
There exist also some differences in the two models. For suitable model parameters in the first model, as the temperature is lowered, the s-wave order condenses inside the d-wave order resulting in the coexisting phase with both orders. However, when the scalar order condenses the first one starts to disappear, and finally only the s-wave condensate is left for sufficiently low temperatures. If one changes the model parameter $m_1^2\leftrightarrow m_2^2$, the inverse is also true: the condensate of d-wave order emerges following the condensate of s-wave order, and then the d-wave condensate finally kills the s-wave order. Those two kinds of coexisting phase are one to one correspondence. In contrast, in the second model, one sees only the first kind of the coexisting phase. What’s more, for the first model, the ratio $\rho_s/\rho$ increases in the s+d coexisting phase as the temperature is lowered, while it decreases in the second case. This gives an obvious evidence that the ratio $\rho_s/\rho$ versus temperature is model dependent.
Coexistence and Competition of Magnetism and Superconductivity {#sect:M&S}
==============================================================
The novel paired mechanism makes p-wave superconductor have many features which are different from the traditional knowledge coming from s-wave superconductor both in theories and experiments. In the usual picture, superconductivity and magnetism are incompatible with each other. Especially, ferromagnetic phase, a spontaneously magnetized phase which has nonzero magnetic moment without external magnetic field and appears when the temperature is lower than a critical one called “Curie temperature", can not coexist with superconductivity at a sample.[^29] This is rooted in the microscopic theory of superconductivity from BCS theory. However, this understanding is broken by p-wave superconductor. The discovery of the superconducting ferromagnet[^30] materials, such as UGe$_2$ [@Lonzarich], URhGe [@Aoki], UCoGe [@Huy] and ZrZn$_2$ [@Pfleiderer], came as a big surprise. In this material, superconductivity is realized well below the Curie temperature, without expelling the ferromagnetic order.
The nature of superconducting state in ferromagnetic materials is currently under debate. For a review of phenomenological theory of ferromagnetic unconventional superconductors with spin-triplet Cooper pairing of electrons, one can see refs. [@D.I.Uzunov; @MacHida; @Nevidomskyy]. However, the microscopic theory about the coexistence of magnetism and superconductivity in strongly interacting heavy electrons is either too complex or insufficiently developed to describe the complicated behaviour. So it is still a fascinating thing to find a suitable theory to describe the coexistence and competition of the ferromagnetism and superconductivity in strong correlated system.
Holographic frame to discuss the coexistence and competition between spontaneously magnetic order phase and superconductivity initiated in refs. [@Amoretti:2013oia; @Iqbal:2010eh]. Because of lacking an individual model to describe spontaneously magnetization and the time reversal symmetry broken, these models cannot give complete features of this topic. A very new idea proposed in ref. [@Cai:2014oca] tries to give an independent model describing spontaneously magnetization in holographic frame, which opens a new direction. We will introduce the main results in this framework. For more details, one can refer to refs. [@Cai:2014oca; @Cai:2014dza; @Cai:2015mja].
The holographic model for ferromagnetism/paramagnetism phase transition
-----------------------------------------------------------------------
Before going on the topic of coexistence of ferromagnetism and superconductivity, let’s first review how to build a holographic ferromagnetism which is independent on superconductivity in ref. [@Cai:2014oca]. This model is realized by adding a real antisymmetric field into Einstein-Maxwell theory in a (3+1) dimensional AdS spacetime, $$\label{LM2}
S=\frac1{2\kappa^2}\int d^4x\sqrt{-g}(\mathcal{R}+\frac6{L^2}-F^{\mu\nu}F_{\mu\nu}+\lambda^2 L_M)£¬$$ where $$\label{LM3}
L_M=-\frac14\nabla^\mu M^{\nu\tau}\nabla_\mu M_{\nu\tau}-\frac{m^2}4M^{\mu\nu}M_{\mu\nu}-\frac12M^{\mu\nu}F_{\mu\nu}-\frac J8 V(M_{\mu\nu}).$$ Here $2\kappa^2=16\pi G$ and $G$ is the Newtonian gravitational constant, $\lambda$ and $J$ are two constants, $m$ is the mass of the real tensor field $M_{\mu\nu}$, $A_\mu$ is the gauge potential of U(1) gauge field. The antisymmetric tensor field $M_{\mu\nu}$ is the effective polarization tensor of the U(1) gauge field strength $F_{\mu\nu}$ with the self-interaction $V(M_{\mu\nu})$ which should be expanded as the even power of $M_{\mu\nu}$. The probe limit corresponds to $\lambda\rightarrow0$. Under this limit, the equation for polarization field decouples from the gauge field and gravity field, $$\label{eomFM}
\nabla^2M_{\mu\nu}-m^2M_{\mu\nu}-J{M_\mu}^\delta {M_\delta}^\tau {M_\tau}_\nu-F_{\mu\nu}=0,$$ with the dyonic Reissner-Nordström (RN) background [@Cai], $$\begin{aligned}
\label{geom}
ds^2 &=& r^2(-f(r)dt^2+dx^2+dy^2)+\frac{dr^2}{r^2f(r)},\nonumber \\
f(r) &=&1-\frac{1+\mu^2+B^2}{r^3}+\frac{\mu^2+B^2}{r^4},\\
A_\mu &=& \mu(1-1/r)dt+Bxdy. \nonumber
$$ Here the horizon radius has been scaled to be unitary. If we only care about the magnetic part of polarization field, then a self-consistent ansatz for polarization field is $M_{\mu\nu}=-p(r)dt\wedge dr+\rho(r)dx\wedge dy$. Taking this ansatz into equation , we have $$\label{rhop}
\begin{split}
\rho''+\frac{f'\rho'}f-\left( \frac{2f'}{rf}+\frac4{r^2}+\frac{m^2}{r^2f}\right)\rho+\frac{J\rho^3}{r^6f}-\frac{B}{r^2f}=0,\\
p''+\left(\frac{f'}{f}+\frac4r\right)p'-\left(\frac2{r^2}+\frac{m^2}{r^2f}\right)p-\frac{Jp^3}{r^2f}-\frac{\mu}{r^4f}=0,
\end{split}$$ where a prime denotes the derivative with respect to $r$. It is interesting to see that these two equations decouple from each other in this case, which makes it to be possible that we can neglect the dynamic of $p(r)$ if we only care about the dynamic of magnetism. At the horizon, we need to impose a regular boundary condition. Near AdS boundary, the linearized equations have following asymptotic solutions for $\rho(r)$, $$\label{asym2}
\rho\sim \rho_+r^{(1+\delta)/2}+\rho_-r^{(1-\delta)/2}-\frac{B}{4+m^2},\\$$ with $\delta=\sqrt{17+4m^2}$. In order to make the theory self-consistent and spontaneous condensation appear, we need following restriction on parameters, $$\label{m2}
-4<m^2<-\frac32, ~~~ \text{and}~~\rho_+=0.$$ According to the action one can derive the magnetic moment from polarization field, which reads $$\label{QN1}
N=-\frac{1}{2}\int_1^\infty dr\frac{\rho}{r^2}.$$ Here we have set the constant $\lambda=1$ in this expression for convenience. This integration converges only when $\rho_+=0$. In the case without external magnetic field, i.e. $B=0$, if there is a solution such that $\rho(r)\neq0$, the magnetic moment then is nonzero, which gives a ferromagnetic phase for dual boundary. Because the action implies transformation for $\rho(r)$ such as $\rho(r)\rightarrow-\rho(r)$ under the time reversal transformation, the condensed phase of $\rho$ gives a time reversal symmetry broken spontaneously, which is necessary for magnetic ordered phase.
Ferromagnetism and p-wave superconductivity
-------------------------------------------
Once the two independent models for ferromagnetic phase transition and p-wave superconductor are in hand, we can combine them to discuss the possibility of coexistence. For example, we can combine the Einstein-Maxwell-complex vector theory for p-wave superconductor with ferromagnetic model. The complete action reads [@Cai:2014dza] $$\label{action1}
S=\int d^4x\sqrt{-g}\left[\mathcal{R}+\frac{6}{L^2}-F_{\mu\nu} F^{\mu \nu}+\lambda^2(\mathcal{L}_{\rho}+\mathcal{L}_{M}+\mathcal{L}_{\rho M})\right],$$ with $$\label{LrhoM}
\begin{split}
&\mathcal{L}_\rho=-\frac{1}{2}\rho_{\mu\nu}^\dagger\rho^{\mu\nu}-m_1^2\rho_\mu^\dagger\rho^\mu+iq\gamma \rho_\mu\rho_\nu^\dagger F^{\mu\nu}-V_\rho,\\
&\mathcal{L}_M=-\frac14\nabla^\mu M^{\nu\tau}\nabla_\mu M_{\nu\tau}-\frac{m_2^2}4M^{\mu\nu}M_{\mu\nu}-\frac{1}2M^{\mu\nu}F_{\mu\nu}-V_M,\\
&\mathcal{L}_{\rho M}=-i\alpha\rho_\mu\rho_\nu^\dagger M^{\mu\nu},\\
&V_\rho=-\frac{\Theta}2 \rho_{[\mu}\rho_{\nu]}^\dagger\rho^\mu\rho^{\dagger\nu}.
\end{split}$$ Here $\alpha\neq0$ and $\Theta$ are two coupling constant. $L_M$ is the Lagrangian for polarization field which is just as the same as . $L_\rho$ is the Lagrangian for complex vector field, which is similar to the one we discussed before. However, there is an additional term $V_\rho$ which describes the magnetic moment interaction of complex vector field. This term is irrelevant for the previous section where we only care about superconductivity but is relevant when we care about spontaneous magnetization.
Under the probe limit $\lambda\rightarrow0$, a self-consistent ansatz of action is, $$\label{VMansatz}
M_{\mu\nu}=-p(r)dt\wedge dr+h(r)dx\wedge dy,~~\rho_\mu=\rho_xdx+i\rho_ydy.$$ Then we can get the equations of motion for complex vector field and polarization field under the background , $$\label{eqcomp1}
\begin{split}
h''+\frac{f'}fh'+\left(\frac{Jh^2}{r^6f}-\frac{2f'}{rf}-\frac4{r^2}-\frac{m_2^2}{fr^2}\right)h-\frac{2c\alpha \rho_x^2}{r^2f}=0,\\
\rho_x''+(\frac{f'}f+\frac2r)\rho_x'+\left(\frac{q^2\phi^2}{r^4f^2}-\frac{\Theta c^2\rho_x^2}{r^4f}-\frac{m_1^2}{fr^2}-\frac{ch\alpha}{fr^4}\right)\rho_x=0,\\
c''+\left(\frac{f'}f+\frac2r+\frac{2\rho_x'}{\rho_x}\right)c'-\frac{(1-c^2)(c\Theta\rho_x^2+\alpha h)}{fr^4}=0,\\
\end{split}$$ where we have defined $c(r)$ as $\rho_y(r)=c(r)\rho_x(r)$. Note that the equation for $p(r)$ decouples from the others. The linearized equations near the AdS boundary give following asymptotic solutions [^31] $$\label{adsbound}
\begin{split}
\rho_x={\rho_x}_+r^{(\delta_1-1)/2}+{\rho_x}_-r^{-(\delta_1+1)/2},\quad c=c_+r^{\delta_1}+c_-,\\
h(r)=h_+r^{(1+\delta_2)/2}+h_-r^{(1-\delta_2)/2},
\end{split}$$ where $\delta_1=\sqrt{1+4m_1^2}$ and $\delta_2=\sqrt{17+4m_2^2}$ with $m_1^2>-1/4,~m_2^2>-4$. As the previous subsection, we should impose the condition $h_+=0$ for the polarization field and ${\rho_x}_+=c_+=0$ for the complex vector field, i.e., we require that the condensation and magnetization would happen spontaneously. The equations have solutions only when $c(r)=0,\pm1$. Because of the equivalent of $\alpha\rightarrow-\alpha$ and $c\rightarrow-c$, we assume $\alpha>0$ without loss of generality. The magnetic moment is defined as the same as , $$\label{TolM}
N=-\int_{r_h}^\infty\frac h{2r^2}dr.$$ According to the dictionary of AdS/CFT, the expectation value of p-wave superconducting order parameter is a complex vector $\overrightarrow{P}$, whose mode is $P=\sqrt{1+c^2}|\rho_{x-}|$. Though the expression of magnetic moment density does’t contain the terms of complex vector field, it is effected by $\rho_\mu$ through the mixture terms in equations .
In the pure p-wave model, the global U(1) and spatial rotation symmetries are broken spontaneously when $\rho_x$ or $\rho_y$ is nonzero without source. Here it is also true. Moreover, there is an additional symmetry breaking. If one notes following rules for time reversal transformation, $$\label{timerev}
h\rightarrow-h,~~~\rho_y\rightarrow-\rho_y,$$ then when $h\neq0$ or $\rho_y=\pm\rho_x\neq0$ (they both lead nonzero magnetic moment), the time reversal symmetry is broken spontaneously, which agrees with the fact that a spontaneously magnetized phase is with a time reversal symmetry broken spontaneously.
Because the complex vector field and polarization field can condense in low temperatures in an AdS RN black hole background respectively, this model gives a wide possibility to investigate the influence between p-wave superconductivity and spontaneous magnetization. We take $T_{sc0}$ and $T_{C0}$ as the critical temperatures of $\rho_x$ and $h$, when $\alpha=0$. Depending on the values of them, the p-wave superconudcting order or ferromagnetism will appear first. The interesting question is whether the other phase transition can still happen.
Coexistence of superconductivity and ferromagnetism
---------------------------------------------------
The first case we will consider is $T_{C0}>T_{sc0}$, i.e., the ferromagnetic phase appears first. The equation for $c$ in equations shows that $c\neq0$ if $h\neq0$. So there isn’t a phase such that $\{h<0,~\rho_x\neq0,~\rho_y=0\}$. When temperature is decreased to lower than $T_{C0}$, five kinds of phases may appear. They are phase A $\{h=\rho_x=\rho_y=0\}$, phase B $\{\rho_x=\rho_y=0,h<0\}$, phase C $\{\rho_x\neq0, h=\rho_y=0\}$, phase $D_1$ $\{\rho_x=\rho_y\neq0, h<0\}$ and phase $D_2$ $\{\rho_x=-\rho_y\neq0, h<0\}$, corresponding to normal phase, pure ferromagnetic phase, pure p-wave superconducting phase and two kinds of superconducting ferromagnetic phases, respectively.
In this case, whether the p-wave superconductivity can appear depends on the sign of interaction of magnetic moment of the complex vector field, i.e., the sign of $\Theta$. The possible phases and the physical favored phase in different temperature regions are summarized in table \[Tab1\]. In the case of $\Theta>0$, there is a critical temperature $T_{sc}$ between $T_{C0}$ and $T_{sc0}$, lower than which, the p-wave superconductivity can appear from ferromagnetic phase and the system will show ferromagnetism and superconductivity both. In addition, the critical temperature for superconductivity is increased rather than decreased by spontaneous magnetization. This promotion is enhanced by increasing of interaction strength between complex vector field and antisymmetric tensor field. Numerical results imply that the magnetism and superconductivity can coexist even in the zero temperature limit. However if $\Theta<0$, the p-wave superconducting state can not appear and the system will only be in a pure ferromagnetic state.
[|c|c|c|c|]{}\
temperature & $T>T_{C0}$& $T_{sc}<T<T_{C0}$&$T<T_{sc}$\
Possible& A & A,B & A, B, $D_1$, $D_2$, C(if $T<T_{sc0}$)\
Physical($\Theta>0$) & A& B& $D_1$\
Physical($\Theta<0$) & A&\
The other case is $T_{sc0}>T_{C0}$, i.e., the case where the p-wave superconducting phase appears first. When $T_{C0}<T<T_{sc0}$, the equations show there may exist three kinds of p-wave superconducting phases. One is the usual p-wave superconducting phase C($\{h=\rho_y=0, \rho_x\neq0\}$), the other two are new superconducting phases denoted as $E_1$ with $\{h<0,\rho_x=\rho_y\neq0\}$ and $E_2$ with $\{h<0,\rho_x=-\rho_y\neq0\}$. The magnetization in two phases $E_1$ and $E_2$ is induced by the p-wave pair rather than been produced spontaneously, which is different from case in phases $D_1$ and $D_2$.
[|c|c|c|c|]{}\
Temperature & $T>T_{sc0}$& $T_{C0}<T<T_{sc0}$&$T<T_{C0}$\
Possible& A & A, $E_1$, C & A, $E_1$, C, B\
Physical &A&\
\
\
[|c|c|c|c|]{}\
Temperature & $T>T_{sc0}$& $T_{C0}<T<T_{sc0}$&$T<T_{C0}$\
Possible& A & A, C & A, C, B\
Physical &A& C& B\
Numerical results show that situations also depend on the sign of $\Theta$. All the results are summarized in table \[Tab2\]. If $\Theta>0$, with decreasing the temperature, the system will transit into phase $E_1$, where p-wave superconductivity with a kind of induced magnetism appears. The superconductivity and magnetism appear both, however, it should better be called a magnetic superconducting phase rather than a ferromagnetic superconducting phase, because the magnetic moment is not spontaneously produced and proportional to $T_{sc0}-T$ rather than $\sqrt{T_{sc0}-T}$ near the critical temperature (see figure \[TG2b\]). If $\Theta<0$, the system will be in the pure p-wave superconducting phase without magnetism if temperature is less than $T_{sc0}$. When temperature is lower than $T_{C0}$, the system will transit into the pure ferromagnetic phase from the p-wave superconducting phase. Therefore the ferromagnetism and superconductivity can not coexist if $\Theta<0$.
![The behaviours of $N$ near the critical temperature in the phases $D_1$(left) and $E_1$(right). Here $m_1^2=-3/16, m_2^2=-3,J=-1, \Theta=1$ and $\alpha=0.1$. In the left one, $q=1.4$. In the right one, $q=1.4$. The plots were taken from ref. [@Cai:2014dza].[]{data-label="TG2b"}](TN1.pdf "fig:"){width="50.00000%"} ![The behaviours of $N$ near the critical temperature in the phases $D_1$(left) and $E_1$(right). Here $m_1^2=-3/16, m_2^2=-3,J=-1, \Theta=1$ and $\alpha=0.1$. In the left one, $q=1.4$. In the right one, $q=1.4$. The plots were taken from ref. [@Cai:2014dza].[]{data-label="TG2b"}](TNP4.pdf "fig:"){width="50.00000%"}
Tables \[Tab1\] and \[Tab2\] show that the ground state near zero temperature limit only depends on the sign of $\Theta$. These can be understood in a physical manner if we pay more attention to this phenomenological parameter in , where it was introduced to describe the self-interaction between the magnetic moments of complex vector field. The case of $\Theta>0$ means that the p-wave pair will attract the one with the same magnetic moment direction and repulse the one with the opposite magnetic moment. Under the influence of spontaneous magnetization, the magnetic moment of p-wave pair will tend to align along the direction of spontaneous magnetization. As a result, p-wave pair and spontaneous magnetization would be enhanced by each other and therefore survive. However, if $\Theta<0$, the p-wave pair will repulse the one with the same magnetic moment direction. So the p-wave pair will align without net magnetism and the system is in a pure p-wave superconducting phase in the region where superconductivity dominates. When $T<T_{C0}$, the ferromagnetism can appear, which tends to make p-wave pairs have same direction. But the p-wave pairs with same magnetic moment direction will repulse each other, which leads that the p-wave pair is not stable and will be de-paired. So the system can only be in the ferromagnetic phase.
Conclusion and Discussion {#sect:conclusion}
=========================
Due to the strong/weak duality characteristic of the holographic correspondence, it provides us with a powerful tool to study the properties of strongly interacting systems by a weakly coupled gravity theory with one extra spatial dimension. Although the underlying dynamics which govern the dual field theory and the gravity are apparently different, as we have shown, in the framework of holography quantum computations in the dual (strongly coupled) field theory can be translated into classical calculations in the bulk, where one can just solve differential equations with suitable boundary conditions. Within this framework, holographic correspondence is considered as a hopeful approach to understand the properties of strongly correlated electron systems.
The bulk gravitational models that we have reviewed are some phenomenological models. In such bottom-up approach, the gravity duals were constructed using the minimal set of fields that captured the essential dynamics. They just involve gravity interacting with an effective U(1) gauge field and a charged field serving as the order parameter. We have a lot of the degrees of freedom to choose the form of interactions as well as the value of couplings.[^32] Nevertheless, one has seen that those simple models would describe dual superconductors rather well. Some interesting features have been uncovered. Let us take holographic p-wave models as an example. For the SU(2) Yang-Mills model , the conductivities are strongly anisotropic in a manner which is suggestive of a gap with nodes. The low-lying excitations of the normal state have a relaxation time growing rapidly as the temperature is lowered, which agrees with the absence of impurity scattering. For the second model , it has been found that the vector condensate can be induced by an applied magnetic field, and the condensation of the charged vector operator forms a vortex lattice structure in the spatial directions perpendicular to the magnetic field. Going beyond the probe approximation, the model displays a rich phase structure. In terms of temperature and chemical potential, the complete phase diagrams have been constructed for the conducting phase, insulating phase and their corresponding superconducting phases and some new phase boundaries are revealed. The Maxwell-vector model is a generalization of the SU(2) model in the sense that the vector field has a general mass and gyromagnetic ratio. The third model realizes a p-wave superconducting phase by involving a charged two-form in the bulk. The p-wave states exhibit a helical structure and some of them display the phenomenon of pitch inversion as the temperature is decreased. The ground state of the condensed phase has zero entropy density and exhibits an emergent scaling symmetry in the IR.
It is clear that the key ingredient in constructing a gravitational dual of a superconductor is to find an instability which breaks a U(1) symmetry, e.g., at low temperatures and causes a condensate to form spontaneously. One may ask whether those phenomenological bulk duals of superconductors are just a Ginzburg-Landau description. The answer is exactly no. Let us stress two key differences. First, the instability in the Ginzburg-Landau model must be put in by hand, while it arises naturally in holographic setup. Second, the Ginzburg-Landau model is only valid near the transition point, whereas the gravitational description can characterize the whole dynamics. For a given bulk action, scanning through values of model parameters corresponds to scanning through many different dual field theories. In that sense, a simple holographic model has a kind of universality, i.e., the results may be true for a large class of dual field theories, quite insensitive to the details of their dynamics. Another confusion is that we realized the spontaneous breaking of a continuous U(1) symmetry in $(2+1)$ dimensions at finite temperature, in apparent contradiction to the Coleman-Mermin-Wagner theorem. The cure is that the large $N$ limit evades the theorem as fluctuations are suppressed. It would be interesting to discuss the effect of bulk quantum corrections which correspond to $1/N$ corrections in the dual field theory [@Anninos:2010sq]. Finally, although the hair breaks a local U(1) symmetry in the bulk, according to the dictionary, the dual system consists of a condensate breaking a global U(1) symmetry. On the other hand, the onset of superconductivity is characterized by the condensation of a composite charged operator spontaneously breaking U(1) gauge symmetry. So strictly speaking, what one has realized is a dual theory of superfluid [@Herzog:2008he; @Brihaye:2011vk; @Arean:2010wu; @Wu:2014bba] rather than superconductor. However, in the limit that the U(1) symmetry is “weakly gauged" one can still view the dual theory describing a superconductor. [^33]
Throughout this brief summary we have been mainly concerned with static and homogeneous case and focused on some basic aspects. This is a rapidly devolving field, due to the limitation of length, we are not able to give more details for many interesting developments, such as introduction of momentum dissipation (to break translational symmetry) [@Flauger:2010tv; @Horowitz:2013jaa; @Dias:2013bwa; @Zeng:2014uoa; @Koga:2014hwa; @Ling:2014laa; @Arean:2013mta; @Erdmenger:2015qqa; @Kim:2015dna], construction of holographic Josephson Junction [@Horowitz:2011dz; @Wang:2011rva; @Siani:2011uj; @Wang:2011ri; @Wang:2012yj; @Rozali:2013pla; @Cai:2013sua; @Takeuchi:2013kra; @Kiritsis:2011zq; @Li:2014xia], and investigation on dynamics for far-from equilibrium state [@Murata:2010dx; @Bhaseen:2012gg; @Sonner:2014tca; @Bai:2014tla; @Adams:2012pj; @Gao:2012aw; @Li:2013fhw; @Garcia-Garcia:2013rha; @Chesler:2014gya; @Du:2014lwa]. The analysis for the most part has been done numerically. To explore the properties of holographic superconductors using analytical techniques can be found, for example, in refs. [@Herzog:2010vz; @Siopsis:2010uq; @Zeng:2010zn; @Cai:2011ky; @Momeni:2011iw; @Zeng:2012zza; @Pan:2012jf; @Gangopadhyay:2012gx; @Huang:2013sca; @Banerjee:2012vk; @Momeni:2013bca; @Lu:2013tza]. Optimistically the growing literature based on holographic duality might shed some light on the understanding of mysterious phenomena and eventually microscopic origins of strongly correlated superconductivity.
The applications of the holographic correspondence are still going on. It was written by G. T. Horowitz and J. Polchinski [@Horowitz:2006ct] that we find it difficult to believe that nature does not make use of it, but the precise way in which it does so remains to be discovered. In addition to holographic superconductors, the holographic approach has been used to understand some other aspects of condensed matter physics, including (non-)Fermi liquids [@Lee:2008xf; @Liu:2009dm; @Cubrovic:2009ye; @Faulkner:2009wj; @Faulkner:2010tq], quantum Hall effect [@Davis:2008nv; @Fujita:2009kw; @Bergman:2010gm], strange metals [@Hartnoll:2009ns; @Faulkner:2010da; @Kim:2010zq; @Davison:2013txa], topological insulators [@HoyosBadajoz:2010ac; @Ryu:2010fe; @Karch:2010mn], Hubbard model [@Fujita:2014mqa] and so on. A major application using holographic duality is to describe quantum chromodynamics (QCD), especially for the quark gluon plasma produced in particle accelerators. It is referred as AdS/QCD or holographic QCD, which has been widely studied [@Polchinski:2001tt; @BoschiFilho:2002ta; @Erlich:2005qh; @deTeramond:2005su; @Babington:2003vm; @Kruczenski:2003uq; @Gubser:2006bz; @Shuryak:2005ia; @Gursoy:2008bu; @Herzog:2006ra; @Colangelo:2010pe; @Chen:2009kx; @Huang:2007fv; @vanderSchee:2014qwa]. Another emerging subject is the fluid/gravity correspondence, which translates problems in fluid dynamics into problems in general relativity [@Bhattacharyya:2008jc; @Rangamani:2009xk; @Bredberg:2011jq]. Readers who are interested in those exciting achievements are encouraged to consult those relevant references.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by the National Natural Science Foundation of China ( No.11035008, No.11375247, No.11205226 and No.11435006 ). L Li was supported in part by European Union’s Seventh Framework Programme under grant agreements (FP7-REGPOT-2012-2013-1) no 316165, the EU-Greece program “Thales" MIS 375734 and was also co-financed by the European Union (European Social Fund, ESF) and Greek national funds through the Operational Program “Education and Lifelong Learning" of the National Strategic Reference Framework (NSRF) under “Funding of proposals that have received a positive evaluation in the 3rd and 4th Call of ERC Grant Schemes".
[99]{}
J. Bardeen, L. N. Cooper and J. R. Schrieffer, “Theory Of Superconductivity,” Phys. Rev. [**108**]{}, 1175 (1957). E. W. Carlson, V. J. Emery, S. A. Kivelson and D. Orgad, “Concepts in High Temperature Superconductivity,” \[arXiv:cond-mat/0206217\].
J. M. Maldacena, “The large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys. [**2**]{}, 231 (1998) \[Int. J. Theor. Phys. [**38**]{}, 1113 (1999)\] \[arXiv:hep-th/9711200\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, “Gauge theory correlators from non-critical string theory,” Phys. Lett. B [**428**]{}, 105 (1998) \[arXiv:hep-th/9802109\]. E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. [**2**]{}, 253 (1998) \[arXiv:hep-th/9802150\].
X. L. Qi, “Exact holographic mapping and emergent space-time geometry,” arXiv:1309.6282 \[hep-th\].
G. T. Horowitz, “Introduction to Holographic Superconductors,” Lect. Notes Phys. [**828**]{}, 313 (2011) \[arXiv:1002.1722 \[hep-th\]\].
C. P. Herzog, “Lectures on Holographic Superfluidity and Superconductivity,” J. Phys. A [**42**]{}, 343001 (2009) \[arXiv:0904.1975 \[hep-th\]\].
N. Iqbal, H. Liu and M. Mezei, “Lectures on holographic non-Fermi liquids and quantum phase transitions,” arXiv:1110.3814 \[hep-th\]. D. Musso, “Introductory notes on holographic superconductors,” arXiv:1401.1504 \[hep-th\].
S. A. Hartnoll, C. P. Herzog and G. T. Horowitz, “Building a Holographic Superconductor,” Phys. Rev. Lett. [**101**]{}, 031601 (2008) \[arXiv:0803.3295 \[hep-th\]\].
S. A. Hartnoll, C. P. Herzog and G. T. Horowitz, “Holographic Superconductors,” JHEP [**0812**]{}, 015 (2008) \[arXiv:0810.1563 \[hep-th\]\]. T. Nishioka, S. Ryu and T. Takayanagi, “Holographic Superconductor/Insulator Transition at Zero Temperature,” JHEP [**1003**]{}, 131 (2010) \[arXiv:0911.0962 \[hep-th\]\]. J. -W. Chen, Y. -J. Kao, D. Maity, W. -Y. Wen and C. -P. Yeh, “Towards A Holographic Model of D-Wave Superconductors,” Phys. Rev. D [**81**]{}, 106008 (2010) \[arXiv:1003.2991 \[hep-th\]\]. F. Benini, C. P. Herzog, R. Rahman and A. Yarom, “Gauge gravity duality for d-wave superconductors: prospects and challenges,” JHEP [**1011**]{}, 137 (2010) \[arXiv:1007.1981 \[hep-th\]\].
A. P. Mackenzie and Y. Maeno, “p-wave superconductivity," Physica. B [**280**]{}, 148-153 (2000).
S. S. Gubser and S. S. Pufu, “The Gravity dual of a p-wave superconductor,” JHEP [**0811**]{}, 033 (2008) \[arXiv:0805.2960 \[hep-th\]\].
R. G. Cai, S. He, L. Li and L. F. Li, “A Holographic Study on Vector Condensate Induced by a Magnetic Field,” JHEP [**1312**]{}, 036 (2013) \[arXiv:1309.2098 \[hep-th\]\]. R. -G. Cai, L. Li and L. -F. Li, “A Holographic P-wave Superconductor Model,” JHEP [**1401**]{}, 032 (2014) \[arXiv:1309.4877 \[hep-th\]\]. F. Aprile, D. Rodriguez-Gomez and J. G. Russo, “p-wave Holographic Superconductors and five-dimensional gauged Supergravity,” JHEP [**1101**]{}, 056 (2011) \[arXiv:1011.2172 \[hep-th\]\].
A. Donos and J. P. Gauntlett, “Holographic helical superconductors,” JHEP [**1112**]{}, 091 (2011) \[arXiv:1109.3866 \[hep-th\]\].
A. Donos and J. P. Gauntlett, “Helical superconducting black holes,” Phys. Rev. Lett. [**108**]{}, 211601 (2012) \[arXiv:1203.0533 \[hep-th\]\].
F. Denef and S. A. Hartnoll, “Landscape of superconducting membranes,” Phys. Rev. D [**79**]{}, 126008 (2009) \[arXiv:0901.1160 \[hep-th\]\].
S. S. Gubser, C. P. Herzog, S. S. Pufu and T. Tesileanu, “Superconductors from Superstrings,” Phys. Rev. Lett. [**103**]{}, 141601 (2009) \[arXiv:0907.3510 \[hep-th\]\]. J. P. Gauntlett, J. Sonner and T. Wiseman, “Holographic superconductivity in M-Theory,” Phys. Rev. Lett. [**103**]{}, 151601 (2009) \[arXiv:0907.3796 \[hep-th\]\]. S. S. Gubser, S. S. Pufu and F. D. Rocha, “Quantum critical superconductors in string theory and M-theory,” Phys. Lett. B [**683**]{}, 201 (2010) \[arXiv:0908.0011 \[hep-th\]\]. J. P. Gauntlett, J. Sonner and T. Wiseman, “Quantum Criticality and Holographic Superconductors in M-theory,” JHEP [**1002**]{}, 060 (2010) \[arXiv:0912.0512 \[hep-th\]\]. S. Kalyana Rama, S. Sarkar, B. Sathiapalan and N. Sircar, “Strong Coupling BCS Superconductivity and Holography,” Nucl. Phys. B [**852**]{}, 634 (2011) \[arXiv:1104.2843 \[hep-th\]\].
N. Bobev, A. Kundu, K. Pilch and N. P. Warner, “Minimal Holographic Superconductors from Maximal Supergravity,” JHEP [**1203**]{}, 064 (2012) \[arXiv:1110.3454 \[hep-th\]\].
V. L. Ginzburg and L. D. Landau, “On the Theory of superconductivity,” Zh. Eksp. Teor. Fiz. [**20**]{}, 1064 (1950).
M. Cyrot, “Ginzburg-Landau theory for superconductors”, Rep. Prog. Phys. [**36**]{}, 103 (1973).
S. Weinberg, “Superconductivity for Particular Theorists,” Progress of Theoretical Physics Supplement, 86, 43 (1986).
S. Kachru, X. Liu and M. Mulligan, “Gravity duals of Lifshitz-like fixed points,” Phys. Rev. D [**78**]{}, 106005 (2008) \[arXiv:0808.1725 \[hep-th\]\]. D. T. Son, “Toward an AdS/cold atoms correspondence: A Geometric realization of the Schrodinger symmetry,” Phys. Rev. D [**78**]{}, 046003 (2008) \[arXiv:0804.3972 \[hep-th\]\]. K. Balasubramanian and J. McGreevy, “Gravity duals for non-relativistic CFTs,” Phys. Rev. Lett. [**101**]{}, 061601 (2008) \[arXiv:0804.4053 \[hep-th\]\].
C. Charmousis, B. Gouteraux, B. S. Kim, E. Kiritsis and R. Meyer, “Effective Holographic Theories for low-temperature condensed matter systems,” JHEP [**1011**]{}, 151 (2010) \[arXiv:1005.4690 \[hep-th\]\]. B. Gouteraux and E. Kiritsis, “Generalized Holographic Quantum Criticality at Finite Density,” JHEP [**1112**]{}, 036 (2011) \[arXiv:1107.2116 \[hep-th\]\]. L. Huijse, S. Sachdev and B. Swingle, “Hidden Fermi surfaces in compressible states of gauge-gravity duality,” Phys. Rev. B [**85**]{}, 035121 (2012) \[arXiv:1112.0573 \[cond-mat.str-el\]\]. X. Dong, S. Harrison, S. Kachru, G. Torroba and H. Wang, “Aspects of holography for theories with hyperscaling violation,” JHEP [**1206**]{}, 041 (2012) \[arXiv:1201.1905 \[hep-th\]\]. B. S. Kim, “Schródinger Holography with and without Hyperscaling Violation,” JHEP [**1206**]{}, 116 (2012) \[arXiv:1202.6062 \[hep-th\]\].
P. Breitenlohner and D. Z. Freedman, “Stability in Gauged Extended Supergravity,” Annals Phys. [**144**]{}, 249 (1982). A. W. Peet and J. Polchinski, “UV / IR relations in AdS dynamics,” Phys. Rev. D [**59**]{}, 065011 (1999) \[hep-th/9809022\].
V. Balasubramanian and P. Kraus, “A Stress tensor for Anti-de Sitter gravity,” Commun. Math. Phys. [**208**]{}, 413 (1999) \[hep-th/9902121\]. M. Bianchi, D. Z. Freedman and K. Skenderis, “How to go with an RG flow,” JHEP [**0108**]{}, 041 (2001) \[hep-th/0105276\].
M. Bianchi, D. Z. Freedman and K. Skenderis, “Holographic renormalization,” Nucl. Phys. B [**631**]{}, 159 (2002) \[hep-th/0112119\].
D. T. Son and A. O. Starinets, “Minkowski space correlators in AdS/CFT correspondence: Recipe and applications,” JHEP [**0209**]{}, 042 (2002) \[hep-th/0205051\].
C. P. Herzog and D. T. Son, “Schwinger-Keldysh propagators from AdS/CFT correspondence,” JHEP [**0303**]{}, 046 (2003) \[hep-th/0212072\]. K. Skenderis and B. C. van Rees, “Real-time gauge/gravity duality,” Phys. Rev. Lett. [**101**]{}, 081601 (2008) \[arXiv:0805.0150 \[hep-th\]\].
I. R. Klebanov and E. Witten, “AdS / CFT correspondence and symmetry breaking,” Nucl. Phys. B [**556**]{}, 89 (1999) \[hep-th/9905104\].
E. Witten, “Multitrace operators, boundary conditions, and AdS / CFT correspondence,” hep-th/0112258. M. Berkooz, A. Sever and A. Shomer, “’Double trace’ deformations, boundary conditions and space-time singularities,” JHEP [**0205**]{}, 034 (2002) \[hep-th/0112264\]. S. W. Hawking, “Breakdown of Predictability in Gravitational Collapse," Phys. Rev. D [**14**]{}, 2460 (1976).
S. W. Hawking, “Information loss in black holes,” Phys. Rev. D [**72**]{}, 084013 (2005) \[hep-th/0507171\]. S. W. Hawking, “Information Preservation and Weather Forecasting for Black Holes,” arXiv:1401.5761 \[hep-th\].
O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri and Y. Oz, “Large N field theories, string theory and gravity,” Phys. Rept. [**323**]{}, 183 (2000) \[hep-th/9905111\]. S. A. Hartnoll, “Lectures on holographic methods for condensed matter physics,” Class. Quant. Grav. [**26**]{}, 224002 (2009) \[arXiv:0903.3246 \[hep-th\]\].
J. McGreevy, “Holographic duality with a view toward many-body physics,” Adv. High Energy Phys. [**2010**]{}, 723105 (2010) \[arXiv:0909.0518 \[hep-th\]\].
J. Casalderrey-Solana, H. Liu, D. Mateos, K. Rajagopal and U. A. Wiedemann, “Gauge/String Duality, Hot QCD and Heavy Ion Collisions,” arXiv:1101.0618 \[hep-th\].
A. Adams, L. D. Carr, T. Schäfer, P. Steinberg and J. E. Thomas, “Strongly Correlated Quantum Fluids: Ultracold Quantum Gases, Quantum Chromodynamic Plasmas, and Holographic Duality,” New J. Phys. [**14**]{}, 115009 (2012) \[arXiv:1205.5180 \[hep-th\]\].
S. Sachdev, “What can gauge-gravity duality teach us about condensed matter physics?,” Ann. Rev. Condensed Matter Phys. [**3**]{}, 9 (2012) \[arXiv:1108.1197 \[cond-mat.str-el\]\].
S. Franco, A. Garcia-Garcia and D. Rodriguez-Gomez, “A General class of holographic superconductors,” JHEP [**1004**]{}, 092 (2010) \[arXiv:0906.1214 \[hep-th\]\].
F. Aprile and J. G. Russo, “Models of Holographic superconductivity,” Phys. Rev. D [**81**]{}, 026009 (2010) \[arXiv:0912.0480 \[hep-th\]\]. F. Aprile, S. Franco, D. Rodriguez-Gomez and J. G. Russo, “Phenomenological Models of Holographic Superconductors and Hall currents,” JHEP [**1005**]{}, 102 (2010) \[arXiv:1003.4487 \[hep-th\]\]. Y. Liu and Y. W. Sun, “Holographic Superconductors from Einstein-Maxwell-Dilaton Gravity,” JHEP [**1007**]{}, 099 (2010) \[arXiv:1006.2726 \[hep-th\]\].
S. Chen, Q. Pan and J. Jing, “Holographic superconductor models in the non-minimal derivative coupling theory,” Chin. Phys. B [**21**]{}, 040403 (2012) \[arXiv:1012.3820 \[gr-qc\]\].
Y. Peng, Q. Pan and B. Wang, “Various types of phase transitions in the AdS soliton background,” Phys. Lett. B [**699**]{}, 383 (2011) \[arXiv:1104.2478 \[hep-th\]\]. F. Bigazzi, A. L. Cotrone, D. Musso, N. P. Fokeeva and D. Seminara, “Unbalanced Holographic Superconductors and Spintronics,” JHEP [**1202**]{}, 078 (2012) \[arXiv:1111.6601 \[hep-th\]\].
A. Dey, S. Mahapatra and T. Sarkar, “Generalized Holographic Superconductors with Higher Derivative Couplings,” JHEP [**1406**]{}, 147 (2014) \[arXiv:1404.2190 \[hep-th\]\]. D. Arean and J. Tarrio, “Bifundamental Superfluids from Holography,” arXiv:1501.02804 \[hep-th\].
S. S. Gubser, “Breaking an Abelian gauge symmetry near a black hole horizon,” Phys. Rev. D [**78**]{}, 065034 (2008) \[arXiv:0801.2977 \[hep-th\]\]. T. Faulkner, G. T. Horowitz and M. M. Roberts, “Holographic quantum criticality from multi-trace deformations,” JHEP [**1104**]{}, 051 (2011) \[arXiv:1008.1581 \[hep-th\]\].
G. T. Horowitz and M. M. Roberts, “Holographic Superconductors with Various Condensates,” Phys. Rev. D [**78**]{}, 126008 (2008) \[arXiv:0810.1077 \[hep-th\]\].
G. T. Horowitz and M. M. Roberts, “Zero Temperature Limit of Holographic Superconductors,” JHEP [**0911**]{}, 015 (2009) \[arXiv:0908.3677 \[hep-th\]\]. K. Maeda and T. Okamura, “Characteristic length of an AdS/CFT superconductor,” Phys. Rev. D [**78**]{}, 106006 (2008) \[arXiv:0809.3079 \[hep-th\]\].
K. Maeda, M. Natsuume and T. Okamura, “Vortex lattice for a holographic superconductor,” Phys. Rev. D [**81**]{}, 026002 (2010) \[arXiv:0910.4475 \[hep-th\]\].
T. Faulkner, G. T. Horowitz, J. McGreevy, M. M. Roberts and D. Vegh, “Photoemission ‘experiments’ on holographic superconductors,” JHEP [**1003**]{}, 121 (2010) \[arXiv:0911.3402 \[hep-th\]\]. A. Bagrov, B. Meszena and K. Schalm, “Pairing induced superconductivity in holography,” JHEP [**1409**]{}, 106 (2014) \[arXiv:1403.3699 \[hep-th\]\].
G. T. Horowitz and B. Way, “Complete Phase Diagrams for a Holographic Superconductor/Insulator System,” JHEP [**1011**]{}, 011 (2010) \[arXiv:1007.3714 \[hep-th\]\].
R. G. Cai, S. He, L. Li and Y. L. Zhang, “Holographic Entanglement Entropy in Insulator/Superconductor Transition,” JHEP [**1207**]{}, 088 (2012) \[arXiv:1203.6620 \[hep-th\]\]. R. G. Cai, S. He, L. Li and L. F. Li, “Entanglement Entropy and Wilson Loop in Stúckelberg Holographic Insulator/Superconductor Model,” JHEP [**1210**]{}, 107 (2012) \[arXiv:1209.1019 \[hep-th\]\]. W. Yao and J. Jing, “Holographic entanglement entropy in insulator/superconductor transition with Born-Infeld electrodynamics,” JHEP [**1405**]{}, 058 (2014) \[arXiv:1401.6505 \[hep-th\]\].
P. Basu, J. He, A. Mukherjee and H. H. Shieh, “Hard-gapped Holographic Superconductors,” Phys. Lett. B [**689**]{}, 45 (2010) \[arXiv:0911.4999 \[hep-th\]\].
M. M. Roberts and S. A. Hartnoll, “Pseudogap and time reversal breaking in a holographic superconductor,” JHEP [**0808**]{}, 035 (2008) \[arXiv:0805.3898 \[hep-th\]\]. A. Akhavan and M. Alishahiha, “P-Wave Holographic Insulator/Superconductor Phase Transition,” Phys. Rev. D [**83**]{}, 086003 (2011) \[arXiv:1011.6158 \[hep-th\]\]. M. Ammon, J. Erdmenger, P. Kerner and M. Strydom, “Black Hole Instability Induced by a Magnetic Field,” Phys. Lett. B [**706**]{}, 94 (2011) \[arXiv:1106.4551 \[hep-th\]\]. Y. -Y. Bu, J. Erdmenger, J. P. Shock and M. Strydom, “Magnetic field induced lattice ground states from holography,” JHEP [**1303**]{}, 165 (2013) \[arXiv:1210.6669 \[hep-th\]\]. M. Ammon, J. Erdmenger, M. Kaminski and P. Kerner, “Superconductivity from gauge/gravity duality with flavor,” Phys. Lett. B [**680**]{}, 516 (2009) \[arXiv:0810.2316 \[hep-th\]\]. P. Basu, J. He, A. Mukherjee and H. -H. Shieh, “Superconductivity from D3/D7: Holographic Pion Superfluid,” JHEP [**0911**]{}, 070 (2009) \[arXiv:0810.3970 \[hep-th\]\]. M. Ammon, J. Erdmenger, M. Kaminski and P. Kerner, “Flavor Superconductivity from Gauge/Gravity Duality,” JHEP [**0910**]{}, 067 (2009) \[arXiv:0903.1864 \[hep-th\]\]. M. Ammon, J. Erdmenger, V. Grass, P. Kerner and A. O’Bannon, “On Holographic p-wave Superfluids with Back-reaction,” Phys. Lett. B [**686**]{}, 192 (2010) \[arXiv:0912.3515 \[hep-th\]\]. R. -G. Cai, Z. -Y. Nie and H. -Q. Zhang, “Holographic Phase Transitions of P-wave Superconductors in Gauss-Bonnet Gravity with Back-reaction,” Phys. Rev. D [**83**]{}, 066013 (2011) \[arXiv:1012.5559 \[hep-th\]\]. R. -G. Cai, Z. -Y. Nie and H. -Q. Zhang, “Holographic p-wave superconductors from Gauss-Bonnet gravity,” Phys. Rev. D [**82**]{}, 066007 (2010) \[arXiv:1007.3321 \[hep-th\]\]. L. A. Pando Zayas and D. Reichmann, “A Holographic Chiral $p_x + ip_y$ Superconductor,” Phys. Rev. D [**85**]{}, 106012 (2012) \[arXiv:1108.4022 \[hep-th\]\].
R. -G. Cai, S. He, L. Li and Y. -L. Zhang, “Holographic Entanglement Entropy on P-wave Superconductor Phase Transition,” JHEP [**1207**]{}, 027 (2012) \[arXiv:1204.5962 \[hep-th\]\]. R. E. Arias and I. S. Landea, “Backreacting p-wave Superconductors,” JHEP [**1301**]{}, 157 (2013) \[arXiv:1210.6823 \[hep-th\]\]. R. -G. Cai, L. Li, L. -F. Li and R. -K. Su, “Entanglement Entropy in Holographic P-Wave Superconductor/Insulator Model,” JHEP [**1306**]{}, 063 (2013) \[arXiv:1303.4828 \[hep-th\]\].
D. Djukanovic, M. R. Schindler, J. Gegelia and S. Scherer, “Quantum electrodynamics for vector mesons,” Phys. Rev. Lett. [**95**]{}, 012001 (2005) \[hep-ph/0505180\]. J. A. Young and S. A. Bludman, “Electromagnetic Properties of a Charged Vector Meson,” Phys. Rev. [**131**]{}, 2326 (1963).
Rasolt. Mark, “Superconductivity in high magnetic fields,” Phys. Rev. Lett. [**58**]{}, 1482 (1987).
M. Rasolt and Z. Tesanovic, “Theoretical aspects of superconductivity in very high magnetic fields,” Rev. Mod. Phys. [**64**]{}, 709 (1992).
F. Levy, I. Sheikin, B. Grenier, A. Huxley, “Magnetic Field-induced Superconductivity in the Ferromagnet URhGe,” Science **309** 1343 (2005).
S. Uji, H. Shinagawa, T. Terashima, T. Yakabe, Y. Teral, M. Tokumoto, A. Kobayashi, H. Tanaka, H. Kobayashi, “Magnetic-field-induced superconductivity in a two-dimensional organic conductor,” Nature **410** 908 (2010).
R. G. Cai, L. Li, L. F. Li and Y. Wu, “Vector Condensate and AdS Soliton Instability Induced by a Magnetic Field,” JHEP [**1401**]{}, 045 (2014) \[arXiv:1311.7578 \[hep-th\]\].
M. N. Chernodub, “Superconductivity of QCD vacuum in strong magnetic field,” Phys. Rev. D [**82**]{}, 085011 (2010) \[arXiv:1008.1055 \[hep-ph\]\]. M. N. Chernodub, “Spontaneous electromagnetic superconductivity of vacuum in strong magnetic field: evidence from the Nambu–Jona-Lasinio model,” Phys. Rev. Lett. [**106**]{}, 142003 (2011) \[arXiv:1101.0117 \[hep-ph\]\]. R. -G. Cai, L. Li, L. -F. Li and R. -Q. Yang, “Towards Complete Phase Diagrams of a Holographic P-wave Superconductor Model,” JHEP [**1404**]{}, 016 (2014) \[arXiv:1401.3974 \[gr-qc\]\].
V. P. Maslov, “Zeroth-Order Phase Transitions," Mathematical Notes [**76**]{}, 697 (2004).
H. B. Zeng, “Possible Anderson localization in a holographic superconductor,” Phys. Rev. D [**88**]{}, 126004 (2013) \[arXiv:1310.5753 \[hep-th\]\]. H. B. Zeng and H. Q. Zhang, “Zeroth Order Phase Transition in a Holographic Superconductor with Single Impurity,” arXiv:1411.3955 \[hep-th\]. P. Chaturvedi and G. Sengupta, “p-wave Holographic Superconductors from Born-Infeld Black Holes,” arXiv:1501.06998 \[hep-th\]. A. A. Kordyuk, “Iron-based superconductors: Magnetism, superconductivity, and electronic structure," Low. Temp. Phys. [**38**]{}, 888 (2012).
A. Chubukov, “Pairing Mechanism in Fe-Based Superconductors," Annu. Rev. Condens. Matter Phys. [**3**]{}, 357 (2012).
H. Q. Yuan [*et al.*]{}, “Observation of Two Distinct Superconducting Phases in CeCu$_2$Si$_2$," Science [**302**]{}, 2104 (2003).
P. Fulde and R. A. Ferrell, “Superconductivity in a Strong Spin-Exchange Field,” Phys. Rev. [**135**]{}, A550 (1964). A. I. larkin and Y. N. Ovchinnikov, “Nonuniform state of superconductors,” Zh. Eksp. Teor. Fiz. [**47**]{}, 1136 (1964) \[Sov. Phys. JETP [**20**]{}, 762 (1965)\].
A. Donos, J. P. Gauntlett and C. Pantelidou, “Competing p-wave orders,” Class. Quant. Grav. [**31**]{}, 055007 (2014) \[arXiv:1310.5741 \[hep-th\]\]. P. J. Hirschfeld, W. O. Putikka and D. J. Scalapino, “Microwave conductivity of d-wave superconductors,” Phys. Rev. Lett. [**71**]{}, 3705 (1993).
F. Benini, C. P. Herzog and A. Yarom, “Holographic Fermi arcs and a d-wave gap,” Phys. Lett. B [**701**]{}, 626 (2011) \[arXiv:1006.0731 \[hep-th\]\]. J. W. Chen, Y. S. Liu and D. Maity, “$d+id$ Holographic Superconductors,” JHEP [**1105**]{}, 032 (2011) \[arXiv:1103.1714 \[hep-th\]\]. K. -Y. Kim and M. Taylor, “Holographic d-wave superconductors,” JHEP [**1308**]{}, 112 (2013) \[arXiv:1304.6729 \[hep-th\]\]. M. R. Norman “The Challenge of Unconventional Superconductivity,” Science [**332**]{}, 196 (2011).
E. Berg, E. Fradkin, S. A. Kivelson and J. M. Tranquada, “Striped superconductors: how spin, charge and superconducting orders intertwine in the cuprates," New J. Phys. [**11**]{}, 115004 (2009).
J. Zaanen, “A Modern, but way too short history of the theory of superconductivity at a high temperature," \[arXiv:1012.5461 \[cond-mat.supr-con\]\].
R. G. Cai, L. Li, L. F. Li and Y. Q. Wang, “Competition and Coexistence of Order Parameters in Holographic Multi-Band Superconductors,” JHEP [**1309**]{}, 074 (2013) \[arXiv:1307.2768 \[hep-th\]\]. Z. Y. Nie, R. G. Cai, X. Gao and H. Zeng, “Competition between the s-wave and p-wave superconductivity phases in a holographic model,” JHEP [**1311**]{}, 087 (2013) \[arXiv:1309.2204 \[hep-th\]\]. I. Amado, D. Arean, A. Jimenez-Alba, L. Melgar and I. Salazar Landea, “Holographic s+p Superconductors,” Phys. Rev. D [**89**]{}, 026009 (2014) \[arXiv:1309.5086 \[hep-th\]\]. L. F. Li, R. G. Cai, L. Li and Y. Q. Wang, “Competition between s-wave order and d-wave order in holographic superconductors,” JHEP [**1408**]{}, 164 (2014) \[arXiv:1405.0382 \[hep-th\]\]. M. Nishida, “Phase Diagram of a Holographic Superconductor Model with s-wave and d-wave,” JHEP [**1409**]{}, 154 (2014) \[arXiv:1403.6070 \[hep-th\]\]. P. Basu, J. He, A. Mukherjee, M. Rozali and H. -H. Shieh, “Competing Holographic Orders,” JHEP [**1010**]{}, 092 (2010) \[arXiv:1007.3480 \[hep-th\]\]. D. Musso, “Competition/Enhancement of Two Probe Order Parameters in the Unbalanced Holographic Superconductor,” JHEP [**1306**]{}, 083 (2013) \[arXiv:1302.7205 \[hep-th\]\].
Y. Liu, K. Schalm, Y. W. Sun and J. Zaanen, “Bose-Fermi competition in holographic metals,” JHEP [**1310**]{}, 064 (2013) \[arXiv:1307.4572 \[hep-th\]\]. A. Amoretti, A. Braggio, N. Maggiore, N. Magnoli and D. Musso, “Coexistence of two vector order parameters: a holographic model for ferromagnetic superconductivity,” JHEP [**1401**]{}, 054 (2014) \[arXiv:1309.5093 \[hep-th\]\]. W. Y. Wen, M. S. Wu and S. Y. Wu, “A Holographic Model of Two-Band Superconductor,” Phys. Rev. D [**89**]{}, 066005 (2014) \[arXiv:1309.0488 \[hep-th\]\]. A. Donos, J. P. Gauntlett, J. Sonner and B. Withers, “Competing orders in M-theory: superfluids, stripes and metamagnetism,” JHEP [**1303**]{}, 108 (2013) \[arXiv:1212.0871 \[hep-th\]\].
M. Silaev and E. Babaev, “Microscopic derivation of two-component Ginzburg-Landau model and conditions of its applicability in two-band systems," Phys. Rev. B [**85**]{}, 134514 (2012) \[arXiv:1110.1593 \[cond-mat\]\].
A. A. Shanenko, M. V. Milošević, F. M. Peeters and A. V. Vagov, “Extended Ginzburg-Landau Formalism for Two-Band Superconductors,” Phys. Rev. Lett. [**106**]{}, 047005 (2011) \[arXiv:1101.0971 \[cond-mat.supr-con\]\].
A. Vagov, A. A. Shanenko, M. V. Milošević, V. M. Axt, and F. M. Peeters, “Two-band superconductors: Extended Ginzburg-Landau formalism by a systematic expansion in small deviation from the critical temperature,” Phys. Rev. B [**86**]{}, 144514 (2012) \[arXiv:1207.6297 \[cond-mat.supr-con\]\].
J. Carlstrom, E. Babaev and M. Speight, “Type-1.5 superconductivity in multiband systems: Effects of interband couplings,” Phys. Rev. B [**83**]{}, 174509 (2011) \[arXiv:1009.2196 \[cond-mat.supr-con\]\].
C. Buzea and T. Yamashita, “Review of superconducting properties of MgB$_{2}$,” Superconductors, Science & Technology, Vol. 14, No. 11 (2001) R115-R146 \[arXiv:cond-mat/0108265 \[cond-mat.supr-con\]\].
P. J. Hirschfeld, M. M. Korshunov and I. I. Mazin, “Gap symmetry and structure of Fe-based superconductors,” Rep. Prog. Phys. [**74**]{}, 124508 (2011) \[arXiv:1106.3712 \[cond-mat.supr-con\]\].
David C. Johnston, “The Puzzle of High Temperature Superconductivity in Layered Iron Pnictides and Chalcogenides,” Advances in Physics [**59**]{}, 803-1061 (2010) \[arXiv:1005.4392 \[cond-mat.supr-con\]\].
G. R. Stewart, “Superconductivity in Iron Compounds,” Rev. Mod. Phys. [**83**]{} 1589-1652 (2011) \[arXiv:1106.1618 \[cond-mat.supr-con\]\].
Z. Y. Nie, R. G. Cai, X. Gao, L. Li and H. Zeng, “Phase transitions in a holographic s+p model with backreaction,” arXiv:1501.00004 \[hep-th\]. P. Goswami, B. Roy, “Axionic superconductivity in three dimensional doped narrow gap semiconductors,” Phys. Rev. B [**90**]{}, 041301(R)(2014) \[arXiv:1307.3240 \[cond-mat.supr-con\]\].
G. G. Lonzarich, S. S. Saxena, P. Agarwal, K. Ahilan, F. M. Grosche, R. K. W. Haselwimmer, M. J. Steiner, E. Pugh, et al., “Superconductivity on the border of itinerant-electron ferromagnetism in UGe2," Nature [**406**]{} 587 (2000).
D. Aoki, A. Huxley, E. Ressouche, D. Braithwaite, J. Flouquet, J-P. Brison, E. Lhotel and C. Paulsen, “Coexistence of superconductivity and ferromagnetism in URhGe," Nature [**413**]{} 613 (2001).
N. T. Huy, A. Gasparini, et al., “Superconductivity on the border of weak itinerant ferromagnetism in UCoGe," Phys. Rev. Lett. [**99**]{}, 067006 (2007).
C. Pfleiderer, M. Uhlarz, S. M. Hayden, R. Vollmer, H. Lohneysen, N. Bernhoeft and G. Lonzarich, “ Coexistence of superconductivity and ferromagnetism in the d-band metal ZrZn$_2$," Nature [**412**]{}, 58 (2001).
D. I. Uzunov, “Theory of ferromagnetic unconventional superconductors with spin-triplet electron pairing,” \[arXiv:1204.1007v2\[cond-mat\]\].
K. Machida, T. Ohmi, “Theory of Ferromagnetic Superconductivity," Phys. Rev. Lett [**86**]{}, 850 (2001) \[arXiv:cond-mat/0008245\].
Nevidomskyy H. Andriy, “Coexistence of Ferromagnetism and Superconductivity Close to a Quantum Phase Transition: The Heisenberg- to Ising-type Crossover," Phys. Rev. Lett. [**94**]{},097003 (2005) \[arXiv:cond-mat/0412247\].
N. Iqbal, H. Liu, M. Mezei and Q. Si, “Quantum phase transitions in holographic models of magnetism and superconductors,” Phys. Rev. D [**82**]{}, 045002 (2010) \[arXiv:1003.0010 \[hep-th\]\].
R. -G. Cai and R. -Q. Yang, “Paramagnetism-Ferromagnetism Phase Transition in a Dyonic Black Hole,” Phys. Rev. D [**90**]{}, 081901 (2014) \[arXiv:1404.2856 \[hep-th\]\].
R. G. Cai and R. Q. Yang, “Coexistence and competition of ferromagnetism and $p$-wave superconductivity in holographic model,” Phys. Rev. D [**91**]{}, no. 2, 026001 (2015) \[arXiv:1410.5080 \[hep-th\]\].
R. G. Cai, R. Q. Yang and F. V. Kusmartsev, “A holographic model for antiferromagnetic quantum phase transition induced by magnetic field,” arXiv:1501.04481 \[hep-th\].
R. -G. Cai and Y. -Z. Zhang, “Black plane solutions in four-dimensional space-times,” Phys. Rev. D [**54**]{}, 4891 (1996) \[gr-qc/9609065\]. D. Anninos, S. A. Hartnoll and N. Iqbal, “Holography and the Coleman-Mermin-Wagner theorem,” Phys. Rev. D [**82**]{}, 066008 (2010) \[arXiv:1005.1973 \[hep-th\]\].
C. P. Herzog, P. K. Kovtun and D. T. Son, “Holographic model of superfluidity,” Phys. Rev. D [**79**]{}, 066002 (2009) \[arXiv:0809.4870 \[hep-th\]\]. Y. Brihaye and B. Hartmann, “Holographic superfluid/fluid/insulator phase transitions in 2+1 dimensions,” Phys. Rev. D [**83**]{}, 126008 (2011) \[arXiv:1101.5708 \[hep-th\]\]. D. Arean, M. Bertolini, C. Krishnan and T. Prochazka, “Type IIB Holographic Superfluid Flows,” JHEP [**1103**]{}, 008 (2011) \[arXiv:1010.5777 \[hep-th\]\]. Y. B. Wu, J. W. Lu, W. X. Zhang, C. Y. Zhang, J. B. Lu and F. Yu, “Holographic $p$-wave superfluid,” Phys. Rev. D [**90**]{}, no. 12, 126006 (2014) \[arXiv:1410.5243 \[hep-th\]\].
O. Domenech, M. Montull, A. Pomarol, A. Salvio and P. J. Silva, “Emergent Gauge Fields in Holographic Superconductors,” JHEP [**1008**]{}, 033 (2010) \[arXiv:1005.1776 \[hep-th\]\]. X. Gao, M. Kaminski, H. B. Zeng and H. Q. Zhang, “Non-Equilibrium Field Dynamics of an Honest Holographic Superconductor,” JHEP [**1211**]{}, 112 (2012) \[arXiv:1204.3103 \[hep-th\]\].
R. Flauger, E. Pajer and S. Papanikolaou, “A Striped Holographic Superconductor,” Phys. Rev. D [**83**]{}, 064009 (2011) \[arXiv:1010.1775 \[hep-th\]\].
G. T. Horowitz and J. E. Santos, “General Relativity and the Cuprates,” arXiv:1302.6586 \[hep-th\].
Ó. J. C. Dias, G. T. Horowitz, N. Iqbal and J. E. Santos, “Vortices in holographic superfluids and superconductors as conformal defects,” JHEP [**1404**]{}, 096 (2014) \[arXiv:1311.3673 \[hep-th\]\].
H. B. Zeng and J. P. Wu, “Holographic superconductors from the massive gravity,” Phys. Rev. D [**90**]{}, 046001 (2014) \[arXiv:1404.5321 \[hep-th\]\].
J. i. Koga, K. Maeda and K. Tomoda, “Holographic superconductor model in a spatially anisotropic background,” Phys. Rev. D [**89**]{}, no. 10, 104024 (2014) \[arXiv:1401.6501 \[hep-th\]\]. Y. Ling, P. Liu, C. Niu, J. P. Wu and Z. Y. Xian, “Holographic Superconductor on Q-lattice,” arXiv:1410.6761 \[hep-th\]. D. Arean, A. Farahi, L. A. Pando Zayas, I. S. Landea and A. Scardicchio, “Holographic superconductor with disorder,” Phys. Rev. D [**89**]{}, no. 10, 106003 (2014) \[arXiv:1308.1920 \[hep-th\]\]. J. Erdmenger, B. Herwerth, S. Klug, R. Meyer and K. Schalm, “S-Wave Superconductivity in Anisotropic Holographic Insulators,” arXiv:1501.07615 \[hep-th\]. K. Y. Kim, K. K. Kim and M. Park, “A Simple Holographic Superconductor with Momentum Relaxation,” arXiv:1501.00446 \[hep-th\].
G. T. Horowitz, J. E. Santos and B. Way, “A Holographic Josephson Junction,” Phys. Rev. Lett. [**106**]{}, 221601 (2011) \[arXiv:1101.3326 \[hep-th\]\].
Y. Q. Wang, Y. X. Liu and Z. H. Zhao, “Holographic Josephson Junction in 3+1 dimensions,” arXiv:1104.4303 \[hep-th\]. M. Siani, “On inhomogeneous holographic superconductors,” arXiv:1104.4463 \[hep-th\]. Y. Q. Wang, Y. X. Liu and Z. H. Zhao, “Holographic p-wave Josephson junction,” arXiv:1109.4426 \[hep-th\]. Y. Q. Wang, Y. X. Liu, R. G. Cai, S. Takeuchi and H. Q. Zhang, “Holographic SIS Josephson Junction,” JHEP [**1209**]{}, 058 (2012) \[arXiv:1205.4406 \[hep-th\]\]. M. Rozali and A. Vincart-Emard, “Chiral Edge Currents in a Holographic Josephson Junction,” JHEP [**1401**]{}, 003 (2014) \[arXiv:1310.4510 \[hep-th\]\]. R. G. Cai, Y. Q. Wang and H. Q. Zhang, “A holographic model of SQUID,” JHEP [**1401**]{}, 039 (2014) \[arXiv:1308.5088 \[hep-th\]\]. S. Takeuchi, “Holographic Superconducting Quantum Interference Device,” arXiv:1309.5641 \[hep-th\].
E. Kiritsis and V. Niarchos, “Josephson Junctions and AdS/CFT Networks,” JHEP [**1107**]{}, 112 (2011) \[Erratum-ibid. [**1110**]{}, 095 (2011)\] \[arXiv:1105.6100 \[hep-th\]\].
H. F. Li, L. Li, Y. Q. Wang and H. Q. Zhang, “Non-relativistic Josephson Junction from Holography,” JHEP [**1412**]{}, 099 (2014) \[arXiv:1410.5578 \[hep-th\]\].
K. Murata, S. Kinoshita and N. Tanahashi, “Non-equilibrium Condensation Process in a Holographic Superconductor,” JHEP [**1007**]{}, 050 (2010) \[arXiv:1005.0633 \[hep-th\]\]. M. J. Bhaseen, J. P. Gauntlett, B. D. Simons, J. Sonner and T. Wiseman, “Holographic Superfluids and the Dynamics of Symmetry Breaking,” Phys. Rev. Lett. [**110**]{}, 015301 (2013) \[arXiv:1207.4194 \[hep-th\]\]. J. Sonner, A. del Campo and W. H. Zurek, “Universal far-from-equilibrium Dynamics of a Holographic Superconductor,” arXiv:1406.2329 \[hep-th\].
X. Bai, B. H. Lee, L. Li, J. R. Sun and H. Q. Zhang, “Time Evolution of Entanglement Entropy in Quenched Holographic Superconductors,” JHEP [**1504**]{}, 066 (2015) \[arXiv:1412.5500 \[hep-th\]\].
A. Adams, P. M. Chesler and H. Liu, “Holographic Vortex Liquids and Superfluid Turbulence,” Science [**341**]{}, 368 (2013) \[arXiv:1212.0281 \[hep-th\]\]. X. Gao, A. M. Garcia-Garcia, H. B. Zeng and H. Q. Zhang, “Normal modes and time evolution of a holographic superconductor after a quantum quench,” JHEP [**1406**]{}, 019 (2014) \[arXiv:1212.1049 \[hep-th\]\].
W. -J. Li, Y. Tian and H. -b. Zhang, “Periodically Driven Holographic Superconductor,” JHEP [**1307**]{}, 030 (2013) \[arXiv:1305.1600 \[hep-th\]\].
A. M. García-García, H. B. Zeng and H. Q. Zhang, “A thermal quench induces spatial inhomogeneities in a holographic superconductor,” JHEP [**1407**]{}, 096 (2014) \[arXiv:1308.5398 \[hep-th\]\].
P. M. Chesler, A. M. Garcia-Garcia and H. Liu, “Far-from-equilibrium coarsening, defect formation, and holography,” arXiv:1407.1862 \[hep-th\]. Y. Du, C. Niu, Y. Tian and H. Zhang, “Holographic Vortex Pair Annihilation in Superfluid Turbulence,” arXiv:1412.8417 \[hep-th\]. C. P. Herzog, “An Analytic Holographic Superconductor,” Phys. Rev. D [**81**]{}, 126009 (2010) \[arXiv:1003.3278 \[hep-th\]\]. G. Siopsis and J. Therrien, “Analytic Calculation of Properties of Holographic Superconductors,” JHEP [**1005**]{}, 013 (2010) \[arXiv:1003.4275 \[hep-th\]\]. H. B. Zeng, X. Gao, Y. Jiang and H. S. Zong, “Analytical Computation of Critical Exponents in Several Holographic Superconductors,” JHEP [**1105**]{}, 002 (2011) \[arXiv:1012.5564 \[hep-th\]\]. R. G. Cai, H. F. Li and H. Q. Zhang, “Analytical Studies on Holographic Insulator/Superconductor Phase Transitions,” Phys. Rev. D [**83**]{}, 126007 (2011) \[arXiv:1103.5568 \[hep-th\]\].
D. Momeni, E. Nakano, M. R. Setare and W. Y. Wen, “Analytical study of critical magnetic field in a holographic superconductor,” Int. J. Mod. Phys. A [**28**]{}, 1350024 (2013) \[arXiv:1108.4340 \[hep-th\]\]. X. X. Zeng, X. M. Liu and W. B. Liu, “Analytic treatment on stimulated holographic superconductors,” Int. J. Mod. Phys. A [**27**]{}, 1250010 (2012).
Q. Pan, J. Jing, B. Wang and S. Chen, “Analytical study on holographic superconductors with backreactions,” JHEP [**1206**]{}, 087 (2012) \[arXiv:1205.3543 \[hep-th\]\]. S. Gangopadhyay and D. Roychowdhury, “Analytic study of properties of holographic p-wave superconductors,” JHEP [**1208**]{}, 104 (2012) \[arXiv:1207.5605 \[hep-th\]\]. W. H. Huang, “Analytic Study of First-Order Phase Transition in Holographic Superconductor and Superfluid,” Int. J. Mod. Phys. A [**28**]{}, 1350140 (2013) \[arXiv:1307.5614 \[hep-th\]\]. R. Banerjee, S. Gangopadhyay, D. Roychowdhury and A. Lala, “Holographic s-wave condensate with nonlinear electrodynamics: A nontrivial boundary value problem,” Phys. Rev. D [**87**]{}, no. 10, 104001 (2013) \[arXiv:1208.5902 \[hep-th\]\]. D. Momeni, M. Raza and R. Myrzakulov, “Analytical coexistence of s, p, s + p phases of a holographic superconductor,” arXiv:1310.1735 \[hep-th\]. J. W. Lu, Y. B. Wu, P. Qian, Y. Y. Zhao and X. Zhang, “Lifshitz Scaling Effects on Holographic Superconductors,” Nucl. Phys. B [**887**]{}, 112 (2014) \[arXiv:1311.2699 \[hep-th\]\].
G. T. Horowitz and J. Polchinski, “Gauge/gravity duality,” In \*Oriti, D. (ed.): Approaches to quantum gravity\* 169-186 \[gr-qc/0602037\].
S. S. Lee, “A Non-Fermi Liquid from a Charged Black Hole: A Critical Fermi Ball,” Phys. Rev. D [**79**]{}, 086006 (2009) \[arXiv:0809.3402 \[hep-th\]\]. H. Liu, J. McGreevy and D. Vegh, “Non-Fermi liquids from holography,” Phys. Rev. D [**83**]{}, 065029 (2011) \[arXiv:0903.2477 \[hep-th\]\]. M. Cubrovic, J. Zaanen and K. Schalm, “String Theory, Quantum Phase Transitions and the Emergent Fermi-Liquid,” Science [**325**]{}, 439 (2009) \[arXiv:0904.1993 \[hep-th\]\].
T. Faulkner, H. Liu, J. McGreevy and D. Vegh, “Emergent quantum criticality, Fermi surfaces, and AdS(2),” Phys. Rev. D [**83**]{}, 125002 (2011) \[arXiv:0907.2694 \[hep-th\]\]. T. Faulkner and J. Polchinski, “Semi-Holographic Fermi Liquids,” JHEP [**1106**]{}, 012 (2011) \[arXiv:1001.5049 \[hep-th\]\].
J. L. Davis, P. Kraus and A. Shah, “Gravity Dual of a Quantum Hall Plateau Transition,” JHEP [**0811**]{}, 020 (2008) \[arXiv:0809.1876 \[hep-th\]\].
M. Fujita, W. Li, S. Ryu and T. Takayanagi, “Fractional Quantum Hall Effect via Holography: Chern-Simons, Edge States, and Hierarchy,” JHEP [**0906**]{}, 066 (2009) \[arXiv:0901.0924 \[hep-th\]\]. O. Bergman, N. Jokela, G. Lifschytz and M. Lippert, “Quantum Hall Effect in a Holographic Model,” JHEP [**1010**]{}, 063 (2010) \[arXiv:1003.4965 \[hep-th\]\].
S. A. Hartnoll, J. Polchinski, E. Silverstein and D. Tong, “Towards strange metallic holography,” JHEP [**1004**]{}, 120 (2010) \[arXiv:0912.1061 \[hep-th\]\]. T. Faulkner, N. Iqbal, H. Liu, J. McGreevy and D. Vegh, “From Black Holes to Strange Metals,” arXiv:1003.1728 \[hep-th\]. B. S. Kim, E. Kiritsis and C. Panagopoulos, “Holographic quantum criticality and strange metal transport,” New J. Phys. [**14**]{}, 043045 (2012) \[arXiv:1012.3464 \[cond-mat.str-el\]\].
R. A. Davison, K. Schalm and J. Zaanen, “Holographic duality and the resistivity of strange metals,” Phys. Rev. B [**89**]{}, 245116 (2014) \[arXiv:1311.2451 \[hep-th\]\].
C. Hoyos-Badajoz, K. Jensen and A. Karch, “A Holographic Fractional Topological Insulator,” Phys. Rev. D [**82**]{}, 086001 (2010) \[arXiv:1007.3253 \[hep-th\]\]. S. Ryu and T. Takayanagi, “Topological Insulators and Superconductors from String Theory,” Phys. Rev. D [**82**]{}, 086014 (2010) \[arXiv:1007.4234 \[hep-th\]\]. A. Karch, J. Maciejko and T. Takayanagi, “Holographic fractional topological insulators in 2+1 and 1+1 dimensions,” Phys. Rev. D [**82**]{}, 126003 (2010) \[arXiv:1009.2991 \[hep-th\]\]. M. Fujita, S. Harrison, A. Karch, R. Meyer and N. M. Paquette, “Towards a Holographic Bose-Hubbard Model,” arXiv:1411.7899 \[hep-th\].
J. Polchinski and M. J. Strassler, “Hard scattering and gauge / string duality,” Phys. Rev. Lett. [**88**]{}, 031601 (2002) \[hep-th/0109174\]. H. Boschi-Filho and N. R. F. Braga, “QCD / string holographic mapping and glueball mass spectrum,” Eur. Phys. J. C [**32**]{}, 529 (2004) \[hep-th/0209080\]. J. Erlich, E. Katz, D. T. Son and M. A. Stephanov, “QCD and a holographic model of hadrons,” Phys. Rev. Lett. [**95**]{}, 261602 (2005) \[hep-ph/0501128\]. G. F. de Teramond and S. J. Brodsky, “Hadronic spectrum of a holographic dual of QCD,” Phys. Rev. Lett. [**94**]{}, 201601 (2005) \[hep-th/0501022\].
J. Babington, J. Erdmenger, N. J. Evans, Z. Guralnik and I. Kirsch, “Chiral symmetry breaking and pions in nonsupersymmetric gauge / gravity duals,” Phys. Rev. D [**69**]{}, 066007 (2004) \[hep-th/0306018\]. M. Kruczenski, D. Mateos, R. C. Myers and D. J. Winters, “Towards a holographic dual of large N(c) QCD,” JHEP [**0405**]{}, 041 (2004) \[hep-th/0311270\]. S. S. Gubser, “Drag force in AdS/CFT,” Phys. Rev. D [**74**]{}, 126005 (2006) \[hep-th/0605182\]. E. Shuryak, S. -J. Sin and I. Zahed, “A Gravity dual of RHIC collisions,” J. Korean Phys. Soc. [**50**]{}, 384 (2007) \[hep-th/0511199\]. U. Gursoy, E. Kiritsis, L. Mazzanti and F. Nitti, “Deconfinement and Gluon Plasma Dynamics in Improved Holographic QCD,” Phys. Rev. Lett. [**101**]{}, 181601 (2008) \[arXiv:0804.0899 \[hep-th\]\]. C. P. Herzog, “A Holographic Prediction of the Deconfinement Temperature,” Phys. Rev. Lett. [**98**]{}, 091601 (2007) \[hep-th/0608151\]. P. Colangelo, F. Giannuzzi and S. Nicotri, “Holography, Heavy-Quark Free Energy, and the QCD Phase Diagram,” Phys. Rev. D [**83**]{}, 035015 (2011) \[arXiv:1008.3116 \[hep-ph\]\]. H. -Y. Chen, K. Hashimoto and S. Matsuura, “Towards a Holographic Model of Color-Flavor Locking Phase,” JHEP [**1002**]{}, 104 (2010) \[arXiv:0909.1296 \[hep-th\]\]. M. Huang, S. He, , Q. S. Yan and Y. Yang, “Confront Holographic QCD with Regge Trajectories,” Eur. Phys. J. C [**66**]{}, 187 (2010) \[arXiv:0710.0988 \[hep-ph\]\].
W. van der Schee, “Gravitational collisions and the quark-gluon plasma,” arXiv:1407.1849 \[hep-th\].
S. Bhattacharyya, V. E. Hubeny, S. Minwalla and M. Rangamani, “Nonlinear Fluid Dynamics from Gravity,” JHEP [**0802**]{}, 045 (2008) \[arXiv:0712.2456 \[hep-th\]\]. M. Rangamani, “Gravity and Hydrodynamics: Lectures on the fluid-gravity correspondence,” Class. Quant. Grav. [**26**]{}, 224003 (2009) \[arXiv:0905.4352 \[hep-th\]\].
I. Bredberg, C. Keeler, V. Lysov and A. Strominger, “From Navier-Stokes To Einstein,” JHEP [**1207**]{}, 146 (2012) \[arXiv:1101.2451 \[hep-th\]\].
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: E-mail: [email protected]
[^4]: E-mail: [email protected]
[^5]: Loosely speaking, $N^2$ can be considered as the degrees of freedom in the dual field theory, and $\lambda$ as the characteristic strength of interactions. An elementary introduction to this correspondence can be found in the next section.
[^6]: Holographic superconductor models constructed in the top-down approach can be found, for example, in refs. [@Denef:2009tp; @Gubser:2009qm; @Gauntlett:2009dn; @Gubser:2009gp; @Gauntlett:2009bh; @KalyanaRama:2011ny; @Bobev:2011rv].
[^7]: To have a wide description of the Ginzburg-Landau theory see ref. [@cyrot:1973] and references therein.
[^8]: The simple examples are Lifshitz symmetry [@Kachru:2008yh] and Schrödinger symmetry [@Son:2008ye; @Balasubramanian:2008dm], while more generic cases are those with generalized Lifshitz invariance and hyperscaling violation [@Charmousis:2010zz; @Gouteraux:2011ce; @Huijse:2011ef; @Dong:2012se], and the associated Schrödinger cousins [@Kim:2012nb]. However, those take us outside the best understood AdS/CFT framework. We shall focus on the most well defined case involving the bulk geometry with the asymptotically AdS behaviour.
[^9]: Note that $AdS_{d+1}$ spacetime is stable even when the mass squared $m^2$ of scalar field is negative provided $m^2L^2\geq m^2_{BF}L^2=-d^2/4$ [@Breitenlohner:1982jf]. The lower bound $m^2_{BF}=-d^2/4L^2$ is often called Breitenloner-Freedman (BF) bound.
[^10]: An intrinsically real-time holographic prescription was first proposed by the authors of ref. [@Son:2002sd] by essentially analytically continuing the Euclidean prescription. It has been justified by a holographic version of the Schwinger-Keldysh formalism [@Herzog:2002pc; @Skenderis:2008dh].
[^11]: In fact, it has been shown that even more general quantisations are possible, like double trace deformation, see, for example, refs. [@Witten:2001ua; @Berkooz:2002ug].
[^12]: The model is a s-wave one since the condensed field is a scalar field dual to a scalar operator in the field theory side. This model can be straightforwardly generalized to other spacetime dimensions. Holographic s-wave superconductors with generalised couplings have also been considered in a number of works [@Franco:2009yz; @Aprile:2009ai; @Aprile:2010yb; @Liu:2010ka; @Chen:2010hi; @Peng:2011gh; @Bigazzi:2011ak; @Dey:2014xxa; @Arean:2015wea].
[^13]: In the probe limit, the concrete value of the charge $q$ does not play an essential role. Without loss of generality, we take $q$ to be one.
[^14]: We only consider the standard quantization here, regarding the leading coefficient as the source of dual operator. An alternative way inducing spontaneous symmetry breaking in holographic superconductors is to introduce double trace deformation [@Faulkner:2010gj].
[^15]: It has been shown that the conductivity is directly related to the reflection coefficient with the frequency given the incident energy [@Horowitz:2009ij]. The key point is that even as $T\rightarrow0$ there is still tunneling through the barrier provided by the effective potential. Therefore, a nonzero conductivity at small frequencies will always exist, and hence there is no hard gap in the optical conductivity at zero temperature. To obtain a superconductor with a hard gap, one might consider non-minimally coupled scalars in the bulk.
[^16]: Its precise form can be found in below.
[^17]: The ration $\rho_s/\rho$ versus temperature in the left plot is reminiscent of the temperature dependence of the superfluid of liquid He II as measured from in the torsional oscillation disk stack experiment. However, we find $\rho_s$ goes to zero linearly here, while the experiment gives a critical exponent about $0.67$.
[^18]: However, by considering the back reaction to the metric in the SU(2) model , it has been shown that the conductivity in the $y$ direction has a “hard gap" at zero temperature, i.e., the real part of the conductivity is zero for an excitation frequency less than the gap frequency [@Basu:2009vv].
[^19]: The terminology “pseudogap” here is to denote a well defined gap in the dissipative conductivity at low frequencies in which the conductivity is not identically zero.
[^20]: In order to satisfy the equations of motion with the given ansatz, $\theta$ can only be chosen as $\theta_+=\frac{\pi}{2}+2n\pi$ or $\theta_-=-\frac{\pi}{2}+2n\pi$ with $n$ an arbitrary integer. Here and below the upper signs correspond to the $\theta_+$ case and the lower to the $\theta_-$ case.
[^21]: The $m^2$ has a lower bound as $m^2=-1$ with ${\Delta}_+={\Delta}_-=1$. In that case, there exists a logarithmic term in the asymptotical expansion of $\rho_x$. One has to treat such a term as the source set to be zero to avoid the instability induced by this term [@Horowitz:2008bn]. We will always consider the case with $m^2 >-1$.
[^22]: In the theory of superfluidity and superconductivity, a discontinuity of the free energy was discussed theoretically and an exactly solvable model for such phase transition was given in ref. [@Maslv:2004]. The zeroth order transition was also observed in holographic superconductors in refs. [@Zeng:2013yoa; @Zeng:2014dra; @Chaturvedi:2015hra].
[^23]: For more details about this result, please see ref. [@Donos:2013woa].
[^24]: The main reason for this choice is to obtain real scaling dimensions. For other values of $m, e$ which can avoid complex scaling dimensions will give similar results [@Donos:2013woa].
[^25]: Note that one can set $\bar f_0=k=1$ by scaling $t$ and $x_1$.
[^26]: As the temperature is lowered, the pitch $(2\pi/k)$ first increases, becoming divergent (i.e., $k=0$) at some particular temperature, then changes sign and finally decreases in magnitude to a value $k<0$ at $T=0$.
[^27]: In fact, the previous d-wave model does not consider this aspect.
[^28]: Ref. [@Nishida:2014lta] also discussed the following model, but with a coupling between the scalar field and the tensor field, and studied the phase structure in terms of the coupling parameter and temperature with fixed charges of two orders. In the folloowing discussion, there is no direct interaction between scalar and tensor fields and the model parameter is the ratio of two fields. Note that in paper [@Nishida:2014lta], when the coupling $\eta= 0$, there also exists a coexisting phase under the model parameters $m^2_1 = -2, m^2_2 = 0$ and $q_2 = 1.95$. Both results are consistent with each other in that case.
[^29]: However, under special conditions superconductivity may coexist with antiferromagnetic order, where neighboring electron spins arrange in an antiparallel configuration. Since antiferromagnets don’t have net magnetism, we won’t involve them here.
[^30]: We will use “superconducting ferromagnet" to denote the materials whose Curie temperature is higher than superconducting transition temperature and “ferromagnetic superconductor" to denote the opposite case.
[^31]: The asymptotic solution of $c(r)$ depends on the source free condition of $\rho_x$. When $\rho_{x+}\neq0$, asymptotic solution of $c(r)$ becomes $c=c_++c_-r^{-\delta_1}$.
[^32]: In principle, the arbitrary can be fixed by embedding the bulk model into some low energy effective theory of string/M theory.
[^33]: In fact, most of the condensed matter theories do not include dynamical photons, as their effects are usually small. For example, in the BCS theory electromagnetic field is often introduced as an external field. The possibility of introducing dynamical gauge fields in holographic superconductors was discussed in refs. [@Domenech:2010nf; @Gao:2012yw].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Initial data for the spherically symmetric Einstein-Vlasov system is constructed whose past evolution is regular and whose future evolution contains a black hole. This is the first example of initial data with these properties for the Einstein-matter system with a “realistic” matter model. One consequence of the result is that there exists a class of initial data for which the ratio of the Hawking mass $\open{m}=\open{m}(r)$ and the area radius $r$ is arbitrarily small everywhere, such that a black hole forms in the evolution. This result is in a sense analogous to the result [@Cu3] for a scalar field. Another consequence is that there exist black hole initial data such that the solutions exist for all Schwarzschild time $t\in (-\infty,\infty)$.'
author:
- |
Håkan Andréasson\
Mathematical Sciences\
University of Gothenburg\
Mathematical Sciences\
Chalmers University of Technology\
S-41296 Göteborg, Sweden\
email: [email protected]
title: Black hole formation from a complete regular past for collisionless matter
---
Introduction
============
An important question in the study of gravitational collapse is to identify physically admissible initial data, and it is natural to require that the past evolution of the data is regular. However, most of the existing mathematical results which ensure a regular past also ensure a regular future, cf. [@RR; @CK], which rules out the study of the formation of black holes. The exceptions being the classical result for dust [@OS], where some classes of solutions have a regular past, and the recent result [@Ds2] for a scalar field. In the latter work, which in part rests on the studies [@Cu2; @Cu3], initial data whose past evolution is regular and whose future evolution forms a black hole is constructed. Now, neither dust nor a scalar field are realistic matter models in the sense that they are used by astrophysicists. Dust is a perfect fluid where the pressure is assumed to be zero, and a scalar field is merely a toy model. Thus, there is so far no example of a solution to the Einstein-matter system for a realistic matter model possessing a regular past and a singular future.
In this work we consider collisionless matter governed by the Vlasov equation, cf. [@A1] and [@R1] for an introduction. Although this is a simple matter model, it has rich dynamics and many features that are desirable of a realistic matter model. Indeed, it allows for anisotropic pressure, there is a large number of stable and unstable spherically symmetric and axially symmetric stationary solutions [@AKR4; @Rn1], there is numerical support that time periodic solutions exist [@AR2], it behaves as Type I matter in critical collapse [@AR2; @OC; @RRS2], and it is used by astrophysicists [@BT]. The following theorem is the main result of the present paper.
\[Theorem1\] There exists a class of initial data ${\cal J}$ for the spherically symmetric Einstein-Vlasov system with the property that black holes form in the future time direction and in the past time direction spacetime is causally geodesically complete.
We will see that a consequence of this result is that for any $\epsilon>0$, initial data can be constructed with the property that the ratio $\open{m}/r$ of the initial Hawking mass $\open{m}=\open{m}(r)$, and the area radius $r$, is less than $\epsilon$ everywhere, such that a black hole forms in the evolution. We formulate this as a corollary.
\[Corollary1\] Given $\epsilon>0$, there exists a class ${\cal J}_r$ of initial data for the spherically symmetric Einstein-Vlasov system which satisfy $$\sup_r \frac{\open{m}\,(r)}{r}\leq \epsilon,$$ for which black holes form in the evolution.
This result improves the main result of [@AKR2] and is analogous to the result [@Cu3] in the case of a scalar field where conditions on the data are given which ensure the formation of black holes. These conditions give no lower bound on $2\open{m}/r$ but involve other restrictions. Another consequence of our result is the following corollary.
\[Corollary2\] There exists a class ${\cal J}_s$ of black hole initial data for the spherically symmetric Einstein-Vlasov system such that the corresponding solutions exist for all Schwarzschild time $t\in (-\infty,\infty)$.
In the future time direction this corollary was shown in [@AKR2], the improvement here is that the solutions exist on the entire real line.
The present result relies in part on the previous studies [@A2], [@AKR1] and [@AKR2], which now will be reviewed. In [@AKR1] global existence in a maximal time gauge is shown for a particular class of initial data where the particles are moving rapidly outwards. One of the restrictions imposed on the initial data is that $$\label{ratiomtg}
\sup_{r}\frac{2\open{m}(r)}{r}<k_0,$$ where the constant $k_0$ is roughly $1/10.$ The situation considered in [@AKR2] is in a sense the reverse since the initial data is such that the particles move rapidly inwards and the quantity $\sup_{r} 2\open{m}/r$ is required to be close to one. The main result of [@AKR2] is that data of this kind guarantee the formation of black holes in the evolution. The analysis in [@AKR2] is carried out in Schwarzschild coordinates, i.e. in a polar time gauge. Now, particles that move inward in the future time direction move outward in the past time direction. It is thus natural to try to combine these two results with the goal of constructing solutions with a regular past and a singular future. The conditions on the ratio $2\open{m}/r$ are clearly very different in [@AKR1] compared to [@AKR2], and moreover, the Cauchy hypersurfaces are different since a maximal time gauge and a polar time gauge are imposed in the respective cases. The main reason why a maximal time gauge is used in [@AKR1] is due to the difficulties related to the so called pointwise terms in the characteristic equations in Schwarzschild coordinates. In [@A2] the problem of global existence for general initial data is investigated under conditional assumptions on the solutions. The analysis along characteristics is applied to a modified quantity for which the problems with the pointwise terms in Schwarzschild coordinates do not appear.
In the present work we combine the strategies in [@A2] and [@AKR1] and show global existence for rapidly outgoing particles in Schwarzschild coordinates. In particular the result in [@AKR1] is improved by showing that the restriction (\[ratiomtg\]) can be relaxed, and for sufficiently fast moving particles $2\open{m}/r$ is allowed to be arbitrarily close to one. By combining this result with the result in [@AKR2] we are then able to construct data whose past is regular and whose future contains a black hole.
The outline of the paper is as follows. In the next section the spherically symmetric Einstein-Vlasov system is introduced. Global existence for rapidly outgoing particles is shown in section 3 for two different sets of initial data which are adapted to Corollary 1 and Corollary 2 respectively. Finally, in section 4 the proofs of Theorem \[Theorem1\], Corollary 1 and Corollary 2 are given.
The Einstein-Vlasov system
==========================
For an introduction to the Einstein-Vlasov system and kinetic theory we refer to [@A1] and [@R1], and for a careful derivation of the system given below we refer to [@R]. In Schwarzschild coordinates the spherically symmetric metric takes the form $$ds^{2}=-e^{2\mu(t,r)}dt^{2}+e^{2\lambda(t,r)}dr^{2}
+r^{2}(d\theta^{2}+\sin^{2}{\theta}d\varphi^{2}).$$ The Einstein equations read $$\begin{aligned}
&\displaystyle e^{-2\lambda}(2r\lambda_{r}-1)+1=8\pi r^2\rho,&\label{ee1}\\
&\displaystyle e^{-2\lambda}(2r\mu_{r}+1)-1=8\pi r^2 p,&\label{ee2}\\
&\displaystyle\lambda_{t}=-4\pi re^{\lambda+\mu}j,&\label{ee3}\\
&\displaystyle e^{-2\lambda}(\mu_{rr}+(\mu_{r}-\lambda_{r})(\mu_{r}+
\frac{1}{r}))-e^{-2\mu}(\lambda_{tt}+\lambda_{t}(\lambda_{t}-\mu_{t}))=
8\pi p_T.&\label{ee4}\end{aligned}$$ The indices $t$ and $r$ denote partial derivatives. The Vlasov equation for the density function $f=f(t,r,w,L)$ is given by $$\partial_{t}f+e^{\mu-\lambda}\frac{w}{E}\partial_{r}f
-(\lambda_{t}w+e^{\mu-\lambda}\mu_{r}E-
e^{\mu-\lambda}\frac{L}{r^3E})\partial_{w}f=0,\label{vlasov}$$ where $$E=E(r,w,L)=\sqrt{1+w^{2}+L/r^{2}}.\label{E}$$ Here $w\in (-\infty,\infty)$ can be thought of as the radial component of the momentum variables, and $L\in [0,\infty)$ is the square of the angular momentum. The matter quantities are defined by $$\begin{aligned}
\rho(t,r)&=&\frac{\pi}{r^{2}}
\int_{-\infty}^{\infty}\int_{0}^{\infty}Ef(t,r,w,L)\;dwdL,\label{rho}\\
p(t,r)&=&\frac{\pi}{r^{2}}\int_{-\infty}^{\infty}\int_{0}^{\infty}
\frac{w^{2}}{E}f(t,r,w,L)\;d
wdL,\label{p}\\
j(t,r)&=&\frac{\pi}{r^{2}}
\int_{-\infty}^{\infty}\int_{0}^{\infty}wf(t,r,w,L)\;dwdL,\label{j}\\
p_T(t,r)&=&\frac{\pi}{2r^{4}}\int_{-\infty}^{\infty}\int_{0}^{\infty}\frac{L}{E}f(t,r,w,L)\;
dwdL.\label{q}\end{aligned}$$ Here $\rho,p,j$ and $p_T$ are the energy density, the radial pressure, the current and the tangential pressure respectively. The following boundary conditions are imposed to ensure asymptotic flatness $$\lim_{r\rightarrow\infty}\lambda(t,r)=\lim_{r\rightarrow\infty}\mu(t,r)=0,\label{bdryaf}$$ and a regular centre requires that $$\label{regcentre}
\lambda(t,0)=0.$$ We point out that the Einstein equations are not independent and that e.g. the equations (\[ee3\]) and (\[ee4\]) follow by (\[ee1\]), (\[ee2\]) and (\[vlasov\]).
As initial data it is sufficient to prescribe a density function $\open{f}=\open{f}(r,w,L)\geq 0$ such that $$\label{notsinit}
\int_0^r 4\pi\eta^2\open{\rho}\,(\eta)\,d\eta < \frac{r}{2}.
$$ Here we denote by $\open{\rho}$ the energy density induced by the initial distribution function $\open{f}$. This condition ensures that no trapped surfaces are present initially. Given $\open{f},$ equations (\[ee1\]) and (\[ee2\]) can be solved to give $\lambda$ and $\mu$ at $t=0.$ We will only consider initial data such that $\open{f}(r,\cdot,L)=0$ if $r\leq\epsilon$, for some $\epsilon>0$, or if $L\geq L_+,$ for some $L_+>0.$ If the initial data is $C^1([\epsilon,\infty[,]-\infty,\infty[,[0,\infty[)$ we say that it is regular.
Let us write down a couple of facts about the system (\[ee1\])-(\[bdryaf\]). A solution to the Vlasov equation can be written $$f(t,r,w,L)=\open{f}(R(0,t,r,w,L),W(0,t,r,w,L),L),
\label{repre}$$ where $R$ and $W$ are solutions of the characteristic system $$\begin{aligned}
\frac{dR}{ds}&=&e^{(\mu-\lambda)(s,R)}\frac{W}{E(R,W,L)},\label{char1}\\
\frac{dW}{ds}&=&-\lambda_{t}(s,R)W-e^{(\mu-\lambda)(s,R)}\mu_{r}(s,R)E(R,W,L)
\nonumber\\
& &+e^{(\mu-\lambda)(s,R)}\frac{L}{R^3E(R,W,L)},\label{char2}\end{aligned}$$ such that $(R(s,t,r,w,L),W(s,t,r,w,L),L)=(r,w,L)$ when $s=t$. This representation shows that $f$ is nonnegative for all $t\geq 0,$ $\|f\|_{\infty}=\|\open{f}\|_{\infty},$ and that $f(t,r,w,L)=0$ if $L>L_{+}.$ The Hawking mass $m$ of the sphere of area radius $r$ is given by $$m(t,r)=4\pi\int_{0}^{r}\eta^{2}\rho(t,\eta)d\eta,\label{m}$$ and by integrating (\[ee1\]) we find $$e^{-2\lambda(t,r)}=1-\frac{2m(t,r)}{r}.\label{e2-lambda}$$ A fact that we will need is that $$\mu+\lambda\leq 0.$$ This is easily obtained by adding the equations (\[ee1\]) and (\[ee2\]), which gives $$\lambda_{r}+\mu_{r}\geq 0,$$ and then using the boundary conditions (\[bdryaf\]). Furthermore, from (\[e2-lambda\]) we get that $\lambda\geq 0,$ and it follows that $\mu\leq 0.$ We also introduce the notations $\hat{\mu}$ and $\check{\mu}.$ From equation (\[ee2\]) and (\[bdryaf\]) we have $$\label{muhatcheck}
\mu(t,r)=-\int_r^{\infty}\frac{m(t,\eta)}{\eta^2}e^{2\lambda}\,d\eta-\int_r^{\infty}4\pi \eta pe^{2\lambda}\, d\eta=:\hat{\mu}+\check{\mu}.$$ We will need an expression for $\hat{\mu}_t$. By (\[ee3\]) and (\[e2-lambda\]) it follows that $m_t(t,r)=-4\pi r^2 j(t,r)e^{\mu-\lambda}$, and we obtain $$\label{dtmuhat}
\hat{\mu}_t(t,r)=\int_r^{\infty}4\pi j(t,\eta) e^{(\mu+\lambda)(t,\eta)}e^{2\lambda(t,\eta)}d\eta.$$ An important quantity is the ADM mass $M$, given by $$M=4\pi\int_{0}^{\infty}r^{2}\rho(t,r)dr.\label{adm}$$ The fact that it is conserved follows by using (\[ee3\]) and (\[e2-lambda\]).
The following result is given in [@A2] but since the proof is short we include it for completeness. By a regular solution we mean a solution which is launched by regular initial data with compact support.
Let $(f,\mu,\lambda)$ be a regular solution to the Einstein-Vlasov system. Then $$\begin{aligned}
&\displaystyle\int_{0}^{\infty}4\pi r(\rho+p)e^{2\lambda}e^{\mu+\lambda}dr\leq 1,&\label{expr}\\
&\displaystyle\int_{0}^{\infty}(\frac{m}{r^2}+4\pi rp)e^{2\lambda}e^{\mu}dr\leq
1.&\label{expr2}\end{aligned}$$
**Proof.** Using the boundary condition (\[bdryaf\]) we get $$\begin{aligned}
1\geq
1-e^{\mu+\lambda}(t,0)&=&\int_{0}^{\infty}\frac{d}{dr}e^{\mu+\lambda}dr\\
&=&\int_{0}^{\infty}(\mu_{r}+\lambda_{r})e^{\mu+\lambda}dr.\end{aligned}$$ The right hand side equals (\[expr\]) by equations (\[ee1\]) and (\[ee2\]) which completes the first part of the lemma. The second part follows by considering $e^{\mu}$ instead of $e^{\mu+\lambda}.$
$\Box$
Finally, we note that in [@RR] and [@A2] local existence theorems are proved for compactly supported and non-compactly supported initial data respectively, and it will be used below that solutions exist on some time interval $[0,T[.$
Global existence for outgoing matter
====================================
In order to understand the mechanism behind the global existence result for outgoing matter we recall the example in [@AKR1] and consider the much simpler Vlasov-Poisson system which is the Newtonian limit of the Einstein-Vlasov system. Due to the spherical symmetry the maximal force experienced by a particle at distance $r$ from the origin is $-M/r^2$ in the Vlasov-Poisson case, where $M>0$ is the total mass of the particle ensemble. Hence along any particle trajectory $$\frac{d}{dt} \left(\frac{1}{2} w^2 - \frac{M}{r}\right) =
w \dot w + \frac{M}{r^2}\dot r = w\,\left(\dot w + \frac{M}{r^2}\right)
\geq 0,$$ as long as its radial velocity $\dot r = w= x\cdot v/r \geq 0$. Hence $$\frac{1}{2} w^2(t) - \frac{M}{r(t)} \geq \frac{1}{2} w^2(0) - \frac{M}{r(0)}$$ and $$\frac{1}{2} w^2(t) \geq \frac{1}{2} w^2(0) - \frac{M}{r(0)}$$ on any time interval on which $w(t)$ remains non-negative. Now let $w_1>0$ be an initial lower bound for the radial velocities of the particles in the ensemble, $r_1>0$ an initial lower bound for their distance from the origin, and assume that $$W_1 := \frac{1}{2} w_1^2 - \frac{M}{r_1} > 0.$$ Then as long as a particle is moving outward, $$w(t) > W_1,\ r(t) > r_1 + W_1 t.$$ This implies that all the particles keep moving outward for all future time.
Let us turn back to the spherically symmetric Einstein-Vlasov system. The notation below is adapted to the notation in [@AKR2] since the aim is to show that the initial data we construct overlap with the initial data in [@AKR2]. Two different sets of initial data, adapted to Corollary 1 and 2, will be considered and two similar results on global existence will be shown below; Theorem 2 and Theorem 3. Let us point out that if the only goal had been to improve the global existence result for rapidly outgoing particles in [@AKR1] then we could have considered a simpler class of initial data analogous to the data in [@AKR1], cf. Remark 2 below.
Let $0<r_0<r_1$ be given and put $M=r_1/2$. Let $\open{f}_s$ be data of a steady state supported in $[r'_0,r_0]$ and let $$\label{M-checkM}
M_\mathrm{in}:=\int_{r'_0}^{r_0}4\pi r^2\open{\rho}_s(r)dr,$$ where $\open{\rho}_s$ is induced by $\open{f}_s$. The results in [@A4] guarantee that such steady states exist if $r'_0$ is sufficiently small, and moreover that $$\label{moverrss}
\sup_{0\leq r\leq r_0}\frac{2\open{m}(r)}{r}<\frac89.$$ This implies in particular that $2M_{\mathrm{in}}/r_0< 8/9$ so that $M>9M_{\mathrm{in}}/8$. Let $M_\mathrm{out}:=M-M_{\mathrm{in}}$. Let $R_1>r_1$ be such that $$\label{mediumstrip}
R_1-r_1<\frac{r_1-r_0}{6},$$ and define $$R_0:=\frac{1}{2}(r_1+R_1).$$ Let $L_+>0$ and let $W_{*}>0$ be such that $$\label{Lplusc}
|W_{*}|\geq 1+\frac{\sqrt{L_{+}}}{R_{0}}.$$ Let $W_{-}>0$ satisfy $$\label{mainc}
|W_{-}|\,e^{\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}}
(1-\frac{2M}{R_{0}})^{3/2}\geq 3 |W_{*}|.$$ We remark that since $W_*>0$ and $W_->0$ in this section, the modulus is superfluous but it will be needed below. The same remark applies to the time variable which is non-negative in this section but which will be non-positive below and we therefore in some situations write the modulus of the time variable, cf. (\[supp-as\]). We are now in a position to specify the initial data. Let $\open{f}=\open{f}_s+\open{f}_m\,$ be initial data of ADM mass $M$ such that $$\supp \open{f}_m \subset [R_{0}, R_1]\times [W_-, \infty[\times [0, L_+]\,,$$ and $$\label{checkM}
\int_{R_0}^{R_1}4\pi r^2\open{\rho}_m(r)dr=M_\mathrm{out},$$ where $\open{\rho}_m$ is the induced energy density by $\open{f}_m$. *Remark 1:* Note that the condition (\[checkM\]) can be arranged by first choosing $h_m$ such that $\supp h_m\subset [R_{0}, R_1]\times [W_-, \infty[\times [0, L_+]$ and then choosing an amplitude $A\in {\mathbb R}_+$ such that $f_m:=Ah_m$ satisfies (\[checkM\]).
Before stating the main result in this section we define $$\kappa_*:=\frac{|W_*|}{\sqrt{1+W_*^2+L_+/R_0^2}}(1-\frac{2M}{R_0})e^{-\frac{M}{R_0(1-\frac{2M}{R_0})}}.$$
\[Theorem2\] Assume that $r'_0,r_0,r_1,M_{\mathrm{in}},M, L_+, R_{0}, R_1, W_*, W_{-}$ and $\open{f}$ are given as above, and consider a solution $f$ of the system (\[ee1\])-(\[ee4\]), launched by $\open{f},$ on its maximal existence interval $[0,T[$. Then $T=\infty$, and $$\label{supp-as}
\supp f_m(t)
\subset [R_0+|t\,\kappa_*|, \infty[\times [W_*, \infty[\times [0, L_+],\,$$ and the resulting spacetime is future causally geodesically complete.
*Remark 2:* The initial data in the theorem is adapted to match the initial data constructed in [@AKR2]. However, it is important to note that the presence of the steady state given by $f_s$ is not needed if the only aim is to construct initial data for proving global existence and geodesic completeness. Hence, by taking $f_s=0$ and disregard the parameters $r'_0,r_0$ and $M_{\mathrm{in}}$ and thus consider a simpler class of initial, then Theorem 2 can be directly compared with the result in [@AKR1], and it can be seen to be an improvement of this result.
*Proof:* We first notice by following the arguments in [@AKR2], that the only way the matter in the outer region $r\geq R_0$ can affect the static solution is via the metric function $\mu$. By dropping the time derivatives in the Vlasov equation we see that in the remaining equation the factor $e^{\mu-\lambda}$ can be canceled. Hence, the static Einstein-Vlasov system is formulated in terms of the quantities $f,\,\lambda$ and $\mu_r$ and not $\mu$ itself. Therefore $f,\,\lambda$ and $\mu_r$ remain time independent for $r\leq r_0$. The arguments of the proof will therefore mainly concern the outer matter given by $f_m$. In particular we will see that the outer matter which initially is moving outwards will continue to move outwards, and therefore there is no direct interaction with the steady state. However, in the last part of the proof which concerns causal geodesic completeness the steady state will have an influence.
Let $[0,t_1[$ be the maximal time interval such that for $t\in [0,t_1[$ and $(r,w,L)\in \supp f_m(t)$, $w>W_*$. By continuity $t_1>0$. Suppose that $t_1\in ]0,T[$, then we must have $w=W_{*}$ for some $w\in \supp f_m(t_1)$, but we will show that $w>W_{*}$ for all $w\in \supp f_m(t_1)$. Thus $t_1=T$ and since the matter stays strictly away from $r=0$ it follows that $T=\infty$ in view of [@A2] or [@RRS].
Consider a characteristic $(R(s),W(s),L)$ with $R(0)\in [R_0,R_1]$ and let $$G(t):=E(R(t),W(t),L)+W(t).$$ Note that $G>0$. We have by (\[char1\]) and (\[char2\]) that $$\begin{aligned}
\frac{dG(s)}{ds}&=&-\left[\lambda_t(s,R(s))\frac{W(s)}{E(R(s),W(s),L)}+\mu_r(s,R(s)) e^{(\mu-\lambda)(s,R(s))}\right]G(s)\nonumber\\
& &+\frac{Le^{(\mu-\lambda)(s,R(s))}}{R(s)^3E(R(s),W(s),L)}.\label{charG}\end{aligned}$$ Below we will often suppress the arguments but it should be clear that $R=R(s),\,\mu_r=\mu_r(s,R(s))$ etc. The quantity $G$, which was first introduced in [@A2], is not suitable for the purpose here and the quantity that we will consider is $G(t)e^{\hat{\mu}(t,R(t))}(1-2M/R(t))$. We have $$\begin{aligned}
\label{ddsobject}
\frac{d}{ds}(Ge^{\hat{\mu}}(1-\frac{2M}{R}))&=&-\left[\lambda_t\frac{W}{E}+\mu_r e^{\mu-\lambda}\right]
Ge^{\hat{\mu}}(1-\frac{2M}{R})\nonumber\\
& &+\frac{Le^{\mu-\lambda}}{R^3E}e^{\hat{\mu}}(1-\frac{2M}{R})
+[\hat{\mu}_r\frac{W}{E}e^{\mu-\lambda}+\hat{\mu}_t]Ge^{\hat{\mu}}(1-\frac{2M}{R})\nonumber\\
& &+\frac{2M}{R^2}\frac{W}{E}e^{\mu-\lambda}Ge^{\hat{\mu}}\nonumber\\
&=&-\left[\lambda_t\frac{W}{E}+\check{\mu}_r e^{\mu-\lambda}-\hat{\mu}_t\right]
Ge^{\hat{\mu}}(1-\frac{2M}{R})\nonumber\\
& &+\frac{Le^{\mu-\lambda}}{R^3E}e^{\hat{\mu}}(1-\frac{2M}{R})
-\frac{(1+\frac{L}{R^2})}{E}\frac{m}{R^2}e^{\mu+\lambda}e^{\hat{\mu}}(1-\frac{2M}{R})\nonumber\\
& &+\frac{2M}{R^2}\frac{W}{E}e^{\mu-\lambda}Ge^{\hat{\mu}}.\end{aligned}$$ Here we used that $$\hat{\mu}_r\frac{W}{E}e^{\mu-\lambda}G-\hat{\mu}_r e^{\mu-\lambda}G=-\frac{(1+\frac{L}{R^2})}{E}\hat{\mu}_re^{\mu-\lambda}
=-\frac{(1+\frac{L}{R^2})}{E}\frac{m}{R^2}e^{\mu+\lambda}.$$ Consider the second last term in (\[ddsobject\]). In view of (\[e2-lambda\]) we have $$1-\frac{2M}{R(t)}\leq e^{-2\lambda(t,R(t))},$$ and we get $$-\frac{(1+\frac{L}{R^2})}{E}\frac{m}{R^2}e^{\mu+\lambda}e^{\hat{\mu}}(1-\frac{2M}{R})
\geq -\frac{(1+\frac{L}{R^2})}{E}\frac{m}{R^2}e^{\mu-\lambda}e^{\hat{\mu}}.$$ Now, since $W\geq W_{*}$ on $[0,t_1],$ and since $R(t)\geq R_0$, we get in view of (\[Lplusc\]) that $$\begin{aligned}
&-&\frac{(1+\frac{L}{R^2})}{E}\frac{m}{R^2}e^{\mu-\lambda}e^{\hat{\mu}}
+\frac{2M}{R^2}\frac{W}{E}e^{\mu-\lambda}Ge^{\hat{\mu}}\nonumber\\
& &\geq \Big(-\frac{(1+\frac{L}{R^2})}{E}
+\frac{2WG}{E}\Big)\frac{m}{R^2}e^{\mu-\lambda}e^{\hat{\mu}}\geq 0.\end{aligned}$$ Hence we have $$\label{ddsclean}
\frac{d}{ds}\Big(Ge^{\hat{\mu}}(1-\frac{2M}{R})\Big)\geq
-\left[\lambda_t\frac{W}{E}+\check{\mu}_r e^{\mu-\lambda}-\hat{\mu}_t\right]
Ge^{\hat{\mu}}(1-\frac{2M}{R}).$$ This implies that $$\begin{aligned}
\displaystyle& &G(t_1)e^{\hat{\mu}(t_1,R(t_1))}(1-\frac{2M}{R(t_1)})\nonumber\\
\displaystyle& &\geq e^{-\int_{0}^{t_1}
\left[\lambda_t(s,R(s))\frac{W}{E}
+\check{\mu}_r(s,R(s))e^{(\mu-\lambda)(s,R(s))}-\hat{\mu}_t(s,R(s))\right]ds}
G(0)e^{\hat{\mu}(0,R(0))}(1-\frac{2M}{R(0)}).\nonumber\\
\label{Gineqmain}\end{aligned}$$ Let $\gamma$ be the curve $$\gamma:=\{(t,r):0\leq t\leq t_1,\, r=R(t)\}.$$ The time integral in (\[Gineqmain\]) can be written as $$\int_{\gamma}e^{(-\mu+\lambda)(t,r)}\lambda_{t}(t,r)\,dr+
\Big(e^{(\mu-\lambda)(t,r)}\check{\mu}_{r}(t,r)-\hat{\mu}_t(t,r)\Big)\,dt.\label{Hcurveweak}$$ We will apply Green’s formula in the plane to this curve integral. Let $R_{\infty}\geq R_1+t_1$, so that $f(\cdot,R_{\infty},\cdot)=0$ for $t\in [0,t_1]$. Let $\Gamma$ be the closed curve $$\Gamma=\gamma+C_{t_1}+C_{\infty}+C_0,$$ where $$C_{t_1}:=\{(t,r):t=t_1,\;R(t_1)\leq r\leq R_{\infty}\},$$ $$C_{0}:=\{(t,r):t=0,\;R(0)\leq r\leq R_{\infty}\},$$ and $$C_{\infty}:=\{(t,r):t_1\geq t\geq 0,\; r=R_{\infty}\}.$$ We have $$\begin{aligned}
&\displaystyle\oint_{\Gamma} e^{-\mu+\lambda}\lambda_{t}\,dr+
(e^{\mu-\lambda}\check{\mu}_{r}-\hat{\mu}_t)\,dt&\nonumber\\
&\displaystyle =\int\int_{\Omega}\partial_{t}\left(e^{-\mu+\lambda}
\lambda_{t}\right)-\partial_{r}
\left( e^{\mu-\lambda}\check{\mu}_{r}-\hat{\mu}_t\right)drdt&\nonumber\\
&\displaystyle =\int\int_{\Omega}\partial_{t}\left(e^{-\mu+\lambda}
\lambda_{t}\right)-\partial_{r}
\left( e^{\mu-\lambda}\mu_{r}-\hat{\mu}_t\right)drdt
+\int\int_{\Omega}\partial_{r}
\left(e^{\mu+\lambda}\frac{m}{r^{2}}\right)drdt.&\nonumber\\\end{aligned}$$ By using (\[ee4\]) and (\[dtmuhat\]) this identity can be written $$\begin{aligned}
& &\phantom{GH} \displaystyle\oint_{\Gamma} e^{-\mu+\lambda}\lambda_{t}\,dr+
(e^{\mu-\lambda}\check{\mu}_{r}-\hat{\mu}_t)\,dt\nonumber\\
& &\displaystyle =\int\int_{\Omega}e^{\mu+\lambda}\left(\frac{2m}{r^{3}}
-4\pi(\rho-p)-8\pi p_T-4\pi e^{2\lambda}j\right)\,drdt\nonumber\\
& &\phantom{G} \displaystyle +\int\int_{\Omega}e^{\mu+\lambda}\left((\mu_r+\lambda_r)\frac{m}{r^2}+4\pi\rho-\frac{2m}{r^3}\right)drdt\nonumber\\
& &\displaystyle=\int\int_{\Omega}4\pi e^{\mu+\lambda}\left[
(\rho+p)e^{2\lambda}\frac{m}{r}+p-2p_T-je^{2\lambda}\right] drdt.\nonumber\\\end{aligned}$$ Here we used that $\mu_r+\lambda_r=4\pi r(\rho+p)e^{2\lambda}$. Since $w\geq W_{*}>0$ on $[0,t_1]$ we have in view of (\[p\]) and (\[j\]) that $j>p$, and we have $$p(1+e^{2\lambda}\frac{m}{r})=pe^{2\lambda}(1-\frac{2m}{r}+\frac{m}{r})\leq pe^{2\lambda}.$$ Hence, by dropping the term involving $p_T$ due to sign, we thus obtain $$\begin{aligned}
\label{rectangle}
& &\oint_{\Gamma} e^{-\mu+\lambda}\lambda_{t}\,dr+
(e^{\mu-\lambda}\check{\mu}_{r}-\hat{\mu}_t)\,dt\leq \int\int_{\Omega}4\pi e^{\mu+\lambda}
\rho e^{2\lambda}\frac{m}{r}\, drdt\nonumber\\
& &\leq \int\int_{\Lambda}4\pi e^{\mu+\lambda}
\rho e^{2\lambda}\frac{m}{r}drdt.\end{aligned}$$ Here $\Lambda=\{(t,r):0\leq t\leq t_1,\; R_{0}\leq r < \infty\},$ and the last inequality follows since the integrand is nonnegative and $\Omega\subset\Lambda$. Next we estimate $\rho$ in terms of $j$. Since for $w\in \supp f_m(t),\, 0\leq t\leq t_1$, we have $w\geq W_{*}\geq 1+\sqrt{L_+}/R_0$, we get for $r\geq R_0$, $$\begin{aligned}
\rho(t,r)
&\leq&
\frac{\pi}{r^2}\int_{-\infty}^{\infty}\int_0^{\infty} f\,dL\,dw
+ \frac{\pi}{r^2}\int_{-\infty}^{\infty}\int_0^{\infty} w f\,dL\,dw\nonumber\\
&&
{}+
\frac{\pi}{r^2}\int_{-\infty}^{\infty}\int_0^{\infty} \frac{\sqrt{L}}{r} f\,dL\,dw\nonumber\\
&\leq&
3\,\frac{\pi}{r^2}\int_{-\infty}^{\infty}\int_0^{\infty} w f\,dL\,dw =3\,j(t,r).\label{rho3j}\end{aligned}$$ We estimate the right hand side in (\[rectangle\]) by using the Vlasov equation from which it follows that $$\label{matteridentity}
\frac{\partial}{\partial t}\left(r^{2}e^{2\lambda}\rho(t,r)\right)=
-\frac{\partial}{\partial r}\left(r^{2}e^{\mu+\lambda}j\right)-re^{\mu+\lambda}2je^{2\lambda}
\frac{m}{r}.$$ Since $j(t,R_{0})=0$ and $j(t,R_{\infty})=0$, this gives $$\begin{aligned}
\int_{R_{0}}^{R_{\infty}}r^{2}e^{2\lambda(t_1,r)}\rho(t_1,r)dr&-&\int_{R_{0}}^{R_{\infty}}r^{2}e^{2\lambda(0,r)}\rho(0,r)dr\\
&=& -\int\int_{\Lambda}re^{\mu+\lambda}2je^{2\lambda}
\frac{m}{r}drdt.\end{aligned}$$ Thus, we get $$\label{Lambdaest}
\int\int_{\Lambda}re^{\mu+\lambda}2je^{2\lambda}
\frac{m}{r}drdt\leq \int_{R_{0}}^{\infty}r^{2}e^{2\lambda}\rho(0,r)dr\leq\frac{1}{4\pi}\frac{M}{1-\frac{2M}{R_{0}}},$$ where the last inequality follows in view of (\[e2-lambda\]) and (\[adm\]). Using inequality (\[rho3j\]) we therefore have $$\begin{aligned}
4\pi\int\int_{\Lambda}e^{\mu+\lambda}\rho e^{2\lambda}
\frac{m}{r}drdt&\leq& 4\pi\int\int_{\Lambda}e^{\mu+\lambda}3je^{2\lambda}
\frac{m}{r}drdt\nonumber\\
&\leq& 4\pi\frac{1}{R_{0}} \int\int_{\Lambda}re^{\mu+\lambda}3je^{2\lambda}
\frac{m}{r}drdt\nonumber\\
&\leq&\frac32\frac{M}{R_{0}(1-\frac{2M}{R_{0}})}.\label{domainest}\end{aligned}$$ In order to obtain an estimate for (\[Hcurveweak\]) it remains to estimate the boundary terms since $$\int_{\gamma}...=\oint_{\Gamma}...\,-\,\int_{C_{t_1}}...\,-\,\int_{C_{\infty}}...\,-\,\int_{C_0}...$$ First we notice that the curve integral along $C_{\infty}$ vanishes since both $p$ and $j$, which determine $\check{\mu}_r$ and $\hat{\mu}_t$, vanishes for $r=R_{\infty}$. Since $j\geq 0$ we have that $\lambda_t\leq 0$, which implies that the integral along $C_0$ can be dropped due to sign noticing the orientation of $C_0$. The term along $C_{t_1}$ can be estimated as follows. $$\begin{aligned}
& &\Big|\int_{C_{t_1}}e^{(-\mu+\lambda)(t,r)}\lambda_{t}(t,r)dr+
\Big(e^{(\mu-\lambda)(t,r)}\check{\mu}_{r}(t,r)-\hat{\mu}_t(t,r)\Big)dt\Big|\nonumber\\
& &=\Big|\int_{R(t_1)}^{R_{\infty}}e^{(-\mu+\lambda)(t_1,r)}\lambda_{t}(t_1,r)\,dr\Big|\leq
\int_{R_0}^{\infty}4\pi r\,e^{2\lambda}|j(t_1,r)|\,dr\nonumber\\
& &\leq \frac{M}{R_{0}(1-\frac{2M}{R_{0}})}.\label{jbdryest}\end{aligned}$$ In the first inequality we used (\[ee3\]) and in the second we used (\[e2-lambda\]), the fact that $|j|\leq\rho$, and (\[adm\]). We have thus obtained the estimate $$\int_{\gamma} e^{-\mu+\lambda}\lambda_{t}\,dr+
(e^{\mu-\lambda}\check{\mu}_{r}-\hat{\mu}_t)\,ds\leq \frac{5M}{2R_{0}(1-\frac{2M}{R_{0}})}.$$ Inserting this into the main inequality we get $$\label{ineqt1}
G(t_1)e^{\hat{\mu}(t_1,R(t_1))}(1-\frac{2M}{R(t_1)})
\geq e^{\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}}
G(0)e^{\hat{\mu}(0,R(0))}(1-\frac{2M}{R(0)}).\nonumber$$ Noticing that $\hat{\mu}$ is monotone in $r$ and nonpositive, and that $R(0)\geq R_0$, we obtain the inequality $$\begin{aligned}
G(t_1)
&\geq& e^{\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}}
G(0)e^{\hat{\mu}(0,R_{0})}(1-\frac{2M}{R_{0}})\nonumber\\
&\geq& e^{\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}}
G(0)\sqrt{\frac{R_{0}-2M}{R_{0}}}(1-\frac{2M}{R_{0}}).\label{ineqGt1}\end{aligned}$$ Here we made use of the estimate $$\label{hatmuest}
\hat{\mu}(t,R_0)
\geq -\int_{R_0}^{\infty}\frac{M\,d\eta}{\eta^2(1-\frac{2M}{\eta})}=\frac12\log{\Big(1-\frac{2M}{R_0}\Big)}.$$ We have that $G(0)>W(0)\geq W_-$, and in view of (\[Lplusc\]) we also have $3W(t)\geq G(t)$ on $[0,t_1]$. We now use the condition (\[mainc\]) and obtain $$3W(t_1)\geq G(t_1)> 3W_{*}.\nonumber$$ Thus $W(t_1)>W_{*}$, and necessarily we have $t_1=T$. As was pointed out in the beginning of the proof, since matter stay strictly away from the centre of symmetry, $T=\infty$, cf. [@A2] or [@RRS]. Let us next show that (\[supp-as\]) holds. From the characteristic equation (\[char1\]) we can conclude that $R(t)\geq R_0+|\kappa_* t|$ in view of the estimates $$e^{-\lambda(t,R(t))}=\Big(1-\frac{2m(t,R(t))}{R(t)}\Big)^{1/2}\geq \Big(1-\frac{2M}{R_0}\Big)^{1/2},$$ and $\mu(t,R(t))=\hat{\mu}(t,R(t))+\check{\mu}(t,R(t))\geq\hat{\mu}(t,R_0)+\check{\mu}(t,R_0)$ where $\hat{\mu}(t,R_0)$ is estimated by (\[hatmuest\]) and $$\check{\mu}(t,R_0)=-\int_{R_0}^{\infty}4\pi\eta p e^{2\lambda(t,\eta)}\,d\eta \geq -\frac{M}{R_0(1-\frac{2M}{R_0})}.$$ The latter inequality is analogous to (\[jbdryest\]). We remark that the estimate $R(t)\geq R_0+|\kappa_* t|$ is rough and can be improved by using arguments from [@AR1]. In order to complete the proof of the theorem we have to show that any causal geodesic is complete. We follow the argument in [@AKR1] and introduce the coordinates $$x^0 = t,\ x^1 = r\sin\theta\cos\phi,\ x^2 = r\sin\theta\sin\phi,\
x^3 = r \cos\theta.$$ In these coordinates the metric becomes $$g_{00} = -e^{2\mu},\ g_{0a} = 0,\
g_{ab} = \delta_{ab} + (e^{2\lambda}-1)\frac{x_a x_b}{r^2},$$ where Latin indices $a,b$ run from $1$ to $3$ and $x_a = \delta_{ab}x^b$. Let us now consider an arbitrary future directed, time-like or null geodesic, i.e., a solution $(x^\gamma (\tau),p^\gamma(\tau))$ of the geodesic equations $$\begin{aligned}
\frac{dx^{\gamma}}{d\tau} & = & p^{\gamma}, \\
\frac{dp^{\gamma}}{d\tau} & = & - \Gamma^{\gamma}_{\delta\epsilon}\,p^{\delta}p^\epsilon,\end{aligned}$$ where Greek indices $\gamma,\delta,\epsilon$ run from $0$ to $3$, $\Gamma^{\gamma}_{\delta\epsilon}$ are the Christoffel symbols, and $$p^0 >0,\ g_{\gamma\delta} p^\gamma p^\delta = - m^2 \leq 0.$$ We notice that before we had $m=1$ which means that the particles in our system have rest mass $1$, but for causal geodesic completeness we need to consider any $m\geq 0$. Such a geodesic exists on a maximally extended interval $[0,\tau_+[$, and future geodesic completeness means that $\tau_+=\infty$ for all such geodesics.
The following relations between the variables $r$, $w$, $L$, and $p^\gamma$ hold: $$\begin{aligned}
E
&=& e^{\mu} \, p^0,\\
w
&=& \frac{x_a p^a}{r}\,e^{\lambda},\\
\frac{L}{r^2}
&=&
\delta_{ab} p^a p^b - \left(\frac{x_a p^a}{r}\right)^2,\end{aligned}$$ where we now re-define $$E=E(r,w,L)=\sqrt{m^2+w^{2}+L/r^{2}}.$$ Since $dt/d\tau = p^0 > 0$ we can re-parameterize the geodesic by coordinate time $t\in [0,t_+[$. We remark that since there is no matter in the region $r\leq r'_0$ the arguments in [@A2; @RRS], which apply to any causal geodesic, imply that $t_+=\infty$. This will nevertheless be shown below. The arguments in [@A2; @RRS] are however not sufficient to conclude that $\tau_+=\infty$ since $E$ and $p^0$ may grow in time. We now show that $E$ and $p^0$, in the present situation, are bounded for any causal geodesic which then implies that $\tau_+=\infty$.
We call the domains $[0,r'_0]$, $[r'_0,R_0]$, and $[R_0,\infty[$ the inner vacuum region, the steady state region, and the outer matter region respectively.
Our strategy is to first show that a geodesic with sufficiently large $E$, for which $R\leq R_0$, will at a later time travel outwards for all times. We will then apply results from [@A2].
*Remark 3:* The analysis given below rests in part on the results in the previous part of the proof that rapidly outgoing characteristics will continue to move outwards. However, since we have good control of the metric and the matter in the domain $R(t)\leq R_0+\kappa_* t$ there are alternative approaches to the one given below to obtain a bound on $E$.
Consider a causal geodesic $(R(t),W(t),L)$. We now re-define the quantities $W_*, W_-$ and $\kappa_*$ to adapt these to the causal geodesic we now consider. Let $$W^g_*:=\max\{1+\frac{\sqrt{L_{+}}}{R_{0}},m+\frac{\sqrt{L}}{{r'}_{0}}\},$$ and let $W^g_->0$ satisfy $$W^g_-\,e^{\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}}
(1-\frac{2M}{R_{0}})^{3/2}\geq 3 |W^g_{*}|.$$ Define $$\kappa^g_*:=\frac{|W^g_*|}{\sqrt{1+{W^g_*}^2+L/{r'}_0^2}}(1-\frac{2M}{R_0})e^{-\frac{M}{R_0(1-\frac{2M}{R_0})}}.$$ We have that along a geodesic $$\label{Etdot}
\frac{dE}{dt}=-\Big[\lambda_t\frac{W}{E}+\mu_r e^{\mu-\lambda}\Big]W,$$ $$\label{Wtdot}
\frac{dW}{dt}=-\Big[\lambda_t W+\mu_r e^{\mu-\lambda}E-e^{\mu-\lambda}\frac{L}{r^3E}\Big],$$ and $$\label{Rtdot}
\frac{dR}{dt}=\frac{W}{E}e^{(\mu-\lambda)(t,R(t))}.$$ To control solutions of equation (\[Rtdot\]) we need to estimate $e^{\mu-\lambda}$. We have in view of (\[moverrss\]), $$\begin{aligned}
\label{hatmuest2}
\hat{\mu}(t,r'_0)&\geq&-\int_{r'_0}^{r_0}\frac{m(t,r)\,dr}{r^2(1-\frac{2m(t,r)}{r})}
-\int_{r_0}^{R_0}\frac{M_{\mathrm{in}}\,dr}{r^2(1-\frac{2M_{\mathrm{in}}}{r})}-\int_{R_0}^{\infty}\frac{M\,dr}{r^2(1-\frac{2M}{r})}\nonumber\\
&\geq&-4\log{\frac{r_0}{r'_0}}+\frac12\log{\Big(1-\frac{2M_{\mathrm{in}}}{r_0}\Big)}+\frac12\log{\Big(1-\frac{2M}{R_0}\Big)}.\nonumber \end{aligned}$$ Moreover, since the steady state is a given regular solution we have that $p$ is uniformly bounded on $r\in [r'_0,r_0]$. This gives $$\begin{aligned}
\check{\mu}(t,r)&=&-\int_{r'_0}^{r_0}4\pi r\,e^{2\lambda}p\,dr
-\int_{r_0}^{R_{0}}4\pi r\,e^{2\lambda}p\,dr-\int_{R_{0}}^{\infty}4\pi r\,e^{2\lambda}p\,dr\nonumber
\\
&\geq&-C-\frac{M_{\mathrm{in}}}{r_{0}(1-\frac{2M_{\mathrm{in}}}{r_{0}})}-\frac{M}{R_{0}(1-\frac{2M}{R_{0}})},\label{jbdryest2}\end{aligned}$$ in analogy with estimate (\[jbdryest\]) and where (\[moverrss\]) was used to bound $e^{2\lambda}$ on $[r'_0,r_0]$. Since $e^{-\lambda}=\sqrt{1-2m(t,r)/r}$ is strictly positive we have that for some constant $C_0>0$ $$\label{eml}
e^{(\mu-\lambda)(t,r)}\geq C_0 \mbox{ for }r\in [0,\infty[.$$ Define $$P:=\frac{C_0}{\sqrt{m^2+1+L/{r'}_0^2}}.$$ It follows that a geodesic $(R,W,L)$ with $|W|\geq 1$ and $R\geq r'_0$ satisfies $$\label{speedo}
\Big|\frac{dR}{dt}\Big|\geq P.$$ Since the aim is to show that $E$ is bounded we can assume that for some $t_0\in [0,t_+[$ $$\label{t0}
E(t_0)\geq Y_1e^{Y_2}+m+\frac{\sqrt{L}}{r'_0},$$ where $Y_1\geq W^g_-$ and where $Y_2\geq 0$ will be specified below. Note that this condition implies that $$|W(t_0)|\geq Y_1e^{Y_2}.$$ We will consider the cases $R(t_0)\geq R_0$; $r'_0\leq R(t_0)\leq R_0$; and $R(t_0)<r'_0$, and the subcases that $W(t_0)$ is positive or negative. Consider the first case with $W(t_0)$ positive. In this case we can directly refer to the arguments in the previous part of the proof, using the adapted quantities introduced above, noticing that $G(t_0)\geq E(t_0)\geq W^g_-$, to conclude that the geodesic will travel outwards for all times with $R(t)\geq R_0+\kappa^g_*\,t$. If instead $W(t_0)$ is negative we can assume without loss of generality that the geodesic will have $R(t)=R_0$ at some time $t$. This follows in view of (\[speedo\]) since the geodesic must otherwise have $|W(t)|\leq 1$ at some time $t$ but then $E(t)$ is bounded by $\sqrt{m^2+1+L/R_0^2}$ and the argument can be restarted at some later time $t_0$ for which (\[t0\]) holds. Hence we are in the situation of the second case which we will treat below.
Before continuing with the remaining cases we need some auxiliary results. The right hand side of (\[Wtdot\]) depends locally on $r$ except for the function $\mu$. Now $\mu\leq 0$ and since the steady state situated in $r'_0\leq r\leq r_0$ is a given regular static solution, it follows that there is a constant $\Gamma>0$, such that for $|W|\geq 1$ we have $$\label{GammaW}
\Big|\frac{dW}{dt}\Big|\leq\Gamma|W|,\;\mbox{for }r\in [r'_0,R_0].$$ Let $$q:=\frac{\Gamma(R_0-r'_0)}{P}.$$ We choose $Y_2$ such that $Y_2\geq 2q+1.$ We return to the remaining cases. We note that in view of (\[Etdot\]), $E$ is constant in the inner vacuum region. Hence, in the case where $R(t_0)\leq r'_0$, we can without loss of generality assume that $W(t_0)$ is positive and $R(t_0)=r'_0$, since if $W<0$ the geodesic will continue inwards for some time, with constant $E$, reach a turning point where $W=0$ if $L>0$, and then it will travel outwards, with positive $W$, and it cannot change sign until it hits the steady state region. We remark that if $L=0$ the geodesic will pass through the origin and the “turning point” will be $r=0$ where the sign of $W$ changes. Assume now that $R(t)\leq R_0$ on the time interval $I:=[t_0,t_0+\Delta t]$ where $\Delta t=(R_0-r'_0)/P$. We then have in view of (\[Wtdot\]) that $$\log{W(t)}\geq \log{Y_1}+Y_2-\Gamma\Delta t\geq \log{Y_1}+q+1,$$ for $t\in I$. This implies in particular, since $W(t)\geq 1$, that $$\frac{|W(t)|\,e^{\mu-\lambda}}{\sqrt{m^2+W(t)^2+L/R(t)^2}}\geq P$$ on $I$. In view of (\[Rtdot\]) the geodesic has thus necessarily crossed the region $r'_0\leq r\leq R_0$ within the time interval $\Delta t$. Hence $R(t)=R_0$ with $W(t)\geq W^g_-$ at some $t\in I$. We can then repeat the arguments in the first case to conclude that the geodesic continues to travel outwards. We turn to the case where $R(t_0)\in [r'_0,R_0]$. In the case that $W(t_0)$ is positive the argument from the previous case applies. If $W(t_0)$ is negative we have by the same argument that on the time interval $I:=[t_0,t_0+\Delta t]$ the maximum change for $\log{|W|}$ is $q$ so that $R(t)=r'_0$ at some $t$ with $\log{|W(t)|}\geq \log{Y_1}+q+1$. Thus we are back in the case where the geodesic is in the interior vacuum region and the geodesic will turn and have $R(t)=r'_0$ at some later time $t$ with reversed sign on $W(t)$ so that $\log{W(t)}\geq Y_1+q+1$ and we can repeat the previous argument.
To conclude we have reduced the situation so that we only need to consider geodesics that satisfy $R(t)\geq R_0+\kappa^g_* (t-t_2)$ for $t\geq t_2$ for some $t_2\geq 0$. In the case of a timelike geodesic associated to a particle upper bounds on $G=E+W$ and $H:=E-W$, and thus on $E$, are obtained in [@A2] under the assumption that there is no matter in the domain $\{(t,r):r\leq \epsilon\}$. These bounds are however time dependent. The arguments in [@A2] do not depend on the rest mass $m$ and are thus unchanged for a causal geodesic. If we apply these arguments in the present situation, where the geodesic satisfies $R(t)\geq R_0+\kappa^g_* (t-t_2)$, the bounds (4.18) and (4.19) in [@A2] can in fact be seen to be time independent. Indeed, we need to reconsider the estimates in [@A2] leading to these bounds. For simplicity we put $t_2=0$. Let us here only consider the bound for $G$ since the bound for $H$ follows analogously. We thus reconsider the estimates of (4.14), (4.15) and (4.17) in [@A2]. For the inequality (4.14) in [@A2] we now get $$\big|\int_{C_t}4\pi rje^{2\lambda}dr\big|\leq \frac{1}{R_0+|\kappa^g_* t|}\frac{1}{\Big(1-\frac{2M}{R_0+|\kappa^g_* t|}\Big)}\int_0^{\infty}
4\pi r^2\rho\, dr\leq C(M,R_0).$$ Since $$\frac{L}{E}\leq R\sqrt{L},$$ the bound (4.15) in [@A2] is replaced by $$\frac{Le^{\mu-\lambda}}{R^3E}\leq \frac{\sqrt{L}}{(R_0+|\kappa^g_* t|)^2}.$$ Finally, for the term (4.17) in [@A2] we get $$\int\int_{\Omega} 8\pi(\rho-p)e^{\mu+\lambda}\,dtdr\leq\int_0^t\int_{R_0+|\kappa^g_* s|}^{\infty} \frac{8\pi r^2\rho}{(R_0+|\kappa^g_* s|)^2}\,drds\leq C(M,R_0).$$ These estimates turn (4.18) in [@A2] into the time independent bound $$G(t)\leq C(M,R_0).$$ By an analogous argument it follows that also $H(t)\leq C(M,R_0)$, and therefore $E(t)\leq C(M,R_0)$. This shows that $E$ is bounded for any causal geodesic. Now, since $\mu$ is bounded from below it follows that $p^0$ and $p^a$ are bounded and thus $t_+=\infty$, and since $dt/d\tau=p^0$ we have that $\tau_+=\infty$, and the proof of the theorem is complete.
$\Box$
For the proof of Corollary \[Corollary1\] a slightly different set of initial data is needed. The reason is that there is no result in the literature which says that there are steady states of the Einstein-Vlasov system for which $m/r$ is arbitrarily small everywhere. By numerical simulations,using the code developed in the work [@AR3], we find nevertheless evidence that this is true. It is on the other hand known that there are steady states for which $2m/r$ can be arbitrarily close to $8/9$, cf. [@A4] and [@A3].
The set up below is similar to our previous set up with the difference that the inner matter which is supported in $[r'_0,r_0]$ is not given by a steady state.
Let $0<r'_0<r_0<r_1$ be given and put $M=r_1/2$. Let $L_+>0$ and let $M_{\mathrm{out}}<M$ be such that $$\label{compactin}
\frac{M-M_{\mathrm{out}}}{r'_0}<\frac12.$$ Put $M_{\mathrm{in}}:=M-M_{\mathrm{out}}$. Let $R_1>r_1$ be such that $$\label{mediumstrip}
R_1-r_1<\frac{r_1-r_0}{6},$$ and define $$R_0:=\frac{1}{2}(r_1+R_1).$$ Let $W_{*}>0$ satisfy $$\label{Lplusc2}
|W_{*}|\geq 1+\frac{\sqrt{L_{+}}}{r'_{0}}.$$ Let $\open{f}=\open{f}_{\mathrm{in}}+\open{f}_{\mathrm{out}}$ be such that $$\label{cn1}
\supp{\open{f}_{\mathrm{in}}}\subset [r'_0,r_0]\times ]W_{1},\infty]\times [0,L_+],$$ $$\label{cn2}
\supp{\open{f}_{\mathrm{out}}}\subset [R_0,R_1]\times ]W_2,\infty]\times [0,L_+],$$ where $W_1>0$ and $W_2>0$ satisfy $$\label{mainc1}
|W_{1}|\,e^{-\Big(\frac{5M_{\mathrm{in}}}{2r'_{0}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})}+\frac{5M}{2R_0(1-\frac{2M}{R_0})}\Big)}
(1-\frac{2M_{\mathrm{in}}}{r'_{0}})^{3/2}(1-\frac{2M}{R_0})^{1/2}\geq 3 |W_{2}|,$$ $$\label{mainc2}
|W_{2}|\,e^{-\Big(\frac{3M_{\mathrm{in}}}{2r'_0(1-\frac{2M_{\mathrm{in}}}{r'_0})}+\frac{-5M}{2R_{0}(1-\frac{2M}{R_{0}})}\Big)}
(1-\frac{2M}{R_{0}})^{3/2}\geq 3 |W_{*}|,$$ and such that $$\label{MinMout}
M_{\mathrm{in}}=\int_{r'_0}^{r_0}4\pi r^2\open{\rho}_{\mathrm{in}}\,dr,\;\;
M_{\mathrm{out}}=\int_{R_0}^{R_1}4\pi r^2\open{\rho}_{\mathrm{out}}\,dr.$$ Define $$\kappa'_{*}=\frac{|W_*|}{\sqrt{1+W_*^2+L_+/{r'}_0^2}}A^{3/2}B,$$ where $$A:=\min\{1-\frac{2M_{\mathrm{in}}}{r_0'},1-\frac{2M}{R_0}\},$$ and $$B:=e^{-\Big(\frac{M_{\mathrm{in}}}{r'_0(1-\frac{2M_{\mathrm{in}}}{r'_0})}+
\frac{M}{R_0(1-\frac{2M}{R_0})}\Big)}.$$
\[Theorem3\] Assume that $r_0', r_0, r_1, R_{0}, R_1, L_+, M_{\mathrm{out}}, W_*, W_{1}, W_{2}$ and $\open{f}$ are given as above, and consider a solution $f$ of the system (\[ee1\])-(\[ee4\]), launched by $\open{f},$ on its maximal existence interval $[0,T[$. Then $T=\infty$, and $$\label{supp-as2}
\supp f(t)\subset [r'_0+|t\,\kappa'_*|, \infty[\times
[W_*, \infty[\times [0, L_+],\,$$ and the resulting spacetime is future causally geodesically complete.
*Proof:* The proof follows to a large extent the previous proof and the set up is identical. For a characteristic originating from $[R_0,R_1]$, corresponding to the outer matter, we consider the same quantity $$\label{originalquantity}
G(t)e^{\hat{\mu}(t,R(t))}(1-\frac{2M}{R(t)})$$ as above. We follow the steps identically until we reach inequality (\[rectangle\]) where we replace $\Lambda$ by $\Lambda':=\{(t,r):0\leq t\leq t_1, r'_0\leq r\leq \infty\}$. The reason for this modification is that there is no matter at $r=r'_0$ and this fact guarantees that the boundary terms which result from the identity (\[matteridentity\]) are zero as before. The estimate (\[Lambdaest\]) is now replaced by $$\begin{aligned}
\int\int_{\Lambda'}re^{\mu+\lambda}2je^{2\lambda}\frac{m}{r}\,drdt &\leq&\int_{r'_0}^{\infty}r^2e^{2\lambda}\rho(0,r)\, dr\nonumber\\
& &\leq \int_{r'_0}^{R_0}r^2e^{2\lambda}\rho(0,r)\, dr+\int_{R_0}^{\infty}r^2e^{2\lambda}\rho(0,r)\, dr\nonumber\\
& & \leq \frac{1}{4\pi}\frac{M_\mathrm{in}}{1-\frac{2M_{\mathrm{in}}}{r'_0}}+\frac{1}{4\pi}\frac{M}{1-\frac{2M}{R_0}}.\label{primedomainest}\end{aligned}$$ Using this estimte in (\[domainest\]) we get $$\begin{aligned}
4\pi\int\int_{\Lambda}e^{\mu+\lambda}\rho e^{2\lambda}
\frac{m}{r}drdt&\leq& 4\pi\int\int_{\Lambda}e^{\mu+\lambda}3je^{2\lambda}
\frac{m}{r}drdt\\
&\leq& 4\pi\frac{1}{R_{0}} \int\int_{\Lambda}re^{\mu+\lambda}3je^{2\lambda}
\frac{m}{r}drdt\nonumber\\
&\leq&\frac32\frac{M_{\mathrm{in}}}{r'_{0}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})}+\frac32\frac{M}{R_{0}(1-\frac{2M}{R_{0}})},\label{set}\end{aligned}$$ in place of (\[domainest\]). The remaining estimates are unchanged and the condition (\[mainc2\]) then ensures that the characteristic we consider satisfies $|W(t_1)|>|W_*|$.
For a characteristic originating from the interval $[r'_0,r_0]$ we define the time $t_{R_0}$ such that if the characteristic reaches $r=R_0$ in the time interval $[0,t_1]$ then this happens at $t=t_{R_0}$. Note that on the time interval $[0,t_1]$ all particles move outwards so the characteristic can only cross $r=R_0$ once. On $[0,t_{R_0}]$ we consider the quantity $$\label{innerquantity}
G(t)e^{\hat{\mu}(t,R(t))}(1-\frac{2M_{\mathrm{in}}}{R(t)})$$ instead of (\[originalquantity\]). The influence from the outer matter can only enter through the metric coefficient $\mu$ and the actual value of $\mu$ plays no role in the proof. We can thus follow the steps in the proof above replacing $t_1$ with $t_{R_0}$, $M$ with $M_{\mathrm{in}}$ and $R_0$ with $r'_0$ until we reach inequality (\[rectangle\]) where we replace $\Lambda$ by $\Lambda':=\{(t,r):0\leq t\leq t_{R_0}, r'_0\leq r\leq \infty\}$. We then again use the estimate (\[set\]) to replace (\[domainest\]). The estimate (\[jbdryest\]) of the boundary term is slightly changed in this case since the domain of integration is now $[r'_0,\infty]$. The estimate (\[jbdryest\]) now becomes $$\begin{aligned}
& &\Big|\int_{C_{t_{R_0}}}e^{(-\mu+\lambda)(t,r)}\lambda_{t}(t,r)dr+
\Big(e^{(\mu-\lambda)(t,r)}\check{\mu}_{r}(t,r)-\hat{\mu}_t(t,r)\Big)dt\Big|\nonumber\\
& &\leq\int_{r'_0}^{R_{0}}4\pi r\,e^{2\lambda}|j(t_{R_0},r)|\,dr+\int_{R_{0}}^{\infty}4\pi r\,e^{2\lambda}|j(t_{R_0},r)|\,dr\nonumber
\\
& &\leq \frac{M_{\mathrm{in}}}{r'_{0}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})}+\frac{M}{R_{0}(1-\frac{2M}{R_{0}})}.\label{jbdryest2}\end{aligned}$$ Similarly the estimate (\[hatmuest\]) also gives two terms by the same splitting, cf. (\[hatmuest\]). Hence (\[ineqt1\]) becomes $$\begin{aligned}
\label{ineqtR0}
G(t_{R_0})e^{\hat{\mu}(t_{R_0},R(t_{R_0}))}(1-\frac{2M_{\mathrm{in}}}{R(t_{R_0})})
&\geq& e^{-\Big(\frac{5M_{\mathrm{in}}}{2r'_{0}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})}+\frac{5M}{2R_0(1-\frac{2M}{R_0})}\Big)}\nonumber\\
& &\times G(0)e^{\hat{\mu}(0,R(0))}(1-\frac{2M_{\mathrm{in}}}{R(0)}),\nonumber\end{aligned}$$ and inequality (\[ineqGt1\]) is then replaced by $$\begin{aligned}
G(t_{R_0})
&\geq& e^{-\Big(\frac{5M_{\mathrm{in}}}{2r'_{0}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})}+\frac{5M}{2R_0(1-\frac{2M}{R_0})}\Big)}\nonumber\\
& &\times G(0)(1-\frac{2M}{R_0})^{1/2}(1-\frac{2M_{\mathrm{in}}}{r'_{0}})^{3/2}.\nonumber\end{aligned}$$ Here the estimate (\[hatmuest\]) was modified by splitting the integration domain in the subintervals $[r'_0,R_0]$ and $[R_0,\infty]$ as above which yields $$\label{hatmuest2}
\hat{\mu}(0,r'_0)
\geq\frac12\log{\Big(1-\frac{2M_{\mathrm{in}}}{r'_0}\Big)}+\frac12\log{\Big(1-\frac{2M}{R_0}\Big)}.$$ We use condition (\[mainc1\]) and obtain $$W(t_{R_0})>W_{2},$$ by using the arguments in the previous case.
On the remaining time interval $]t_{R_0},t_1]$ we again consider the quantity $G(t)e^{\hat{\mu}(t,R(t))}(1-2M/R(t))$ as for a characteristic originating from $[R_0,R_1]$ and repeat the arguments in that situation. Note here that $R_1$ is in this case replaced by $R_1+t_{R_0}$ but this has no influence on the argument. Thus the first statement of Theorem \[Theorem3\] holds for this class of initial data. We next consider statement (\[supp-as2\]). If we let $$A:=\min\{1-\frac{2M_{\mathrm{in}}}{r'_0},1-\frac{2M}{R_0}\},$$ we have that $e^{-\lambda(t,r)}\geq A^{1/2}$. We also have $$\label{muhatinandout}
e^{\hat{\mu}(t,r)}\geq (1-\frac{2M_{\mathrm{in}}}{r'_0})^{1/2}(1-\frac{2M}{R_0})^{1/2},$$ in view of inequality (\[hatmuest2\]). Similarly we get $$e^{\check{\mu}(t,r)} \geq e^{-\Big(\frac{M_{\mathrm{in}}}{r'_0(1-\frac{2M_{\mathrm{in}}}{r'_0})}+\frac{M}{R_0(1-\frac{2M}{R_0})}\Big)}=:B,$$ analogously to (\[jbdryest2\]). Thus if we define $$\kappa'_*:=\frac{|W_*|}{\sqrt{1+W_{*}^{2}}+L_+/{r'}_{0}^{2}}A^{3/2}B,$$ it follows that (\[supp-as2\]) holds. The proof of causal geodesic completeness follows the steps in the previous proof but is easier since in this case there is no steady state present.
$\Box$
Proof of Theorem 1
==================
In this section we prove Theorem \[Theorem1\], Corollary \[Corollary1\] and Corollary \[Corollary2\]. The results of Theorem \[Theorem2\] and Theorem \[Theorem3\] are time reversible in the following sense: by taking initial data as specified in Theorem \[Theorem2\] or in Theorem \[Theorem3\] but with reversed momenta, disregarding the steady state, so that the outgoing particles are ingoing, then global existence *to the past* holds and (\[supp-as\]) and (\[supp-as2\]) become $$\label{supp-aspast}
\supp f(t)\subset [a_1+|a_2 t|, \infty[\,\times\, ]-\infty,W_*]\,\times\, [0, L_+],$$ where $a_1$ and $a_2$ are equal to $R_0$ and $\kappa_*$ respectively $r'_0$ and $\kappa'_*$, and spacetime is past causally geodesically complete. We denote such initial data by ${\cal I}_s$ and ${\cal I}_r$ respectively. We show that there is a sub class of ${\cal I}_s$ and of ${\cal I}_r$ which satisfy the conditions in [@AKR2]. These conditions guarantee that black holes form to the future, cf. also [@AKR3] where an additional argument is given to match the definition of a black hole in [@Cu1]. We recall the set up and the conditions on the initial data in [@AKR2]. There are two slightly different initial data sets in [@AKR2] which both guarantee the formation of black holes and for our purpose any of these will do. Let us here give the details of Case (i) on p.688 in [@AKR2].
Let $0<r_0<r_1$ be given, put $M=r_1/2$, and fix $0<M_\mathrm{out}<M$ such that $$\label{icnts}
\frac{2(M-M_\mathrm{out})}{r_0}<\frac{8}{9}.$$ Let $R_1>r_1$ be such that $$\label{mediumstrip}
R_1-r_1<\frac{r_1-r_0}{6},$$ and define $$R_0:=\frac{1}{2}(r_1+R_1).$$ Denote by $\open{\rho}\,$ the energy density induced by the initial distribution function $\open{f}$. It is required that all the matter in the outer region $[r_0, \infty[$ is initially located in the strip $[R_0,R_1]$, with $M_\mathrm{out}$ being the corresponding fraction of the ADM mass $M$, i.e., $$\label{checkM}
\int_{r_0}^{\infty}4\pi r^2\open{\rho}(r)dr = \int_{R_0}^{R_1}4\pi r^2\open{\rho}(r)dr=M_\mathrm{out}.$$ Furthermore, the remaining fraction $M_{\mathrm{in}}=M-M_\mathrm{out}$ should be initially located within the ball of area radius $r_0$, i.e., $$\label{M-checkM}
\int_{0}^{r_0}4\pi r^2\open{\rho}(r)dr=M_\mathrm{in}.$$ If the inner matter is chosen to be a steady state then the solution exists for all $r\in [0,\infty[$ and for all times but generally this is not required and we define the set $$\label{ddef}
D:=\{(t,r) \in [0,\infty[^2 \mid r \geq \gamma^+(t)\},$$ where $\gamma^+$ is an outgoing radial null geodesic originating from $r=r_0>0$, i.e., $$\label{gamma+}
\frac{d \gamma^+}{ds}(s)=e^{(\mu-\lambda)(s,\gamma^+(s))},\;\gamma^+(0)=r_0.$$ In this set up we use the following expression for the Hawking mass $$m(t,r)=M-\int_r^{\infty}4\pi \eta^2\rho(t,\eta)\,d\eta,$$ instead of (\[m\]). The analysis in [@AKR2] is restricted to $D$ since the dynamics of the inner matter is not essential to conclude that a black hole forms. We define $$\label{Gammadef}
\Gamma(r_1,R_1) := \sqrt{\frac{R_1-r_1}{R_1+r_1}},$$ and we require that the parameter $W_-<0$ satisfies $$\label{Wminuscondition}
\Gamma(r_1,R_1)^2|W_-|^2\geq\frac{10}{d},$$ where $$\label{dcond}
d=\min\left\{\frac12,\frac{r_0}{12 R_1},\frac{r_1-r_0}{300 R_1}\right\}.$$ We impose the [**General support condition:**]{} For all $(r,w,L) \in \supp \fn\,$ the following holds: $$r \in ]0,r_0] \cup [R_0,R_1],$$ and if $r\in [R_0,R_1]$ then $w \leq W_-$ and also $$\label{hypoL}
0< L\leq \frac{3L}{\eta}\,\open{m}(\eta) +\eta\,\open{m}(\eta),\ \eta\in [r_0,R_1].$$ One of the two classes of initial data specified in [@AKR2], which guarantee the formation of black holes, can now be given. Let $$\begin{aligned}
\label{I1def}
{\cal{I}}_B := \Bigl\{ \fn
&\mid&
\fn \ \mbox{is regular, satisfies (\ref{checkM}), (\ref{M-checkM}),
the general support condition,}\nonumber \\
&&
\mbox{and for}\ (r,w,L)\in \supp \fn\ \mbox{with}\ r\in [R_0,R_1],
\sqrt{L}/r_0 \leq 1
\Bigr\}.
\ \label{condL1t1}\end{aligned}$$ Corollary 2.3 in [@AKR2] shows that if the inner matter is a steady state then the solution exist for all Schwarzschild time. We denote the subset of ${\cal I}_B$ for which the inner matter is a steady state by ${\cal I}_{B}^s$. The initial data sets ${\cal I}_s$ and ${\cal I}_r$ satisfy to a large extent the conditions given above and we are now in a position to define the subsets ${\cal J}_s\subset {\cal I}_s$ and ${\cal J}_r\subset {\cal I}_r$ which satisfy the claims of Theorem \[Theorem1\]. Let $$\label{L+}
L_+=\min\{r_0^2,r_0 M_{\mathrm{in}}\}.$$ The set ${\cal J}_s$ is defined by $$\label{Jdef}
{\cal{J}}_s := \Big\{ \fn\in {\cal I}_{s}\mid L_+\; \mbox{satisfies } (\ref{L+})\; \mbox{and } W_-\; \mbox{satisfies}\, (\ref{Wminuscondition})\Big\}.\nonumber$$ Note that there is no conflict between the conditions (\[mainc\]) and (\[Wminuscondition\]) by taking $|W|_-$ sufficiently large. It is clear that ${\cal J}_s\subset {\cal I}_{B}^s$. Next we consider the initial data set where both the inner and outer matter move rapidly inwards. We define $$\label{J2}
{\cal J}_r:=\Big\{\open{f}=\open{f}_{\mathrm{in}}+\open{f}_{\mathrm{out}}\in {\cal I}_r\mid
L_+\, \mbox{satisfies } (\ref{L+})\, \mbox{and }W_2\; \mbox{satisfies } (\ref{Wminuscondition})\Big\}.\nonumber$$ Since this set is a subset of ${\cal I}_B$ the proof of the theorem is complete.
$\Box$
*Proof of Corollary 1:* Let $\epsilon>0$ be given and choose initial data in ${\cal J}_r.$ By evolving the data to the past we have by (\[supp-as2\]) that the matter is supported in the domain $\{(t,r):r\geq r'_0+|\kappa'_* t|\}$. Thus, by evolving to time $t=-T,$ where $T$ is sufficiently large so that $$\sup_{r}\frac{m(t,r)}{r}\leq \frac{M}{r'_0+|\kappa'_*|T}\leq \epsilon,$$ the claim follows by taking as initial data the solution at $t=-T$.
$\Box$
*Proof of Corollary 2:* Consider initial data in ${\cal J}_s$. The statement then follows by combining Corollary 2.3 in [@AKR2] and Theorem \[Theorem2\].
$\Box$
[**Acknowledgment:**]{} The author is grateful for discussions with Gerhard Rein and Alan Rendall.
[AAAA]{} , The Einstein-Vlasov System/Kinetic Theory. [*Living Rev. Relativity*]{} [**14**]{}, (2011), 4.
, Regularity results for the spherically symmetric Einstein-Vlasov system. [*Ann. Henri Poincaré*]{} [**11**]{}, 781–803 (2010).
, Sharp bounds on $2m/r$ of general spherically symmetric static objects, [*J. Differential Equations*]{}[**245**]{}, 2243–2266 (2008).
, On static shells and the Buchdahl inequality for the spherically symmetric Einstein-Vlasov system. [*Comm. Math. Phys.*]{} [**274**]{}, 409-425 (2007).
, Global existence for the spherically symmetric Einstein-Vlasov system with outgoing matter. [*Comm. Partial Differential Eqns.*]{}, 656–668 (2008).
, The formation of black holes in spherically symmetric gravitational collapse. [*Math. Ann.* ]{} [**350**]{}, 683–705 (2011).
, Gravitational collapse and the formation of black holes for the spherically symmetric Einstein-Vlasov system. [*Quart. Appl. Math.*]{}, [**68**]{}, 17–42 (2010).
, Existence of axially symmetric static solutions of the Einstein-Vlasov system. [*Commun. Math. Phys.* ]{} [**308**]{}, 23–47 (2011).
, The asymptotic behaviour in Schwarzschild time of Vlasov matter in spherically symmetric gravitational collapse. [*Math. Proc. Camb. Phil. Soc.,*]{} [**149**]{}, 173–188 (2010).
, A numerical investigation of the stability of steady states and critical phenomena for the spherically symmetric Einstein-Vlasov system. [*Class. Quantum Grav.* ]{} [**23**]{}, 3659–3677 (2006).
, On the steady states of the spherically symmetric Einstein-Vlasov system. [*Class. Quantum Grav.* ]{} [**24**]{}, 1809–1832 (2007).
, [*Galactic Dynamics*]{}, Princeton University Press 1987.
On the global initial value problem and the issue of singularities. [*Class. Quantum Grav.,* ]{} [**16**]{}, A23–A35 (1999).
Bounded variation solutions of the spherically symmetric Einstein-scalar field equations. [*Comm. Pure Appl. Math.,* ]{} [**46**]{}, 1131–1220 (1993).
The formation of black holes and singularities in spherically symmetric gravitational collapse. [*Comm. Pure Appl. Math.* ]{} [**44**]{}, 339-373 (1991).
Princeton Mathematical Series [**41**]{}. Princeton University Press 1993.
Black hole formation from a complete regular past. [*Commun. Math. Phys.* ]{}[**289**]{}, 579–596 (2009).
, Critical phenomena at the threshold of black hole formation for collisionless matter in spherical symmetry. [*Phys. Rev. D*]{}, [**65**]{}, 024007, 1–10, (2001).
, On continued gravitational contraction, [*Phys. Rev.*]{} [**56**]{}, 455–459 (1939).
The Vlasov-Einstein system with surface symmetry. [*Habilitationsschrift,* ]{}Munich, (1995).
Collisionless Kinetic Equations from Astrophysics – The Vlasov–Poisson System, in Dafermos, C.M. and Feireisl, E., eds., [*Handbook of Differential Equations: Evolutionary Equations, Vol. 3*]{}, pp. 383–476, (Elsevier/North-Holland, Amsterdam, 2006).
Global existence of solutions of the spherically symmetric Vlasov-Einstein system with small initial data. [*Commun. Math. Phys.,* ]{} [**150**]{}, 561–583 (1992). Erratum: [*Commun. Math. Phys.,* ]{} [**176**]{}, 475–478 (1996).
A regularity theorem for solutions of the spherically symmetric Vlasov-Einstein system. [*Commun. Math. Phys.* ]{} [**168**]{}, 467–478 (1995).
Critical collapse of collisionless matter: A numerical investigation. [*Phys. Rev. D*]{}, [**58**]{}, 044007, 1–8, (1998).
An introduction to the Einstein-Vlasov system. [*Banach Center Publ.,* ]{} [**41**]{}, 35–68 (1997).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Open source software projects evolve thanks to a group of volunteers that help in their development. Thus, the success of these projects depends on their ability to attract (and keep) developers. We believe the openness of a project, i.e., how easy is for a new user to actively contribute to it, can help to make a project more attractive. To explore the openness of a software project, we propose three metrics focused on: (1) the distribution of the project community, (2) the rate of acceptance of external contributions and (3) the time it takes to become an official collaborator of the project. We have adapted and applied these metrics to a subset of GitHub projects, thus giving some practical findings on their openness.'
author:
- |
Valerio Cosentino, Javier Luis Cánovas Izquierdo and Jordi Cabot\
\
bibliography:
- 'msrChallenge.bib'
date: 17 January 2014
title: Three Metrics to Explore the Openness of GitHub projects
---
Introduction {#sec:introduction}
============
Data and Approach {#sec:dataset}
=================
Community Composition {#sec:op1}
=====================
External Contribution Analysis {#sec:op2}
==============================
Time to Become Collaborator {#sec:op3}
===========================
Related Work {#sec:relatedWork}
============
Conclusion and Future Work {#sec:conclusion}
==========================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Systems containing few Fermions (e.g., electrons) are of great current interest. Fluorescence occurs when electrons drop from one level to another without changing spin. Only electron gases in a state of equilibrium are considered. When the system may exchange electrons with a large reservoir, the electron-gas fluorescence is easily obtained from the well-known Fermi-Dirac distribution. But this is not so when the number of electrons in the system is prevented from varying, as is the case for isolated systems and for systems that are in thermal contact with electrical insulators such as diamond. Our accurate expressions rest on the assumption that single-electron energy levels are evenly spaced, and that energy coupling and spin coupling between electrons are small. These assumptions are shown to be realistic for many systems. Fluorescence from short, nearly isolated, quantum wires is predicted to drop abruptly in the visible, a result not predicted by the Fermi-Dirac distribution. Our exact formulas are based on restricted and unrestricted partitions of integers. The method is considerably simpler than the ones proposed earlier, which are based on second quantization and contour integration.'
address:
- 'Mas Liron, F30440 Saint Martial, France'
- 'Centre d’Électronique et de Micro-optoélectronique de Montpellier [@UMR], Université Montpellier II, F34095 Montpellier, France'
- 'Département de Mathématiques et Informatique Appliquées, Université Paul Valéry, F34199 Montpellier, France'
author:
- Jacques Arnaud
- Laurent Chusseau
- 'Fabrice Philippe [@also]'
title: 'Fluorescence from few electrons[^1]'
---
Introduction
============
A number of remarkable experiments involving few electrons in semiconductors and free-space, metal particles, and spin-1/2 atoms at low temperatures, have been recently reported [@Zory], [@Pell]. Only electrons are considered below. These collections of electrons may be isolated or in thermal contact with the environment, but, in any event, the number of particles is constant. The Fermi-Dirac (FD) distribution holds when electrons may be freely exchanged with a large reservoir (grand-canonical ensemble), but is inaccurate for the systems considered. The present paper provides simple and accurate formulas for electron occupancy and fluorescence for evenly-spaced single-electron energy levels. Spontaneous emission is supposed to be weak enough not to perturb importantly the system state of equilibrium [@fai]. Only vanishingly small Coulomb interaction between electrons [@den] is considered. The time required for the system to reach equilibrium is not needed because averagings are performed over unlimited time scales. Quantum Optics effects, such as resonance fluorescence or super-radiance, will not be considered.
The assumption of evenly-spaced single-electron energy levels is not as restrictive as one may think at first. Consider indeed one-dimensional devices such as the quantum wires employed in modern laser diodes [@Zory]. If the wire is uniform over its length and the valence and conduction bands are parabolic in the energy-momentum space, the energy spacing, $\epsilon$, between adjacent levels is not a constant. However, the variations of $\epsilon$ may be neglected near the Fermi level as long as the temperature is not too high. For zero-band-gap semiconductors such as Pb$_{0.84}$Sn$_{0.16}$Te, the energy-momentum relationship is linear rather than parabolic, and $\epsilon$ is exactly a constant. Level spacings in small irregular metal particles (with a size on the order of $10~$nm) are nearly uniform as a consequence of the mechanism of level “repulsion”. The probability that adjacent levels be separated by $\epsilon$ is, for the appropriate ensemble, of the form: $\epsilon ^{4} \exp(-\epsilon
^{2})$, a sharply peaked function of $\epsilon$ [@den]. As is well-known, the Landau levels that describe electron motion in uniform magnetic fields [@Pell] are evenly spaced. These levels are highly degenerate, but the coupling between degenerate states (expressing the drift of electrons through the magnetic field lines) may be neglected over some period of time. Likewise, two or three-dimensional harmonic oscillators, modeling for example the confinement of electrons in traps, exhibit degenerate evenly-spaced levels. Our approach may be generalized to degenerate levels. It is appropriate to mention also that the density of states of (two-dimensional) quantum wells with parabolic bands is a constant within a sub-band. This implies that the energy-level spacing is constant on the average, though not exactly.
The amount of light spontaneously emitted by electronic systems depends on the optical-mode density of state, which is different for free-space, low-dimensional structures, or photonic band gap materials. It is not the purpose of the present paper to discuss such problems. Because all comparisons are made at the same optical wavelength, terms depending on wavelength only (essentially the optical-mode density of state) may be dropped. The quantity that we calculate is the probability that the system exhibits an electron at level $k$ and a hole (no electron) at level $k'$, with $k-k'=\hbar
\omega/\epsilon \equiv{d}$, where $\hbar$ denotes the Planck constant divided by $2\pi$ and $\omega$ the angular optical frequency of observation. In pure semiconductors, electron-momentum conservation entails that transitions may occur only symmetrically with respect to the Fermi level, implying that $k = (1+d)/2$, $k' = (1-d)/2$, with odd $d$, if $k=0$ labels the zero-temperature top electron. We will particularly consider the case where this condition holds.
Let us recall that, in Statistical Mechanics, isolated systems are described by microcanonical ensembles, systems that may exchange energy but not particles with a reservoir are described by canonical ensembles, and systems that may exchange both energy and particles with a reservoir are described by grand-canonical ensembles. The Fermi-Dirac (FD) distribution [@Kittel] is applicable to finite systems only in the latter case. That is, the FD distribution is invalid for isolated systems and for systems that are in contact with an electrical insulator such as diamond. In the present context, “finite” (or “small”, or “mesoscopic”) means that $k_{B}T$, where $T$ denotes the absolute electron gas temperature, is not necessarily large compared with the adjacent-level energy spacings $\epsilon$. In grand canonical ensembles, the fluorescence is proportional to the *product* of the probabilities that the upper level be occupied and that the lower level be empty. In canonical ensembles, it turns out that fluorescence is proportional to the *difference* between lower and upper levels occupancies. This is apparently a new result.
Exact formulas for level occupancy in finite single-spin systems with evenly-spaced levels in contact with a heat bath have been reported before [@den], [@AJP], [@Fab1], [@Schon], [@Fab2]. Our method [@AJP] consists of first enumerating the microstates of isolated systems. Subsequent averaging provides expressions for the canonical occupancies. This method is considerably simpler than the second-quantization methods and integral-transformation formulas employed in [@den], [@Schon]. The present paper generalizes the results reported in [@AJP] to account for the fact that some electrons may change spin in the course of time. Simple, exact formulas for fluorescence are obtained for isolated systems and canonical ensembles. For arbitrary level energies and the canonical ensemble, known recurrence formulas [@Borr] are satisfied by our more special, but explicit, forms.
The FD distribution is derived in Section 2 to set up the notation and to recall why, in grand-canonical ensembles, spin is properly accounted for by a two-fold degeneracy. This method of accounting for the electron spin, however, is invalid in microcanonical or canonical ensembles because unpaired electrons may be in either one of two distinguishable states. Finite systems that exchange energy but not particles with a reservoir are considered in Section 3 and 4. In Section 3, the electron spin is supposed to be strictly preserved and, for simplicity, it is supposed that there are as many electrons with spin up and with spin down. Formulas valid for single-spin electrons need only be multiplied by factors of 2 in that situation. In Section 4, electrons are allowed to change spin in the course of time (but not during a spontaneous emission event). For the sake of clarity, only essential formulas relating to canonical ensembles are given in the main text, detailed derivations and intermediate results being relegated to Appendices A and B.
The purpose of Appendix \[sec.AppA\] is to explain why the total number $W$ of microstates of isolated systems is related to the partitions of integers, for single-spin-state electrons. It is shown that the number $m(k)$ of microstates whose level $k$ is occupied is simply related to $W$. The number $m(k;k')$ of microstates whose level $k$ is occupied and level $k'$ is empty is shown to be simply related to $m(k)$, and thus to $W$. Averaging, with the Boltzmann factor as a weight, provides the corresponding formulas in the canonical ensemble. Appendix \[sec.AppB\] explains how the possibility that electrons may change spin in the course of time is accounted for.
The Fermi-Dirac Distribution {#FD}
============================
The probability $p$ that a system in thermal *and* electrical contact with a large medium contains $N^+$ electrons with spin up, $N^-$ electrons with spin down, and energy $U$, is proportional to the corresponding number of medium states (subscripts $m$) conveniently written as $\exp \left( S_{m} ( N^+_{m}, N^-_{m}, U_{m} ) \right)$ with $S_{m}$ the system entropy. If the system-medium contact is very weak, energies as well as particle numbers add up. A first-order expansion of $S_{m}$ with respect to its arguments then gives $$p ( N^+ , N^- , U ) = C \exp(-\alpha N^+ -\alpha N^- -\beta U)
\label{e1}$$ where $C$ denotes a constant and $$\begin{aligned}
\beta & = & \frac{\partial S_{m}}{\partial U_{m}} \nonumber \\
\alpha & \equiv & - \beta \mu = \frac{\partial S_{m}}{\partial
N^+_{m}} = \frac{\partial S_{m}}{\partial N^-_{m}}\label{e2}\end{aligned}$$ Here, $\beta\equiv1/k_{B}T$, where $T$ denotes the temperature, and $\mu$ the Fermi level. A single $\mu$ value occurs because the medium behavior is the same for electrons of opposite spins.
It follows from Eq. (\[e1\]) that the probabilities $p^{(0)}$ that a nondegenerate state of energy $\epsilon_{k}$ is unoccupied, $p^{(1)}$ that it is occupied by an electron of either spin, and $p^{(2)}$ that it is occupied by two electrons, are respectively $$\begin{aligned}
p^{(0)} & = & C \nonumber \\
p^{(1)} & = & C z \nonumber \\
p^{(2)} & = & C z^2
\label{e3}\end{aligned}$$ with $z \equiv \exp ( - \alpha - \beta \epsilon_{k} )$. Normalization ($p^{(0)} + 2 p^{(1)} + p^{(2)} = 1$) gives $C=1/(1+z)^2$. The occupancy (average number of electrons) $n_{FD} = 2 p^{(1)} + 2
p^{(2)}$ of level $k$ is therefore $$\begin{aligned}
n_{FD} ( q ; k ) & = & \frac{2}{z^{-1} + 1} \nonumber \\
& = & \frac{2}{\exp ( \beta ( \epsilon_{k} - \mu ) ) + 1}
\nonumber \\
& = & \frac{2}{q^{\frac{1}{2} - k} + 1}
\label{e4}\end{aligned}$$ where $q \equiv \exp(-\beta)$ denotes the Boltzmann factor. In the last expression it is assumed that $\epsilon_{k} = k$, where $k$ denotes any relative integer. The separation $\epsilon$ between adjacent-level energies is taken as the energy unit, with a typical value for $1~\mu$m–long quantum wires $\epsilon = 1\;$meV. In the last expression in Eq. (\[e4\]), $k=1$ labels the level just above the top electron at $T = 0\;$K, and we have set $\mu = \frac{1}{2}$. Note that for large $k$–values, $n_{FD} ( q ; k ) \approx 2
q^{k-1/2}$.
The average system energy added on top of the $T = 0\;$K energy is obtained by summing up the occupancy over all levels, and subtracting a similar sum for the $T = 0\;$K distribution. The result is [@AJP] $$r_{FD} (q) =\sum_{j=1,3\ldots}\frac {j}{q^{-j/2}+1}\label
{rFD}$$
In grand canonical ensembles, occupancies at different levels are independent. For a single spin state, this means that the probability that level $k$ is occupied and level $k'$ is empty is the product of level $k$ occupancy and \[$1-$level $k'$ occupancy\]. When the two spin-states are considered, we obtain $$\label{e5}
L_{FD} ( q ; k, k') = 2\frac{ n_{FD}( q, k)}{2} \left( 1 -
\frac{n_{FD} ( q, k')}{2} \right)$$ Fluorescence may indeed occur for 8 out of the 16 possibilities of occupancy of levels $k$ and $k'$ (no electron, spin-up electron, spin-down electron, or two electrons for each of the two levels). Because occupancies are independent, the sum of the probabilities that fluorescence events occur is indeed found to be given by Eq.(\[e5\]).
If the electron-momentum conservation law is enforced, we have $k =
(1+d)/2$, $k' = (1-d)/2$, and the FD fluorescence in Eq. (\[e5\]) reads after rearranging $$L_{FD} ( q ; d) = \frac{2}{{\left( q^{-d/2} + 1 \right)}^2}
\label{e6}$$ Thus, the fluorescence in the grand-canonical ensemble is given by a simple function of temperature $T$ and angular optical frequency of observation $\omega$. We will see that the canonical ensemble fluorescence is given by a simple series.
Fluorescence without spin flip
==============================
For single-spin electrons, the occupancy in isolated systems is given by a simple formula reported in [@AJP]. The proof, omitted in [@AJP], is given in Appendix \[sec.AppA\] of the present paper (see Eq. (\[eq:ni\])). If the energy added to the system is denoted by $r$, the number $W(r)$ of configurations of the system is equal here to the number $p(r)$ of *partitions* of $r$. Indeed, microstates may be obtained by shifting electrons upward from their $T
= 0\;$K locations by non-increasing steps that sum up to $r$. Let us recall that a partition of $r$ is a set of positive integers summing up to $r$. For example $(2,1,1)$ is a partition of $4$. The number $p(4)$ of partitions of $4$ equals $5$. By convention, $p(0)=1$ and $p(r)=0$ if $r<0$.
To illustrate the difference existing between the exact result and the Fermi-Dirac distribution, let us note that, for any microstate, the energy separation between the top electron and the lowest hole cannot exceed $r\epsilon $, where $\epsilon$ denotes as before the adjacent-level energy spacing. Accordingly, the fluorescence drops abruptly at an angular optical frequency $\omega = r\epsilon /\hbar$. When $r\gg 1$, a system temperature $T$ may be defined [@AJP]: $k_{B}T\approx \epsilon \sqrt{6r}/\pi$. As an example, room-temperature isolated systems with $\epsilon=1$ meV should not emit visible light according to the exact formula, while some faint light is expected according to the FD distribution.
In the present section, it is supposed that the electron spins are preserved in the course of time, and that there are as many electrons with spin up and spin down, for simplicity. It then suffices to multiply the expressions for single-spin electron occupancy, average energy, and fluorescence, given in Appendix \[sec.AppA\], by factors of 2.
The occupancy reads, according to Eq. (\[eq:A12\]) $$\label{e8}
n_{u}(q;k) = - 2\sum_{j=1,2\ldots}(-1)^{j} q^{j k + j (j -
1)/2}$$ Note that, for large $k$–values, $n_{u}(q;k) \approx 2 q^k$ so that $n_{u}(q;k)/n_{FD}(q;k)=\sqrt{q}$, if the expression for $n_{FD}(q;k)$ in Eq. (\[e4\]) is used.
The average added energy reads [@AJP] $$\label{ru}
r_{u} (q)= 2 \sum_{j=1,2\ldots}\frac{j}{q^{-j}-1}$$
We first compare in Fig. \[fig:1\]a the occupancy in isolated systems with the FD occupancy. The former is obtained by multiplying $n_{i}(r;k)$ given in Eq. (\[eq:ni\]) by a factor of 2 to account for the two spin states. The FD occupancy is given in Eq.(\[e4\]), with $q$ expressed in terms of the average energy $r$ with the help of Eq. (\[rFD\]). Note that, below 0.1, isolated-system occupancies are *smaller* than FD occupancies.
Consider next the case where the system is in contact with a heat bath. The ratio of the exact occupancy in Eq. (\[e8\]) (where $q$ is expressed in terms of the average energy $r$ with the help of Eq.(\[ru\])) and the FD occupancy (Eqs. (\[e4\]), (\[rFD\])), is represented in Fig. \[fig:1\]b as a function of the FD occupancy for various values of the average energy $r$ (namely, $r=6$, 60, and 600). Below 0.1, the exact occupancy *exceeds* the FD occupancy. Figure \[fig:2\]a shows that, when the comparison is made at equal temperatures $T$ (instead of equal average energies), the opposite occurs. In Fig. \[fig:2\], we have chosen to represent the occupancy ratios as functions of the level number $k$ (referred to the Fermi level) instead of the FD occupancy.
To evaluate fluorescence, we need to know the number $m(r;k;k')$ of microstates of added energy $r$ having an electron at level $k$ and none at level $k'=k-d$, where $d \equiv \hbar \omega/\epsilon $. Appendix \[sec.AppA\] shows that this quantity is easily expressible in terms of the numbers $m(r;k)$ of microstates having an electron at level $k$ (irrespectively of other state occupancies). Averaging the result over $r$ with $q^r$ as a weight to account for energy fluctuations in the presence of a heat bath, the fluorescence is found to be (see Eq. (\[L\])) $$\label{e9}
L_{u}(q;k,k') = \frac{n_{u}(q;k') - n_{u}(q;k)}{q^{-d}-1}$$ where $n_{u}(q;k)$ is given in Eq. (\[e8\]).
If the law of electron-momentum conservation is enforced, the fluorescence reads $$\label{Lu}
L_{u}(q;d)= L_{u}(q;\frac{1+d}{2},\frac{1-d}{2})$$
Recall that $q \equiv \exp(- \epsilon/ k_{B} T)$ (where $\epsilon$ is typically $1\;$meV and room temperature corresponds to $k_{B}T =
26\;$meV), and $d \equiv \hbar \omega/\epsilon$. The fluorescence ratio: $L_{u}(q;d)/L_{FD}(q;d)$ according to Eq. (\[e6\]) and Eq.(\[Lu\]), is represented in Fig. \[fig:3\]a as a function of $d$ for different temperatures. It is interesting that canonical and grand-canonical fluorescences almost coincide at small wavelengths even though the occupancies are quite different in that limit.
Fluorescence with spin flip
===========================
Electron spins are now allowed to vary in the course of time (but not during a fluorescence event). The numbers $N^+$ and $N^-$ of electrons with spin up and spin down, respectively, may fluctuate, but their sum $N^+ + N^- \equiv N$ remains constant if the system is isolated or in contact with an electrically-insulating heat sink such as diamond. The occupancy and fluorescence for coupled-spin electrons is derived from previous expressions through a succession of partitionings and averagings. Because the details are lengthy, they are relegated to Appendix \[sec.AppB\]. Remarkably, many summations can be performed in closed form so that the final result is simple.
The occupancy reads $$n_{c}(q;k) = \frac{\sum_{j= - \infty}^{+ \infty}q^{j^2}
n_{u}(q;k+j)}{\sum_{j=- \infty}^{+ \infty}q^{j^2}}
\label{e10}$$ where $n_{u}(q;k)$ is given in Eq. (\[e8\]). Comparisons with the FD distributions are exemplified in Figs. \[fig:1\]d and \[fig:2\]b. With the help of theta functions [@Berndt] the average added energy may be written as a simple sum (see Eq.(\[eq:B5\])) $$\label{e11}
r_{c} (q) =2 \sum_{j=1,2\ldots}\left(
\frac{j}{q^{-j} -1} + \frac{(-1)^j j}{q^j - q^{-j}} \right)$$
The fluorescence reads (see Eqs. (\[A13\]) and (\[eq:B6\])) $$\label{e12}
L_{c}(q;d) = \frac{\sum_{j = -\infty}^{+\infty}q^{j^2}
L_{u}(q;j+\frac{1+d}{2},j+\frac{1-d}{2})} {\sum_{j=
-\infty}^{+\infty}q^{j^2}}$$ where $L_{u}(q;k,k')$ is given in Eq. (\[e9\]). Fluorescence is illustrated in Fig. \[fig:3\]b. Note that the exact result is closer to the FD result when electrons are allowed to change spin in the course of time.
Conclusion
==========
When a (possibly small) system is in thermal and electrical contact with a large medium such as a piece of copper, the average number of electrons occupying some energy level (occupancy) is twice the value given by Fermi-Dirac (FD) formula. The fluorescence (light spontaneously emitted without electron-spin flip), defined in terms of the probability that an electron at level $k$ may drop to level $k-\hbar\omega/\epsilon $ ($\omega$ denotes the angular optical frequency of observation), is also easily obtained. But when the system is isolated, or in thermal contact with an electrical insulator, electron occupancies are given by different expressions. Because modern electronics often employ short quantum wires supported by diamond heat sinks, it is important to have at our disposal precise expressions for occupancy and fluorescence in such situations. The expressions obtained in this paper were illustrated by comparison with the FD results. We considered the case where the electron spins are strictly maintained in the course of time (Section 3), and the case where spin-flip is allowed (Section 4). We found, for example, that small FD occupancies should be multiplied by approximately $\exp ( -
\epsilon / 2 k_{B} T ) \approx 0.22$ if $\epsilon = 1\;$meV and $T=4\;$K, a factor that differs significantly from unity. But, unexpectedly, the fluorescence turns out to be given rather accurately by the FD distribution.
Our mathematical approach is based on a direct enumeration of the microstates, and the results are expressed in terms of the number of partitions of integers. This method is considerably simpler than those previously reported for similar models, both conceptually and algebraically. A computer simulation has given results that are in very good agreement with the analytical formulas reported in this paper.
It is our intention to report in the future analytical and numerical results relating to mesoscopic laser-diode light fluctuations. A preliminary step consists of considering single-mode cavities incorporating the electron gas at thermal equilibrium, with one or two bands of states (for a single band, see Appendix \[sec.AppA\] of the present paper). The intraband Auger effect and the stimulated transitions may be introduced at that stage. Next, the probability that low-lying electrons be promoted to high-lying levels by the action of a (quiet or fluctuating) pump, and the probability that light quanta be absorbed, are introduced. At low power, our simulation gives output light fluctuations that agree very well with elementary laser-noise theory predictions [@chuss]. At high power, new effects (temperature fluctuations, spectral-hole burning, statistical fluctuations of the optical gain $\ldots$) occur that are difficult to handle analytically [@Arn]. The analytical formulas reported in this paper are helpfull to assess the accuracy of the simulation in special situations.
Occupancy and fluorescence for single-spin-state electrons. Arbitrary energy levels {#sec.AppA}
===================================================================================
We are only concerned in the main text with one-electron energy levels $\{ \epsilon_{k} \} = \mathbb{Z}$. Rigorous occupancy formulas are obtained in the present Appendix by considering first arbitrary $\{ \epsilon_{k} \}$. Eventually the number of levels is allowed to go to infinity.
Consider an isolated system whose nondegenerate one-electron level energies are, in increasing order, $\epsilon_{1}$, $\epsilon_{2}$, …$\epsilon_{k}$, …$\epsilon_{B}$, with $N \le B$ single-spin electrons. According to the Pauli principle each level may be occupied by only $0$ or $1$ electron. The system energy $U$ is therefore the sum of $N$ of the $\epsilon_{k}$. The purpose of this Appendix is to evaluate:
1. The number $W(N,U)$ of possible ways of obtaining some given $U$ (number of microstates, or “statistical weight”).
2. The number $m(N,U;k)$ of microstates whose level $k$ is occupied. The occupancy $n(N,U;k)$ of level $k$ is defined as $m(N,U;k)/W(N,U)$.
3. The number $m(N,U;k;k')$ of microstates whose level $k$ is occupied and level $k'$ is empty. The fluorescence $L(N,U;k;k')$ emitted by electrons dropping from level $k$ to level $k'$ is defined as $m(N,U;k;k')/W(N,U)$.
These evaluations will be presented in reversed order. Let us first relate the number $m(N,U;k;k')$ of microstates whose level $k$ is occupied and level $k'$ is empty, to the numbers $m(N,U;k)$ defined above. For each microstate, let the electron at level $k$ be transferred to the lower empty level $k'$. The number $N$ of electrons is unaffected but the total energy gets reduced from $U$ to $U-d$ where $d=\epsilon_{k}-\epsilon_{k'}>0$, and the roles of $k$ and $k'$ are reversed. The following equality $$m(N,U;k;k') = m(N,U-d;k';k)
\label{m3}$$ therefore holds. Now notice that, $$\begin{aligned}
\lefteqn{m(N,U;k) - m(N,U;k;k') =} \nonumber \\
& & m(N,U;k') - m(N,U;k';k)
\label{m4}\end{aligned}$$ because the two sides of the above equation count microstates whose levels $k$ and $k'$ are *both* occupied. When the expression in Eq. (\[m4\]) is introduced in Eq. (\[m3\]) iteration gives a recurrence relation for $m(N,U;k;k')$ that reads $$\begin{aligned}
\lefteqn{m(N,U;k;k') =} \nonumber \\
& & \sum_{j=1,2\ldots } \left( m(N,U-jd;k') - m(N,U-jd;k)
\right)
\label{m6}\end{aligned}$$ The above series terminates when the total energy vanishes, i.e., when $jd$ exceeds $U$.
Consider next the $m(N,U;k)$ microstates whose levels $k$ are occupied, and remove these electrons. The same number of microstates is obtained, with $N-1$ electrons, total energy $U-\epsilon_{k}$, and no electron at level $k$. The number of these new microstates may be written as the difference between the *total* number of microstates and the number of microstates whose level $k$ is *occupied*. We have therefore the identity $$\label{m}
m(N,U;k) = W(N-1,U-\epsilon_{k}) - m(N-1,U-\epsilon_{k};k)$$ After a sufficient number of iterations, either the energy or the number of electrons becomes negative and the last term vanishes. The quantity $m(N,U;k)$ may therefore be written as a finite sum $$\label{m1}
m(N,U;k) = - \sum_{j=1,2\ldots }(-1)^{j} W(N-j,U-j\epsilon _{k})$$ The series terminates when $j$ exceeds either $N$ or $U /
\epsilon_{k}$.
We have the obvious identity (the number of occupied states for the whole set of microstates being written in two different manners) $$N W(N,U)=\sum_{k \geq 1}m(N,U;k)
\label{A6}$$
A system in contact with a heat bath at temperature $T$ is described by the canonical ensemble. Let us define as in the main text $q
\equiv \exp ( - \beta )$, where $\beta \equiv 1 / k_{B} T$. The so-called partition function $Z(N,q)$ is the sum over $U$ of $q^{U}
W(N,U)$, and the average energy is $( q / Z ) \partial Z(N,q) /
\partial q$. When both sides of Eq. (\[A6\]) are multiplied by $q^U$ and summed over $U$ we obtain, using Eq. (\[m1\]) $$\begin{aligned}
Z(N,q) & = & \frac{1}{N} \sum_{k=1,2\ldots} \sum_{U}
q^{U} m(N,U;k) \nonumber \\
& = & - \frac{1}{N} \sum_{j=1,2\ldots} (-1)^j
\sum_{k=1,2\ldots} q^{j \epsilon_{k}} \nonumber \\
& & \times \sum_{U} q^{U - j \epsilon_{k}} W(N-j,U-j
\epsilon_{k}) \nonumber \\
& = & - \frac{1}{N} \sum_{j=1,2\ldots N} (-1)^j Z(1,q^j)
Z(N-j,q)
\label{eq:A7}\end{aligned}$$
Indeed, for a single electron ($N = 1$), $U$ may only take one of the $\epsilon_{k}$ values and the statistical weight $W$ is unity. It follows that $Z(1,q)$ is the sum over $k$ from 1 to $B$ of $q^{\epsilon_{k}}$. Note that $Z(0,q)=1$. Equation (\[eq:A7\]) was obtained earlier [@Borr2] from a less direct proof.
The occupancy (average number of electrons) $n(q;k)$ of level $k$, is equal to the sum over $U$ of $q^{U} m(N,U;k)$, divided by $Z(N,q)$, where $m(N,U;k)$ is given in Eq. (\[m1\]) and $Z(N,q)$ is defined in Eq. (\[eq:A7\]) from a recurrence relation. Thus $$\label{nu1}
n(q;k) = \frac{- \sum_{j=1,2\ldots}(-1)^{j}
q^{j\epsilon_{k}}Z(N-j,q)}{Z(N,q)}$$ Expression (\[nu1\]) was reported before [@Borr].
The probability that level $k$ be occupied and that level $k'$ be empty at temperature $T$ is similarly obtained by summing $q^{U}
m(N,U;k;k')$ over $U$, and dividing the result by $Z(N,q)$, where $m(N,U;k;k')$ is given in Eq. (\[m6\]). The result of the summation may be expressed as the *difference* of the lower and upper occupancies, according to $$\label{L}
L(q;k,k') = \frac{n(q;k') -
n(q;k)}{q^{\epsilon_{k'}-\epsilon_{k}}-1}$$ where the occupancy $n(q;k)$ is given in Eq. (\[nu1\]). Thus the fluorescence is equal to the difference between the occupancies at $k'$ and $k$ multiplied by the Bose function.
Let us now specialize the above formulas for the case where $\epsilon_{k}=k$, $k=1\dots B$. Considering the displacement of the $N$ electrons from their least-energy locations ($k=1$ to $N$) beginning to the one on top, we observe that $W(N,U)$ is the number $p(P;N,r)$ of partitions of the *added energy* $r \equiv U - N (
N+1 ) / 2$ into at most $N$ parts, none of them exceeding $P\equiv
B-N$. Note that the numbers $p(P;N,r)$ may be obtained from a recurrence relation [@Andrew], $p(P;N,r)-p(P;N-1,r)=p(P-1;N,r-N)$. Let us now change slightly our notation, letting $k=0$ denote the top electron in the least energy configuration, and let us employ the added energy $r$ instead of the total energy $U$ as an argument. Equation (\[m1\]) reads $$\begin{aligned}
\lefteqn{m(N,r;k) = - \sum_{j=1,2\ldots }(-1)^{j}} \nonumber \\
&& \times p ( P+j ; N-j , r-jk- j(j-1)/2 )
\label{A10}\end{aligned}$$
If $r$ does not exceed $N$ and $P$, it is intuitive that $p(P;N,r) =
p(r)$, where $p(r)$ denotes the number of unrestricted partitions of $r$. Equation (\[A10\]) then simplifies to $$\label{A11}
m(r;k) = - \sum_{j=1,2\ldots }(-1)^{j} p( r - j k -
j(j-1)/2 )$$ This expression was reported (for the first time to our knowledge), in [@AJP]. If $N$ and $P$ are infinite ($\{ \epsilon_{k} \} =
\mathbb{Z}$), Eq. (\[A11\]) holds for any finite value of $r$ and the corresponding single-spin-state occupancy of an isolated system is $$n_{i}(r;k) = m(r;k) / p(r)
\label{eq:ni}$$ Averaging the numerator and denominator of above expression with $q^r$ as a weight with $r$ from 0 to $\infty$, gives the canonical occupancy [@den] $$n(q;k) = - \sum_{j=1,2\ldots}(-1)^{j} q^{jk+j(j-1)/2}
\label{eq:A12}$$
Finally the expression in Eq. (\[L\]) simplifies in the present situation to $$\label{A13}
L(q;k,k') = \frac{n(q;k') - n(q;k)}{q^{k'-k}-1}$$
We have set up a Monte Carlo simulation program that enables us to recover previous analytical expressions. For the case of isolated systems, a constant probability per unit time is ascribed to level-changing events that preserve energy. The system eventually reaches a state of equilibrium with an electron distribution very close to the one derived from previous recurrence formulas. The Fermi-Dirac distribution is obtained in the limit of large $B$-values, with temperatures and Fermi levels that depend on the energy given initially to the system. Our computer program may handle single-electron level distributions that could be difficult to analyze theoretically (for example, two bands of states).
When the system is in thermal contact with a heat bath, electrons at level $k$ are ascribed a probability per unit time, $p$, of being promoted to level $k+1$ and a probability, $qp$, of being demoted to level $k-1$, provided these levels are empty. Strictly speaking, these prescriptions rest on an Einstein-type model of solids that supposes that the atoms are vibrating at frequency $\omega_{phonon} =
\epsilon /\hbar$, where $\epsilon$ denotes as before the electronic level spacing. But the details of the thermalization model turn out to be rather unimportant. The computer program enabled us to reproduce the theoretical results with great accuracy. For example, when $B=100$, $\epsilon=1$meV, and $T=$100K, the numerical distribution fits the Fermi-Dirac distribution with a discrepancy not exceeding 0.2%.
When the electron gas is enclosed in a single-mode cavity, the probability that the cavity contains $m$ light quanta is proportional to $W(r-md)$, where $W(r)$ denotes the statistical weight of the electron gas for an added energy $r$, see Appendix \[sec.AppB\] of [@AJP]. If, initially, only the highest levels are occupied, we obtain exactly, from the recurrence relation satisfied by $p(P;N,r)$, ${\mathrm{variance}}(m)/{\mathrm{average}}(m)=(B+1)/6$.
Occupancy and fluorescence for two-spin-states electrons {#sec.AppB}
========================================================
In the present Appendix we restrict ourselves to energy levels $\epsilon_{k} = k$, with $k=1,2\ldots$, the origin of the energy being set at $k=0$. Electrons are allowed to change spin in the course of time. We first consider an isolated system with constant numbers of spin-up and spin-down electrons ($N^{+}$ and $N^{-}$, respectively), electrons of different spins being allowed to exchange energy. Next, spin-flip is allowed. Averaging, with the Boltzmann factor as a weight, provides occupancies for the case where the systems are in contact with a heat bath. Occupancies in these various situations are illustrated in Fig. \[fig:1\], again by comparison with the FD distribution.
Consider first an isolated system with $N^{+}$ spin-up electrons and $N^{-}$ spin-down electrons, and suppose that the two sub-systems may exchange energy but that spin flip is not allowed. Setting $N^{+}+N^{-} \equiv 2 N$ and $N^{+}-N^{-} \equiv 2 n$, the system least energy is $$\begin{aligned}
U_{0} & = & \frac{N^{+} ( N^{+}+1 )}{2} + \frac{N^{-} ( N^{-}+1
)}{2} \nonumber \\
& = & N(N+1)+n^2
\label{eq:B1}\end{aligned}$$
If $r \ge n^2$ denotes the energy added to the system on top of $N(N+1)$, the remaining energy $s=r-n^2$ splits into $r_{1}$ in sub-system 1 and $r_{2}=s-r_{1}$ in sub-system 2.
We have shown in Appendix \[sec.AppA\] that, for the case presently considered and in the limit $N \to \infty$, the number of microstates for single-spin electrons is the number of partitions $p(r)$ of the excess energy $r$. The number of microstates relating to one particular splitting of $s$ is therefore $p(r_{1}) p(r_{2})$. Accordingly, occupancies are obtained by averaging single-spin state occupancies shifted by $\pm n$, with a probability law proportional to $p(r_{1}) p(s-r_{1})$ with $r_{1}$ running from 0 to $s$. $$\begin{aligned}
\lefteqn{n_{e}(r;k) =} \nonumber \\
&& \frac{\sum_{r_{1}} \left( m(r_{1};k-n) p(s-r_{1}) + m(s-r_{1};k+n)
p(r_{1}) \right)}{\sum_{r_{1}}p(s-r_{1}) p(r_{1})} \nonumber \\
&&
\label{eq:B2}\end{aligned}$$ where $m(r;k)$ is given by Eq. (\[A11\]). Figure \[fig:1\]c compares $n_{e}(r;k)$ in Eq. (\[eq:B2\]) to $n_{FD}$ in Eqs.(\[e4\]), (\[rFD\]) for the case where $n=0$ ($N^{+}=N^{-}$ or $r=s$) and various values of the added energy.
When spin flip is allowed, $N^{+}+N^{-}=2 N$ remains fixed, but $n
\equiv (N^{+}-N^{-}) / 2$ may take any value that does not exceed $\sqrt{r}$, where $r$ denotes as before the energy added on top of $N(N+1)$. If $r=6$ for example, five values of $n$ are permitted, namely $n = 0$, $n = \pm 1$ and $n = \pm 2$. It thus suffices to sum the numerator and denominator of Eq. (\[eq:B2\]) over permissible values of $n$. The occupancy reads $$\begin{aligned}
\lefteqn{n_{s}(r;k) =} \nonumber \\
&& \frac{\sum_{n}\sum_{r_{1}}\left( m(r_{1};k-n) p(s-r_{1}) +
m(s-r_{1};k+n) p(r_{1}) \right)}{\sum_{n}\sum_{r_{1}}p(s-r_{1})
p(r_{1})} \nonumber \\
&&
\label{eq:B3}\end{aligned}$$ where the sum over $r_{1}$ is from 0 to $s=r-n^2$, and $m(r;k)$ is given Eq. (\[A11\]). The two terms in the numerator give equal contributions. Figure \[fig:1\]e compares $n_{s}(r;k)$ as given in Eq. (\[eq:B3\]) with the FD distribution for various values of the added energy.
When the system is in contact with a heat bath at temperature reciprocal $\beta$, $r$ fluctuates with a probability law $q^r$ where, as before, $q \equiv \exp (-\beta)$. Accordingly, occupancies are obtained by multiplying the numerator and denominator of the previous expression in Eq. (\[eq:B3\]) by $q^r$ and summing over $r$ from 0 to $\infty$. The denominator gives the sum over states $$\begin{aligned}
Z^{\star}(q) & = & \sum_{r}q^r
\sum_{n}\sum_{r_{1}}p(r_{1})p(r-n^2-r_{1}) \nonumber \\
& = & {\left( \sum_{r}q^r p(r) \right)}^2 \sum_{n}q^{n^2}
\label{eq:B4}\end{aligned}$$ which may be written as an infinite product [@Berndt]. The average added energy reads $$\begin{aligned}
r_{c}(q) & = & \frac{q}{Z^{\star}}
\frac{{\mathrm{d}}Z^{\star}}{{\mathrm{d}}q} \nonumber \\
& = & \sum_{j=1,2\ldots}\left( \frac{2 j}{q^{-j}-1} + \frac{(-1)^j
2 j}{q^j - q^{-j}} \right)
\label{eq:B5}\end{aligned}$$
The expression for the occupancy may be reduced to a double sum, which coincides with the one given in Eq. (\[e8\]), shifted by $\pm n$ and weighted by $q^{n^2}$. This final result is intuitive since unbalancing between $N^{+}$ and $N^{-}$ increments the energy by $n^2$. We have $$n_{c}(q;k) = \frac{\sum_{n}\left( n(q;k-n) + n(q;k+n) \right)
q^{n^2}}{\sum_{n}q^{n^2}}
\label{eq:B6}$$ where the sums over $n$ run from $-\infty$ to $+\infty$, and $n(q;k)$ is given in Eq. (\[eq:A12\]). The two terms in the numerator give the same contributions. The occupancy $n_{c}(q;k)$ may be expressed as a function of the average energy and $k$ with the help of Eq.(\[eq:B5\]). Comparison with the FD distribution is in Fig.\[fig:1\]e.
Similarly the fluorescence is obtained by shifting $k$ by $n$ with a weight factor $q^{n^2}$.
[99]{}
CEMII is Unité Mixte de Recherche n¡5507 associée au Centre National de la Recherche Scientifique.
Also with LIRMM, 161 rue Ada, F34392 Montpellier, France.
P.S. [Zory]{} (ed.), Quantum-well lasers, Academic Press, Boston (1993), see chapter 10.
S. [Pell]{} and G. [Gabrielse]{} “Observing the quantum limit of an electron cyclotron: QND measurements of quantum jumps between Fock states”, Physical Review Letters **83**, 1287 (1999).
B. [Fain]{} “Stationary spontaneous emission up to fourth-order approximation” Quantum and Semiclassical Optics, **8**, 755 (1996).
R. [Denton]{}, B. [M[ü]{}hlschlegel]{} and D.J. [Scalapino]{}, “Thermodynamic properties of electrons in small metal particles”, Physical Review B **7**, 3589-3607 (1973).
C. [Kittel]{} and H. [Kroemer]{}, “Thermal Physics”, W.H. Freeman and Co., San Francisco (1980).
J. [Arnaud]{}, J.M. [Bo[é]{}]{}, L. [Chusseau]{} and F. [Philippe]{} “Illustration of the Fermi-Dirac statistics”, American Journal of Physics **67**, 215-221 (1999).
F. [Philippe]{} and J.M. [Bo[é]{}]{}, “Partitions and the Fermi-Dirac distribution”, Journal of Combinatorial Theory A (to appear).
K. [Schönhammer]{} and V. [Meden]{}, “Fermion-boson transmutation and comparison of statistical ensembles in one dimension”, American Journal of Physic **64**, 1168-1176 (1996).
F. [Philippe]{}, “Fermions in finite canonical ensemble: comparison between two formalisms” Journal of Physic A **33**, L93-L95 (2000).
P. [Borrmann]{}, J. [Harting]{}, O. [M[ü]{}lken]{} and E.R. [Hilf]{}, “Calculation of thermodynamical properties of finite Bose-Einstein systems”, Physical Revieaw A **60**, 1519 (1999), see Eq. (12).
G.E. Andrew, The theory of partitions in: Encyclopedia of mathematics and its applications, G.C. Rota (ed.), Addison-Wesley, Reading, Mass. (1976).
B.C. Berndt, “Ramanujan theory of theta functions” in: Theta functions: from the classical to the modern, M. Ram Murty (ed.), CRM Proceedings and lecture notes, AMS, Providence (1993).
J. Arnaud, “Optically-pumped semiconductor squeezed-light generation”, Optical and Quantum Electronics **27**, 225 (1995).
P. Borrmann and G. Franke, “Recursion formulas for quantum statistical partition functions”, J. Chem. Phys. **98**, 2484-2485 (1993).
L. Chusseau and J. Arnaud, “Sub-poissonian light from mesoscopic laser diodes. A simulation”, (unpublished).
[^1]: Published in *Physical Review B*, **62** (2000) 13482–13489
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a covariant gauge we implicitly assume that the Green’s function propagates information from one point of the space-time to another, so that the Green’s function is responsible for the dynamics of the relativistic particle. In the light front form, which in principle is a change of coordinates, one would expect that this feature would be preserved. In this manner, the fermionic field propagator can be split into a propagating piece and a non-propagating (“contact”) term. Since the latter (“contact”) one does not propagate information, and therefore, assumedly with no harm to the field dynamics we wanted to know what would be the impact of dropping it off. To do that, we investigated its role in the Ward identity in the light front.'
author:
- 'J.H.O.Sales'
- 'A.T. Suzuki'
- 'J.D. Bolzan'
title: 'QED$_{4}$ Ward Identity for fermionic field in the light-front'
---
11.10.Gh, 03.65.-w, 11.10.-z
Introduction
============
One of the most important concepts in quantum field theories is the question of renormalizability. In QED (Quantum Electrodynamics) specifically, the electric charge renormalization is guaranteed solely by the renormalization of the photon propagator. This result is a consequence of the so-called Ward identity, demonstrated by J.C.Ward in 1950 [@Ward1; @Ward2; @Ward3]. The importance of this result can be seen and emphasized in the fact that without the validity of such an identity, there would be no guarantee that the renormalized charge of different fermions (electrons, muons, etc.) would be the same. In other words, without such identity, charges of different particles must have different renormalization constants, a feature not so gratifying nor elegant. Moreover, without the Ward identity, renormalizability would have to be laboriously checked order by order in perturbation theory.
What the Ward identity does is to relate the vertex function of the theory with the derivative of the self-energy function of the electron, and this important correlation is expressed in terms of equality between the renormalization constants, namely, $Z_{1}=Z_{2}$, where $Z_{1}$and $Z_{2}$ are the renormalization constants related to the vertex function and the fermionic propagator respectively. Since the renormalized electric charge is given in terms of the bare electric charge via the product $%
e_{R}=Z_{3}^{1/2}Z_{2}Z_{1}^{-1}e_{0}$, it follows immediately that $%
e_{R}=Z_{3}^{1/2}e_{0}$, i.e., electric charge renormalization depends solely on the renormalization of the photon propagator.
We know that light-front dynamics is plagued with singularities of all sorts and because of this the connection between the covariant quantities and light-front quantities cannot be so easily established. If we want to describe our theory in terms of the light-front coordinates or variables, we must take care of the boundary conditions that fields must obey. Thus, a simple projection from the covariant quantities to light-front quantities via coordinate transformations is bound to be troublesome. This can be easily seen in our checking of the QED Ward identity in the light-front, where the fermionic propagator does bear an additional term proportional to $%
\gamma ^{+}(p^{+})^{-1}$ oftentimes called contact term in the literature, which, of course, is conspicuously absent in the covariant propagator. This term, as we will see, is crucial to the Ward identity in the light-front. The covariant propagating term solely projected onto the light-front coordinates therefore violates Ward identity, and therefore breaks gauge invariance. Such result is obviously wrong and unwarranted.
The outline of our paper is as follows: We begin by considering the standard derivation for the covariant case Ward identity and show explicitly that the fermionic propagator there cannot be analytically regularized, otherwise Ward identity cannot be achieved. Then we explicitly construct our fermionic propagator in terms of the light-front coordinates, with the proper contact term in it and in the following section we deal with the checking of the Ward identity proper. Finally, the next two sections are devoted to the concluding remarks and Appendix; in the latter we define our light-cone coordinates convention and notation and include explicit calculations showing that without the contact term in the fermionic propagator, Ward identity is not satisfied, and thus gauge invariance is violated.
The Ward Identity
=================
There are several ways to write down the Ward identity for fermions, and one of them is inferred from manipulations of their propagator, namely, $S(p)$. Multiplying by its inverse, we get the identity
$$S(p)S^{-1}(p)=I\text{,}$$
Deriving both sides with respect to $p^{\mu }$ we get
$$\frac{\partial S(p)}{\partial p^{\mu }}S^{-1}(p)+S(p)\frac{\partial S^{-1}(p)%
}{\partial p^{\mu }}=0$$
which leads to
$$\frac{\partial S(p)}{\partial p^{\mu }}S^{-1}(p)=-S(p)\frac{\partial
S^{-1}(p)}{\partial p^{\mu }}$$
Finally, multiplyng both sides from the left by the propagator itself
$$\frac{\partial S(p)}{\partial p^{\mu }}=-S(p)\frac{\partial S^{-1}(p)}{%
\partial p^{\mu }}S(p) \label{passo}$$
Now, using $\displaystyle {S(p)=\frac{i}{p\!\!\!/-m}}$ it follows that its inverse is $\displaystyle {S^{-1}(p)=-i(p\!\!\!/-m)}$. Deriving this last expression with respect to $p^{\mu }$ we get
$$\frac{\partial S^{-1}(p)}{\partial p^{\mu }}=-i\gamma _{\mu }$$ which inserted into (\[passo\]) leads to the differential form of the Ward Identity, namely,
$$\frac{\partial S(p)}{\partial p^{\mu }}=iS(p)\gamma _{\mu }S(p)\text{.}
\label{ward}$$
Here it is important to stress that if the propagator were raised to a power as in the analytic regularization scheme, i.e. if it had the form $\displaystyle{S(p)= \frac{i}{\left (p\!\!\!/-m \right)^{\sigma }}}$, with $\sigma \neq 1$, the identity (\[ward\]) would *not* be fulfilled.
Fermion propagator in the Light-Front
=====================================
With the light-front coordinate transformations given in Appendix A, we can find the corresponding fermionic propagator, beginning with the term p/, as in (\[p barra fl\]):
$$p\!\!\!/=p_{\mu }\gamma ^{\mu }=\left( \gamma ^{+}p^{-}+\gamma
^{-}p^{+}\right) -\left( \overrightarrow{\gamma }_{\perp }\cdot
\overrightarrow{p}_{\perp }\right) \text{,}$$
then
$$S(p)=\frac{i}{\left[ \left( \gamma ^{+}p^{-}+\gamma
^{-}p^{+}\right) -\left( \overrightarrow{\gamma }_{\perp }\cdot
\overrightarrow{p}_{\perp }\right) -m\right] }\text{,}$$
or, in another way, using $S(p)=\displaystyle{\frac{i\left( p\!\!\!/+m\right) }{p^{2}-m^{2}
}}$,
$$\begin{aligned}
&&\bigskip \left. S(p)=\frac{i\left[ \left( \gamma ^{+}p^{-}+\gamma
^{-}p^{+}\right) -\left( \overrightarrow{\gamma }_{\perp }\cdot
\overrightarrow{p}_{\perp }\right) +m\right] }{p^{+}\left(
p^{-}-p_{on}\right) }\text{,}\right. \nonumber \\
&& \nonumber \\
&& \left. S(p)=\frac{i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left(
p^{-}-p_{on}\right) }+\frac{i\gamma ^{+}}{2p^{+}}\text{,}\right.
\label{ward fl}\end{aligned}$$
where
$$p\!\!\!/_{on}=\left( \gamma ^{+}p_{on}+\gamma ^{-}p^{+}\right) -\left(
\overrightarrow{\gamma }_{\perp }\cdot \overrightarrow{p}_{\perp }\right)
\text{ ,}$$
$$p_{on}=\frac{p_{\perp }^{2}+m^{2}}{2p^{+}}\text{.}$$
The Ward Identity on the Light-Front
====================================
There are two manners to test if the propagator (\[ward fl\]) on the light-front satisfy the Ward identity (\[ward\]). The simplest and most direct one is to do the derivatives $\displaystyle{\frac{\partial S_{_{{}}}^{-1}(p)}{
\partial p^{\mu }}}$ for each component and put them in (\[passo\]):
$$\frac{\partial S_{{}}^{-1}(p)}{\partial p^{+}}=-i\gamma ^{-}$$
$$\frac{\partial S_{{}}^{-1}(p)}{\partial p^{-}}=-i\gamma ^{+}$$
$$\frac{\partial S_{{}}^{-1}(p)}{\partial p_{1,2}}=-i\gamma ^{1,2}\text{.}
\label{s-1 resultados}$$
$$\frac{\partial S(p)}{\partial p^{+}}=iS(p)\gamma ^{-}S(p)$$
$$\frac{\partial S(p)}{\partial p^{-}}=iS(p)\gamma ^{+}S(p)$$
$$\frac{\partial S(p)}{\partial p_{1,2}}=iS(p)\gamma ^{1,2}S(p),
\label{s fl resultados}$$
where $p_{1,2}=p_{\perp }$ and $\gamma ^{1,2}=\gamma ^{\perp }$ are the transversal or perpendicular components.
Comparing (\[s fl resultados\]) and (\[ward\]), one verifies that the Ward identity is satisfied if one includes the necessary factors due to the change of coordinate system, or, in other words, considering the Jacobian determinant of this transformation.
The second manner to test the identity on the light-front is working explicitly with all the figures of (\[ward\]). The details are presented in Appendix B, and below we put the principal results:
$$\frac{\partial S(p)}{\partial p^{+}}=iS(p)\gamma ^{-}S(p)=\frac{%
-ip^{-}\left( p\!\!\!/+m\right) }{2\left[ p^{+}\left( p^{-}-p_{on}\right) %
\right] ^{2}}+\frac{i\gamma ^{-}}{2p^{+}\left( p^{-}-p_{on}\right) }\text{,}
\label{derivada +}$$
$$\frac{\partial S(p)}{\partial p^{-}}=iS(p)\gamma ^{+}S(p)=\frac{-i\left(
p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right) ^{2}}\text{,}
\label{derivada -}$$
$$\frac{\partial S(p)}{\partial p_{1,2}}=iS(p)\gamma ^{1,2}S(p)=\frac{%
ip_{1,2}\left( p\!\!\!/+m\right) }{2\left[ p^{+}\left( p^{-}-p_{on}\right) %
\right] ^{2}}+\frac{i\gamma ^{1,2}}{2p^{+}\left( p^{-}-p_{on}\right) }\text{,%
} \label{derivada transversal}$$
that is, again one corroborates the relations (\[s fl resultados\]). An important point here is that, using the simplified propagator $\displaystyle{S(p)=\frac{
i\left( p\!\!\!/_{on}+m\right) }{p^{+}\left( p^{-}-p_{on}\right) }}$ as some authors do, the Ward Identity is *not* fulfilled, as shown in Appendix C.
Conclusions
===========
We have shown here that the Ward identity for the fermionic field in the light-front is preserved to guarantee that the charge renomalization constant depends solely on the photon renormalization constant, as it is expected. However, one important point emerges in our computation, and that is that the Ward identity in the light-front is valid provided the fermionic field propagator bears the relevant “contact” term piece, which is absent in the covariant propagator and its straightforward projection into light-front variables.
Our computation has demonstrated once again the significance of the light-front zero-mode contribution that the so-called “contact” term bears in it, without which Ward identity would be violated. Although the zero-mode term does not carry physical information, its non-vanishing contribution nonetheless is crucial to the validity of the Ward identity in the light-front formalism. In other words, “contact” term may not carry information from one space-time point to another in the light front, but contains relevant physical information needed to ensure the Ward identity, and therefore, for the correct charge renormalization.
Appendix
========
Light-front Coordinates
-----------------------
The Light-front is characterized by the null-plane $x^{+}=t+z=0$, which is its time coordinate. All of the coordinates are set regarding this plane, and one has new definitions of the scalar product, for example. The basic relations on the light-front are
$$\begin{aligned}
&&\left. x^{+}=\frac{1}{\sqrt{2}}\left( x^{0}+x^{3}\right) \right.
\nonumber \\
&&\left. x^{-}=\frac{1}{\sqrt{2}}\left( x^{0}-x^{3}\right) \right.
\nonumber \\
&&\left. \overrightarrow{x}_{\perp }=x^{1}\overrightarrow{i}+x^{2}%
\overrightarrow{j}\right. \text{,} \label{light front]}\end{aligned}$$
so, the scalar product is given by
$$a^{\mu }b_{\mu }=\left( a^{+}b^{-}+a^{-}b^{+}\right) -\overrightarrow{a}%
_{\perp }\cdot \overrightarrow{b}_{\perp }\text{.} \label{produto escalar]}$$
Using (\[produto escalar\]\]), one can write the product $p\!\!\!/$ on the light-front:
$$p\!\!\!/=p_{\mu }\gamma ^{\mu }=\left( \gamma ^{+}p^{-}+\gamma
^{-}p^{+}\right) -\left( \overrightarrow{\gamma }_{\perp }\cdot
\overrightarrow{p}_{\perp }\right) \text{.} \label{p barra fl}$$
Checking the Ward Identity
--------------------------
In this Appendix, we show the details of the algebra necessary to arrive at (\[derivada +\]-\[derivada transversal\]). In the first place, we list the numerous properties that Dirac gama matrices in the light-front obey and should be used:
$$\begin{array}{cc}
\left. \gamma ^{+}\gamma ^{+}=\gamma ^{-}\gamma ^{-}=0\right. & \left.
\gamma ^{1}\gamma ^{\pm }\gamma ^{2}+\gamma ^{2}\gamma ^{\pm }\gamma
^{1}=0\right. \\
\left. \gamma ^{+}\gamma ^{-}\gamma ^{+}=2\gamma ^{+}\right. & \left.
\gamma ^{\mp }\gamma ^{\pm }\gamma ^{1,2}+\gamma ^{1,2}\gamma ^{\pm }\gamma
^{\mp }=2\gamma ^{1,2}\right. \\
\left. \gamma ^{-}\gamma ^{+}\gamma ^{-}=2\gamma ^{-}\right. & \left.
\left\{ \left( \gamma _{\perp }p_{\perp }\right) ,\gamma ^{\pm }\right\}
=0\right. \\
\left. \gamma ^{1}\gamma ^{\pm }\gamma ^{1}=\gamma ^{\pm }\right. & \left.
\left\{ \gamma ^{+},\gamma ^{-}\right\} =2I\right. \\
\left. \gamma ^{2}\gamma ^{\pm }\gamma ^{2}=\gamma ^{\pm }\right. & \left.
\left( \gamma _{\perp }p_{\perp }\right) \gamma ^{\pm }\left( \gamma _{\perp
}p_{\perp }\right) =\left( p_{\perp }\right) ^{2}\gamma ^{\pm }\right. \\
\left. \left\{ \gamma ^{\pm },\gamma ^{1,2}\right\} =0\right. & \left.
\left\{ \left( \gamma _{\perp }p_{\perp }\right) ,\gamma ^{1,2}\right\}
=2p_{1,2}\right. \\
\left. \gamma ^{\pm }\gamma ^{1,2}\gamma ^{\pm }=0\right. & \left. \gamma
^{\pm }\gamma ^{\mp }\left( \gamma _{\perp }p_{\perp }\right) +\left( \gamma
_{\perp }p_{\perp }\right) \gamma ^{\mp }\gamma ^{\pm }=2\left( \gamma
_{\perp }p_{\perp }\right) \right. \\
\left. \gamma ^{1}\gamma ^{1}=\gamma ^{2}\gamma ^{2}=-I\right. & \left.
\gamma ^{\pm }\gamma ^{1,2}\left( \gamma _{\perp }p_{\perp }\right) +\left(
\gamma _{\perp }p_{\perp }\right) \gamma ^{1,2}\gamma ^{\pm }=2\gamma ^{\pm
}p_{1,2}\right. \\
\left. \gamma ^{\pm }\gamma ^{1}\gamma ^{2}+\gamma ^{2}\gamma ^{1}\gamma
^{\pm }=0\right. & \left. \gamma ^{+}\gamma ^{1,2}\gamma ^{-}+\gamma
^{-}\gamma ^{1,2}\gamma ^{+}=-2\gamma ^{1,2}\right. \\
\left. \left\{ \gamma ^{1},\gamma ^{2}\right\} =0\right. & \left. \left(
\gamma _{\perp }p_{\perp }\right) \gamma ^{1,2}\left( \gamma _{\perp
}p_{\perp }\right) =\mp \left( p_{1}\right) ^{2}\gamma ^{1,2}\pm \left(
p_{2}\right) ^{2}\gamma ^{1,2}-2p_{1}p_{2}\gamma ^{2,1}\right.
\end{array}
\label{gamas}$$
Next, some useful relations:
$$\frac{\partial p_{on}}{\partial p^{+}}=-\frac{p_{\perp }^{2}+m^{2}}{2\left(
p^{+}\right) ^{2}}=-\frac{p_{on}}{p^{+}} \label{derivada pon}$$
$$\frac{\partial p\!\!\!/_{on}}{\partial p^{+}}=-\gamma ^{+}\frac{p_{on}}{p^{+}%
}+\gamma ^{-} \label{derivada p barra on}$$
$$\frac{\partial p_{on}}{\partial p_{1}}=\frac{p_{1}}{p^{+}}.$$
Remembering that the fermion propagator is $S(p)=\displaystyle\frac{i\left(
p\!\!\!/_{on}+m\right) }{p^{+}\left( p^{-}-p_{on}\right) }+\frac{i\gamma ^{+}
}{2p^{+}}$, the plus component derivative is
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{i\left( \frac{\partial
p\!\!\!/_{on}}{\partial p^{+}}\right) }{2p^{+}\left( p^{-}-p_{on}\right) }-%
\frac{i\left( p\!\!\!/_{on}+m\right) }{2\left( p^{+}\right) ^{2}\left(
p^{-}-p_{on}\right) }+\frac{i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left(
p^{-}-p_{on}\right) ^{2}}\left( \frac{\partial p_{on}}{\partial p^{+}}%
\right) -\frac{i\gamma ^{+}}{2\left( p^{+}\right) ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-i\gamma ^{+}\frac{p_{on}}{p^{+}}%
+i\gamma ^{-}}{2p^{+}\left( p^{-}-p_{on}\right) }-\frac{i\left(
p\!\!\!/_{on}+m\right) }{2\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right)
}-\frac{ip_{on}\left( p\!\!\!/_{on}+m\right) }{2\left( p^{+}\right)
^{2}\left( p^{-}-p_{on}\right) ^{2}}-\frac{i\gamma ^{+}}{2\left(
p^{+}\right) ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-i\gamma ^{+}\left( p^{-}\right)
^{2}-i\gamma ^{-}p^{+}p_{on}+i\left( \gamma _{\perp }p_{\perp }\right)
p^{-}-imp^{-}}{2\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-ip^{-}\left( p\!\!\!/+m\right) +%
\frac{i}{2}p^{+}\gamma ^{-}\left( p^{-}-p_{on}\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-ip^{-}\left( p\!\!\!/+m\right) }{%
2\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}+\frac{i\gamma ^{-}}{%
2p^{+}\left( p^{-}-p_{on}\right) }. \label{resultado derivada positiva}$$
Now, calculating the term $iS(p)\gamma ^{-}S(p)$ and exploiting the properties of the gamma functions, we have
$$=-i\left\{ \frac{\left( p\!\!\!/_{on}+m\right) \gamma ^{-}\left(
p\!\!\!/_{on}+m\right) }{4\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}%
}+\frac{\left( p\!\!\!/_{on}+m\right) \gamma ^{-}\gamma ^{+}}{4\left(
p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }+\frac{\gamma ^{+}\gamma
^{-}\left( p\!\!\!/_{on}+m\right) }{4\left( p^{+}\right) ^{2}\left(
p^{-}-p_{on}\right) }+\frac{\gamma ^{+}\gamma ^{-}\gamma ^{+}}{4\left(
p^{+}\right) ^{2}}\right\}$$
$$=-i\left\{ \frac{p\!\!\!/_{on}\gamma ^{-}p\!\!\!/_{on}+m\left\{
p\!\!\!/_{on},\gamma ^{-}\right\} +m^{2}\gamma ^{-}}{4\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}+\frac{p\!\!\!/_{on}\gamma ^{-}\gamma
^{+}+\gamma ^{+}\gamma ^{-}p\!\!\!/_{on}+m\left\{ \gamma ^{+},\gamma
^{-}\right\} }{4\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }+\frac{%
\gamma ^{+}}{2\left( p^{+}\right) ^{2}}\right\}$$
$$=-i\left\{ \frac{2\gamma ^{+}\left( p_{on}\right) ^{2}-2p_{on}\left( \gamma
_{\perp }p_{\perp }\right) +\left( p_{\perp }\right) ^{2}\gamma
^{-}+2mp_{on}+m^{2}\gamma ^{-}}{4\left[ p^{+}\left( p^{-}-p_{on}\right) %
\right] ^{2}}+\frac{4\gamma ^{+}p_{on}-2\left( \gamma _{\perp }p_{\perp
}\right) +2m}{4\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }+\frac{%
\gamma ^{+}}{2\left( p^{+}\right) ^{2}}\right\}$$
$$=-i\left\{ \frac{2p^{-}p\!\!\!/-2\gamma ^{-}p^{+}\left( p^{-}-p_{on}\right)
+2mp^{-}}{4\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}\right\}$$
$$=\frac{-ip^{-}\left( p\!\!\!/+m\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}+\frac{i\gamma ^{-}}{2p^{+}\left(
p^{-}-p_{on}\right) }\text{.} \label{resultado produto positivo}$$
One can the see that, from (\[resultado derivada positiva\]) and (\[resultado produto positivo\]), $\displaystyle\frac{\partial S(p)}{\partial p^{+}}
=iS(p)\gamma ^{-}S(p)$.
For the minus component, the derivative is very simple,
$$\frac{\partial S(p)}{\partial p^{-}}=-\frac{i\left( p\!\!\!/_{on}+m\right) }{%
2p^{+}\left( p^{-}-p_{on}\right) ^{2}} \label{resultado derivada negativa}$$
And the term $iS(p)\gamma ^{+}S(p)$,
$$=-i\left\{ \frac{\left( p\!\!\!/_{on}+m\right) \gamma ^{+}\left(
p\!\!\!/_{on}+m\right) }{4\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}%
}\right\}$$
$$=-i\left\{ \frac{2\gamma ^{-}\left( p^{+}\right) ^{2}-2p^{+}\left( \gamma
_{\perp }p_{\perp }\right) +2p^{+}p_{on}\gamma ^{+}+2mp^{+}}{4\left[
p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}\right\}$$
$$=\frac{-i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right)
^{2}} \label{resultado produto negativo}$$
and again one has $\displaystyle\frac{\partial S(p)}{\partial p^{-}}=iS(p)\gamma ^{+}S(p)$.
Finally, the derivative of the transversal components:
$$\frac{\partial S(p)}{\partial p_{1}}=\frac{i\left( \frac{\partial
p\!\!\!/_{on}}{\partial p_{1}}\right) }{2p^{+}\left( p^{-}-p_{on}\right) }+%
\frac{i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right) ^{2}}%
\left( \frac{\partial p_{on}}{\partial p_{1}}\right)$$
$$\frac{\partial S(p)}{\partial p_{1}}=\frac{i\left( \gamma ^{+}\frac{p_{1}}{%
p^{+}}+\gamma ^{1}\right) }{2p^{+}\left( p^{-}-p_{on}\right) }+\frac{%
ip_{1}\left( p\!\!\!/_{on}+m\right) }{\left( p^{+}\right) ^{2}\left(
p^{-}-p_{on}\right) ^{2}}$$
$$\frac{\partial S(p)}{\partial p_{1}}=\frac{ip_{1}\left[ \gamma
^{+}p^{-}+\gamma ^{-}p^{+}-\left( \gamma _{\perp }p_{\perp }\right) +m\right]
+i\gamma ^{1}p^{+}\left( p^{-}-p_{on}\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}$$
$$\frac{\partial S(p)}{\partial p_{1}}=\frac{ip_{1}\left( p\!\!\!/+m\right) }{2%
\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}+\frac{i\gamma ^{1}}{%
2p^{+}\left( p^{-}-p_{on}\right) }\text{.}
\label{resultado derivada transversal}$$
The term $iS(p)\gamma ^{1}S(p)$ is very laborious and almost all of the gamma matrices properties must be used:
$$=-i\left\{ \frac{p\!\!\!/_{on}\gamma ^{1}p\!\!\!/_{on}+m\left\{
p\!\!\!/_{on},\gamma ^{1}\right\} +m^{2}\gamma ^{1}}{4\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}+\frac{p\!\!\!/_{on}\gamma ^{1}\gamma
^{+}+\gamma ^{+}\gamma ^{1}p\!\!\!/_{on}+m\left\{ \gamma ^{1},\gamma
^{-}\right\} }{4\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }+\frac{%
\gamma ^{+}\gamma ^{1}\gamma ^{+}}{4\left( p^{+}\right) ^{2}}\right\}$$
$$\begin{aligned}
&&\left. =-i\left\{ \frac{-2\gamma ^{1}p^{+}p_{on}-2\gamma
^{+}p_{1}p_{on}-2\gamma ^{-}p^{+}p_{1}-\gamma ^{1}\left( p_{1}\right)
^{2}+\gamma ^{1}\left( p_{2}\right) ^{2}-2\gamma ^{2}p_{1}p_{2}+m^{2}\gamma
^{1}-2mp_{1}}{4\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}+\right.
\right. \\
&&\left. \left. +\frac{2\gamma ^{+}p_{1}-2\gamma ^{1}p^{+}}{4\left(
p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }\right\} \right. \end{aligned}$$
$$=-i\left\{ \frac{-2p_{1}\left[ \gamma ^{+}p^{-}+\gamma ^{-}p^{+}-\left(
\gamma _{\perp }p_{\perp }\right) +m\right] +\gamma ^{1}\left( p_{1}\right)
^{2}+\gamma ^{1}\left( p_{2}\right) ^{2}-2\gamma ^{1}p^{+}p^{-}+m^{2}\gamma
^{1}}{4\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}\right\}$$
$$=-i\left[ \frac{-2p_{1}\left( p\!\!\!/+m\right) +\gamma ^{1}\left(
p_{1}^{2}+p_{2}^{2}+m^{2}-2p^{+}p^{-}\right) }{4\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}\right]$$
$$=-i\left[ \frac{-2p_{1}\left( p\!\!\!/+m\right) -2\gamma ^{1}\left(
p^{+}p^{-}-p^{+}p_{on}\right) }{4\left[ p^{+}\left( p^{-}-p_{on}\right) %
\right] ^{2}}\right]$$
$$=\frac{ip_{1}\left( p\!\!\!/+m\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}+\frac{i\gamma ^{1}}{2p^{+}\left(
p^{-}-p_{on}\right) } \label{resultado produto transversal}$$
and from (\[resultado derivada transversal\]) and (\[resultado produto transversal\]), one has $\displaystyle\frac{\partial S(p)}{\partial p_{1,2}}=iS(p)\gamma
^{1,2}S(p)$.
The Ward Identity for the propagator without contact term
---------------------------------------------------------
Here we work on the Ward Identity for the simplified propagator $S(p)=\displaystyle\frac{
i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right) }$.
For the minus component, the derivative is the same as the one obtained before,
$$\frac{\partial S(p)}{\partial p^{-}}=\frac{-i\left( p\!\!\!/_{on}+m\right) }{
2p^{+}\left( p^{-}-p_{on}\right) ^{2}}\text{;}
\label{derivada negativa modificada}$$
and the term $iS(p)\gamma ^{+}S(p)$ is equal too, because in the other case, the terms $\displaystyle\frac{i\gamma ^{+}}{2p^{+}}$ do not contribute due to the property $\gamma ^{+}\gamma ^{+}=0$:
$$iS(p)\gamma ^{+}S(p)=\frac{-i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left(
p^{-}-p_{on}\right) ^{2}}\text{,} \label{produto negativo modificado}$$
so, for the negative component, the Ward identity *is* satisfied
$$\frac{\partial S(p)}{\partial p^{-}}=iS(p)\gamma ^{+}S(p)\text{.}$$
For the plus component, one has
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-i\gamma ^{+}\frac{p_{on}}{p^{+}}%
+i\gamma ^{-}}{2p^{+}\left( p^{-}-p_{on}\right) }-\frac{i\left(
p\!\!\!/_{on}+m\right) }{2\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right)
}-\frac{ip_{on}\left( p\!\!\!/_{on}+m\right) }{2\left( p^{+}\right)
^{2}\left( p^{-}-p_{on}\right) ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{ip^{+}p^{-}\gamma
^{-}-p^{+}p_{on}\gamma ^{-}-p^{-}p_{on}\gamma ^{+}+\left( p_{on}\right)
^{2}\gamma ^{+}-p^{-}\left( p\!\!\!/_{on}+m\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}$$
$$\frac{\partial S(p)}{\partial p^{+}}=\frac{-ip^{-}\left(
p\!\!\!/_{on}+m\right) }{2\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}%
}+\frac{i\gamma ^{-}}{2p^{+}\left( p^{-}-p_{on}\right) }-\frac{i\gamma
^{+}p_{on}}{2\left( p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }\text{;}
\label{derivada positiva modificada}$$
And the term $iS(p)\gamma ^{-}S(p)$,
$$i\left[ \frac{i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left(
p^{-}-p_{on}\right) }\right] \gamma ^{-}\left[ \frac{i\left(
p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right) }\right] =-i\frac{%
\left( p\!\!\!/_{on}+m\right) \gamma ^{-}\left( p\!\!\!/_{on}+m\right) }{4%
\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}$$
$$\left. =-i\left\{ \frac{2\gamma ^{+}\left( p_{on}\right) ^{2}\gamma
^{+}-2p_{on}\left( \gamma _{\perp }p_{\perp }\right) +\left( p_{\perp
}\right) ^{2}\gamma ^{-}+2mp_{on}+m^{2}\gamma ^{-}}{4\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}\right\} \right.$$
$$=-i\left\{ \frac{2p^{-}\left( p\!\!\!/_{on}+m\right) +\left[ -2\gamma
^{+}p_{on}-2\gamma ^{-}p^{+}+2\left( \gamma _{\perp }p_{\perp }\right) -2m%
\right] \left( p^{-}-p_{on}\right) }{4\left[ p^{+}\left( p^{-}-p_{on}\right) %
\right] ^{2}}\right\}$$
$$=\frac{-ip^{-}\left( p\!\!\!/_{on}+m\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}+\frac{i\gamma ^{-}}{2p^{+}\left(
p^{-}-p_{on}\right) }+\frac{i\gamma ^{+}p_{on}}{2\left( p^{+}\right)
^{2}\left( p^{-}-p_{on}\right) }-\frac{\left[ \left( \gamma _{\perp
}p_{\perp }\right) -m\right] }{2\left( p^{+}\right) ^{2}\left(
p^{-}-p_{on}\right) }\text{;} \label{produto positivo modificado}$$
and because of the presence of the last term and the wrong signal of the third, one has
$$\frac{\partial S(p)}{\partial p^{+}}\neq iS(p)\gamma ^{-}S(p)\text{.}$$
For the transversal components, the derivative is the same as obtained before,
$$\frac{\partial S(p)}{\partial p_{1}}=\frac{ip_{1}\left( p\!\!\!/+m\right) }{2%
\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}-\frac{i\gamma ^{1}}{%
2p^{+}\left( p^{-}-p_{on}\right) }\text{.}
\label{derivada transversal modificada}$$
And the term $iS(p)\gamma ^{1}S(p)$,
$$i\left[ \frac{i\left( p\!\!\!/_{on}+m\right) }{2p^{+}\left(
p^{-}-p_{on}\right) }\right] \gamma ^{1}\left[ \frac{i\left(
p\!\!\!/_{on}+m\right) }{2p^{+}\left( p^{-}-p_{on}\right) }\right] =-i\frac{%
\left( p\!\!\!/_{on}+m\right) \gamma ^{-}\left( p\!\!\!/_{on}+m\right) }{4%
\left[ p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}$$
$$\begin{aligned}
&&\left. =-i\left\{ \frac{-2p_{1}\left[ \gamma ^{+}p^{-}+\gamma
^{-}p^{+}-\left( \gamma _{\perp }p_{\perp }\right) +m\right] +2\gamma
^{+}p^{-}p_{1}+\gamma ^{1}\left( p_{1}\right) ^{2}+}{4\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}\right. \right. \\
&&\left. \left. \frac{-2\gamma ^{1}p^{+}p_{on}-2\gamma
^{+}p_{1}p_{on}+\gamma ^{1}\left( p_{2}\right) ^{2}+m^{2}\gamma ^{1}}{4\left[
p^{+}\left( p^{-}-p_{on}\right) \right] ^{2}}\right\} \right. \end{aligned}$$
$$=\frac{ip_{1}\left( p\!\!\!/+m\right) }{2\left[ p^{+}\left(
p^{-}-p_{on}\right) \right] ^{2}}-\frac{\gamma ^{+}p_{1}}{2\left(
p^{+}\right) ^{2}\left( p^{-}-p_{on}\right) }\text{,}
\label{produto transversal modificado}$$
then, comparing (\[derivada transversal modificada\]) and (\[produto transversal modificado\]), one has $\displaystyle\frac{\partial S(p)}{\partial p_{1,2}}
\neq iS(p)\gamma ^{1,2}S(p)$.
**Acknowledgments:** J.D. Bolzan thanks CNPq for financial support. J.H.O. Sales thanks FAPESP and the hospitality of the Institute for Theoretical Physics, UNESP, where part of this work has been performed.
[10]{}
J.C.Ward, Physical Review, [**77**]{}, (1950) 293-293
J.C.Ward, Physical Review, [**78**]{}, (1950) 182-182
Y. Takahashi, Nuovo Cimento, [**6**]{}, ser. 10, (1957) 370
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'While stochastic gradient descent (SGD) is a workhorse in machine learning, the learning properties of many practically used variants are hardly known. In this paper, we consider least squares learning and contribute filling this gap focusing on the effect and interplay of multiple passes, mini-batching and averaging, and in particular tail averaging. Our results show how these different flavors of SGD can be combined to achieve optimal learning errors, hence providing practical insights.'
author:
- 'Nicole Mücke[^1] , Gergely Neu[^2] and Lorenzo Rosasco[^3]'
bibliography:
- 'bib\_SGD.bib'
title:
- 'Beating SGD Saturation with Tail-Averaging and Minibatching'
- 'Beating SGD Saturation with Tail-Averaging and Minibatching'
---
Introduction
============
Stochastic gradient descent (SGD) provides a simple and yet stunningly efficient way to solve a broad range of machine learning problems. Our starting observation is that, while a number of variants including multiple passes over the data, mini-batching and averaging are commonly used, their combination and learning properties are studied only partially. The literature on convergence properties of SGD is vast, but usually only one pass over the data is considered, see, e.g., [@NemJudLan09]. In the context of nonparametric statistical learning, which we consider here, the study of one-pass SGD was probably first considered in [@SmaYao06] and then further developed in a number of papers (e.g., [@YinPon08; @YaoTar14; @orabona]). Another line of work derives statistical learning results for one pass SGD with averaging from a worst-case sequential prediction analysis [@RSS12; @HK14; @RaShaSri11]. The idea of using averaging also has a long history going back to at least the works of [@R88] and [@PolJu92], see also [@ShaZha13] and references therein. More recently, averaging was shown to lead to larger, possibly constant, step-sizes, see [@BacMou13; @DieuBa16; @DieFlaBac17]. A different take on the role of (weighted) averaging was given in [@NeuRos18], highlighting a connection with ridge regression, a.k.a. Tikhonov regularization. A different flavor of averaging called *tail averaging* for one-pass SGD was considered in [@JKKNS18] in a parametric setting. The role of minibatching has also being considered and shown to potentially lead to linear parallelization speedups, see e.g. [@Cotter11] and references therein. Very few results consider the role of multiple passes for learning. Indeed, this variant of SGD is typically analyzed for the minimization of the empirical risk, rather than the actual population risk, see for example [@Ber97]. To the best of our knowledge the first paper to analyze the learning properties of multipass SGD was [@RosVil15], where a cyclic selection strategy was considered. Other results for multipass SGD were then given in [@HarRecSin16] and [@LinCamRos16]. Our starting point are the results in [@LinRos17] where optimal results for multipass SGD where derived considering also the effect of mini-batching. Following the approach in this latter paper, multipass SGD with averaging was analyzed by [@PillRudBa18] with no minibatching.
In this paper, we develop and improve the above results on two fronts. On the one hand, we consider for the first time the role of multiple passes, mini-batching and averaging at once. On the other hand, we further study the beneficial effect of tail averaging. Both mini-batching and averaging are known to allow larger step-sizes. Our results show that their combination allows even more aggressive parameter choices. At the same time averaging was shown to lead to slower convergence rates in some cases. In a parametric setting, averaging prevents linear convergence rates [@BacMou13; @DieFlaBac17]. In a nonparametric setting, it prevents exploiting the possible regularity in the solution [@DieuBa16], a phenomenon called [*saturation*]{} [@engl96]. In other words, uniform averaging can prevent optimal rates in a nonparametric setting. Our results provide a simple explanation to this effect, showing it has a purely deterministic nature. Further, we show that tail averaging allows to bypass this problem. These results parallel the findings of [@JKKNS18], showing similar beneficial effects of tail-averaging and minibatching in the finite-dimensional setting. Following [@LinRos17], our analysis relies on the study of batch gradient descent and then of the discrepancy between batch gradient and SGD, with the additional twist that it also considers the role of tail-averaging. The rest of the paper is organized as follows. In Section \[LS\_learn\], we describe the least-squares learning problem that we consider, as well as the different SGD variants we analyze. In Section \[appet\], we collect a number of observations shedding light on the role of uniform and tail averaging. In Section \[sec:main\], we present and discuss our main results. In Section \[numerics\] we illustrate our results via some numerical simulations. Proofs and technical results are deferred to the appendices.
Least Squares Learning with SGD {#LS_learn}
===============================
An appetizer: Averaging and Gradient Descent Convergence {#appet}
========================================================
Main Results and Discussion {#sec:main}
===========================
Numerical Illustration {#numerics}
======================
[**Acknowledgments**]{}\
\
NM is supported by the German Research Foundation under DFG Grant STE 1074/4-1. L. R. acknowledges the financial support of the AFOSR projects FA9550-17-1-0390 and BAA-AFRL-AFOSR-2016-0007 (European Office of Aerospace Research and Development), and the EU H2020-MSCA-RISE project NoMADS - DLV-777826.
Analysis {#sec:analysis}
========
[^1]: Institute for Stochastics and Applications, University of Stuttgart, [*[email protected]*]{}
[^2]: Universitat Pompeu Fabra, Barcelona, Spain, [ *[email protected]*]{}
[^3]: LCSL, Massachusetts Institute of Technology & Istituto Italiano di Tecnologia & DIBRIS, Universita’ degli Studi di Genova, [*[email protected]*]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide the gauge-invariant expression for large-scale cosmic microwave background temperature fluctuations at second-order in perturbation theory. It enables to unambiguously define the nonlinearity parameter $f_{\rm NL}$ which is used by experimental collaborations to pin down the level of Non-Gaussianity in the temperature fluctuations. Furthermore, it contains a [*primordial*]{} term encoding all the information about the Non-Gaussianity generated at primordial epochs and about the mechanism which gave rise to cosmological perturbations, thus neatly disentangling the primordial contribution to Non-Gaussianity from the one caused by the post-inflationary evolution.'
address:
- '$^{(1)}$[*Astronomy Centre, University of Sussex, Falmer, Brighton, BN1 9QH, U.K.*]{}'
- '$^{(2)}$[*Dipartimento di Fisica ‘Galileo Galilei’, Università di Padova, via Marzolo 8, I-35131, Padova, Italy*]{}'
- '$^{(3)}$[*INFN, Sezione di Padova, via Marzolo 8, I-35131, Padova, Italy*]{}'
author:
- 'Nicola Bartolo $^{(1)}$, Sabino Matarrese $^{(2,3)}$ and Antonio Riotto$^{(3)}$'
title: 'Gauge-Invariant Temperature Anisotropies and Primordial Non-Gaussianity'
---
[H]{}
Inflation has become the dominant paradigm to understand the initial conditions for the density perturbations in the early Universe which are the seeds for the Large-Scale Structure (LSS) and for the Cosmic Microwave Background (CMB) temperature anisotropies [@lrreview]. In the inflationary picture, primordial density and gravity-wave fluctuations are created from quantum fluctuations “redshifted” out of the horizon during an early period of superluminal Universe expansion. Despite the simplicity of the inflationary paradigm, the mechanism by which cosmological curvature (adiabatic) perturbations are generated is not yet established. In the standard slow-roll inflationary scenario associated to one-single field, the inflaton, density perturbations are due to fluctuations of the inflaton itself when it slowly rolls down along its potential. In the curvaton mechanism [@curvaton] the final curvature perturbation $\zeta$ is produced from an initial isocurvature mode associated with the quantum fluctuations of a light scalar (other than the inflaton), the curvaton, whose energy density is negligible during inflation. Recently, other mechanisms for the generation of cosmological perturbations have been proposed, the inhomogeneous reheating scenario [@gamma1], ghost-inflation [@ghost] and the D-cceleration scenario [@dacc], to mention a few. A precise measurement of the spectral index $n_\zeta$ of comoving curvature perturbations will provide a powerful constraint to slow-roll inflation models and the standard scenario for the generation of cosmological perturbations which predicts $|n_\zeta-1|$ significantly below unity. However, alternative mechanisms generically also predict a value of $n_\zeta$ very close to unity. Thus, even a precise measurement of the spectral index will not allow us to efficiently discriminate among them. On the other hand, the lack gravity-wave signals in CMB anisotropies will not give us any information about the perturbation generation mechanism, since alternative mechanisms predict an amplitude of gravity waves far too small to be detectable by future experiments aimed at observing the $B$-mode of the CMB polarization.
There is, however, a third observable which will prove fundamental in providing information about the mechanism chosen by Nature to produce the structures we see today. It is the deviation from a Gaussian statistics, [*i.e.*]{}, the presence of higher-order connected correlation functions of CMB anisotropies. Since for every scenario there exists a well defined prediction for the strength of Non-Gaussianity (NG) and its shape as a function of the parameters, testing the NG level of primordial fluctuations is one of the most powerful probes of inflation [@review] and is crucial to discriminate among different – but otherwise indistinguishable - mechanisms. For instance, the single-field slow-roll inflation model itself produces negligible NG, and the dominant contribution comes from the evolution of the ubiquitous second-order perturbations after inflation, which is potentially detectable with future observations of CMB temperature and polarization anisotropies. This effect [*must exist*]{} regardless of the inflationary models, setting the minimum NG level of cosmological perturbations. Therefore, if we do not find any evidence for this ubiquitous NG, then it will challenge our understanding of the evolution of cosmological perturbations at a deeper level.
Motivated by the extreme relevance of pursuing NG in the CMB anisotropies, in this Letter we provide the exact expression for large-scale CMB temperature fluctuations at second order in perturbation theory. This expression has various virtues. First, it is gauge-invariant. Second, from it one can unambiguously extract the exact definition of the nonlinearity parameter $f_{\rm NL}$ which is used by the experimental collaborations to pin down the level of NG in the temperature fluctuations. Third, it contains a “primordial” term encoding all the information about the NG generated in primordial epochs, namely during or immediately after inflation, and depends upon the various fluctuation generation mechanisms. As such, the expression neatly disentangles the primordial contribution to the NG from that arising after inflation. Finally, the expression applies to all scenarios for the generation of cosmological perturbations.
In order to obtain our gauge-independent formula for the temperature anisotropies we first perturb a spatially flat Robertson-Walker background. Here we follow the formalism of Ref. [@MMB] expanding metric perturbations in a first and a second-order part as $$\begin{aligned}
\label{metric1}
g_{00}&=&-a^2 \left( 1+2 \phi^{(1)}+\phi^{(2)} \right)\, ,
g_{0i}=a^2 \left( \hat{\omega}_i^{(1)}+\frac{1}{2}
\hat{\omega}_i^{(2)} \right)\, ,
\nonumber \\
g_{ij}&=&a^2\left[
(1 -2 \psi^{(1)} - \psi^{(2)})\delta_{ij}+
\left( \hat{\chi}^{(1)}_{ij}+\frac{1}{2}\hat{\chi}^{(2)}_{ij} \right)\right]
\, ,\end{aligned}$$ where the scale factor $a(\eta)$ is a function of the conformal time $\eta$. The functions $\phi^{(r)}, \hat{\omega}_i^{(r)},
\psi^{(r)}$ and $\hat{\chi}^{(r)}_{ij}$, where $(r)=(1,2)$, stand for the $r$th-order perturbations of the metric. It is standard use to split the perturbations into the so-called scalar, vector and tensor parts, according to their transformation properties with respect to the $3$-dimensional space with metric $\delta_{ij}$, where scalar parts are related to a scalar potential, vector parts to transverse (divergence-free) vectors and tensor parts to transverse trace-free tensors. Thus $\phi$ and $\psi$ are scalar perturbations, and for instance, $\hat{\omega}_i^{(r)}=\partial_i\omega^{(r)}+\omega_i^{(r)}$, where $\omega^{(r)}$ is the scalar part and $\omega^{(r)}_i$ is a transverse vector, [*i.e.*]{} $\partial^i\omega^{(r)}_i=0$. The metric perturbations will transform according to an infinitesimal change of coordinates. From now on we limit ourselves to a second-order time shift $\eta\rightarrow
\eta-\alpha_{(1)}+\frac{1}{2} ({\alpha'_{(1)}}\alpha_{(1)}
-\alpha_{(2)})$, where a prime denotes differentiation w.r.t. conformal time. In general a gauge corresponds to a choice of coordinates defining a slicing of spacetime into hypersurfaces (at fixed time $\eta$) and a threading into lines (corresponding to fixed spatial coordinates ${\bf x}$), but in this Letter only the former is relevant so that gauge-invariant can be taken to mean independent of the slicing [@mw]. For example, under the time shift, the first-order spatial curvature perturbation $\psi^{(1)}$ transforms as $
\psi^{(1)}\rightarrow\psi^{(1)}-{\mathcal H} \,\alpha_{(1)}$ (here ${\mathcal H}=a'/a$), while $\phi^{(1)} \rightarrow \phi^{(1)}+\alpha'_{(1)}
+{\mathcal H} \alpha^{(1)}$, $\hat{\omega}_i^{(1)} \rightarrow
\hat{\omega}_i^{(1)}-\partial_i \alpha^{(1)}$, and the traceless part of the spatial metric $\hat{\chi}^{(1)}_{ij}$ turns out to be gauge-invariant. At second order in the perturbations we just give some useful examples like the transformation of the energy density and the curvature perturbation [@MMB] $\delta^{(2)} \rho\rightarrow\delta^{(2)} \rho +\rho^\prime\alpha_{(2)} +
\alpha_{(1)}\left(\rho^{\prime\prime}\alpha_{(1)}
+\rho^\prime\alpha_{(1)}^\prime+2\delta^{(1)} \rho^\prime\right)$ and $\psi^{(2)}\rightarrow\psi^{(2)}
+2\alpha_{(1)}
\left(\psi^{(1)\prime}+
2{\mathcal H}\psi^{(1)}\right)
-\left({\mathcal H}'+2{\mathcal H}^2\right)\alpha^2_{(1)}
-{\mathcal H} \alpha_{(1)} \alpha_{(1)}^{\prime}
- \frac{1}{3}\left(2\hat{\omega}^{i}_{(1)}-
\alpha^{,i}_{(1)}\right)\alpha^{(1)}_{,i} -{\mathcal H} \alpha_{(2)}\,$ . In particular, there exists an extension at second order of the well-known gauge-invariant variable $\zeta^{(1)}=-\psi^{(1)}-{\mathcal H} \frac{\delta^{(1)} \rho}{\rho'}$ (the curvature perturbation on uniform density hypersurfaces). It is given by $\zeta=\zeta^{(1)}+(1/2)\zeta^{(2)}$, where [@lw; @mw] $$\begin{aligned}
\label{qqq}
&-&\zeta^{(2)}=
\psi^{(2)}+{\mathcal H}\frac{\delta^{(2)}\rho}{\rho^\prime}
-2{\mathcal H}\frac{\delta^{(1)}\rho^\prime}{\rho^\prime}
\frac{\delta^{(1)}\rho}{\rho^\prime}
-2\frac{\delta^{(1)}\rho}{\rho^\prime}{\psi}^{(1)\prime}
\nonumber \\
&-&4 {\mathcal H} \frac{\delta^{(1)}\rho}{\rho^\prime}{\psi}^{(1)}
+\left(\frac{\delta^{(1)} \rho}{\rho'}\right)^2
\left({\mathcal H} \frac{\rho^{\prime\prime}}{\rho^\prime}-
{\mathcal H}^\prime-2{\mathcal H}^2\right) \, .\end{aligned}$$ The key point here is that the gauge-invariant comoving curvature perturbation $\zeta^{(2)}$ remains [*constant*]{} on superhorizon scales after it has been generated and possible isocurvature perturbations are no longer present. Therefore, $\zeta^{(2)}$ provides all the necessary information about the “primordial” level of NG generated either during inflation, as in the standard scenario, or immediately after it, as in the curvaton scenario. Different scenarios are characterized by different values of $\zeta^{(2)}$, while the post-inflationary nonlinear evolution due to gravity is common to all of them [@BMR2; @BMR3; @BMR4; @review]. For example, in standard single-field inflation $\zeta^{(2)}$ is generated during inflation and its value is $\zeta^{(2)}=2\left(
\zeta^{(1)} \right)^2+{\cal O}\left(n_\zeta-1\right) $ [@ABMR; @BMR2].
We now construct in a gauge-invariant way temperature anisotropies at second order. Temperature anisotropies beyond the linear regime have been calculated in Refs. [@T2nd], following the photons path from last-scattering to the observer in terms of perturbed geodesics. The linear temperature anisotropies read [@T2nd] $$\label{T1}
\frac{\Delta T^{(1)}}{T}=\phi^{(1)}_{\mathcal E} -v^{(1)i}_{\mathcal E}e_i
+ \tau^{(1)}_{\mathcal E}
-\int_{\lambda_{\mathcal O}}^{\lambda_{\mathcal E}} d\lambda A^{(1) \prime} \;,$$ where $A^{(1)}\equiv \psi^{(1)}+\phi^{(1)}+\hat{\omega}^{(1)}_i e^i-
\frac{1}{2}\hat{\chi}^{(1)}_{ij}e^i e^j$, the subscript ${\mathcal E}$ indicates that quantities are evaluated at last-scattering, $e^i$ is a spatial unit vector specifying the direction of observation and the integral is evaluated along the line-of-sight parametrized by the affine parameter $\lambda$. Eq. (\[T1\]) includes the intrinsic fractional temperature fluctuation at emission $\tau_{\mathcal E}$, the Doppler effect due to emitter’s velocity $v^{(1)i}_{\mathcal E}$ and the gravitational redshift of photons, including the Integrated Sachs-Wolfe (ISW) effect. We omitted monopoles due to the observer ${\mathcal O}$ ([*e.g.*]{} the gravitational potential $\psi^{(1)}_{\mathcal O}$ evaluated at the event of observation), which, being independent of the angular coordinate, can be always recast into the definition of temperature anisotropies [@HwangNoh]. Notice however that the physical meaning of each contribution in Eq. (\[T1\]) is not gauge-invariant, as the different terms are gauge-dependent. However, it is easy to show that the whole expression (\[T1\]) is gauge-invariant. Since the temperature $T$ is a scalar, the intrinsic temperature fluctuation transforms as $\tau_{\mathcal E}^{(1)}\rightarrow\tau_{\mathcal E}^{(1)}
+(T'/T)\alpha_{(1)}=\tau_{\mathcal E}^{(1)}
-{\mathcal H}\alpha_{(1)}$, having used the fact that the temperature scales as $T \propto a^{-1}$. Notice, instead, that the velocity $v^{(1)i}_{\mathcal E}$ does not change. Therefore, using the transformations of metric perturbations we find $$\begin{aligned}
\label{proof2}
\frac{\Delta T^{(1)}}{T} &\rightarrow &\frac{\Delta T^{(1)}}{T}+\alpha'_{(1)}
-\int_{\eta_{\mathcal O}}^{\eta_{\mathcal E}}
d\eta \frac{d \alpha'_{(1)}}{d \eta}
=\frac{\Delta T^{(1)}}{T} + {\mathcal O}\, , \nonumber \\
&& \end{aligned}$$ where we have used the fact that the integral is evaluated along the line-of-sight which can be parametrized by the background geodesics $x^{(0) \mu}=
\left( \lambda, (\lambda_{\mathcal O}-\lambda_{\mathcal E}) e^i \right)$ (with $d \lambda/d \eta=1$), and the decomposition for the total derivative along the path for a generic function $f(\lambda,x^i(\lambda))$, $f'=\frac{\partial f}{\partial \lambda}=
\frac{d f}{d \lambda} + \partial_i f e^i$. Eq. (\[proof2\]) shows that the expression (\[T1\]) for first-order temperature anisotropies is indeed gauge-invariant (up to monopole terms related to the observer ${\mathcal O}$). Temperature anisotropies can be easily written in terms of particular combinations of perturbations which are manifestly gauge-invariant. For the gravitational potentials we consider the gauge-invariant definitions $\psi^{(1)}_{\rm GI}=\psi^{(1)}-{\mathcal H} \omega^{(1)}$ and $\phi^{(1)}_{\rm GI}=\phi^{(1)}+{\mathcal H} \omega^{(1)}+\omega^{(1)'}$. For the $(0-i)$ component of the metric and the traceless part of the spatial metric we define $\omega_i^{(1) \rm GI}=\omega_i^{(1)}$ and $
\hat{\chi}^{(1) \rm GI}_{ij}=\hat{\chi}^{(1)}_{ij}$. For the matter variables we use a gauge-invariant intrinsic temperature fluctuation $\tau^{(1)}_{\rm GI}=\tau^{(1)}- {\mathcal H}\omega^{(1)}$, while the velocity itself is gauge-invariant $v^{(1)i}_{\rm GI}=v^{(1)i}$ under time shifts. Following the same steps leading to Eq. (\[proof2\]) one gets the linear temperature anisotropies in Eq. (\[T1\]) in terms of these gauge-invariant quantities $$\begin{aligned}
\label{T1GI}
& &\frac{\Delta T^{(1)}_{\rm GI}}{T}=\phi^{(1)}_{\rm GI} -v^{(1)i}_{\rm GI}e_i
+ \tau^{(1)}_{\rm GI} -\int_{\lambda_{\mathcal O}}^{\lambda_{\mathcal E}}
d \lambda\, A^{(1) \prime}_{\rm GI}\, ,\end{aligned}$$ where $A^{(1)}_{\rm GI}=
\phi^{(1)}_{\rm GI}+\psi^{(1)}_{\rm GI}+\omega_i^{(1) \rm GI}
e_i-\frac{1}{2}\hat{\chi}^{(1) \rm GI}_{ij} e^i
e^j$ and we omitted the subscript ${\mathcal E}$. For the primordial fluctuations we are interested in the large-scale modes set by the curvature perturbation $\zeta^{(1)}$. Defining a gauge-invariant density perturbation $\delta^{(1)}\rho_{\rm GI}=
\delta^{(1)} \rho+\rho' \omega^{(1)}$, we write the curvature perturbation as $\zeta^{(1)}_{\rm GI}=-\psi^{(1)}_{\rm GI}-
{\mathcal H} (\delta^{(1)} \rho_{\rm GI}/\rho')$. Since for adiabatic perturbations in the radiation ($\gamma$) and matter ($m$) eras $(1/4)(\delta^{(1)} \rho_\gamma/\rho_{\gamma}) = (1/3)
(\delta^{(1)} \rho_m/\rho_{m})$, one can write the intrinsic temperature fluctuation as $\tau^{(1)}=(1/4)(\delta^{(1)} \rho_\gamma/\rho_{\gamma})=
-{\mathcal H} (\delta^{(1)} \rho/\rho')$ and a gauge-invariant definition is $\tau^{(1)}_{\rm GI}=-{\mathcal H} (\delta^{(1)}
\rho_{\rm GI}/\rho')$. In the large-scale limit, from Einstein equations, in the matter era $\phi^{(1)}_{\rm GI}=\psi^{(1)}_{\rm GI}=-\frac{3}{5}
\zeta^{(1)}_{\rm GI}$. Thus we obtain the large-scale limit of temperature anisotropies (\[T1GI\]) $\frac{\Delta T^{(1)}_{\rm GI}}{T}=
2\psi^{(1)}_{\rm GI} +\zeta^{(1)}_{\rm GI}=
\psi^{(1)}_{\rm GI}/3$, i.e. the usual Sachs-Wolfe effect.
At second order, the procedure is similar to the one described so long, though more lengthy and cumbersome. We only provide the reader with the main steps to get the final expression. The second-order temperature fluctuations in terms of metric perturbations read [@T2nd] $$\begin{aligned}
\label{T2}
& &\frac{\Delta T^{(2)}}{T}=\frac{1}{2}\phi^{(2)}_{\mathcal{E}}
-\frac{1}{2}\left( \phi_{\mathcal{E}}^{(1)} \right)^2
-\frac{1}{2}v^{(2)i}_{\mathcal E}e_i
+ \frac{1}{2}\tau^{(2)}_{\mathcal E}
-I_2(\lambda_{\mathcal{E}}) \nonumber\\
&+&\left(I_1(\lambda_{\mathcal{E}})+v^{(1)i}_{\mathcal{E}} e_i\right)
\left(-\phi^{(1)}_{\mathcal{E}}-\tau^{(1)}_{\mathcal E}
+v^{(1)i}_{\mathcal{E}} e_i
+I_1(\lambda_{\mathcal{E}})\right)\nonumber\\
&+&x^{(1)0}_{\mathcal{E}} A ^{(1)'}_{\mathcal{E}}+(x^{(1)j}_{\mathcal{E}}
+x^{(1)0}_{\mathcal{E}} e^j)\left(\phi^{(1)}_{,j}-v^{(1)}_{i,j} e^i+
\tau^{(1)}_{,j}\right)_{\mathcal{E}} \nonumber\\
&-&\frac{1}{2}v^{(1)}_{{\mathcal{E}}i} v^{(1)i}_{\mathcal{E}}
+\phi^{(1)}_{\mathcal{E}}\tau^{(1)}_{\mathcal E}+
\frac{\partial \tau^{(1)}}{\partial d^i}d^{(1)i}
-v^{(1)i}_{\mathcal{E}} e_i \phi^{(1)}_{\mathcal{E}}
\nonumber\\
&+&v^{(1)}_{{\mathcal{E}}i}\left(-\hat{\omega}^{(1)i}_{\mathcal{E}}
- I_1^i(\lambda_{\mathcal {E}})\right)\, . \end{aligned}$$ Here $I_2$ is the second-order ISW [@T2nd] $I_2(\lambda_{\mathcal{E}})=\int_{\lambda_{\mathcal{O}}}
^{\lambda_{\mathcal{E}}} d\lambda [\frac{1}{2}A^{(2)'}
-(\hat{\omega}^{(1)'}_i-\hat{\chi}^{(1)'}_{ij} e^j)
(k^{(1)i}+e^i k^{(1)0})+2 k^{(1)0} A^{(1)'} +2 \psi^{(1)'} A^{(1)}
+x^{(1)0} A^{(1)''}+x^{(1)i} A^{(1)'}_{,i}]$, where $A^{(2)}\equiv \psi^{(2)}+\phi^{(2)}+\hat{\omega}^{(2)}_i e^i-
\frac{1}{2}\hat{\chi}^{(2)}_{ij}e^i e^j$, while $k^{(1)0}(\lambda)=-2 \phi^{(1)}
-\hat{\omega}^{(1)i} e_i +I_1(\lambda)$ and $k^{(1)i}(\lambda)=-2 \phi^{(1)} e^i-
\hat{\omega}^{(1)i}+\hat{\chi}^{(1)ij} e_j
- I_1^i(\lambda)$ are the photon wave vectors, with $I_1(\lambda)$ given by the integral in Eq. (\[T1\]) and $I_1^i(\lambda)$ is obtained from the same integral replacing the time derivative with a spatial gradient. Finally in Eq. (\[T2\]) $x^{(1)0}(\lambda)=
\int_{\lambda_{\mathcal{O}}}^{\lambda}d\lambda'
\left[-2 \phi^{(1)}-\hat{\omega}^{(1)}_i e^i+(\lambda-\lambda')
A^{(1)'}\right]$ and $x^{(1)i}(\lambda)=
-\int_{\lambda_{\mathcal{O}}}^{\lambda}d\lambda'
\left[2 \psi^{(1)} e^i+\hat{\omega}^{(1)i}-\hat{\chi}^{(1)ij} e_j
+(\lambda-\lambda')A^{(1),i}\right]$ are the geodesics at first order, and $d^{(1)i}=e^i-\frac{e^i-k^{(1)i}}{|e^i-k^{(1)i}|}$ is the direction of the photon emission. As usual we have omitted the monopole terms due to the observer. Using the transformation rules of Ref. [@MMB], it is possible to check that the expression (\[T2\]) is gauge-invariant. We can express the second-order anisotropies in terms of explicitly gauge-invariant quantities, whose definition proceeds as for the linear case, by choosing the shifts $\alpha^{(r)}$ such that $\omega^{(r)}=0$. For example, we consider the gauge-invariant gravitational potential [@BMR4] $$\begin{aligned}
\label{PhiGI}
&\phi^{(2)}_{\rm GI}&=\phi^{(2)}+\omega^{(1)}\left[2\left(
\psi^{(1)'}+2\frac{a'}{a}\psi^{(1)}\right)+\omega^{(1) \prime \prime}
+ 5 \frac{a'}{a}\omega^{(1) '}\right.\nonumber \\
&+&\left. \left( {\mathcal H}'+2 {\mathcal H}^2 \right)
\omega^{(1)}\right]
+2\omega^{(1)'}\left(2\psi^{(1)}+\omega^{(1)'}\right)+
\frac{1}{a} \left( a\alpha^{(2)} \right)' \, , \nonumber \\\end{aligned}$$ where $\alpha^{(2)}=\omega^{(2)}+\omega^{(1)}\omega^{(1)'}
+\nabla^{-2}\partial^i[-4\psi^{(1)}\partial_i\omega^{(1)}-2
\omega^{(1)'}\partial_i\omega^{(1)}]$. Expressing the second-order temperature anisotropies (\[T2\]) in terms of our gauge-invariant quantities and taking the large-scale limit we find $\Delta T^{(2)}_{\rm GI}/T=(1/2)\phi^{(2)}_{\rm GI}
-(1/2)\left( \phi_{\rm GI}^{(1)} \right)^2
+ (1/2)\tau^{(2)}_{\rm GI}+\phi^{(1)}_{\rm GI} \tau^{(1)}_{\rm GI}$ (having dropped the subscript ${\mathcal E}$), and the gauge-invariant intrinsic temperature fluctuation at emission is $\tau^{(2)}_{\rm GI}=(1/4)
(\delta^{(2)} \rho^{\rm GI}_\gamma/\rho_\gamma)
-3( \tau^{(1)}_{\rm GI})^2$. We have dropped those terms which represent integrated contributions and other second-order small-scale effects that can be distinguished from the large-scale part through their peculiar scale dependence. At this point we make use of Einstein’s equations. We take the expression for $\zeta^{(2)}$ in Eq. (\[qqq\]), and we use the $(0-0)$ component and the traceless part of the $(i-j)$ Einstein’s equation at second order (see Eqs. (153) and (155) of Ref. [@review]). Thus, on large scales we find that the temperature anisotropies are given by $$\label{main}
\frac{\Delta T^{(2)}_{\rm GI}}{T}=
\frac{1}{18} \left( \phi^{(1)}_{\rm GI} \right)^2
-\frac{{\mathcal K}}{10}-\frac{1}{10} \left[ \zeta^{(2)}_{\rm GI}-
2 \left( \zeta^{(1)}_{\rm GI} \right)^2 \right]\, ,$$ where we have defined a kernel ${\mathcal K}=
10 \nabla^{-4} \partial_i \partial^j
(\partial^i \psi^{(1)} \partial_j
\psi^{(1)}) -\nabla^{-2}
( \frac{10}{3} \partial^i \psi^{(1)} \partial_i \psi^{(1)} )$. Eq. (\[main\]) is the main result of this Letter. It clearly shows that there are two contributions to the final nonlinearity in the large-scale temperature anisotropies. The contribution, $[\zeta^{(2)}_{\rm GI}-
2 ( \zeta^{(1)}_{\rm GI} )^2]$, comes from the “primordial” conditions set during or after inflation. They are encoded in the curvature perturbation $\zeta$ which remains constant once it has been generated. The remaining part of Eq. (\[main\]) describes the post-inflation processing of the primordial non-Gaussian signal due to the nonlinear gravitational dynamics, including also second-order corrections at last scattering to the Sachs-Wolfe effect [@T2nd]. Thus, the expression in Eq. (\[main\]) allows to neatly disentangle the primordial contribution to NG from that coming from that arising after inflation. While the nonlinear evolution after inflation is the same in each scenario, the primordial content will depend on the particular mechanism generating the perturbations. We parametrize the primordial NG in the terms of the conserved curvature perturbation (in the radiation or matter dominated epochs) $\zeta^{(2)}=2a\left(\zeta^{(1)}\right)^2$, where $a$ depends on the physics of a given scenario. For example, in the curvaton case $a=(3/4r)-r/2$, where $r \approx (\rho_\sigma/\rho)_{\rm D}$ is the relative curvaton contribution to the total energy density at curvaton decay [@review]. In the minimal picture for the inhomogeneous reheating scenario, $a=1/4$. For the other scenarios we refer the reader to Ref.[@review]. From Eq. (\[main\]) we can extract the nonlinearity parameter $f_{\rm NL}$ which is usually adopted to phenomenologically parametrize the NG level of cosmological perturbations and has become the standard quantity to be observationally constrained by CMB experiments [@ks; @k]. The definition of $f_{\rm NL}$ adopted in the analyses performed in Refs. [@ks; @k] goes through the conventional Sachs-Wolfe formula $\Delta T/T= - \Phi/3$ where $\Phi$ is Bardeen’s potential [@Bardeen80], which is conventionally expanded as (up to a constant offset, which only affects the temperature monopole) $\Phi = \Phi_{\rm L} + f_{\rm NL} * \left(\Phi_{\rm L}\right)^2$, with $\Phi_{\rm L} = - \phi^{(1)}_{\rm GI}$. Here the $\star$ product reminds the fact that the nonlinearity parameter might have a non-trivial scale dependence [@review]. Therefore, using $\zeta^{(1)}=-\frac{5}{3}
\psi^{(1)}_{\rm GI}$ during matter domination, from Eq. (\[main\]) we read the nonlinearity parameter in momentum space $$\label{f_NL}
f_{\rm NL}({\bf k}_1,{\bf k}_2)
=-\left[ \frac{5}{3} \left(1-a \right)
+\frac{1}{6}-\frac{3}{10} {\mathcal K}
\right]+1\,$$ where ${\mathcal K}=10\, ({\bf k}_1 \cdot {\bf k}_3)
({\bf k}_2 \cdot {\bf k}_3)/k^4 -\frac{10}{3}
{\bf k}_1 \cdot {\bf k}_2/k^2$ with ${\bf k}_3+{\bf k}_1+{\bf k_2}=0$ and $k=\left|
{\bf k}_3\right|$. In fact the formula (\[f\_NL\]) already accounts for an additional nonlinear effect entering in the CMB angular $3$-point function from the angular averaging performed with a perturbed line element $d \Omega (1-2 \psi^{(1)}_{\rm GI})$ [@review], implying a $+1$ shift in $f_{\rm NL}$. In particular within the standard scenario where cosmological perturbations are due to the inflaton the primordial contribution to NG is given by $a=1-\frac{1}{4} (n_{\zeta}-1)$ [@ABMR; @BMR2], where the spectral index is expressed in terms of the usual slow-roll parameters as $n_{\zeta}-1=-6 \epsilon +2 \eta$ [@lrreview]. The nonlinearity parameter from inflation now reads $$f^{\rm inf}_{\rm NL}=-\frac{5}{12} (n_{\zeta}-1)
+\frac{5}{6}+\frac{3}{10} {\mathcal K}\, .$$ Therefore the main NG contribution comes from the post-inflation evolution of the second-order perturbations which give rise to order-one coefficients, while the primordial contribution is proportional to $|n_{\zeta}-1|\ll 1$. This is true even in the “squeezed” limit first discussed by Maldacena [@Maldacena], where one of the wavenumbers is much smaller than the other two, *e.g.* $k_1 \ll k_{2,3}$ and ${\cal K}\rightarrow 0$.
[*Acknowledgments.*]{} We thank J. Peebles for spurring our efforts in disentangling the primordial (inflationary) NG in CMB anisotropies.
[99]{}
D. H. Lyth and A. Riotto, Phys. Rept. 314 (1999) 1.
K. Enqvist and M. S. Sloth, Nucl. Phys. B 626 (2002) 395; D. Lyth and D. Wands, Phys. Lett. B 524 (2002) 5; T. Moroi and T. Takahashi, Phys. Lett. B 522 (2001) 215 \[Erratum-ibid. B 539 (2002) 303\].
G. Dvali, A. Gruzinov and M. Zaldarriaga, Phys. Rev. D 69 (2004) 023505.
N. Arkani-Hamed, H. C. Cheng, M. A. Luty and S. Mukohyama, arXiv:hep-th/0312099.
E. Silverstein and D. Tong, arXiv:hep-th/0310221.
N. Bartolo, E. Komatsu, S. Matarrese and A. Riotto, arXiv:astro-ph/0406398, to apperar in Phys. Rept. S. Matarrese, S. Mollerach and M. Bruni, Phys. Rev. D 58 (1998) 043504.
K. A. Malik and D. Wands, Class. Quant. Grav. 21 (2004) L65.
D. H. Lyth and D. Wands, Phys. Rev. D 68 (2003) 103515.
N. Bartolo, S. Matarrese and A. Riotto, JHEP 0404 (2004) 006.
N. Bartolo, S. Matarrese and A. Riotto, Phys. Rev. D 69 (2004) 043503.
N. Bartolo, S. Matarrese and A. Riotto, JCAP 0401 (2004) 003.
V. Acquaviva, N. Bartolo, S. Matarrese and A. Riotto, Nucl. Phys. B [**667**]{} (2003) 119.
T. Pyne and S. M. Carroll, Phys. Rev. D 53 (1996) 2920; S. Mollerach and S. Matarrese, Phys. Rev. D 56 (1997) 4494.
J. Hwang and H. Noh, Phys. Rev. D 59 (1999) 067302.
E. Komatsu and D. N. Spergel, Phys. Rev. D 63 (2001) 063002.
E. Komatsu, et al., Astrophys. J. Suppl. 148 (2003) 119.
J. M. Bardeen, Phys. Rev. D 22, (1980) 1882.
J. Maldacena, JHEP 0305 (2003) 013.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce a hybrid stochastic estimator to design stochastic gradient algorithms for solving stochastic optimization problems. Such a hybrid estimator is a convex combination of two existing biased and unbiased estimators and leads to some useful property on its variance. We limit our consideration to a hybrid SARAH-SGD for nonconvex expectation problems. However, our idea can be extended to handle a broader class of estimators in both convex and nonconvex settings. We propose a new single-loop stochastic gradient descent algorithm that can achieve $\BigO{\max\set{\sigma^3\varepsilon^{-1},\sigma\varepsilon^{-3}}}$-complexity bound to obtain an $\varepsilon$-stationary point under smoothness and $\sigma^2$-bounded variance assumptions. This complexity is better than $\BigO{\sigma^2\varepsilon^{-4}}$ often obtained in state-of-the-art SGDs when $\sigma < \BigO{\varepsilon^{-3}}$. We also consider different extensions of our method, including constant and adaptive step-size with single-loop, double-loop, and mini-batch variants. We compare our algorithms with existing methods on several datasets using two nonconvex models.'
author:
- |
Quoc Tran-Dinh$^{\dagger}$, Nhan H. Pham$^{\dagger}$, Dzung T. Phan$^{\ddagger}$, *and* Lam M. Nguyen$^{\ddagger}$\
$^{\dagger}$Department of Statistics and Operations Research\
The University of North Carolina at Chapel Hill, Chapel Hill, NC27599, USA.\
`[email protected], [email protected]`\
- |
$^{\ddagger}$IBM Research, Thomas J. Watson Research Center\
Yorktown Heights, NY10598, USA.\
`[email protected], [email protected]`
title: Hybrid Stochastic Gradient Descent Algorithms for Stochastic Nonconvex Optimization
---
Introduction {#sec:intro}
============
Consider the following stochastic nonconvex optimization problem of the form: $$\label{eq:ncvx_prob}
\min_{x\in\R^p}\Big\{ f(x) := \Exps{\xi}{f(x;\xi)} \Big\},$$ where $f(\cdot;\cdot) : \R^p\times \Omega \to \R$ is a stochastic function defined such that for each $x\in\R^p$, $f(x;\cdot)$ is a random variable in a given probability space $(\Omega, \Probn)$, while for each realization $\xi\in\Omega$, $f(\cdot;\xi)$ is smooth on $\R^p$; and $\Exps{\xi}{f(x;\xi)}$ is the expectation of $f(x;\xi)$ w.r.t. $\xi$ over $\Omega$.
#### Our goals and assumptions:
Since is nonconvex, our goal in this paper is to develop a new class of stochastic gradient algorithms to find an $\varepsilon$-approximate stationary point $\widetilde{x}_T$ of such that $\Exp{\norms{\nabla{f}(\widetilde{x}_T)}^2} \leq \varepsilon^2$ under mild assumptions as stated in Assumption \[as:A1\].
\[as:A1\] The objective function $f$ of satisfies the following conditions:
- (**Boundedness from below**) There exists a finite lower bound $f^{\star} := \inf_{x\in\R^p}f(x) > -\infty$.
- (**$L$-average smoothness**) The function $f(\cdot;\xi)$ is $L$-average smooth on $\R^p$, i.e. there exists $L\in (0, +\infty)$ such that $$\label{eq:L_smooth}
\Exps{\xi}{\norm{\nabla{f}(x;\xi) - \nabla{f}(y;\xi)}^2} \leq L^2\norms{x - y}^2,~~\forall x, y\in\R^p.
\vspace{-1.5ex}$$
- (**Bounded variance**) There exists $\sigma \in (0, \infty)$ such that $$\label{eq:bounded_variance2}
\Exps{\xi}{\norms{\nabla f(x;\xi) - \nabla{f}(x)}^2} \leq \sigma^2, ~~~\forall x\in\R^p.$$
These assumptions are very standard in stochastic optimization methods [@ghadimi2013stochastic; @Nemirovski2009a]. The $L$-average smoothness of $f$ is weaker than the smoothness of $f$ for each realization $\xi\in\Omega$. Note that our methods described in the sequel are also applicable to the finite-sum problem $\min_{x} \set{f(x) = \frac{1}{n}\sum_{i=1}^nf_i(x)}$ as long as the above assumptions hold. However, we do not specify our methods to solve this problem. In this case, $\sigma$ in can be replaced by other alternatives, e.g., $\sigma_n^2 := \frac{1}{n}\sum_{i=1}^n\left[\norms{\nabla{f}_i(x)}^2 - \norms{\nabla{f}(x)}^2\right]$.
#### Our key idea:
Different from existing methods, we introduce a convex combination of a biased and unbiased estimator of the gradient $\nabla{f}$ of $f$, which we call a *hybrid stochastic gradient* estimator. While the biased estimator exploited in this paper is SARAH in [@Nguyen2017_sarah], the unbiased one can be any unbiased estimator. SARAH is a recursive biased and variance reduced estimator for $\nabla{f}$. Combining it with an unbiased estimator allows us to reduce the bias and variance of the hybrid estimator. In this paper, we only focus on the standard stochastic estimator as an unbiased candidate.
#### Related work:
Under Assumption \[as:A1\], problem covers a large number of applications in machine learning and data sciences. The stochastic gradient descent (SGD) method was first studied in [@RM1951], and becomes extremely popular in recent years. [@Nemirovski2009a] seems to be the first work showing the convergence rates of robust SGD variants in the convex setting, while [@Nemirovskii1983] provides an intensive complexity analysis for many optimization algorithms, including stochastic methods. Variance reduction methods have also been widely studied, see, e.g. [@allen2016katyusha; @chambolle2017stochastic; @SAGA; @johnson2013accelerating; @nitanda2014stochastic; @reddi2016stochastic; @schmidt2017minimizing; @shalev2013stochastic; @Xiao2014]. In the nonconvex setting, [@ghadimi2013stochastic] seems to be the first algorithm achieving $\BigO{\sigma^2\varepsilon^{-4}}$-complexity bound. Other researchers have also made significant progress in this direction, including [@allen2017natasha2; @allen2018neon2; @SVRG++; @fang2018spider; @lei2017non; @Nguyen2018_iSARAH; @Nguyen2019_SARAH; @Nguyen2017_sarahnonconvex; @reddi2016proximal; @wang2018spiderboost; @zhou2018stochastic]. A majority of these works, including [@allen2017natasha2; @allen2018neon2; @SVRG++; @lei2017non; @reddi2016proximal], rely on SVRG estimator in order to obtain better complexity bounds. Hitherto, the complexity of SVRG-based methods remains worse than the best-known results, which is obtained in [@fang2018spider; @pham2019proxsarah; @wang2018spiderboost] via the SARAH estimator. However, as discussed in [@pham2019proxsarah; @wang2018spiderboost], the method called SPIDER in [@fang2018spider; @lei2017non] does not practically perform well due to small step-size and its dependence on the reciprocal of the estimator’s norm. [@wang2018spiderboost] amends this issue by using a large constant step-size, but requires large mini-batch and does not consider the single sample case and single loop variants. [@pham2019proxsarah] provides a more general framework to treat composite problems where it covers as special case, but it does not consider the single loop as in SGDs.
#### Our contribution:
To this end, our contribution can be summarized as follows:
- We propose a hybrid stochastic estimator for a stochastic gradient of a nonconvex function $f$ in by combining the SARAH estimator from [@Nguyen2017_sarah] and any unbiased stochastic estimator such as SGD and SVRG. However, we only focus on the SGD estimator in this paper. We prove some key properties of this hybrid estimator that can be used to design new algorithms.
- We exploit our hybrid estimator to develop a single-loop SGD algorithm that can achieve an $\varepsilon$-stationary point $\widetilde{x}_m$ such that $\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \varepsilon^2$ in at most $\BigO{\sigma\varepsilon^{-3} + \sigma^3\varepsilon^{-1}}$ stochastic gradient evaluations. This complexity significantly improves $\BigO{\sigma^2\varepsilon^{-4}}$ of SGD if $\sigma < \BigO{\varepsilon^{-3}}$. We extend our algorithm to a double loop variant, which requires $\BigO{\max\set{ \sigma\varepsilon^{-3}, \sigma^2\varepsilon^{-2}}}$ stochastic gradient evaluations. This is the best-known complexity in the literature for stochastic gradient-type methods for solving .
- We also investigate other variants of our method, including adaptive step-sizes, and mini-batches. In all these cases, our methods achieve the best-known complexity bounds.
Let us emphasize the following points of our contribution. Firstly, although our single-loop method requires three gradients per iteration compared to standard SGDs, it can achieves better complexity bound. Secondly, it can be cast into a variance reduction method where it starts from a “good” approximation $v_0$ of $\nabla{f}(x^0)$, and aggressively reduces the variance. Thirdly, our step-size is $\eta = \BigO{m^{-1/3}}$ which is larger than $\eta = \BigO{m^{-1/2}}$ in SGDs. Fourthly, the step-size of the adaptive variant is increasing instead of diminishing as in SGDs. Finally, our method achieves the same best-known complexity as in variance reduction methods studied in [@fang2018spider; @pham2019proxsarah; @wang2018spiderboost]. We believe that our approach can be extended to other estimators such as SVRG [@johnson2013accelerating] and SAGA [@SAGA], and can be used for Hessians to develop second-order methods as well as to solve convex and composite problems.
#### Paper organization:
The rest of this paper is organized as follows. Section \[sec:stochastic\_estimators\] introduces our new hybrid stochastic estimator for the gradient of $f$ and investigates its properties. Section \[sec:algorithms\] proposes a single-loop hybrid SGD-SARAH algorithm and its complexity analysis. It also considers a double-loop and mini-batch variants with rigorous complexity analysis. Section \[sec:num\_experiments\] provides two numerical examples to illustrate our methods and compares them with state-of-the-art methods. All the proofs and additional experiments can be found in the Supplementary Document.
**Notation:** We work with Euclidean spaces, $\R^p$, equipped with standard inner product $\iprods{\cdot,\cdot}$ and norm $\norm{\cdot}$. For a smooth function $f$ (i.e., $f$ is continuously differentiable), $\nabla{f}$ denotes its gradient. We use $\Ub_{{\mathbf{p}}}(\Sc)$ to denote a distribution on $\Sc$ with probability ${\mathbf{p}}$. If ${\mathbf{p}}$ is uniform, then we simply use $\Ub(\Sc)$. We also use $\BigO{\cdot}$ to present big-O notion in complexity theory, and $\sigma(\cdot)$ to denote a $\sigma$-field.
Hybrid stochastic gradient estimators {#sec:stochastic_estimators}
=====================================
In this section, we propose new stochastic estimators for the gradient of a smooth function $f$.
Let $u_t$ be an unbiased estimator of $\nabla{f}(x_t)$ formed by a realization $\zeta_t$ of $\xi$, i.e. $\Exps{\zeta_t}{u_t} = \nabla{f}(x_t)$. We attempt to develop the following stochastic estimator for $\nabla{f}(x_t)$ in : $$\label{eq:v_estimator}
v_t := \beta_{t-1}v_{t-1} + \beta_{t-1}(\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)) + (1-\beta_{t-1})u_t,$$ where $\xi_t$ and $\zeta_t$ are two independent realizations of $\xi$ on $\Omega$. Clearly, if $\beta_t = 0$, then we obtain a simple unbiased stochastic estimator, and $\beta_t = 1$, we obtain the SARAH estimator in [@Nguyen2017_sarah]. We are interested in the case $\beta_t \in (0, 1)$, in which we call $v_t$ in a **hybrid stochastic** estimator.
Note that we can rewrite $v_t$ as $$v_t := \beta_{t-1}\nabla{f}(x_t;\xi_t) + (1-\beta_{t-1})u_t + \beta_{t-1}(v_{t-1} - \nabla{f}(x_{t-1};\xi_t)).$$ The first two terms are two stochastic gradients estimated at $x_t$, while the third term is the difference $v_{t-1} - \nabla{f}(x_{t-1};\xi_t)$ of the previous estimator and a stochastic gradient at the previous iterate. Here, since $\beta_{t-1} \in (0, 1)$, the main idea is to exploit more recent information than the old ones. In fact, the hybrid estimator $v_t$ covers many other estimators, including SGD, SVRG, and SARAH. We can use one of the following two concrete unbiased estimators $u_t$ of $\nabla{f}(x_t)$ as follows:
- **The SGD estimator:** $u_t := u_t^{\mathrm{sgd}} = \nabla{f}(x_t;\zeta_t)$.
- **The SVRG estimator:** $u_t := u_t^{\mathrm{svrg}} = \nabla{f}(\tilde{x}) + \nabla{f}(x_t;\zeta_t) - \nabla{f}(\tilde{x};\zeta_t)$, where $\nabla{f}(\tilde{x})$ is a full gradient evaluated at a given snapshot point $\tilde{x}$.
However, for the sake of presentation, we only focus on the SGD estimator $u_t := u_t^{\mathrm{sgd}}$.
We first prove the following property of the estimator $v_t$ showing how the variance is estimated.
\[le:key\_estimate10\] Let $v_t$ be defined by . Then $$\label{eq:biased_estimator}
\Exps{(\xi_t,\zeta_t)}{v_t} = \nabla{f}(x_t) + \beta_{t-1}(v_{t-1} - \nabla{f}(x_{t-1})).$$ If $\beta_{t-1} \neq 0$, then $v_t$ is a biased estimator. Moreover, we have $$\label{eq:key_estimate10}
\begin{array}{ll}
\Exps{(\xi_t,\zeta_t)}{\norms{v_t - \nabla{f}(x_t)}^2} &= \beta_{t-1}^2\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2 - \beta_{t-1}^2\norms{\nabla{f}(x_{t-1}) - \nabla{f}(x_t)}^2 \vspace{1ex}\\
& + {~} \beta_{t-1}^2\Exps{\xi_t}{\norms{\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)}^2} \vspace{1ex}\\
& + {~} (1-\beta_{t-1})^2\Exps{\zeta_t}{\norms{u_t - \nabla{f}(x_t)}^2}.
\end{array}$$
From , we can see that $v_t$ remains a biased estimator as long as $\beta_{t-1} \in (0, 1]$. Its biased term is $$\mathrm{Bias}(v_t) = \norms{\Exps{(\xi_t,\zeta_t)}{v_t - \nabla{f}(x_t) \mid \Fc_t}} = \beta_{t-1}\norms{v_{t-1} - \nabla{f}(x_{t-1})} \leq \norms{v_{t-1} - \nabla{f}(x_{t-1})}.$$ This shows that the bias $v_t$ estimator is smaller than the one in the SARAH estimator $v_t^{\textrm{sarah}} := v_{t-1}^{\textrm{sarah}} + \nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)$ from [@Nguyen2017_sarah], which is $\mathrm{Bias}(v^{\mathrm{sarah}}_t) = \norms{v_{t-1}^{\mathrm{sarah}} - \nabla{f}(x_{t-1})}$.
The following lemma bounds the second moment of $v_t - \nabla{f}(x_t)$ with $v_t$ defined in .
\[le:upper\_bound\_new\] Assume that $f(\cdot,\cdot)$ is $L$-smooth and $u_t$ is an SGD estimator. Then, we have the following upper bound on the variance $\Exp{\norms{v_t - \nabla{f}(x_t)}^2}$ of $v_t$: $$\label{eq:vt_variance_bound_new}
\Exp{\norms{v_t - \nabla{f}(x_t)}^2} \leq \omega_t\Exp{\norms{v_0 - \nabla{f}(x^0)}^2} + L^2\sum_{i=0}^{t-1}\omega_{i,t}\Exp{\norms{x_{i+1} - x_{i}}^2} + S_t,$$ where the expectation is taking over all the randomness $\Fc_t := \sigma(v_0, v_1, \cdots, v_t)$, $\omega_{t} := \prod_{i=1}^{t}\beta_{i-1}^2$, $\omega_{i, t} := \prod_{j=i+1}^{t}\beta_{j-1}^2$ for $i=0,\cdots, t$, and $S_{t} := \sum_{i=0}^{t-1}\big(\prod_{j=i+2}^{t}\beta_{j-1}^2\big)(1-\beta_i)^2\sigma_{i+1}^2$ for $t \geq 0$.
Lemmas \[le:key\_estimate10\] and \[le:upper\_bound\_new\] provides two key properties to develop stochastic algorithm in Section \[sec:algorithms\].
Hybrid SARAH-SGD algorithms {#sec:algorithms}
===========================
In this section, we utilize our hybrid stochastic estimator $v_t$ in to develop stochastic gradient methods for solving . We consider three different variants using the hybrid SARAH-SGD estimator.
The generic algorithm framework
-------------------------------
Using $v_t$ defined by , we can develop a new algorithm for solving as in Algorithm \[alg:A1\].
An initial point $x^0$ and parameters $b$, $\beta_t$, and $\eta_t$ (will be specified). \[step:o2\] Generate an unbiased estimator $v_0 := \frac{1}{b}\sum_{\hat{\xi}_i\in\Bc}\nabla{f}(x_0;\hat{\xi}_i)$ at $x_0$ using a mini-batch $\Bc$. \[step:o3\] Update $x_1 := x_0 - \eta_0v_0$. \[step:o4\][**For $t := 1,\cdots,m$ do**]{} \[step:i1\] Generate a proper sample pair $(\xi_t, \zeta_t)$ independently (single sample or mini-batch). \[step:i2\] Evaluate $v_{t} := \beta_{t-1}v_{t-1} + \beta_{t-1}\big(\nabla{f}(x_{t};\xi_t) - \nabla{f}(x_{t-1};\xi_t)\big) + (1-\beta_{t-1})\nabla{f}(x_{t}; \zeta_t)$. \[step:i3\] Update $x_{t+1} :=x_{t} - \eta_t v_{t}$. \[step:o5\] Choose $\widetilde{x}_m$ from $\set{x_0, x_1, \cdots, x_m}$ (at random or deterministic, specified later).
Algorithm \[alg:A1\] looks essentially the same as any SGD scheme with only one loop. The differences are at Step \[step:o2\] with a mini-batch estimator $v_0$ and at Step \[step:i2\], where we use our hybrid gradient estimator $v_t$. In addition, we will show in the sequel that it uses different step-sizes and leads to different variants. Unlike the inner loop of SARAH or SVRG, each iteration of Algorithm \[alg:A1\] requires three individual gradient evaluations instead of two as in these methods. The snapshot at Step \[step:o2\] of Algorithm \[alg:A1\] relies on a mini-batch $\Bc$ of the size $b$, which is independent of $(\xi_t,\zeta_t)$ in the loop $t$.
Convergence analysis
--------------------
We analyze two cases: constant step-size and adaptive step-size. In both cases, $\beta_t$ is fixed for all $t$.
### Convergence of Algorithm \[alg:A1\] with constant step-size $\eta$ and constant $\beta$
Assume that we run Algorithm \[alg:A1\] within $m$ iterations $m\geq 1$. In this case, given $0 < c_1 < \sqrt{b(m+1)}$, we choose $\eta$ and $\beta$ in Algorithm \[alg:A1\] as follows: $$\label{eq:step_size1}
\eta := \frac{2}{L(\sqrt{1 + 4\alpha_m^2} + 1)}~~~~~~\text{with}~~~~~\beta := 1 - \frac{c_1}{\sqrt{b(m+1)}}~~~~\text{and}~~~~\alpha_m^2 := \frac{\beta^2(1-\beta^{2m})}{1-\beta^2}.$$ The following theorem estimates the complexity of Algorithm \[alg:A1\] to approximate an $\varepsilon$-stationary point of , whose proof is given in Subsection \[apdx:th:singe\_loop\_const\_step\] of the supplementary document.
\[th:singe\_loop\_const\_step\] Let $\sets{x_t}$ be the sequence generated by Algorithm \[alg:A1\] using the step-size $\eta$ defined by . Let us choose $\widetilde{x}_m\sim\Ub(\sets{x_t}_{t=0}^m)$. Then
- The step-size $\eta$ satisfies $\eta \geq \underline{\eta} := \frac{2\sqrt{c_1}}{3{{L}}\big[ b(m+1)\big]^{1/4}}$. In addition, we have $$\label{eq:single_SGD}
\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \frac{3b^{1/4}L\big[f(x^0) - f^{\star}\big]}{\sqrt{c_1}(m+1)^{3/4}} + \left(c_1 + \frac{1}{c_1}\right)\frac{ \sigma^2}{\sqrt{b(m+1)}}.$$
- If we choose $b := c_2\sigma^{8/3} (m + 1)^{1/3}$ for any $c_2 > 0$, then to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \varepsilon^2$, we need to choose $$\label{eq:choice_of_m}
m := \bigg\lfloor\tfrac{\sigma}{\varepsilon^3}\Big[\tfrac{3 L c_2^{1/4}}{\sqrt{c_1}}\big[f(x^0) - f^{\star}\big] + \left(c_1 + \tfrac{1}{c_1}\right)\tfrac{1}{\sqrt{c_2}}\Big]^{3/2}\bigg\rfloor = \BigO{\frac{\sigma}{\varepsilon^3}}.$$ In particular, if we choose $c_1 = 1$, then the number of oracle calls is $\Tc_{ge}$ is $$\label{eq:overall_complexity2}
{\!\!\!\!\!\!\!\!}\begin{array}{ll}
\Tc_{ge} &:= \dfrac{\sigma^3}{\varepsilon}\left[ 3 L c_2^{9/4}\big[f(x^0) - f^{\star}\big] + 2c_2^{3/2}\right]^{1/2} {\!\!\!} + \dfrac{3\sigma}{\varepsilon^3}\left[ 3 L c_2^{1/4}\big[f(x^0) - f^{\star}\big] + \frac{2}{\sqrt{c_2}}\right]^{3/2} \vspace{1ex}\\
&= \BigO{\dfrac{\sigma^3}{\varepsilon}+ \dfrac{\sigma}{\varepsilon^3}}.
\end{array}{\!\!\!\!\!\!\!\!\!}$$ Moreover, the step-size $\eta$ satisfies $\eta \geq \underline{\eta} := \frac{2}{3{{L}}c_2^{1/4}\sigma^{2/3}(m+1)^{1/3}} = \BigO{m^{-1/3}}$.
Here, $\Tc_{ge}$ stands for the number of stochastic gradient evaluations of $f$ in . The complexity $\Tc_{ge}$ in can be written as $\Tc_{ge} = \BigO{\max\set{\sigma\varepsilon^{-3}, \sigma^3\varepsilon^{-1}}}$. If $\sigma < \BigO{\frac{1}{\varepsilon}}$, then our complexity is $\Tc_{ge} = \BigO{\sigma\varepsilon^{-3}}$. Even if $\sigma < \BigO{\frac{1}{\varepsilon^3}}$, then our complexity is still better than $\BigO{\sigma^2\varepsilon^{-4}}$ in SGD.
### Convergence of Algorithm \[alg:A1\] with adaptive step-size $\eta_t$ and constant $\beta$
Let $\beta := 1 - \frac{c_1}{\sqrt{b(m+1)}} \in (0, 1)$ be fixed for some $0 < c_1 < \sqrt{b(m+1)}$. Instead of fixing step-size $\eta_t$ as in , we can update it adaptively as $$\label{eq:update_of_eta_t}
\eta_m := \frac{1}{L},~~\text{and}~~\eta_t := \frac{1}{L + L^2\big[\beta^2\eta_{t+1} + \beta^4\eta_{t+2} + \cdots + \beta^{2(m-t)}\eta_m\big]}~~\text{for}~t=0,\cdots, m-1.$$ It can be shown that $0 < \eta_0 < \eta_1 < \cdots < \eta_m$. Interestingly, our step-size is updated in an increasing manner instead of diminishing as in existing SGD-type methods. Moreover, given $m$, we can pre-compute the sequence of these step-sizes $\set{\eta_t}_{t=0}^m$ in advance within $\BigO{m}$ basic operations. Therefore, it does not significantly incur the computational cost of our method.
The following theorem states the convergence of Algorithm \[alg:A1\] under the adaptive update , whose proof is given in Subsection \[apdx:th:singe\_loop\_adapt\_step\] of the supplementary document.
\[th:singe\_loop\_adapt\_step\] Let $\sets{x_t}$ be the sequence generated by Algorithm \[alg:A1\] using the step-size $\eta_t$ defined by . Let $\Sigma_m := \sum_{t=0}^m\eta_t$, and $\widetilde{x}_m\sim\Ub_{{\mathbf{p}}}(\sets{x_t}_{t=0}^m)$ with ${\mathbf{p}}_t := \Prob{\widetilde{x}_m = x_t} = \frac{\eta_t}{\Sigma_m}$. Then
- The sum $\Sigma_m$ is bounded from below as $\Sigma_m \geq \frac{\sqrt{c_1}(m+1)^{3/4}}{2L b^{1/4}}$.
- If we choose $b := c_2\sigma^{8/3} (m + 1)^{1/3}$ for any $c_2 > 0$, then to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \varepsilon^2$, we need to choose $m := \Big\lfloor\frac{\sigma}{\varepsilon^3}\Big[\frac{3 L c_2^{1/4}}{\sqrt{c_1}}\big[f(x^0) - f^{\star}\big] + \big(c_1 + \frac{1}{c_1}\big)\frac{1}{\sqrt{c_2}}\Big]^{3/2}\Big\rfloor = \BigO{\frac{\sigma}{\varepsilon^3}}$. Therefore, the number of stochastic gradient evaluations $\Tc_{ge}$ is at most the same as in .
Note that in the finite sum case, i.e. $\vert\Omega\vert = n$, we set $b := \min \{n, c_2\sigma^{8/3} (m + 1)^{1/3}\}$ in both Theorems \[th:singe\_loop\_const\_step\] and \[th:singe\_loop\_adapt\_step\]. This complexity remains the same as in Theorem \[th:singe\_loop\_const\_step\]. However, the adaptive stepsize $\eta_t$ potentially gives a better performance in practice as we will see in Section \[sec:num\_experiments\].
Algorithm \[alg:A1\] can be considered as a single-loop variance reduction method, which is similar to SAGA [@SAGA], but Algorithm \[alg:A1\] aims at solving the nonconvex problem . It is different from standard SGD methods, where it can be initialized by a mini-batch and then update the estimator using three individual gradients. Therefore, it has the same cost as SGD with mini-batch of size $3$. As a compensation, we obtain an improvement on the complexity bound as in Theorems \[th:singe\_loop\_const\_step\] and \[th:singe\_loop\_adapt\_step\].
Convergence analysis of the double loop variant
-----------------------------------------------
Since the step-size $\eta_t$ depends on $m$, it is natural to run Algorithm \[alg:A1\] with multiple stages. This leads to a double-loop algorithm as SVRG, SARAH, and SPIDER, where Algorithm \[alg:A1\] is restarted at each outer iteration $s$. The detail of this variant is described in Algorithm \[alg:A2\].
An initial point $\widetilde{x}^0$ and parameters $b$, $m$, $\beta_t$, and $\eta_t$ (will be specified). [**For $s := 1, 2, \cdots, S$ do**]{} \[step:o2\] Run Algorithm \[alg:A1\] with an initial point $x_0^{(s)} := \widetilde{x}^{(s-1)}$. \[step:o5\] Set $\widetilde{x}^{(s)} := x_{m+1}^{(s)}$ as the last iterate of Algorithm \[alg:A1\].
To analyze Algorithm \[alg:A2\], we use $x^{(s)}_t$ to represent the iterate of Algorithm \[alg:A1\] at the $t$-th inner iteration within each stage $s$. From , we can see that each stage $s$, the following estimate holds $$\frac{\eta}{2}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \Exp{f(x_0^{(s)})} - \Exp{f(x_{m+1}^{(s)})} + \frac{\eta\sigma^2\sqrt{m+1}}{(1+\beta)\sqrt{b}}.$$ Here, we assume that we fix the step-size $\eta_t = \eta > 0$ for simplicity of analysis. The complexity of Algorithm \[alg:A2\] is given in the following theorem, whose proof is in Supplementary Document \[apdx:th:double\_loop\_convergence\].
\[th:double\_loop\_convergence\] Let $\sets{x^{(s)}_t}_{t=0\to m}^{s=1\to S}$ be the sequence generated by Algorithm \[alg:A2\] using constant step-size $\eta$ in . Then, the following estimate holds $$\label{eq:double_loop_est}
\frac{1}{S(m+1)}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \frac{3 L b^{1/4}}{S (m+1)^{3/4}} \big[f(\widetilde{x}^0) - f^{\star}\big] + \frac{2 \sigma^2}{\sqrt{b (m+1)}}.$$ Let $\widetilde{x}_T \sim \Ub(\sets{x^{(s)}_t}_{t=0\to m}^{s=1\to S})$. If we choose $b := \frac{c_1\sigma^2}{\varepsilon^2}$ and $m + 1 := \frac{c_2\sigma^2}{\varepsilon^2}$ for some constants $c_1 > 0$ and $c_2 > 0$ and $c_1c_2 > 4$, then, to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_T)}^2} \leq \varepsilon^2$, we require at most $$\label{eq:S_iterations}
S := \Bigg\lfloor \frac{3 L c_1^{1/4} \big[f(\widetilde{x}^0) - f^{\star}\big]}{c_2^{3/4} \sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon}\Bigg\rfloor ~~~\text{outer iterations}.$$ Consequently, the total number of stochastic gradient evaluations $\Tc_{ge}$ does not exceed $$\label{eq:Toc_double_loop}
\Tc_{ge} := (b + 3m)S = \frac{3 L (c_1 + 3 c_2)c_1^{1/4} \big[f(\widetilde{x}^0) - f^{\star}\big]\sigma}{c_2^{3/4}\left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^3} = \BigO{\frac{\sigma}{\varepsilon^3}}.$$
Note that the complexity only holds if $\BigO{\frac{\sigma}{\varepsilon^3}} > \frac{c_1\sigma^2}{\varepsilon^2}$. Otherwise, the total complexity is $\BigO{\max\set{\frac{\sigma}{\varepsilon^3}, \frac{\sigma^2}{\varepsilon^2}}}$, where other constants independent of $\sigma$ and $\varepsilon$, and are hidden. Practically, if $\beta$ is very close to $1$, one can remove the unbiased SGD term to save one stochastic gradient evaluation. In this case, our estimator reduces to SARAH but using different step-size. We observed empirically that when $\beta\approx 0.999$, the performance of our methods is not affected if we do so.
Extensions to mini-batch cases {#subsec:extensions}
------------------------------
We consider a mini-batch hybrid stochastic estimator $\hat{v}_t$ for the gradient $\nabla{f}(x_t)$ defined as: $$\label{eq:vhat_t}
\hat{v}_t := \beta_{t-1}\hat{v}_{t-1} + \frac{\beta_{t-1}}{\hat{b}_t}\sum_{i\in\hat{\Bc}_t}(\nabla{f}(x_t;\xi_i) - \nabla{f}(x_{t-1};\xi_i)) + (1-\beta_{t-1})u_t,$$ where $\beta_{t-1}\in [0, 1]$, and $\hat{\Bc}_t$ is a mini-batch of the size $\hat{b}_t$ and independent of the unbiased estimator $u_t$. Note that $u_t$ can also be a mini-batch unbiased estimator. For example, $u_t := \frac{1}{\tilde{b}_t}\sum_{j\in\tilde{\Bc}_t}\nabla{f}(x_t;\zeta_j)$ is a mini-batch SGD estimator with a mini-batch $\tilde{\Bc}_t$ of size $\tilde{b}_t$, where $\tilde{\Bc}_t$ is independent of $\hat{\Bc}_t$.
Using $\hat{v}_t$ defined by , we can design a mini-batch variant of Algorithms \[alg:A1\] to solve . The following corollary is obtained as a result of Theorems \[th:singe\_loop\_const\_step\] for the mini-batch variant of Algorithm \[alg:A1\], whose proof is in Subsection \[apdx:co:mini\_batch\] of the supplementary document.
\[co:mini\_batch\] Let Algorithm \[alg:A1\] be applied to solve using mini-batch update for $v_t$ with $\hat{b}_t = \tilde{b}_t = \hat{b} \geq 1$ fixed, $0 < c_1 < \sqrt{b(m+1)}$, and the step-size $$\label{eq:step_size_batch}
\eta := \frac{2}{L\left(1 + \sqrt{1 + 4\rho\alpha_m^2}\right)}~~~~\text{with}~~~~\alpha^2_m := \frac{\beta^2(1-\beta^{2m})}{1 - \beta^2}~~~\text{and}~~~\beta := 1 - \frac{c_1}{\sqrt{\rho b (m-1)}}.$$ If we choose $b := c_2\sigma^{8/3}\left[\rho(m+1)\right]^{1/3}$ for any $c_2 > 0$, then to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \varepsilon^2$, we need to choose $$\label{eq:choice_of_m2}
m := \left\lfloor \frac{\rho^{1/2}\sigma}{\varepsilon^3}\left[\frac{3 Lc_2^{1/4}}{2{c_1}}\left( f(x^0) - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{2\sqrt{c_2}}\right]^{3/2}\right\rfloor = \BigO{\frac{\rho^{1/2}\sigma}{\varepsilon^3}}.$$ Therefore, the number of oracle calls is $\Tc_{ge}$ is $$\label{eq:T_ge3}
\Tc_{ge} := \BigO{\frac{\rho^{1/2}\sigma^3}{\varepsilon} + \frac{\sigma}{\rho^{1/2}\varepsilon^3}},$$ where $\rho = \rho(\hat{b}) := \frac{n - \hat{b}}{(n-1)\hat{b}}$ if $n := \vert\Omega\vert$ is finite, and $\rho(\hat{b}) := \frac{1}{\hat{b}}$, otherwise. In particular, if we choose $\hat{b} := \frac{\varepsilon^2\sigma^2}{c_3^2}$ for some $0 < c_3 \leq \varepsilon\sigma$, then, the overall complexity $\Tc_{ge}$ is $\Tc_{ge} := \BigO{\left(c_3 + \frac{1}{c_3}\right)\frac{\sigma^2}{\varepsilon^2}}$.
We can also develop a mini-batch variant of Algorithm \[alg:A2\] and estimate its complexity as in Theorem \[th:double\_loop\_convergence\]. For more details, we refer to Subsection \[apdx:th:double\_loop\_convergence\_mini\_batch\] in the supplementary document due to space limit.
Numerical experiments {#sec:num_experiments}
=====================
We verify our algorithms on two numerical examples and compare them with several existing methods: SVRG [@nonconvexSVRG], SVRG+ [@li2018simple], SPIDER [@fang2018spider], SpiderBoost [@wang2018spiderboost], and SGD [@ghadimi2013stochastic]. Due to space limit, the detailed configuration of our experiments as well as more numerical experiments can be found in Supplementary Document \[apdx:subsec:experiments\]. Our numerical experiments are implemented in Python and running on a MacBook Pro. Laptop with 2.7GHz Intel Core i5 and 16Gb memory.
Logistic regression with nonconvex regularizer {#subsec:exam1}
----------------------------------------------
Our first example is the following well-known problem used in may papers including [@wang2018spiderboost]: [$$\label{eq:exam1}
\min_{x\in\R^p}\bigg\{ f(x) := \frac{1}{n}\sum_{i=1}^n \Big[ f_i(x) := \log(1 + \exp(-a_i^Tx)) + \lambda\sum_{j=1}^p\frac{x_i^2}{1 + x_i^2}\Big] \bigg\},
\vspace{-0.5ex}$$]{} where $a_i\in\R^p$ are given for $i=1,\cdots, n$, and $\lambda > 0$ is a regularization parameter. Clearly, problem fits well with $L_{f_i} = \frac{\Vert A\Vert^2}{4} + 2\lambda$. In this experiment, we choose $\lambda = 0.1$ and normalize the data. One can also verify Assumption \[as:A1\] due to the bounded Hessian of $f_i$.
We use three datasets from LibSVM for : `w8a` ($n=49,749, p=300$), `rcv1.binary` ($n=20,242, p = 47,236$), and `real-sim` $(n=72,309, p = 20,958$). We run $8$ different algorithms as follows. Algorithm \[alg:A1\] with constant step-size (Hybrid-SGD-SL) and adaptive step-size (Hybrid-SGD-ASL) using our theoretical step-sizes and , respectively without tuning. Hybrid-SGD-DL is Algorithm \[alg:A2\]. SGD1 is SGD with constant step-size $\eta_t = \frac{0.1}{L}$, and SGD2 is SGD with adaptive step-size $\eta_t = \frac{0.1}{L(1 + \lfloor t/n\rfloor)}$. Since the stepsize of SPIDER depends on an accuracy $\varepsilon$, we choose $\varepsilon = 10^{-1}$ to get a larger step-size. Our first result in the single-sample case (i.e. when $\hat{b} = 1$, not using mini-batch) is plotted in Fig. \[fig:logistic\_reg\] after $20$ epochs.
\
From Fig. \[fig:logistic\_reg\], we observe that Hybrid-SGD-SL has similar convergence behavior as SGD1, but Hybrid-SGD-ASL works better. Hybrid-SGD-DL is the best but has some oscillation. SGD2 works better than SGD1 and is comparable with Hybrid-SGD-SL/ASL in the two last datasets. SVRG performs very poorly due to its small step-size. SVRG+ works much better than SVRG, and is comparable with our methods. SPIDER is also slow even when we have increased its step-size.
Now, we run 3 single-loop algorithms with mini-batch of the size $\hat{b} :=300$. The result is shown in Fig. \[fig:logistic\_reg2\] after $20$ epochs.
Fig. \[fig:logistic\_reg2\] shows similar performance between Hybrid-SGD-SL and ASL and SGD2. Clearly, these theoretical variants of Algorithm \[alg:A1\] are slightly better than the adaptive SGD variant (SGD2), where a careful step-size is used.
Binary classification involving nonconvex loss and Tikhonov’s regularizer {#subsec:exam2}
-------------------------------------------------------------------------
We consider the following binary classification problem studied in [@zhao2010convex] involving nonconvex loss: [$$\label{eq:exam2}
f^{\star} := \min_{x\in\R^p}\Big\{ f(x) := \frac{1}{n}\sum_{i=1}^n\ell(a_i^{\top}x, b_i) + \frac{\lambda}{2}\norms{x}^2 \Big\},
\vspace{-0.5ex}$$]{} where $a_i\in\R^p$ and $b_i \in\set{-1,1}$ are given data for $i=1,\cdots, n$, $\lambda > 0$ is a regularization parameter, and $\ell$ is a nonconvex loss of the forms: $\ell(\tau, s) = \left(1 - \frac{1}{1+\exp(-\tau s)}\right)^2$ (using in two-layer neural networks). One can check that satisfies Assumption \[as:A1\] with $L \approx 0.15405\max_i\norms{a_i}^2 + \lambda$. We choose $\lambda := 0.01$, and test three variants of Algorithm \[alg:A2\]: Hybrid-SGD-DL and compare them with SpiderBoost, SVRG, and SVRG+. Due to space limit, we only plot one experiment in Fig. \[fig:logistic\_reg3\] after $20$ epochs. Additional experiments can be found in Supplementary Document \[apdx:subsec:experiments\].
As we can see from Fig. \[fig:logistic\_reg3\] that Algorithm \[alg:A2\] performs well and is slightly better than SpiderBoost. Note that SpiderBoost simply uses SARAH estimator with constant stepsize $\eta = \frac{1}{2L}$ but with mini-batch of the size $\lfloor\sqrt{n}\rfloor$. It is not surprise that SpiderBoost makes very good progress to decrease the gradient norms. Both SVRG and SVRG+ perform much worse than Hybrid-SGD-DL and SpiderBoost in this test, but SVRG+ is slightly better than SVRG. In our methods, due to the aid of SARAH part, they also make similar progress as SpiderBoost but using different step-sizes.
Conclusion
==========
We have introduced a new hybrid SARAH-SGD estimator for the objective gradient of expectation optimization problems. Under standard assumptions, we have shown that this estimator has a better variance reduction property than SARAH. By exploiting such an estimator, we have developed a new Hybrid-SGD algorithm, Algorithm \[alg:A1\], that has better complexity bounds than state-of-the-art SGDs. Our algorithm works with both constant and adaptive step-sizes. We have also studied its double-loop and mini-batch variants. We believe that our approach can be extended to other choices of unbiased estimators, Hessian estimators for second-order stochastic methods, and adaptive $\beta$.
Appendix: Properties of the hybrid stochastic estimator {#sec:appendix1}
=======================================================
This supplementary document provides the full proof of all the results in the main text. First, we need the following lemma in the sequel.
\[le:adaptive\_step\_size\] Given $L > 0$ and $\omega \in (0, 1)$. Let $\sets{\eta_t}_{t=0}^m$ be the sequence updated by $$\label{eq:update_of_eta_t_proof}
\eta_m := \frac{1}{L},~~~\text{and}~~\eta_t := \frac{1}{L + L^2\big[\omega\eta_{t+1} + \omega^2\eta_{t+2} + \cdots + \omega^{(m-t)}\eta_m\big]},$$ for $t=0,\cdots, m-1$. Then $$\label{eq:stepsize_pros}
0 < \eta_0 < \eta_1 < \cdots < \eta_m = \frac{1}{L},~~~\text{and}~~~\Sigma_m := \sum_{t=0}^m\eta_t \geq \frac{(m+1)\sqrt{1-\omega}}{2L}.$$
First, from it is obvious to show that $0 < \eta_0 < \cdots < \eta_{m-1} = \frac{1}{L(1+\omega)} < \eta_m = \frac{1}{L}$. At the same time, since $\omega \in (0, 1)$, we have $1 \geq \omega \geq \omega^2 \geq \cdots \geq \omega^{m}$. By Chebyshev’s sum inequality, we have $$\label{eq:estimate1}
\begin{array}{ll}
(m-t)\big(\omega\eta_{t+1} + \omega^2\eta_{t+2} + \cdots + \omega^{m-t}\eta_m\big) &\leq \big(\sum_{j=t+1}^m\eta_i\big)\left(\omega + \omega^2 + \cdots + \omega^{m-t}\right) \vspace{1ex}\\
&\leq \frac{\omega}{1-\omega}\big(\sum_{j=t+1}^m\eta_i\big).
\end{array}$$ From the update , we also have $$\label{eq:estimate2}
\left\{\begin{array}{ll}
L^2\eta_0(\omega\eta_1 + \omega^2\eta_2 + \cdots + \omega^{m}\eta_m) &= 1 - L\eta_0 \\
L^2\eta_1(\omega\eta_2 + \omega^2\eta_3 + \cdots + \omega^{m-1}\eta_{m}) &= 1 - L\eta_1 \\
\cdots & \cdots \\
L^2\eta_{m-1}\omega\eta_m &= 1 - L\eta_{m-1} \\
0 &= 1 - L\eta_{m}.
\end{array}\right.$$ Using into , we get $$\left\{\begin{array}{ll}
\frac{\omega L^2}{1-\omega}\eta_0(\eta_0 + \eta_1 + \cdots + \eta_m) &\geq m - mL\eta_0 + \frac{\omega L^2}{1-\omega}\eta_0^2\\
\frac{\omega L^2}{1-\omega}\eta_1(\eta_0 + \eta_1 + \cdots + \eta_{m}) &\geq (m-1) - (m-1)L\eta_1 + \frac{\omega L^2}{1-\omega}(\eta_1\eta_0 + \eta_1^2) \\
\cdots & \cdots \\
\frac{\omega L^2}{1-\omega}\eta_{m-1}(\eta_0 + \eta_1 + \cdots + \eta_m) &\geq 1 - L\eta_{m-1} + \frac{\omega L^2}{1-\omega}(\eta_{m-1}\eta_0 + \cdots + \eta_{m-1}^2) \\
\frac{\omega L^2}{1-\omega}\eta_{m}(\eta_0 + \eta_1 + \cdots + \eta_m) &\geq 1 - L\eta_{m} + \frac{\omega L^2}{1-\omega}(\eta_{m}\eta_0 + \cdots + \eta_{m}^2).
\end{array}\right.$$ Let $\Sigma_m := \sum_{t=0}^m\eta_t$ and $S_m := \sum_{t=0}^m\eta_t^2$. Summing up both sides of the above inequalities, we get $$\frac{\omega L^2}{1-\omega}\Sigma_m^2 \geq \frac{m^2 + m + 2}{2} - L(m\eta_0 + (m-1)\eta_1 + \cdots + \eta_{m-1} + \eta_m) + \frac{\omega L^2}{2(1-\omega)}\big(S_m + \Sigma_m^2\big).$$ Using again Chebyshev’s sum inequality, we have $$m\eta_0 + (m-1)\eta_1 + \cdots + \eta_{m-1} + \eta_m \leq \frac{m^2 + m + 2}{2(m+1)}\left(\sum_{t=0}^m\eta_t\right) = \frac{(m^2 + m + 2)\Sigma_m}{2(m+1)}.$$ Note that $(m+1)S_m \geq \Sigma_m^2$ by Cauchy-Schwarz’s inequality, which shows that $S_m + \Sigma_m^2 \geq \big(\frac{m+2}{m+1}\big)\Sigma_m^2$. Combining three last inequalities, we obtain the following quadratic inequation in $\Sigma_m$ $$\frac{m\omega L^2}{(1-\omega)}\Sigma_m^2 + L(m^2 + m + 2)\Sigma_m - (m+1)(m^2 + m + 2) \geq 0.$$ Solving this inequation with respect to $\Sigma_m > 0$, we obtain $$\begin{array}{ll}
\Sigma_m &\geq \frac{(1-\omega)\big[\sqrt{(m^2 + m + 2)^2 + \frac{4m(m+1)(m^2 + m + 2)\omega}{1-\omega}} - (m^2 + m + 2)\big]}{2\omega m L} \vspace{1ex}\\
&= \frac{2(m+1)}{L\left[1 + \sqrt{1 + \frac{4m(m+1)\omega}{(1-\omega)(m^2 + m+2)}}\right]} \vspace{1ex}\\
&\geq \frac{2(m+1)\sqrt{1-\omega}}{L\left[\sqrt{1-\omega} + \sqrt{1 + 3\omega}\right]} ~~~~~\text{since}~~\frac{m(m+1)}{m^2+m+2} < 1\vspace{1ex}\\
&\geq \frac{2(m+1)\sqrt{1-\omega}}{L(2 + \sqrt{3\omega})} ~~~~~\text{since}~~\sqrt{1+3\omega} + \sqrt{1-\omega} \leq 2 + \sqrt{3\omega}.
\end{array}$$ Since $\omega \in (0, 1)$, we can overestimate this as $\Sigma_m \geq \frac{(m+1)\sqrt{1-\omega}}{2L}$, which proves .
The proof of Lemma \[le:key\_estimate10\]: Properties of the hybrid SARAH estimator {#apdx:le:key_estimate10}
-----------------------------------------------------------------------------------
By taking the expectation of both sides in and using the fact that $\xi_t$ and $\zeta_t$ are independent, we can easily obtain .
To prove , we first write $$\begin{array}{ll}
v_t - \nabla{f}(x_t) &= \beta_{t-1}(v_{t-1} - \nabla{f}(x_{t-1})) + \beta_{t-1}(\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)) \vspace{1ex}\\
& + {~} (1-\beta_{t-1})\big[u_t - \nabla{f}(x_t)\big] + \beta_{t-1}\big[\nabla{f}(x_{t-1}) - \nabla{f}(x_t)\big].
\end{array}$$ In this case, we have $$\begin{array}{ll}
\norms{v_t - \nabla{f}(x_t)}^2 &= \beta_{t-1}^2\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2 + \beta_{t-1}^2\norms{\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)}^2 \vspace{1ex}\\
& + {~} (1-\beta_{t-1})^2\norms{u_t - \nabla{f}(x_t)}^2 + \beta_{t-1}^2\norms{\nabla{f}(x_{t-1}) - \nabla{f}(x_t)}^2 \vspace{1ex}\\
&+ {~} 2\beta_{t-1}^2(v_{t-1} - \nabla{f}(x_{t-1}))^{\top}(\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}^2(v_{t-1} - \nabla{f}(x_{t-1}))^{\top}(\nabla{f}(x_{t-1}) - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(v_{t-1} - \nabla{f}(x_{t-1}))^{\top}(u_t - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t))^{\top}(u_t - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}^2(\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t))^{\top}(\nabla{f}(x_{t-1}) - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(u_t - \nabla{f}(x_t))^{\top}(\nabla{f}(x_{t-1}) - \nabla{f}(x_t)).
\end{array}$$ Let us first take expectation w.r.t. $\xi_t$ conditioned on $\zeta_t$ to obtain $$\begin{array}{ll}
\Exps{\xi_t}{\norms{v_t - \nabla{f}(x_t)}^2 \mid \zeta_t} &= \beta_{t-1}^2\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2 + \beta_{t-1}^2\Exps{\xi_t}{\norms{\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)}^2 \mid \zeta_t} \vspace{1ex}\\
& + {~} (1-\beta_{t-1})^2\norms{u_t - \nabla{f}(x_t)}^2 - \beta_{t-1}^2\norms{\nabla{f}(x_{t-1}) - \nabla{f}(x_t)}^2 \vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(v_{t-1} - \nabla{f}(x_{t-1}))^{\top}(u_t - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(\nabla{f}(x_t) - \nabla{f}(x_{t-1}))^{\top}(u_t - \nabla{f}(x_t))\vspace{1ex}\\
&+ {~} 2\beta_{t-1}(1-\beta_{t-1})(u_t - \nabla{f}(x_t))^{\top}(\nabla{f}(x_{t-1}) - \nabla{f}(x_t)).
\end{array}$$ Now, taking the expectation over $\zeta_t$, and noting that $\Exps{(\xi_t,\zeta_t)}{\cdot} = \Exps{\zeta_t}{\Exps{\xi_t}{\cdot\mid \zeta_t}}$ and $\Exps{\zeta_t}{u_t - \nabla{f}(x_t)} = 0$, we get $$\begin{array}{ll}
\Exps{(\xi_t,\zeta_t)}{\norms{v_t - \nabla{f}(x_t)}^2} &= \beta_{t-1}^2\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2 + \beta_{t-1}^2\Exps{\xi_t}{\norms{\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)}^2} \vspace{1ex}\\
& + {~} (1-\beta_{t-1})^2\Exps{\zeta_t}{\norms{u_t - \nabla{f}(x_t)}^2} - \beta_{t-1}^2\norms{\nabla{f}(x_{t-1}) - \nabla{f}(x_t)}^2,
\end{array}$$ which is exactly .
The proof of Lemma \[le:upper\_bound\_new\]: Bound on the variance of the hybrid estimator {#apdx:le:upper_bound_new}
------------------------------------------------------------------------------------------
We first upper bound by using $\sigma_t^2 := \Exps{\zeta_t}{\norms{u_t - \nabla{f}(x_t)}^2}$ and then taking the full expectation over $\Fc_t := \sigma(v_0, v_1, \cdots, v_t)$ as $$\begin{array}{ll}
\Exp{\norms{v_t - \nabla{f}(x_t)}^2} &\leq \beta_{t-1}^2\Exp{\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2} + \beta_{t-1}^2\Exp{\norms{\nabla{f}(x_t;\xi_t) - \nabla{f}(x_{t-1};\xi_t)}^2} \vspace{1ex}\\
& +{~} (1-\beta_{t-1})^2\sigma_t^2 \vspace{1ex}\\
&\overset{\tiny\eqref{eq:L_smooth}}{\leq} \beta_{t-1}^2\Exp{\norms{v_{t-1} - \nabla{f}(x_{t-1})}^2} + \beta_{t-1}^2L^2\Exp{\norms{x_t - x_{t-1}}^2} + (1-\beta_{t-1})^2\sigma_t^2.
\end{array}$$ If we define $a_t^2 := \Exp{\norms{ v_t - \nabla{f}(x_t)}^2}$, then the above inequality can lead to $$a_t^2 \leq \beta_{t-1}^2a_{t-1}^2 + \beta_{t-1}^2L^2\Exp{\norms{x_{t} - x_{t-1}}^2} + (1-\beta_{t-1})^2\sigma^2_t.$$ Denote $b_{t-1}^2 := \Exp{\norms{x_t-x_{t-1}}^2}$. Then, we have from the last inequality that $$a_t^2 \leq \beta_{t-1}^2a_{t-1}^2 + L^2\beta_{t-1}^2b_{t-1}^2 + (1-\beta_{t-1})^2\sigma^2_t.$$ By induction, this inequality implies $$\begin{array}{ll}
a_t^2 &\leq {~} \beta_{t-1}^2a_{t-1}^2 + L^2\beta_{t-1}^2b_{t-1}^2 + (1-\beta_{t-1})^2\sigma^2_t \vspace{1ex}\\
&\leq {~} \beta_{t-1}^2\big[\beta_{t-2}^2a_{t-2}^2 + L^2\beta_{t-2}^2b_{t-2}^2 + (1-\beta_{t-2})^2\sigma^2\big] + L^2\beta_{t-1}^2b_{t-1}^2 + (1-\beta_{t-1})^2\sigma^2_t \vspace{1ex}\\
&= {~} \beta_{t-1}^2\beta_{t-2}^2a_{t-2}^2 + L^2\beta_{t-1}^2\beta_{t-2}^2b_{t-2}^2 + L^2\beta_{t-1}^2b_{t-1}^2 + \big[(1-\beta_{t-1})^2\sigma^2_t + \beta_{t-1}^2(1-\beta_{t-2})^2\sigma^2_{t-1}\big] \vspace{1ex}\\
&\leq {~} \beta_{t-1}^2\beta_{t-2}^2\big[\beta_{t-3}^2a_{t-3}^2 + L^2\beta_{t-3}^2b_{t-3}^2 + (1-\beta_{t-3})^2\sigma^2_{t-2}\big] \vspace{1ex}\\
& + {~} L^2\beta_{t-1}^2\beta_{t-2}^2b_{t-2}^2 + L^2\beta_{t-1}^2b_{t-1}^2 + \big[(1-\beta_{t-1})^2\sigma^2_t + \beta_{t-1}^2(1-\beta_{t-2})^2\sigma^2_{t-1}\big] \vspace{1ex}\\
&= {~} \beta_{t-1}^2\beta_{t-2}^2\beta_{t-3}^2a_{t-3}^2 + L^2\beta_{t-1}^2\beta_{t-2}^2\beta_{t-3}^2b_{t-3}^2 + L^2\beta_{t-1}^2\beta_{t-2}^2b_{t-2}^2 \vspace{1ex}\\
& + {~} L^2\beta_{t-1}^2b_{t-1}^2 + \big[(1-\beta_{t-1})^2\sigma^2_t + \beta_{t-1}^2(1-\beta_{t-2})^2\sigma^2_{t-1} + \beta_{t-1}^2\beta_{t-2}^2(1-\beta_{t-3})^2\sigma^2_{t-2}\big] \vspace{1ex}\\
& \cdots \cdots \vspace{1ex}\\
&\leq {~} (\beta_{t-1}^2\cdots\beta_0^2)a_0^2 + L^2(\beta_{t-1}^2\cdots\beta_0^2)b_0^2 + L^2(\beta_{t-1}^2\cdots\beta_1^2)b_1^2 + \cdots + L^2\beta_{t-1}^2b_{t-1}^2 \vspace{1ex}\\
&+ {~} \big[(1-\beta_{t-1})^2\sigma^2_t + \beta_{t-1}^2(1-\beta_{t-2})^2\sigma^2_{t-1} + \beta_{t-1}^2\beta_{t-2}^2(1-\beta_{t-3})^2\sigma^2_{t-2} + \cdots \vspace{1ex}\\
&{~~~} + {~} \beta_{t-1}^2\beta_{t-2}^2\cdots\beta_1^2(1-\beta_0)^2\sigma^2_1\big].
\end{array}$$ Here, we use a convention that $\prod_{i = t+1}^t \beta_i^2 = 1$. As a result, it can be rewritten in a compact form as $$\label{eq:key_estimate1d}
a_t^2 \leq \Big(\prod_{i=1}^{t}\beta_{i-1}^2\Big)a_0^2 + L^2\sum_{i=0}^{t-1}\Big(\prod_{j=i+1}^{t}\beta_{j-1}^2\Big)b_i^2 + \sum_{i=0}^{t-1}\Big(\prod_{j=i+2}^{t}\beta_{j-1}^2\Big)(1-\beta_i)^2\sigma^2_{i+1}.$$ Define $\omega_{t} := \prod_{i=1}^{t}\beta_{i-1}^2$, $\omega_{i, t} := \prod_{j=i+1}^{t}\beta_{j-1}^2$, and $S_{t} := \sum_{i=0}^{t-1}s_i = \sum_{i=0}^{t-1}\big(\prod_{j=i+2}^{t}\beta_{j-1}^2\big)(1-\beta_i)^2\sigma^2_{i+1}$ with $s_i := (1-\beta_i)^2\sigma^2_{i+1}\big(\prod_{j=i+2}^{t}\beta_{j-1}^2\big)$. Then, we can rewrite as $$a_t^2 \leq \omega_{t}a_0^2 + L^2\sum_{i=0}^{t-1}\omega_{i,t}b_i^2 + S_{t},$$ which is exactly .
Appendix: Convergence analysis of Algorithm \[alg:A1\] and Algorithm \[alg:A2\] {#apdx:sec:convergence-analysis}
===============================================================================
We provide the full convergence analysis for Algorithm \[alg:A1\] and Algorithm \[alg:A2\] in the single-sample case.
The proof of Lemma \[le:key\_estimate\_of\_convergence\]: One-iteration analysis {#apdx:le:key_estimate_of_convergence}
--------------------------------------------------------------------------------
The following lemma provides a key estimate for convergence analysis of Algorithm \[alg:A1\].
\[le:key\_estimate\_of\_convergence\] Let $\sets{x_t}$ be the sequence generated by Algorithm \[alg:A1\]. Then, under Assumption \[as:A1\], we have the following estimate: $$\label{eq:key_estimate_31}
\begin{array}{ll}
\Exp{f(x_{m+1})} &\leq \Exp{f(x_0)} - \displaystyle\frac{1}{2}\displaystyle\sum_{t=0}^m\eta_t\Exp{\norms{\nabla{f}(x_t)}^2} \\
& + {~} \dfrac{1}{2}\Big(\displaystyle\sum_{t=0}^m\eta_t\omega_t\Big)\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} + \frac{1}{2}\Big(\displaystyle\sum_{t=0}^m\eta_tS_t\Big) + \frac{1}{2}\Tc_m,
\end{array}$$ where $$\label{eq:T_m}
\Tc_m := L^2\sum_{t=1}^m \eta_t \sum_{i=0}^{t-1}\omega_{i,t}\eta_i^2\Exp{\norms{v_i}^2} - \sum_{t=0}^m\big(\eta_t - L\eta_t^2\big)\Exp{\norms{v_t}^2},$$ and $\omega_t$, $\omega_{i,t}$, and $S_t$ are defined in Lemma \[le:upper\_bound\_new\].
First, from the $L$-smoothness of $f$, we have $$\begin{array}{ll}
f(x_{t+1}) &\leq f(x_t) - \eta_t\iprods{\nabla{f}(x_t), v_t} + \frac{L\eta_t^2}{2}\norms{v_t}^2 \vspace{1ex}\\
&= f(x_t) - \frac{\eta_t}{2}\norms{\nabla{f}(x_t)}^2 - \big(\frac{\eta_t}{2} - \frac{L\eta_t^2}{2}\big)\norms{v_t}^2 + \frac{\eta_t}{2}\norms{v_t - \nabla{f}(x_t)}^2.
\end{array}$$ Taking the expectation over the randomness $(\xi_t, \zeta_t)$ of this estimate, we obtain $$\begin{array}{ll}
\Exps{(\xi_t,\zeta_t)}{f(x_{t+1})} &\leq f(x_t) - \frac{\eta_t}{2}\Exps{(\xi_t,\zeta_t)}{\norms{\nabla{f}(x_t)}^2} - \frac{\eta_t}{2}\big(1 - L\eta_t\big)\Exps{(\xi_t,\zeta_t)}{\norms{v_t}^2} \vspace{1ex}\\
& + {~} \frac{\eta_t}{2}\Exps{(\xi_t,\zeta_t)}{\norms{v_t - \nabla{f}(x_t)}^2}.
\end{array}$$ Taking the full expectation over the entire history up to the $t$-th iteration, and then using and noting that $x_t - x_{t-1} = -\eta_{t-1}v_{t-1}$, we obtain $$\label{eq:est5b}
\begin{array}{ll}
\Exp{f(x_{t+1})} &\leq \Exp{f(x_t)} - \frac{\eta_t}{2}\Exp{\norms{\nabla{f}(x_t)}^2} - \frac{\eta_t}{2}\big(1 - L\eta_t\big)q_t^2 + \frac{\eta_t}{2}a_t^2 \vspace{1ex}\\
&\leq \Exp{f(x_t)} - \frac{\eta_t}{2}\Exp{\norms{\nabla{f}(x_t)}^2} - \frac{\eta_t}{2}\big(1 - L\eta_t\big)q_t^2 \vspace{1ex}\\
& + {~} \frac{\eta_t}{2}\Big[\omega_{t}a_0^2 + L^2\sum_{i=0}^{t-1}\omega_{i,t}\eta_i^2q_i^2 + S_{t}\Big],
\end{array}$$ where $q_t^2 := \Exp{\norms{v_t}^2}$ and $a_t^2 := \Exp{\norms{v_t - \nabla{f}(x_t)}^2}$. Here, we use $b_{t-1}^2 := \Exp{\norms{x_t - x_{t-1}}^2} = \eta_{t-1}^2\Exp{\norms{v_{t-1}}^2} = \eta_{t-1}^2q_{t-1}^2$ in the last inequality.
Summing up from $t=0$ to $t=m$, we obtain $$\label{eq:est5c}
\begin{array}{ll}
\Exp{f(x_{m+1})} &\leq \Exp{f(x_0)} - \sum_{t=0}^m\frac{\eta_t}{2}\Exp{\norms{\nabla{f}(x_t)}^2} - \sum_{t=0}^m\frac{\eta_t}{2}\big(1 - L\eta_t\big)q_t^2 \vspace{1ex}\\
& + {~} \frac{1}{2}\left(\sum_{t=0}^m\omega_t\eta_t\right)a_0^2 + \frac{1}{2}\left(\sum_{t=0}^m\eta_tS_t\right) + \frac{L^2}{2}\sum_{t=0}^m\eta_t\sum_{i=0}^{t-1}\omega_{i,t}\eta_i^2q_i^2.
\end{array}$$ Let us define $T_m$ as in , i.e.: $$\Tc_m := L^2\sum_{t=1}^m \eta_t \sum_{i=0}^{t-1}\omega_{i,t}\eta_i^2q_i^2 - \sum_{t=0}^m\eta_t\big(1 - L\eta_t\big)q_t^2.$$ Then, we obtain from the estimate .
The proof of Theorem \[th:singe\_loop\_const\_step\]: Single-loop with constant step-size {#apdx:th:singe_loop_const_step}
-----------------------------------------------------------------------------------------
We analyze the case $\beta_t = \beta \in (0, 1)$ fixed and the step-size $\eta_t = \eta > 0$ fixed. From Lemma \[le:upper\_bound\_new\], we have $\omega_t = \beta^{2t}$, $\omega_{i,t} = \beta^{2(t-i)}$, and $$\begin{array}{ll}
s_t &:= \sum_{i=0}^{t-1}\big(\prod_{j=i+2}^{t}\beta_{j-1}^2\big)(1-\beta_i)^2 \vspace{1ex}\\
& = (1-\beta)^2\big[1 + \beta^2 + \beta^{4} + \cdots + \beta^{2(t-1)}\big] \vspace{1ex}\\
&= (1-\beta)^2\Big[\frac{1-\beta^{2t}}{1-\beta^2}\Big] \vspace{1ex}\\
&< \frac{1-\beta}{1+\beta}.
\end{array}$$ In this case, by convention that $\omega_0 = 1$, we have $$\label{eq:bound_of_st_omegat}
\sum_{t=0}^ms_t < \frac{(1-\beta)(m+1)}{1+\beta}~~~\text{and}~~\sum_{t=0}^m\omega_t = 1 + \frac{\beta^2(1-\beta^{2m})}{1-\beta^2} = \frac{1-\beta^{2(m+1)}}{1-\beta^2} < \frac{1}{1-\beta^2}.$$ Now, to bound the quantity $\Tc_m$ defined by , we note that $$\begin{array}{ll}
\displaystyle\sum_{t=1}^m\sum_{i=0}^{t-1}\beta^{2(t-i)}q_i^2 &= \displaystyle\sum_{i=0}^{0}\beta^{2(1-i)}q_i^2 + \sum_{i=0}^{1}\beta^{2(2-i)}q_i^2 + \sum_{i=0}^{2}\beta^{2(3-i)}q_i^2 + \cdots + \sum_{i=0}^{m-1}\beta^{2(m-i)}q_i^2 \vspace{1ex}\\
&= \beta^2q_0^2 + \big[\beta^4q_0^2 + \beta^2q_1^2\big] + \big[\beta^6q_0^2 + \beta^4q_1^2 + \beta^2q_0^2\big] + \cdots \vspace{1ex}\\
& + {~} \big[\beta^{2m}q_0^2 + \beta^{2(m-1)}q_1 + \cdots + \beta^2q_{m-1}^2\big] \vspace{1ex}\\
&= \beta^2\big[1 + \beta^2 + \cdots + \beta^{2(m-1)}\big]q_0^2 + \beta^2\big[1 + \beta^2 + \cdots + \beta^{2(m-2)}\big]q_1^2 + \cdots \vspace{1ex}\\ &+ {~} \beta^2\big[1 + \beta^2\big]q_{m-2}^2 + \beta^2q_{m-1}^2\vspace{1ex}\\
&= \frac{\beta^2}{1-\beta^2}\Big[ (1-\beta^{2m})q_0^2 + (1-\beta^{2(m-1)})q_1^2 + \cdots + (1-\beta^2)q_{m-1}^2\Big].
\end{array}$$ Using this expression, we can write $\Tc_m$ from as $$\label{eq:T_m_constant}
\begin{array}{ll}
\Tc_m &= \eta\Big[\frac{\beta^2(1-\beta^{2m})L^2\eta^2}{1-\beta^2} - (1-L\eta)\Big]q_0^2 + \eta\Big[\frac{\beta^2(1-\beta^{2(m-1)})L^2\eta^2}{1-\beta^2} - (1-L\eta)\Big]q_1^2 + \cdots \vspace{1ex}\\
&+ {~} \eta\Big[\frac{\beta^2(1-\beta^{2})L^2\eta^2}{1-\beta^2} - (1-L\eta)\Big]q_{m-1}^2 - \eta(1-L\eta)q_{m}^2.
\end{array}$$ To guarantee $\Tc_m \leq 0$, from , we need to choose $$\label{eq:cond_of_eta}
\left\{\begin{array}{ll}
\frac{L^2\eta^2\beta^2(1-\beta^{2m})}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
\frac{L^2\eta^2\beta^2(1-\beta^{2(m-1)})}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
\cdots & \cdots \vspace{1ex}\\
\frac{L^2\eta^2\beta^2(1-\beta^2)}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
- (1 - L\eta) &\leq 0.
\end{array}\right.$$ Clearly, since $1 - \beta^{2(m-i)} \geq 1 - \beta^2$ for $i=0,\cdots, m-1$, if we define $\alpha_m^2 := \frac{\beta^2(1-\beta^{2m})}{1-\beta^2}$, then the condition holds if $L^2\eta^2\alpha^2_m - (1-L\eta) \leq 0$. By tightening this condition, we obtain a quadratic equation $L^2\eta^2\alpha_m^2 - (1-L\eta) = 0$ in $\eta$, which leads to $$\label{eq:step_size}
\eta := \frac{2}{L(\sqrt{1 + 4\alpha_m^2} + 1)}~~~\text{with}~~\alpha_m^2 := \frac{\beta^2(1-\beta^{2m})}{1-\beta^2}.$$ Note that since $\alpha_m^2 \leq \frac{\beta^2}{1-\beta^2}$, we have $\eta \geq \underline{\eta}: = \frac{2\sqrt{1-\beta^2}}{L(\sqrt{1-\beta^2} + \sqrt{1 + 3\beta^2})}$. In that case, by using and , reduces to $$\label{eq:est5c_1}
\begin{array}{ll}
\Exp{f(x_{m+1})} &\overset{\tiny\eqref{eq:bound_of_st_omegat}}{\leq} \Exp{f(x_0)} - \frac{\eta}{2}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \vspace{1ex}\\
& + {~} \frac{\eta(1-\beta^{2(m+1)})}{2(1-\beta^2)}\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} + \Big[\frac{(1-\beta)(m+1)}{1+\beta}\Big]\frac{\eta\sigma^2}{2}.
\end{array}$$ Note that $\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} \leq \frac{\sigma^2}{b}$ and $\Exp{f(x_{m+1})} \geq f^{\star}$, we can further bound as $$\frac{\eta}{2}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \Exp{f(x_0)} - f^{\star} + \frac{\eta\sigma^2}{2(1+\beta)}\Big[\frac{1}{(1-\beta)b} + (1-\beta)(m+1)\Big].$$ Multiplying both sides of this inequality by $\frac{2}{\eta(m+1)}$, and then using the lower bound of $\eta$ from , we obtain $$\label{eq:main_estimate_for_single_loop}
\begin{array}{ll}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} &\leq \frac{2}{\eta(m+1)}\Big[\Exp{f(x_0)} - f^{\star}\Big] + \frac{\sigma^2}{(1+\beta)}\Big[\frac{1}{(1-\beta)b(m+1)} + (1-\beta)\Big] \vspace{1ex}\\
&\leq \frac{L}{(m+1)}\left(\frac{\sqrt{1-\beta^2} + \sqrt{1+3\beta^2}}{\sqrt{1-\beta^2}}\right)\Big[\Exp{f(x_0)} - f^{\star}\Big] \vspace{1ex}\\
& + {~} \frac{\sigma^2}{(1+\beta)}\Big[\frac{1}{(1-\beta)b(m+1)} + (1-\beta)\Big].
\end{array}$$ Let us choose $\beta := 1 - \frac{c_1}{\sqrt{b(m+1)}}$ for some $0 < c_1 < \sqrt{b(m+1)}$. In this case, the last two terms of the right-hand side of become $$\frac{1}{(1-\beta)b(m+1)} + (1-\beta) = \left(c_1 + \frac{1}{c_1}\right)\frac{1}{\sqrt{b(m+1)}}.$$ With this choice of $\beta$, leads to $$\label{eq:single_loop}
\begin{array}{ll}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} &\leq \frac{L}{(m+1)}\left(\frac{\sqrt{1-\beta^2} + \sqrt{1+3\beta^2}}{\sqrt{1-\beta^2}}\right)\Big[\Exp{f(x_0)} - f^{\star}\Big] \vspace{1ex}\\
& + {~} \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{(1+\beta)\sqrt{b(m+1)}}.
\end{array}$$ (a) Since $\beta = 1 - \frac{c_1}{\sqrt{b(m+1)}} < 1$ and $c_1 < \sqrt{b(m+1)}$, we have $$1 - \beta^2 = 1 - \Big(1 - \frac{c_1}{\sqrt{b(m+1)}} \Big)^2 = \frac{2c_1}{\sqrt{b(m+1)}} - \frac{c_1^2}{b(m+1)} = \frac{2c_1\sqrt{b(m+1)} - c_1^2}{b(m+1)} > \frac{c_1}{\sqrt{b(m+1)}},$$ and $\sqrt{1 - \beta^2} + \sqrt{1+3\beta^2} \leq 1 + \sqrt{1+3\beta^2} \leq 3$. On the other hand, from , we have $$\label{eq:LR_lower_bound}
\eta \geq \underline{\eta} = \frac{2\sqrt{1-\beta^2}}{L(\sqrt{1-\beta^2} + \sqrt{1 + 3\beta^2})} \geq \frac{2\sqrt{c_1}}{3{{L}}\big[ b(m+1)\big]^{1/4}}.$$ This proves (a).
Let us define define $f^0 := f(x_0)$. Then, using $\beta < 1$ and into , we get $$\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{3 L b^{1/4}}{\sqrt{c_1}(m+1)^{3/4}}\big[f^0 - f^{\star}\big] + \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{\sqrt{b(m+1)}}.$$ (b) Let us choose $b := c_2\sigma^{8/3}(m+1)^{1/3}$ for some constant $c_2 > 0$. Then the last estimate becomes $$\label{eq:single_loop_final2}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{\sigma^{2/3}}{(m+1)^{2/3}}\left[\frac{3 L c_2^{1/4}}{\sqrt{c_1}}\big[f^0 - f^{\star}\big] + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{\sqrt{c_2}}\right].$$ To guarantee $\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \varepsilon^2$, from we need to set $$\frac{\sigma^{2/3}}{(m+1)^{2/3}}\left[\frac{3 L c_2^{1/4}}{\sqrt{c_1}}\big[f^0 - f^{\star}\big] + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{\sqrt{c_2}}\right] \leq \varepsilon^2.$$ This leads to $m+1 \geq \frac{\sigma}{\varepsilon^3}\left[\frac{3 L c_2^{1/4}}{\sqrt{c_1}}\big[f^0 - f^{\star}\big] + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{\sqrt{c_2}}\right]^{3/2}$. Therefore, we can choose $m$ as shown in .
Finally, if $c_1 = 1$, then the number of stochastic gradient evaluations is $\Tc_{ge}$ is $$\begin{array}{ll}
\Tc_{ge} &= b + 3m = c_2\sigma^{8/3}(m + 1)^{1/3} + \frac{3\sigma}{\varepsilon^3}\left[ 3 L c_2^{1/4}\big[f^0 - f^{\star}\big] + \frac{2}{\sqrt{c_2}}\right]^{3/2} \vspace{1ex}\\
& = \frac{c_2\sigma^3}{\varepsilon}\left[ 3 L c_2^{1/4}\big[f^0 - f^{\star}\big] + \frac{2}{\sqrt{c_2}}\right]^{1/2} + \frac{3\sigma}{\varepsilon^3}\left[ 3 L c_2^{1/4}\big[f^0 - f^{\star}\big] + \frac{2}{\sqrt{c_2}}\right]^{3/2} \vspace{1ex}\\
& = \frac{\sigma^3}{\varepsilon}\left[ 3 L c_2^{9/4}\big[f^0 - f^{\star}\big] + c_2^{3/2}\right]^{1/2} + \frac{3\sigma}{\varepsilon^3}\left[ 3 L c_2^{1/4}\big[f^0 - f^{\star}\big] + \frac{2}{\sqrt{c_2}}\right]^{3/2} \vspace{1ex}\\
& = \BigO{\frac{\sigma}{\varepsilon^3} + \frac{\sigma^3}{\varepsilon}},
\end{array}$$ which proves .
The proof of Theorem \[th:singe\_loop\_adapt\_step\]: Single-loop with adaptive step-size {#apdx:th:singe_loop_adapt_step}
-----------------------------------------------------------------------------------------
First, from Lemma \[le:key\_estimate\_of\_convergence\], we have $$\label{eq:key_estimate_300}
\begin{array}{ll}
\Exp{f(x_{m+1})} &\leq \Exp{f(x_0)} - \displaystyle\frac{1}{2}\displaystyle\sum_{t=0}^m\eta_t\Exp{\norms{\nabla{f}(x_t)}^2} \vspace{1ex}\\
& + \dfrac{1}{2}\Big(\displaystyle\sum_{t=0}^m\eta_t\omega_t\Big)\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} + \frac{1}{2}\Big(\displaystyle\sum_{t=0}^m\eta_tS_t\Big) + \frac{1}{2}\Tc_m,
\end{array}$$ where $$\label{eq:T_m2}
\Tc_m := L^2\sum_{t=1}^m \eta_t \sum_{i=0}^{t-1}\omega_{i,t}\eta_i^2\Exp{\norms{v_i}^2} - \sum_{t=0}^m\big(\eta_t - L\eta_t^2\big)\Exp{\norms{v_t}^2},$$ and $\omega_t$, $\omega_{i,t}$, and $s_t$ are defined in Lemma \[le:upper\_bound\_new\].
If we fix $\beta_t = \beta \in (0, 1)$, then we can show that $\omega_t = \beta^{2t}$, $\omega_{i,t} = \beta^{2(t-i)}$, and $s_t = (1-\beta)^2\Big[\frac{1-\beta^{2t}}{1-\beta^2}\Big] < \frac{1-\beta}{1+\beta}$ as in the proof of Theorem \[th:singe\_loop\_const\_step\].
Now, let $u_i^2 := \Exp{\norms{v_i}^2}$. To bound the quantity $\Tc_m$ defined by , we note that $$\begin{array}{ll}
\displaystyle\sum_{t=1}^m\eta_t\sum_{i=0}^{t-1}\beta^{2(t-i)}\eta_i^2u_i^2 &= \eta_1\displaystyle\sum_{i=0}^{0}\beta^{2(1-i)}\eta_i^2u_i^2 + \eta_2\sum_{i=0}^{1}\beta^{2(2-i)}\eta_i^2u_i^2 \vspace{1ex}\\
& +{~} \eta_3\displaystyle\sum_{i=0}^{2}\beta^{2(3-i)}\eta_i^2u_i^2 + \cdots + \eta_m\displaystyle\sum_{i=0}^{m-1}\beta^{2(m-i)}\eta_i^2u_i^2 \vspace{1ex}\\
&= \beta^2\eta_1\eta_0^2u_0^2 + \eta_2\big[\beta^4\eta_0^2u_0^2 + \beta^2\eta_1^2u_1^2\big] \vspace{1ex}\\
& + {~} \eta_3\big[\beta^6\eta_2^2u_0^2 + \beta^4\eta_1^2u_1^2 + \beta^2\eta_2^2u_2^2\big] + \cdots \vspace{1ex}\\
& + {~} \eta_m\big[\beta^{2m}\eta_0^2u_0^2 + \beta^{2(m-1)}\eta_1^2u_1^2 + \cdots + \beta^2\eta_{m-1}^2u_{m-1}^2\big] \vspace{1ex}\\
&= \beta^2\eta_0^2\big[\eta_1 + \beta^2\eta_2 + \cdots + \beta^{2(m-1)}\eta_m\big]u_0^2 \vspace{1ex}\\
& + {~} \beta^2\eta_1^2\big[\eta_2 + \beta^2\eta_3 + \cdots + \beta^{2(m-2)}\eta_{m}\big]u_1^2 + \cdots \vspace{1ex}\\
&+ {~} \beta^2\eta_{m-2}^2\big[\eta_{m-1} + \beta^2\eta_{m}\big]u_{m-2}^2 + \beta^2\eta_{m-1}^2\eta_m u_{m-1}^2.
\end{array}$$ Using this expression, we can write $\Tc_m$ from as $$\begin{array}{ll}
\Tc_m &= \eta_0\Big[L^2\beta^2\eta_0\big[\eta_1 + \beta^2\eta_2 + \cdots + \beta^{2(m-1)}\eta_m\big] - (1-L\eta_0)\Big]u_0^2 \vspace{1ex}\\
& + {~} \eta_1\Big[ L^2\beta^2\eta_1\big[\eta_2 + \beta^2\eta_3 + \cdots + \beta^{2(m-2)}\eta_{m}\big] - (1-L\eta_1)\Big] + \cdots \vspace{1ex}\\
&+ {~} \eta_{m-1}\Big[ L^2\beta^2\eta_{m-1}\eta_m - (1-L\eta_{m-1})\Big]u_{m-1}^2 - \eta_m(1-L\eta_m)u_{m}^2.
\end{array}$$ To guarantee $\Tc_m \leq 0$, from the last expression of $\Tc_m$, we can impose the following condition: $$\label{eq:cond_of_eta3}
\left\{\begin{array}{ll}
L^2\beta^2\eta_0\big[\eta_1 + \beta^2\eta_2 + \cdots + \beta^{2(m-1)}\eta_m\big] - (1-L\eta_0) &= 0\vspace{1ex}\\
L^2\beta^2\eta_1\big[\eta_2 + \beta^2\eta_3 + \cdots + \beta^{2(m-2)}\eta_{m}\big] - (1-L\eta_1)&= 0\vspace{1ex}\\
\cdots & \cdots \vspace{1ex}\\
L^2\beta^2\eta_{m-1}\eta_m - (1-L\eta_{m-1}) &= 0\vspace{1ex}\\
- (1 - L\eta_{m}) &= 0.
\end{array}\right.$$ The condition leads to the following update of $\eta_t$: $$\eta_m := \frac{1}{L},~~~\text{and}~~\eta_t := \frac{1}{L + L^2\big[\beta^2\eta_{t+1} + \beta^4\eta_{t+2} + \cdots + \beta^{2(m-t)}\eta_m\big]},~~~t=0,\cdots, m-1,$$ which is exactly .
Next, note that $\beta^2 = \Big(1 - \frac{c_1}{\sqrt{b(m+1)}}\Big)^2 = 1 - \frac{2c_1}{\sqrt{b(m+1)}} + \frac{c_1^2}{b(m+1)}$. Therefore, $1-\beta^2 = \frac{2c_1}{\sqrt{b(m+1)}} - \frac{c_1^2}{b(m+1)} \geq \frac{c_1}{\sqrt{b(m+1)}}$, which implies $\sqrt{1-\beta^2} \geq \frac{\sqrt{c_1}}{(b(m+1))^{1/4}}$. Using $\sqrt{1-\omega} = \sqrt{1-\beta^2} \geq \frac{\sqrt{c_1}}{(b(m+1))^{1/4}}$ into of Lemma \[le:adaptive\_step\_size\], we can show that $\Sigma_m \geq \frac{\sqrt{c_1}(m+1)^{3/4}}{2{{L}}b^{1/4}}$ as in the first statement (a) of Theorem \[th:singe\_loop\_adapt\_step\].
Note that $\omega_t = \beta^{2t}$, by the Chebyshev sum inequality, we have $$\sum_{t=0}^m\omega_t\eta_t = \sum_{t=0}^m\beta^{2t}\eta_t \leq \frac{\Sigma_m}{(m+1)}(1 + \beta^2 + \cdots + \beta^{2m}) \leq \frac{\Sigma_m}{(m+1)(1-\beta^2)}.$$ Utilizing this estimate, $\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} \leq \frac{\sigma^2}{b}$, and $S_t \leq \frac{(1-\beta)\sigma^2}{1+\beta}$ into , and noting that $\Tc_m \leq 0$, we have $$\displaystyle\frac{1}{2}\displaystyle\sum_{t=0}^m\eta_t\Exp{\norms{\nabla{f}(x_t)}^2} \leq f(x_0) - \Exp{f(x_{m+1})} + \frac{\Sigma_m\sigma^2}{2(1-\beta^2)b(m+1)} + \frac{(1-\beta)\sigma^2}{2(1+\beta)}\Sigma_m.$$ Since $\Exp{f(x_{m+1})} \geq f^{\star}$, using this into the last estimate, and multiplying the result by $\frac{2}{\Sigma_m}$, we obtain $$\label{eq:key_estimate_301}
\displaystyle\frac{1}{\Sigma_m}\displaystyle\sum_{t=0}^m\eta_t\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{2}{\Sigma_m}[f(x_0) - f^{\star}] + \frac{\sigma^2}{(1+\beta)}\left[\frac{1}{b(m+1)(1-\beta)} + (1-\beta)\right].$$ Since $\left[\frac{1}{b(m+1)(1-\beta)} + (1-\beta)\right] = \left(c_1 + \frac{1}{c_1}\right)\frac{1}{\sqrt{b(m+1)}}$ for $\beta = 1- \frac{c_1}{\sqrt{b(m+1)}}$, leads to $$\label{eq:key_estimate_302}
\displaystyle\frac{1}{\Sigma_m}\displaystyle\sum_{t=0}^m\eta_t\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{4Lb^{1/4}}{\sqrt{c_1}(m+1)^{3/4}}\left[f(x_0) - f^{\star}\right] + \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{\sqrt{b(m+1)}}.$$ The second statement (b) of Theorem \[th:singe\_loop\_adapt\_step\] is proved similarly as in Theorem \[th:singe\_loop\_const\_step\] using , and we omit the details.
The proof of Theorem \[th:double\_loop\_convergence\]: Double-loop with constant step-size {#apdx:th:double_loop_convergence}
------------------------------------------------------------------------------------------
Similar to the proof of in Theorem \[th:singe\_loop\_const\_step\], we have $$\label{eq:single_loop2}
\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \frac{2}{\eta}\Big[\Exp{f(x_0^{(s)})} - \Exp{f(x_{m+1}^{(s)})}\Big] + \frac{2(m+1)\sigma^2}{\sqrt{b(m+1)}},$$ where we use the superscript $s$ to indicate the stage $s$ in Algorithm \[alg:A2\]. Summing up this inequality from $s=1$ to $s = S$, and then multiplying the result by $\frac{1}{(m+1)S}$ and using $\Exp{f(x_{m+1}^{(S)})} \geq f^{\star} > -\infty$, we get $$\label{eq:est5d_1}
\begin{array}{ll}
\frac{1}{S(m+1)}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} & \leq \frac{2}{\eta S(m+1)}\big[f(\tilde{x}^0) - f^{\star}\big] + \frac{2 \sigma^2}{\sqrt{b (m+1)}} \vspace{1ex}\\
& \leq \frac{3 L b^{1/4}}{S (m+1)^{3/4}} \big[f(\tilde{x}^0) - f^{\star}\big] + \frac{2 \sigma^2}{\sqrt{b (m+1)}}.
\end{array}$$ Here, we use the fact that $\eta \geq \frac{2}{3{{L}}\big[ b(m+1)\big]^{1/4}}$ from in the last inequality.
If we choose $b := \frac{c_1\sigma^2}{\varepsilon^2}$ and $m + 1 := \frac{c_2\sigma^2}{\varepsilon^2}$ for some constants $c_1 > 0$ and $c_2 > 0$ and $c_1c_2 > 4$, then, from , to guarantee $\frac{1}{S(m+1)}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \varepsilon^2$, we require $$\begin{aligned}
\frac{3 L b^{1/4}}{S (m+1)^{3/4}} \big[f^0 - f^{\star}\big] + \frac{2 \sigma^2}{\sqrt{b (m+1)}} &= \frac{3 L c_1^{1/4} \sigma^{1/2}}{\varepsilon^{1/2}} \cdot \frac{\varepsilon^{3/2}}{S c_2^{3/4} \sigma^{3/2}} \big[f^0 - f^{\star}\big] + \frac{2 \sigma^2 \varepsilon^2} {\sigma^2 \sqrt{c_1 c_2}} = \varepsilon^2 \\
& \Leftrightarrow \frac{3 L c_1^{1/4} \varepsilon}{S c_2^{3/4} \sigma} \big[f^0 - f^{\star}\big] = \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^2 \\
& \Leftrightarrow S = \frac{3 L c_1^{1/4} \big[f^0 - f^{\star}\big]}{c_2^{3/4} \sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon}.\end{aligned}$$ Consequently, the total complexity is $$\begin{array}{ll}
\Tc_{ge} &:= (b + 3m)S = (c_1 + 3 c_2) \frac{\sigma^2}{\varepsilon^2} \frac{3 L c_1^{1/4} \big[f^0 - f^{\star}\big]}{c_2^{3/4} \sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon} \vspace{1ex}\\
&= \frac{3 L (c_1 + 3 c_2)c_1^{1/4} \big[f^0 - f^{\star}\big]\sigma}{c_2^{3/4}\left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^3} = \BigO{\frac{\sigma}{\varepsilon^3}}.
\end{array}$$ Since we choose $b := \frac{c_1\sigma^2}{\varepsilon^2}$, the final complexity is $\BigO{\max\set{\frac{\sigma}{\varepsilon^3}, \frac{\sigma^2}{\varepsilon^2}}}$, where other constants independent of $\sigma$ and $\varepsilon$ are hidden.
Appendix: The convergence analysis of the mini-batch variants {#apdx:sec:mini_batch}
=============================================================
In this supplementary document, we provide a full analysis of the mini-batch variants of Algorithm \[alg:A1\] and Algorithm \[alg:A2\].
Variance bound of mini-batch hybrid estimators {#apdx:le:upper_bound_new_batch}
----------------------------------------------
For $\hat{v}_t$ defined by , we have the following property.
\[le:key\_pro\_of\_vhat\_t\] The mini-batch gradient estimator $\hat{v}_t$ defined by satisfies $$\label{eq:key_pro_of_vhat_t}
{\!\!\!\!\!}\begin{array}{ll}
\Exps{(\Bc_t, \hat{\Bc}_t)}{\norms{\hat{v}_t - \nabla{f}(x_t)}^2} &= \beta_{t-1}^2\norms{\hat{v}_{t-1} - \nabla{f}(x_{t-1})}^2 - \rho\beta_{t-1}^2\norms{\nabla{f}(x_{t-1}) - \nabla{f}(x_t)}^2 \vspace{1ex}\\
& + {~} \rho\beta_{t-1}^2\Exps{\xi}{\norms{\nabla{f}(x_t;\xi) - \nabla{f}(x_{t-1};\xi)}^2} \vspace{1ex}\\
& + {~} (1-\beta_{t-1})^2\rho\sigma^2,
\end{array}{\!\!\!\!\!}$$ where $\rho = \rho(\hat{b}) := \frac{n - \hat{b}}{(n-1)\hat{b}}$ if $n := \vert\Omega\vert$ is finite, and $\rho(\hat{b}) := \frac{1}{\hat{b}}$, otherwise.
Let $\hat{v}_t$ be defined by . Let $z_t := \frac{1}{b_t}\sum_{i\in\Bc_t}(\nabla{f}_{\xi_i}(x_t) - \nabla{f}_{\xi_i}(x_{t-1}))$, $\bar{z} := \nabla{f}(x_t) - \nabla{f}(x_{t-1})$, $\Delta_t := \hat{v}_t - \nabla{f}(x_t)$, and $\Delta{u}_t := u_t - \nabla{f}(x_t)$. Clearly, we have $$\Exp{z_t} = \bar{z} ~~~\text{and}~~~\Exp{\Delta{u}_t} = 0.$$ Moreover, we can rewrite $\hat{v}_t$ in as $$\Delta_t = \beta_{t-1}\Delta_{t-1} + \beta_{t-1}z_t + (1-\beta_{t-1})\Delta{u}_t - \beta_{t-1}\bar{z}.$$ Therefore, using these two expressions, we can derive $$\label{eq:bound_v_hat_t}
\begin{array}{ll}
\Exp{\norms{\Delta_t}^2} &= \beta_{t-1}^2\norms{\Delta_{t-1}}^2 + \beta_{t-1}^2\Exp{\norms{z_t}^2} + (1-\beta_{t-1})^2\Exp{\norms{\Delta{u}_t}^2} + \beta_{t-1}^2\norms{\bar{z}}^2\vspace{1ex}\\
& + {~} 2\beta_{t-1}^2\iprods{\Delta_{t-1},\Exp{z_t}} + 2\beta_{t-1}(1-\beta_{t-1})\iprods{\Delta_{t-1}, \Exp{\Delta{u}_t}} - 2\beta_{t-1}^2\iprods{\Delta_{t-1},\bar{z}} \vspace{1ex}\\
& + {~} 2\beta_{t-1}(1-\beta_{t-1})\Exp{\iprods{z_t, \Delta{u}_t}} - 2\beta_{t-1}^2\iprods{\Exp{z_t}, \bar{z}} - 2\beta_{t-1}(1-\beta_{t-1})\iprods{\Exp{\Delta{u}_t}, \bar{z}} \vspace{1ex}\\
&= \beta_{t-1}^2\norms{\Delta_{t-1}}^2 + \beta_{t-1}^2\Exp{\norms{z_t}^2} + (1-\beta_{t-1})^2\Exp{\norms{\Delta{u}_t}^2} - \beta_{t-1}^2\norms{\bar{z}}^2.
\end{array}$$ For the finite-sum case, after a few elementary calculations, we can show that $$\Exp{\norms{z_t}^2} = \frac{n(b_t-1)}{(n-1)b_t}\norms{\bar{z}}^2 + \frac{(n-b_t)}{(n-1)b_t}\Exps{\xi}{\norms{\nabla{f}_{\xi}(x_t) - \nabla{f}_{\xi}(x_{t-1})}^2}.$$ For the expectation case, we have $$\Exp{\norms{z_t}^2} = \big(1- \frac{1}{b_t}\big)\norms{\bar{z}}^2 + \frac{1}{b_t}\Exps{\xi}{\norms{\nabla{f}_{\xi}(x_t) - \nabla{f}_{\xi}(x_{t-1})}^2}.$$ In addition, under Assumption \[as:A1\]$($c$)$, we have $\Exp{\norms{\Delta{u}_t}^2} \leq \rho\sigma^2$.
Substituting one of the two last expressions and the bound of $\Exp{\norms{\Delta{u}_t}^2}$ into , we get .
The following analysis is given under fixed mini-batch sizes when we choose $\hat{b}_t = \tilde{b}_t = \hat{b}$. Similar to Lemma \[le:upper\_bound\_new\], we can bound the variance $\Exp{\|\hat{v}_t - \nabla f(x_t)\|^2}$ of the mini-batch hybrid estimator $\hat{v}_t$ from in the following lemma.
\[le:upper\_bound\_new\_batch\] Assume that $f(\cdot,\cdot)$ is $L$-smooth and $u_t$ is an SGD estimator, $\hat{v}$ is given in , $\Bc_t$ and $\hat{\Bc}_t$ are mini-batches of the size $\hat{b}$. Then, we have the following upper bound on the variance $\Exp{\norms{\hat{v}_t - \nabla{f}(x_t)}^2}$: $$\label{eq:vt_variance_bound_new_batch}
\Exp{\|\hat{v}_t - \nabla f(x_t)\|^2} \le \omega_t \Exp{\|\hat{v}_0 - \nabla f(x_0)\|^2} + L^2\rho \sum_{i=0}^{t-1}\omega_{i,t} \Exp{\|x_{i+1} - x_{i}\|^2} + \rho S_t,$$ where the expectation is taking over all the randomness $\Fc_t := \sigma(v_0, v_1, \cdots, v_t)$, $\omega_{t} := \prod_{i=1}^{t}\beta_{i-1}^2$, $\omega_{i, t} := \prod_{j=i+1}^{t}\beta_{j-1}^2$ for $i=0,\cdots, t$, and $S_{t} := \sum_{i=0}^{t-1}\big(\prod_{j=i+2}^{t}\beta_{j-1}^2\big)(1-\beta_i)^2\sigma$ for $t \geq 0$. $\rho = \frac{n - \tilde{b}}{\tilde{b}(n-1)}$ if $|\Omega|$ is finite and $\rho = \frac{1}{\tilde{b}}$ otherwise.
From Lemma \[le:key\_pro\_of\_vhat\_t\], taking the expectation with respect to $\mathcal{F}_t := \sigma(v_0,v_1,\cdots, v_t)$, we have $$\begin{array}{ll}
\Exp{\norms{\hat{v}_t - \nabla{f}(x_t)}^2} &\leq \beta_{t-1}^2\Exp{\norms{\hat{v}_{t-1} - \nabla{f}(x_{t-1})}^2} \vspace{1ex}\\
& + {~} L^2\rho\beta_{t-1}^2 \Exp{\|x_t - x_{t-1}\|^2} + \rho(1-\beta_{t-1})^2\sigma^2.
\end{array}$$ Let $a_t^2 := \Exp{\norms{\hat{v}_t - \nabla{f}(x_t)}^2} $ and $r_t^2 = \Exp{\|x_{t +1}- x_{t}\|^2}$. By following inductive step as in the proof of Lemma \[le:upper\_bound\_new\], we obtain $$\begin{aligned}
a_t^2 &\le \left(\beta_{t-1}^2\cdots \beta_0^2 \right)a_0^2 + L^2\rho\left(\beta_{t-1}^2\cdots \beta_0^2\right)r_0^2 + \cdots + L^2\rho\beta_{t-1}^2r_{t-1}^2\\
&+ {~} \rho\left[\left(\beta_{t-1}^2\cdots\beta_1^2\right)(1-\beta_0)^2 + \cdots + (1-\beta_{t-1})^2 \right]\sigma^2.\end{aligned}$$ Using the definition of $\omega_{t}$, $\omega_{i, t}$, and $S_{t}$ in Lemma \[le:upper\_bound\_new\], the previous inequality becomes $$\begin{aligned}
a_t^2 &\le \omega_{t}a_0^2 + L^2\rho\sum_{i=0}^{t-1}\omega_{i,t}r_i^2 + \rho S_t,\end{aligned}$$ which is the same as .
The proof of Corollary \[co:mini\_batch\]: Single loop with constant step-size and mini-batches {#apdx:co:mini_batch}
-----------------------------------------------------------------------------------------------
Using Lemma \[le:upper\_bound\_new\_batch\] and following the same path of proof of Lemma \[le:key\_estimate\_of\_convergence\], we can show that $$\label{eq:co4_est1}
\begin{array}{ll}
\Exp{f(x_{m+1}} &\leq \Exp {f(x_0)} - \displaystyle\sum_{t=0}^m\frac{\eta}{2}\Exp{\norms{\nabla{f}(x_t)}^2} + \frac{\eta}{2} \left(\displaystyle\sum_{t=0}^m\omega_t\right)\Exp{\norms{\hat{v}_0 - \nabla{f}(x_0)}^2} \vspace{1ex}\\
& + {~} \frac{\rho\eta}{2}\displaystyle\sum_{t=0}^mS_t + \frac{1}{2}\hat{\Tc}_m,
\end{array}$$ where $$\widehat{\mathcal{T}}_m := \rho L^2\eta^3\sum_{t=0}^m \sum_{i=0}^{t-1}\omega_{i,t}\Exp{\norms{\hat{v}_i}^2} - \eta\sum_{t=0}^m\big(1 - L\eta\big)\Exp{\norms{\hat{v}_t}^2}.$$ Clearly, we can rewrite $\widehat{\mathcal{T}}_m$ as $$\begin{array}{ll}
\widehat{\mathcal{T}}_m &= \eta\Big[\displaystyle\frac{\beta^2(1-\beta^{2m})L^2\eta^2\rho}{1-\beta^2} - (1-L\eta)\Big]q_0^2 \vspace{1ex}\\
& + \eta\Big[\frac{\beta^2(1-\beta^{2(m-1)})L^2\eta^2\rho}{1-\beta^2} - (1-L\eta)\Big]q_1^2 + \cdots \vspace{1ex}\\
&+ {~} \eta\Big[\displaystyle\frac{\beta^2(1-\beta^{2})L^2\eta^2\rho}{1-\beta^2} - (1-L\eta)\Big]q_{m-1}^2 - \eta(1-L\eta)q_{m}^2,
\end{array}$$ where $q_t^2 := \Exp{\norms{\hat{v}_t}^2}$. To guarantee $\widehat{\mathcal{T}}_m \le 0$, we need to have $$\left\{\begin{array}{ll}
\displaystyle\frac{L^2\eta^2\rho\beta^2(1-\beta^{2m})}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
\displaystyle\frac{L^2\eta^2\rho\beta^2(1-\beta^{2(m-1)})}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
\cdots & \cdots \vspace{1ex}\\
\displaystyle\frac{L^2\eta^2\rho\beta^2(1-\beta^2)}{1-\beta^2} - (1-L\eta) &\leq 0\vspace{1ex}\\
- (1 - L\eta) &\leq 0.
\end{array}\right.$$ Let $\alpha^2_m := \frac{\beta^2(1-\beta^{2m})}{1 - \beta^2}$. Since $\alpha^2_1 < \alpha^2_2 < \dots < \alpha^2_m$, the last condition holds if $L^2\eta^2\rho\alpha_m^2 - (1 - L\eta) \le 0$. By tightening this condition, we obtain $$\eta := \frac{2}{L\left(1 + \sqrt{1 + 4\rho\alpha_m^2}\right)}~\text{with}~\alpha^2_m := \frac{\beta^2(1-\beta^{2m})}{1 - \beta^2},$$ which is exactly . Since $\alpha_m^2 \le \frac{\beta^2}{1 - \beta^2}$, we have $\eta \geq \underline{\eta}: = \frac{2\sqrt{1-\beta^2}}{L(\sqrt{1-\beta^2} + \sqrt{1 + \beta^2(4\rho - 1)})}$.
Next, we can reuse the following estimates as in the proof of Theorem \[th:singe\_loop\_const\_step\]: $$\begin{array}{rl}
\displaystyle\sum_{t=0}^mS_t &\le \displaystyle\frac{\sigma^2(1-\beta)(m+1)}{1 +\beta}\\
\displaystyle\sum_{t=0}^m \omega_t &= \displaystyle\frac{1 - \beta^{2(m+1)}}{1 - \beta^2} \le \frac{1}{1 - \beta^2}.
\end{array}$$ Combining these estimate into and notting that $\widehat{\Tc}_m \leq 0$ and $\Exp{\norms{v_0 - \nabla{f}(x_0)}^2} \leq \frac{\sigma^2}{b}$, we can show that $$\label{eq:est5c_1_batch}
\begin{array}{ll}
\Exp{f(x_{m+1})} &\overset{\tiny\eqref{eq:bound_of_st_omegat}}{\leq} \Exp{f(x_0)} - \frac{\eta}{2}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \vspace{1ex}\\
& + {~} \frac{\eta\sigma^2}{2(1+\beta)}\Big[\frac{1}{(1-\beta)b} + \rho(1-\beta)(m+1)\Big].
\end{array}$$ Note that $\Exp{f(x_{m+1})} \ge f^{\star} > -\infty$, can be rewritten as $$\label{eq:main_estimate_for_single_loop_batch1}
\frac{\eta}{2}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \Exp{f(x_0)} - f^{\star} + \frac{\eta\sigma^2}{2(1+\beta)}\Big[\frac{1}{(1-\beta)b} + \rho(1-\beta)(m+1)\Big].$$ If we choose $\beta := 1 - \frac{c_1}{\sqrt{\hat{\rho}b(m+1)}}$ for any $0 < c_1 < \sqrt{b(m+1)}$ such that $$\frac{1}{(1-\beta)b(m+1)} + \rho(1-\beta) = \left(c_1 + \frac{1}{c_1}\right)\sqrt{\frac{\rho}{b(m+1)}},$$ then leads to $$\label{eq:single_loop_batch}
\begin{array}{ll}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} &\leq \frac{L}{(m+1)}\left(\frac{\sqrt{1-\beta^2} + \sqrt{1+\beta^2(4\rho - 1)}}{\sqrt{1-\beta^2}}\right)\Big[\Exp{f(x_0)} - f^{\star}\Big] \vspace{1ex}\\
& + {~} \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{(1+\beta)}\sqrt{\frac{\rho}{b(m+1)}}.
\end{array}$$ Since $\beta = 1 - \frac{c_1}{\sqrt{\rho b(m+1)}} < 1$ and if we choose $b$, $\hat{b}$, and $m$ such that $\rho b(m+1) > c_1^2$, we have $$\begin{array}{ll}
1 - \beta^2 &= 1 - \left(1 - \frac{c_1}{\sqrt{\rho b(m+1)}} \right)^2 = \frac{2c_1}{\sqrt{\rho b(m+1)}} - \frac{c_1^2}{\rho b(m+1)} = \frac{2c_1\sqrt{\rho b(m+1)} - c_1^2}{\rho b(m+1)} > \frac{2c_1}{\sqrt{\rho b(m+1)}},
\end{array}$$ and $\sqrt{1 - \beta^2} + \sqrt{1+\beta^2(4\rho - 1)} \leq 1 + \sqrt{1+3\beta^2} \leq 3$ since $\rho \leq 1$. Therefore, we can bound $\eta$ as $$\eta \geq \underline{\eta} \geq \frac{2c_1}{3L\big[ \rho b(m+1)\big]^{1/4}}.$$ Therefore, the inequality can be rewritten as $$\begin{array}{ll}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} &\leq \frac{3 L(\rho b)^{1/4}}{2c_1(m+1)^{3/4}}\left(\Exp{f(x_0)} - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{(1+\beta)}\sqrt{\frac{\rho}{b(m+1)}}.\\
&\leq \frac{3 L(\rho b)^{1/4}}{2c_1(m+1)^{3/4}}\left(\Exp{f(x_0)} - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{2}\sqrt{\frac{\rho}{b(m+1)}}.
\end{array}$$ Let $f^0 := \Exp{f(x_0)}$. We can write the bound as $$\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{3 L(\rho b)^{1/4}}{2c_1(m+1)^{3/4}}\left( f^0 - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{\sigma^2}{2}\sqrt{\frac{\rho}{b(m+1)}}.$$ Let us choose $b := c_2\sigma^{8/3}(\rho(m+1))^{1/3}$ for some $c_2 > 0$. Then, the last inequality leads to $$\label{eq:single_loop_final_batch}
\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t)}^2} \leq \frac{\rho^{1/3}\sigma^{2/3}}{(m+1)^{2/3}}\left[\frac{3 Lc_2^{1/4}}{2c_1}\left( f^0 - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{2\sqrt{c_2}}\right].$$ From , to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_m)}^2} \leq \varepsilon^2$, we need to choose $$\frac{\rho^{1/3}\sigma^{2/3}}{(m+1)^{2/3}}\left[\frac{3 Lc_2^{1/4}}{2c_1}\left( f^0 - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{2\sqrt{c_2}}\right] \leq \varepsilon^2,$$ which leads to $$m + 1 \geq \frac{\rho^{1/2}\sigma}{\varepsilon^3}\left[\frac{3 Lc_2^{1/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + \left(c_1 + \frac{1}{c_1}\right)\frac{1}{2\sqrt{c_2}}\right]^{3/2}.$$ Hence, we can choose $m$ as in .
Finally, let $c_1 = 1$. Then the number of stochastic gradient evaluations $\Tc_{ge}$ is $$\begin{array}{ll}
\Tc_{ge} &= b + 3\hat{b}m \leq b + \frac{3(m+1)}{\rho} \vspace{1ex}\\
&\leq c_2\sigma^{8/3} \left[\rho(m+1)\right]^{1/3} + \frac{3\sigma}{\rho^{1/2}\varepsilon^3}\left[\frac{3 Lc_2^{1/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + \frac{1}{\sqrt{c_2}}\right]^{3/2}\vspace{1ex}\\
&\leq \frac{\rho^{1/2}\sigma^{3}}{\varepsilon}\left[\frac{3 Lc_2^{9/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + c_2^{3/2}\right]^{1/2} + \frac{3\sigma}{\rho^{1/2}\varepsilon^3}\left[\frac{3 Lc_2^{1/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + \frac{1}{\sqrt{c_2}}\right]^{3/2},
\end{array}$$ which proves , where $\rho \le \frac{1}{\hat{b}}$ if $\vert\Omega\vert$ is infinite and $\rho := \frac{n-\hat{b}}{\hat{b}(n-1)}$ if $\vert\Omega\vert$ is finite. In particular, if $\vert\Omega\vert$ is infinite and we choose $\rho := \frac{c_3^2}{\sigma^2\varepsilon^2}$ for some $c_3 \leq \sigma\varepsilon$, then $$\begin{aligned}
\Tc_{ge} &= \frac{c_3\sigma^{2}}{\varepsilon^2}\left[\frac{3 Lc_2^{9/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + c_2^{3/2}\right]^{1/2} + \frac{3\sigma^2}{c_3\varepsilon^2}\left[\frac{3 Lc_2^{1/4}}{2{c_1}}\left( f^0 - f^{\star}\right) + \frac{1}{\sqrt{c_2}}\right]^{3/2}.\end{aligned}$$ Hence, we obtain $\Tc_{ge} = \BigO{\big(c_3 + \frac{1}{c_3}\big)\frac{\sigma^2}{\varepsilon^2}}$.
The mini-batch variant of Algorithm \[alg:A2\] and its complexity {#apdx:th:double_loop_convergence_mini_batch}
-----------------------------------------------------------------
Let us consider a mini-batch variant of Algorithm \[alg:A2\]. Similar to Theorem \[th:double\_loop\_convergence\], we can prove the following result.
\[co:double\_loop\_convergence\_minibatch\] Let $\sets{x^{(s)}_t}_{t=0\to m}^{s=1\to S}$ be the sequence generated by the mini-batch variant of Algorithm \[alg:A2\] using constant step-size $\eta$ defined in with $c_1 := 1$. Then, the following estimate holds $$\label{eq:double_loop_est2}
\frac{1}{S(m+1)}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \frac{3 L \rho(\hat{b}) b^{1/4}}{S (m+1)^{3/4}} \big[f(\tilde{x}^0) - f^{\star}\big] + \frac{2 \sigma^2\sqrt{\rho(\hat{b})}}{\sqrt{b (m+1)}}.$$ Let $\widetilde{x}_T \sim \Ub(\sets{x^{(s)}_t}_{t=0\to m}^{s=1\to S})$. If we choose $b := \frac{c_1\sigma^2}{\varepsilon^2}$ and $\frac{m + 1}{\hat{b}} := \frac{c_2\sigma^2}{\hat{b}^2\varepsilon^2}$ for some constants $c_1 > 0$ and $c_2 > 0$ and $c_1c_2 > 4$, then, to guarantee $\Exp{\norms{\nabla{f}(\widetilde{x}_T)}^2} \leq \varepsilon^2$, we require $$\label{eq:S_iterations2}
S := \frac{3 L c_1^{1/4} \big[f(\widetilde{x}^0) - f^{\star}\big]}{c_2^{3/4} \hat{b}^{3/2}\sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon}.$$ Consequently, the total number of stochastic gradient evaluations $\Tc_{ge}$ does not exceed $$\label{eq:Toc_double_loop2}
\Tc_{ge} := \left(b + 3\lfloor\tfrac{m}{\hat{b}}\rfloor\right)S = \frac{3 L (c_1 + 3 c_2)c_1^{1/4} \big[f(\widetilde{x}^0) - f^{\star}\big]\sigma}{c_2^{3/4}\hat{b}^{3/2}\left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^3} = \BigO{\frac{\sigma}{\varepsilon^3}}.$$
First, similar to the proof of , we have $$\frac{1}{m+1}\displaystyle\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \frac{3 L(\rho b)^{1/4}}{(m+1)^{3/4}}\left( \Exp{f(x_0^{(s)})} - \Exp{f(x_{m+1}^{(s)})}\right) + 2\sigma^2\sqrt{\frac{\rho}{b(m+1)}}.$$ Summing up this inequality from $s = 1$ to $s = S$ and then using $\Exp{f(x_{m+1}^{(S)})} \geq f^{\star}$ and $\widetilde{x}_0 := x_0^{(1)}$, we can show that $$\frac{1}{(m+1)S}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \frac{3 L(\rho b)^{1/4}}{S(m+1)^{3/4}}\left( \Exp{f(\widetilde{x}_0} - f^{\star}\right) + 2\sigma^2\sqrt{\frac{\rho}{b(m+1)}}.$$ If we choose $b := \frac{c_1\sigma^2}{\hat{b}^2\varepsilon^2}$ and $\frac{m + 1}{\hat{b}} := \frac{c_2\sigma^2}{\varepsilon^2}$ for some constants $c_1 > 0$ and $c_2 > 0$ and $c_1c_2 > 4$, then, $\rho(m+1) = \frac{c_2\sigma^2}{\varepsilon^2}$, $b = \frac{c_1\rho^2\sigma^2}{\varepsilon^2}$, and from , to guarantee $\frac{1}{S(m+1)}\displaystyle\sum_{s=1}^S\sum_{t=0}^m\Exp{\norms{\nabla{f}(x_t^{(s)})}^2} \leq \varepsilon^2$, we require $$\begin{aligned}
\frac{3 L (\rho b)^{1/4}}{S (m+1)^{3/4}} \big[f^0 - f^{\star}\big] + \frac{2 \sigma^2\sqrt{\rho}}{\sqrt{b (m+1)}} &= \frac{3 L c_1^{1/4}\rho^{3/4} \sigma^{1/2}}{\varepsilon^{1/2}} \cdot \frac{\rho^{3/4}\varepsilon^{3/2}}{S c_2^{3/4} \sigma^{3/2}} \big[f^0 - f^{\star}\big] + \frac{2\sigma^2 \varepsilon^2} {\sigma^2 \sqrt{c_1 c_2}} = \varepsilon^2 \\
& \Leftrightarrow \frac{3 L c_1^{1/4}\rho^{3/2} \varepsilon}{S c_2^{3/4} \sigma} \big[f^0 - f^{\star}\big] = \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^2 \\
& \Leftrightarrow S = \frac{3 L \rho^{3/2} c_1^{3/4} \big[f^0 - f^{\star}\big]}{c_2^{3/4} \sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon}.\end{aligned}$$ Consequently, the total complexity is $$\begin{array}{ll}
\Tc_{ge} &:= (b + 3\frac{m}{\hat{b}})S \leq (c_1 + 3 c_2) \frac{\sigma^2}{\varepsilon^2} \frac{3 L c_1^{1/4}\rho^{3/2} \big[f^0 - f^{\star}\big]}{c_2^{3/4} \sigma \left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon} \vspace{1ex}\\
&= \frac{3 L (c_1 + 3 c_2)c_1^{1/4} \big[f^0 - f^{\star}\big]\sigma}{c_2^{3/4}\hat{b}^{3/2}\left( 1 - \frac{2}{\sqrt{c_1 c_2}} \right) \varepsilon^3} = \BigO{\frac{\sigma}{\varepsilon^3}}.
\end{array}$$ Since we choose $b\hat{b}^2 := \frac{c_1\sigma^2}{\varepsilon^2}$ which shows that $b \leq \frac{c_1\sigma^2}{\varepsilon^2}$, the final complexity is $\BigO{\max\set{\frac{\sigma}{\varepsilon^3}, \frac{\sigma^2}{\varepsilon^2}}}$, where other constants independent of $\sigma$ and $\varepsilon$ are hidden.
Appendix: Additional numerical experiments {#apdx:subsec:experiments}
==========================================
In this subsection, we provide more numerical examples on two examples we tested in the main text.
Experiment setup
----------------
**Our algorithms:** We implement the following variants of Algorithm \[alg:A1\] and Algorithm \[alg:A2\] in Python:
- **Single-loop algorithms**: We consider different variants of the single-loop algorithm, Algorithm \[alg:A1\]. We denote them by `Hybrid-SGD-SL` for constant step-size variants, and `Hybrid-SGD-ASL` for adaptive step-size variants.
- **Double-loop algorithms**: These are variants of Algorithm \[alg:A2\]. We denote them by `Hybrid-SGD-DL[1-3]` the variants corresponding to different snapshot gradient batch-sizes of $b = n^{2/3}$, $b = 0.1n$, and $b = n$. We also denote `Hybrid-SGD-DL` as the best variants among these three choices of the batch-size for snapshot gradient.
**Competitors:** We also compare our methods with the most state-of-the-art candidates from the literature. We ignore other variants since their complexity bound is worse than ours and they use complicated routines for hyper-parameter selection.
- Stochastic gradient descent (SGD): We test two variants of SGD. The first one, called SGD1, is with constant step-size $\eta_t := \frac{0.1}{L}$. The second variant, called SGD2, is with an adaptive step-size of the form $\eta_t := \frac{\eta_0}{1 + \eta'\lfloor t/n\rfloor}$, where $\eta_0 > 0$ and $\eta' \geq 0$ are carefully tuned to obtain the best performance. In our tests, we use $\eta_0 := \frac{0.1}{L}$ and $\eta' := 1$.
- SVRG: This algorithmic variant is from [@nonconvexSVRG], where its theoretical step-size in the single sample case is $\eta_t := \frac{1}{3nL}$, and in the mini-batch case is $\eta_t :=\frac{1}{3L} $.
- SVRG+: This is a variant of SVRG studied in [@li2018simple]. Its theoretical step-size in the single sample case is $\eta_t := \frac{1}{6nL}$, and in the mini-batch case is $\eta_t := \frac{1}{6L}$.
- SPIDER: SPIDER [@fang2018spider] is a stochastic gradient method using SARAH estimator (also called Stochastic Path-Integrated Differential EstimatoR). This method achieves the best-known complexity as Algorithm \[alg:A2\] but uses very different step-size $\eta_t := \min\set{\frac{\epsilon}{Ln_0\norm{v^k}},\frac{1}{2Ln_0}}$ where $n_0 = \frac{n^{1/2}}{\hat{b}}$ with $\hat{b}$ is a given mini-batch size in the range $[1,\sqrt{n}]$.
- SpiderBoost: SpiderBoost [@wang2018spiderboost] is a modification of SPIDER by using a large constant step-size $\eta_t := \frac{1}{2L}$, but requires to set very specific mini-batch $\hat{b} = \lfloor\sqrt{n}\rfloor$ to achieve the best-known complexity as in Algorithm \[alg:A2\].
**Problems:** We consider three examples: The first one is the logistic regression with non-convex regularizer as in . The second example is a binary classification with non-convex loss as in . **Datasets:** All the datasets used in this paper are downloaded from LibSVM [@CC01a] at\
[https://www.csie.ntu.edu.tw/ cjlin/libsvm/](https://www.csie.ntu.edu.tw/~cjlin/libsvm/). We select 6 datasets: `w8a` ($n=49,749, p=300$), `rcv1.binary` ($n=20,242, p = 47,236$), `real-sim` $(n=72,309, p = 20,958$), `news20.binary` ($n=19,996, p = 1,355,191$), `url_combined` ($n=2,396,130, p = 3,231,961$), and `epsilon` ($n=400,000, p = 2,000$).
Logistic regression with non-convex regularizer
-----------------------------------------------
In this section, we add more numerical examples to solve problem . Together with the convergence of the trainning loss and gradient norms in Fig. \[fig:logistic\_reg\], the training and test accuracies are also plotted in Fig. \[fig:logistic\_acc1\] for three datasets: `w8a`, `rcv1.binary`, and `real-sim`.
As we can observe from Fig. \[fig:logistic\_acc1\], for `w8a`, all the algorithms except for SVRG achieve similar training accuracy as well as test accuracy. SVRG eventually reaches the same accuracy after around 17 epochs. For `rcv1.binary`, HybridSGD variants, SGD2, and SVRG+ have similar training and test accuracies, but SGD2 is more oscillated than the other methods. SGD1 performs worse than our methods in this case. Both SPIDER and SVRG still perform poorly. For `real-sim`, although our methods, SGD1, and SGD2 achieve lower training accuracy, they are able to reach better test accuracy than SVRG+.
In addition, the training and testing accuracies of the mini-batch case are presented in Fig. \[fig:logistic\_acc2\], where the relative residual of the train loss and the gradient norms are shown in Fig. \[fig:logistic\_reg2\]. Again, our methods achieve training and test accuracies consistently with SGD2 in `w8a` and `real-sim`, while having better accuracy in `rcv1.binary`.
We also run SVRG, SVRG+, SpiderBoost, and our double-loop variant (Algorithm \[alg:A2\]) on three datasets: `w8a`, `rcv1.binary`, and `real-sim`. The results are plotted in Fig. \[fig:logistic\_reg4\].
In this experiment, our double-loop variant and SpiderBoost outperform SVRG and SVRG+. Although the step-size of SVRG+ is $\eta = \frac{1}{6L}$ which is smaller than $\frac{1}{3L}$ in SVRG, SVRG+ still performs better than SVRG. SpiderBoost uses a large step-size $\eta = \frac{1}{2L}$ and it indeed performs slightly better than ours in the `w8a` dataset, but is comparable in other two. Note that our step-size $\eta$ is selected based on our theory in Theorem \[th:double\_loop\_convergence\].
Finally, we conduct experiment on three larger datasets: `epsilon`, `url_combined`, and `news20.binary`. Since the the sample sizes are large, we only run mini-batch variants. The results of the single-loop variants are shown in Fig. \[fig:larger\_datasets1\].
We can observe from Fig. \[fig:larger\_datasets1\] that our single loop variants outperform SGD in all three datasets. Note that the performance of the adaptive step-size variant is similar to its fixed step-size one.
The results of the double-loop variants are also shown in Fig. \[fig:larger\_datasets2\].
Clearly, our double-loop variants achieve better performance than SVRG and SVRG+ due to better convergence rate. SpiderBoost is slightly better than ours in the dataset `epsilon` while they are comparable in the last two datasets since we have the same best-known convergence rate as SpiderBoost.
Binary classification involving non-convex loss and Tikhonov’s regularizer {#subsec:example2}
--------------------------------------------------------------------------
We also conduct additional experiments to test our algorithms for solving . We use two different non-convex loss functions as in [@zhao2010convex] apart from the one used in the main text, which are:
- *Nonconvex loss in two-layer neural networks*: $\ell_1(\tau, s) = \left(1 - \frac{1}{1+\exp(-\tau s)}\right)^2$.
- *Logistic difference loss*: $\ell_2(\tau, s) = \log(1 + \exp(\tau s)) - \log(1 + \exp(-\tau s - 1))$.
These functions are smooth and satisfy Assumption \[as:A1\].
Let us first test our algorithms and other methods on three datasets: `w8a`, `rcv1.binary`, and `real-sim` using single-sample setting. The results are plotted in Fig. \[fig:binary\_single2\] and Fig. \[fig:binary\_single3\]. In this test, HybridSGD-DL achieves the best performance followed by HybridSGD-SL and HybridSGD-ASL. SPIDER has decent performance in the last two datasets. SGD variants also have good performance in all datasets while SGD2 is better than its fixed step-size variant. SVRG+ also has comparable performance with SGD2 whereas SVRG cannot achieve fast convergence due to its small step-size.
Next, we test mini-batch variants. On the one hand, we compare our single-loop variants HybridSGD-SL and HybridSGD-ASL with SGD. On the other hand, we compare our double-loop variants with SVRG, SVRG+, and SpiderBoost. The results for solving with loss $\ell_1$ are shown in Fig. \[fig:binary\_batch2\_21\] and Fig. \[fig:binary\_batch2\_22\] whereas Fig. \[fig:binary\_batch2\_31\] and Fig. \[fig:binary\_batch2\_32\] present the results when using loss $\ell_2$.
Additionally, we repeat the experiments on three larger datasets: `epsilon`, `news20.binary`, and `ulr_combined`. The results are shown in Fig. \[fig:binary\_batch4\_21\], \[fig:binary\_batch4\_22\], \[fig:binary\_batch4\_31\], and \[fig:binary\_batch4\_32\].
In this experiment, although SGD2 has faster decrease during the first few epochs, our HybridSGD-SL and HybridSGD-ASL eventually achieve lower training loss and gradient norm in all datasets while reaching similar training and testing accuracies as SGD2.
Regarding the double-loop variants, our HybridSGD-DL once again has better performance than SVRG and SVRG+ while having comparable performance with SpiderBoost in terms of training loss, gradient norm, and accuracies.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The property of desynchronization in an all-to-all network of homogeneous impulse-coupled oscillators is studied. Each impulse-coupled oscillator is modeled as a hybrid system with a single timer state that self-resets to zero when it reaches a threshold, at which event all other impulse-coupled oscillators adjust their timers following a common reset law. In this setting, desynchronization is considered as each impulse-coupled oscillator’s timer having equal separation between successive resets. We show that, for the considered model, desynchronization is an asymptotically stable property. For this purpose, we recast desynchronization as a set stabilization problem and employ Lyapunov stability tools for hybrid systems. Furthermore, several perturbations are considered showing that desynchronization is a robust property. Perturbations on both the continuous and discrete dynamics are considered. Numerical results are presented to illustrate the main contributions.'
author:
- 'Sean Phillips and Ricardo G. Sanfelice[^1]'
bibliography:
- 'Biblio.bib'
- 'RGS.bib'
- 'SAP.bib'
---
Introduction
============
Impulse-coupled oscillators are multi-agent systems with state variables consisting of timers that evolve continuously until a state-dependent event triggers an instantaneous update of their values. Networks of such oscillators have been employed to model the dynamics of a wide range of biological and engineering systems. In fact, impulse-coupled oscillators have been used to model groups of fireflies [@Mirollo.90.SIAMJAM.BiologicalOscillators], spiking neurons [@Pikovsky.ea.03.BiologicalOscillators; @gerstner2002spiking], muscle cells [@Peskin.75.BiologicalOscillators], wireless networks [@hong2010cooperative], and sensor networks [@Liu05adynamic]. With synchronization being a property of particular interest, such complex networks have been found to coordinate the values of their state variables by sharing information only at the times the events/impulses occur [@Mirollo.90.SIAMJAM.BiologicalOscillators; @Abbott1993].
The opposite of synchronization is [*desynchronization*]{}. In simple words, desynchronization in multi-agent systems is the notion that the agents’ periodic actions are separated “as far apart” as possible in time. Desynchronization is similar to clustering or splay-state configurations, and is sometimes referred in the literature as inhibited behavior [@mauroy:037122; @glass1988clocks]. For impulse-coupled oscillators, desynchronization is given as the behavior in which the separation between all of the timers impulses is equal [@Patel.Desync2007]. This behavior has been found to be present in communication schemes in fish [@Benda.Neuron] and in networks of spiking neurons [@Pfurtscheller19991842; @1997Natur.390.70S]. Desynchronization of oscillators has recently been shown to be of importance in the understanding of Parkinson’s disease [@Mabi.Moehlis2010; @Majtanik.Dolan2004], in the design of algorithms that limit the amount of overlapping data transfer and data loss in wireless digital networks [@hong2010cooperative], and in the design of round-robin scheduling schemes for sensor networks [@Liu05adynamic].
Motivated by the applications mentioned above and the lack of a full understanding of desynchronization in multi-agent systems, this paper pertains to the study of the dynamical properties of desynchronization in a network of impulse-coupled oscillators with an all-to-all communication graph. The uniqueness of the approach emerges from the use of hybrid systems tools, which not only conveniently capture the continuous and impulsive behavior in the networks of interest, but also are suitable for analytical study of asymptotic stability and robustness to perturbations.
More precisely, the dynamics of the proposed hybrid system capture the (linear) continuous evolution of the states as well their impulsive/discontinuous behavior due to state triggered events. Analysis of the asymptotic behavior of the trajectories (or solutions) to these systems is performed using the framework of hybrid systems introduced in [@teel2012hybrid; @Goebel.ea.09.CSM]. To this end, we recast the study of desynchronization as a set stabilization problem. Unlike synchronization, for which the set of points to stabilize is obvious, the complexity of desynchronization requires first to determine such a collection of points, which we refer to as the [*desynchronization set*]{}. We propose an algorithm to compute such set of points. Then, using Lyapunov stability theory for hybrid systems, we prove that the desynchronization set is asymptotically stable by defining a Lyapunov-like function as the distance between the state and (an inflated version of) the desynchronization set. In our context, asymptotic stability of the desynchronization set implies that the distance between the state and the desynchronization set converges to zero as the amount of time and the number of jumps get large. Using the proposed Lyapunov-like function and invoking an invariance principle, the basin of attraction is characterized and shown to be the entire state space minus a set of measure zero, which turns out to actually be an exact estimate of the basin of attraction. Furthermore, also exploiting the availability of a Lyapunov-like function, we analytically characterize the time for the solutions to reach a neighborhood of the desynchronization set. In particular, this characterization provides key insight for the design of algorithms used in applications in which desynchronization is crucial, such as wireless digital networks and sensor networks.
The asymptotic stability property of the desynchronization configuration is shown to be robust to several types of perturbations. The perturbations studied here include a generic perturbation in the form of an inflation of the dynamics of the proposed hybrid system model of the network of interest and several kinds of perturbations on the timer rates. Using the tools presented in [@teel2012hybrid; @Goebel.ea.09.CSM], we analytically characterize the effect of these perturbations on the already established asymptotic stability property of the desynchronization set. In particular, these perturbations capture situations where the agents in the network are heterogeneous due to having differing timer rates, threshold values, and update laws. To verify the analytical results, we simulate networks of impulse-coupled oscillators under several classes of perturbations. Specifically, we show numerical results when perturbations affect the update laws and the timer rates.
The remainder of this paper is organized as follows. Section \[sec:hs\] is devoted to hybrid modeling of networks of impulse-coupled oscillators. Section \[sec:AN\] introduces an algorithm to determine the desynchronization set. Section \[sec:lyapunov\] presents the stability results while the time to convergence is characterized in Section \[sec:timetoconverge\]. The robustness results are in Section \[sec:robustness\]. Section \[sec:numerics\] presents numerical results illustrating our results. Final remarks are given in Section \[sec:conclusion\].
[**Notation**]{}
Hybrid System Model of Impulse-Coupled Oscillators {#sec:hs}
==================================================
Mathematical Model {#sec:hybridmodel}
------------------
In this paper, we consider a model of $N$ impulse-coupled oscillators. Each impulse-coupled oscillator has a continuous state ($\tau_i$ for the $i$-th oscillator) defining its internal timer. Once the timer of any oscillator reaches a threshold (${\bar{\tau}}$), it triggers an impulse and is reset to zero. At such an event, all the other impulse-coupled oscillators rescale their timer by a factor given by $(1+ \varepsilon)$ times the value of their timer, where $\varepsilon \in (-1,0)$.[^2] Figure \[fig:example\] shows a trajectory of two impulse-coupled oscillators [with states]{} ${\tau_1}$ and ${\tau_2}$. In this figure, the dark red circles indicate when a timer state has reached the threshold and, thus, resets to zero. The light green circles indicate when an oscillator is externally reset and, hence, decreases its timer by $(1+\varepsilon)$ times its current state.
According to this outline of the model, the dynamics of the impulse-coupled oscillators involve impulses and timer resets, which are treated as true discrete events and instantaneous updates, while the smooth evolution of the timers before/after these events define the continuous dynamics. We follow the hybrid formalism of [@teel2012hybrid; @Goebel.ea.09.CSM], where a hybrid system is given by four objects $(C,f,D,G)$ defining its *data*:
- *Flow set:* a set $C \subset {\mathbb{R}}^{N}$ specifying the points where flows are possible (or continuous evolution).
- *Flow map:* a single-valued map $f: {\mathbb{R}}^{N} \to {\mathbb{R}}^{N}$ defining the flows.
- *Jump set:* a set $D \subset {\mathbb{R}}^{N}$ specifying the points where jumps are possible (or discrete evolution).
- *Jump map:* a set-valued map $G: {\mathbb{R}}^{N} \rightrightarrows{\mathbb{R}}^{N}$ defining the jumps.
A hybrid system capturing the dynamics of the impulse-coupled oscillators is denoted as ${{{\mathcal}H}}_N := (C,f,D,G)$ and can be written in the compact form $${{{\mathcal}H}}_N: \qquad \tau \in {\mathbb{R}}^{N} \qquad \left\{ \begin{array}{llll} \dot{\tau} &=& f(\tau) &\quad \tau \in C \\ \tau^{+} &\in& G(\tau) & \quad \tau \in D \end{array}\right. , \label{eqn:HS}$$ where $N \in {\mathbb{N}}\setminus \{0,1\}$ is the number of impulse-coupled oscillators. The state of ${{{\mathcal}H}}_{N}$ is given by The flow and jump sets are defined to constrain the evolution of the timers. The flow set is defined by $$\startmodif C := P_N, \stopmodif
\label{eqn:flawiest}$$ where $I := \{1,2, \ldots , N\}$ and ${\bar{\tau}}> 0$ is the threshold. During flows, an internal clock gradually increases based on the homogeneous rate, $\omega$. Then, the flow map is defined as with $\omega > 0$ [[ defining the natural frequency of each impulse-coupled oscillator]{}]{}. The impulsive events are captured by a jump [[ set $D$ and a jump]{}]{} map $G$. Jumps occur when the state is in the jump set $D$ defined as $$D := \left\{ \tau \in P_N : \ \exists i \in I \ \mbox{s.t.} \ \tau_i = {\bar{\tau}}\right\} .
\label{eqn:D}$$ From such points, the $i$-th timer is reset to zero and forces a jump of all other timers. Such discrete dynamics are captured by the following jump map: for each $\tau \in D$ define $
G(\tau) = \left[ g_1(\tau) \ \ g_2(\tau)\ \ \ldots \ \ g_N(\tau) \right]^\top,
$ where, for each $i \in I$, $$g_i (\tau) = \left\{ \begin{array}{l} 0 \qquad \qquad \ \ \ \ \mbox{if } \tau_{i} = {\bar{\tau}}, \tau_r < {\bar{\tau}}\ \ \forall r \in I\setminus \{i\} \\
\{0 , \tau_{i}(1+\varepsilon) \} \ \mbox{if } \tau_{i} = {\bar{\tau}}\ \exists r \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}\\
(1+\varepsilon)\tau_{i} \qquad \ \mbox{if } \tau_i < {\bar{\tau}}\ \exists r \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}\end{array} \right. \label{eqn:gi}$$ with parameters $\varepsilon \in (-1,0)$ and ${\bar{\tau}}> 0$; for $\tau \in D$, $g_i$ is not empty. When a jump is triggered, the state $\tau_i$ jumps according to the $i$-th component of the jump map $g_i$. When a state reaches the threshold ${\bar{\tau}}$, it is reset to zero only when all other states are less than that threshold; otherwise, if multiple timers reach the threshold simultaneously, the jump map is set valued to indicate that either $g_i(\tau) = 0$ or $g_i(\tau) = (1+\varepsilon)\tau_{i}$ is possible. This is to ensure that the jump map satisfies the regularity conditions outlined in Section \[sec:ModelForAnalysis\].[^3]
For example, consider the case $N=2$ the hybrid system ${{{\mathcal}H}}_N = (C,f,D,G)$ has state given by $$\tau=\left[ \begin{array}{c} \tau_{1} \\ \tau_{2}\end{array} \right] \in P_2 := [0,{\bar{\tau}}]\times[0,{\bar{\tau}}] .$$ The states $\tau_{1}$ and $\tau_{2}$ are the timers for both of the oscillators. The hybrid system ${{{\mathcal}H}}_2$ has the following data:
$${{{\mathcal}H}}_2 = \left\{\begin{array}{ll} C = P_2, & \qquad f(\tau) = \left[ \begin{array}{c} 1
\\ 1 \end{array}\right] \forall \tau \in C, \\ \noalign{\medskip}
D = \left\{ \tau \in P_2 \ : \ \exists i \in \{1,2\} \ s.t. \ \tau_{i} = {\bar{\tau}}\right\}, & \qquad
G(\tau) = \left[ \begin{array}{c} g_1(\tau) \\ g_2(\tau) \end{array} \right] \forall \tau \in D , \end{array} \right.$$ where the functions $g_1$ and $g_2$ are defined as $$g_1 (\tau) = \left\{ \begin{array}{ll} 0 & \mbox{if } \tau_{1} = {\bar{\tau}}, \tau_2 < {\bar{\tau}}\\ \{0 , \tau_{1}(1+\varepsilon) \} & \mbox{if } \tau_{1} = {\bar{\tau}}, \tau_2 = {\bar{\tau}}\\
(1+\varepsilon)\tau_{1} & \mbox{if } \tau_1 < {\bar{\tau}}, \tau_2 = {\bar{\tau}}\end{array} \right.
\qquad \qquad g_2 (\tau) = \left\{ \begin{array}{ll} 0 & \mbox{if } \tau_{2} = {\bar{\tau}}, \tau_1 < {\bar{\tau}}\\ \{0 , \tau_{2}(1+\varepsilon) \} & \mbox{if } \tau_{2} = {\bar{\tau}}, \tau_1 = {\bar{\tau}}\\
(1+\varepsilon)\tau_{2} & \mbox{if } \tau_2 < {\bar{\tau}}, \tau_1 = {\bar{\tau}}\end{array} \right. .$$
Basic Properties of ${{{\mathcal}H}}_{N}$ {#sec:BasicCond}
-----------------------------------------
\[sec:ModelForAnalysis\]
### Hybrid Basic Conditions
To apply analysis tools for hybrid systems in [@teel2012hybrid], which will be summarized in Section \[sec:analysis\], the data of the hybrid system ${{{\mathcal}H}}_{N}$ must meet certain mild conditions. These conditions, referred to as the [*hybrid basic conditions*]{}, are as follows:
1. $C$ and $D$ are closed sets in ${\mathbb{R}}^N$.
2. $f: {\mathbb{R}}^N\to{\mathbb{R}}^N$ is continuous on $C$.
3. $G :{\mathbb{R}}^N{\rightrightarrows}{\mathbb{R}}^N$ is an outer semicontinuous[^4] set-valued mapping, locally bounded on $D$, and such that $G(x)$ is nonempty for each $x \in D$.
\[lem:BasicConds\] ${{{\mathcal}H}}_{N}$ satisfies the hybrid basic conditions.
Note that satisfying the hybrid basic conditions [[ implies that ${{{\mathcal}H}}_N$ is well-posed [@teel2012hybrid Theorem 6.30], which automatically gives robustness to vanishing state disturbances; see [@teel2012hybrid; @Goebel.ea.09.CSM]. Section \[sec:robustness\] considers different types of perturbations that ${{{\mathcal}H}}_N$ can withstand. ]{}]{}
### Solutions to ${{{\mathcal}H}}_N$
\[lem:solutions2HS\] From every point in $C \cup D$, there exists a solution and every maximal solution to ${{{\mathcal}H}}_{N}$ is complete and bounded.
Due to the jump map $G$, if the elements of the solution are initially equal (denote this set as ${\mathcal{S}}:= \{\tau \in P_N : \exists i,r \in I, i \neq r, \tau_{i} = \tau_r \}$) it is possible for them to remain equal for all time. Furthermore, it is also possible for solutions to be initialized on the jump set such that one element is at the threshold and another is equal to zero then after the jump they will be equal, e.g. let $\tau_{1} = {\bar{\tau}}$, $\tau_{2} = 0$ then $\tau_{1}^{+} = \tau_{2}^{+} = 0$. We denote this set as ${\mathcal{G}}:= \{\tau \in D\setminus {\mathcal{S}}: \exists i,r \in I, i \neq r, \tau_{i} = 0, \tau_r = {\bar{\tau}}\}$. The next result considers solutions initialized on the set ${\mathcal{X}}:= {\mathcal{S}}\cup {\mathcal{G}}$.
\[lem:defX\] For each $\tau(0,0) \in {\mathcal{X}}$, there exists a solution $\tau$ to ${{{\mathcal}H}}_{N}$ from $\tau(0,0)$ such that, for some $M \in \{0,1\}$, $\tau(t,j) \in {\mathcal{S}}$ for all $t+j \geq M$, $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$.
Consider a solution $\tau$ to the hybrid system ${{{\mathcal}H}}_N$ with initial condition $\tau(0,0) \in \startmodif {\mathcal{S}}\stopmodif$. Due to the flow map for each state being equal, $\tau$ remains in ${\mathcal{S}}$ during flows. Furthermore, at points $\tau \in {\mathcal{S}}\cap D$, the jump map $G$ is set valued by the definition of $g_i$ in . From these points, $G(\tau) \cap {\mathcal{S}}\neq \emptyset$. In fact, for each $\tau(0,0) \in {\mathcal{S}}$, there exists at least one solution such that $\tau(t,j) \in {\mathcal{S}}$ for all $t + j \geq 0$, with $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$. Consider the case of solutions initialized at $\tau(0,0) \in \startmodif {\mathcal{G}}\stopmodif$ (Note that $\tau(0,0) \in D$). It follows that for some $r \in I$, $\tau_r(0,0) = {\bar{\tau}}$ and $g_r(\tau(0,0)) = 0$. Therefore, after the initial jump, we have that $G(\tau(0,0)) \cap {\mathcal{S}}\neq \emptyset,$ by which using previous arguments implies that $\tau(t,j) \in {\mathcal{S}}$ for all $t + j \geq 1$.
Furthermore, there is a distinct ordering to the jumps. If $\tau$ is such that $\tau_i \neq \tau_r$ for all $i\neq r$ then the ordering of each $\tau_i$ is preserved after $N$ jumps. More specifically, we have the following result.
\[lem:ordering\] For every solution $\tau$ to ${{{\mathcal}H}}_N$ with $\tau(0,0) \notin {\mathcal{X}}$, if at $(t_{j},j) \in {\mathop{\rm dom}\nolimits}\tau$ we have for some sequence of nonrepeated elements $\{i_m\}^N_{m = 1}$ of $I$ (that is, a reordering of the elements of the set $I = \{1,2,\ldots,N\}$) then, after $N$ jumps, it follows that
Let $\tau$ be a solution to ${{{\mathcal}H}}_{N}$ from $P_{N}\setminus {\mathcal{X}}$. There exists a sequence $i_{k}$ of distinct elements with $i_{k} \in I$ for each $k \in I$, such that $0 \leq \tau_{i_{1}}(t,j) < \tau_{i_{2}}(t,j) < \ldots < \tau_{i_{N}}(t,j) \leq {\bar{\tau}}$ over $[t_{0},t_{1}]\times \{0\}$. After the jump at $(t,j) = (t_{1},0)$ we have $0 = \tau_{i_{N}}(t,j+1) < \tau_{i_{1}}(t,j+1) < \tau_{i_{2}}(t,j+1) < \ldots < \tau_{i_{N-1}}(t,j+1) < {\bar{\tau}}$. Continuing this way for each jump, it follows that after $N-1$ more jumps, the solution is such that $0 \leq \tau_{i_{1}}(t_{N},j+N) < \tau_{i_{2}}(t_{N},j+N) < \ldots < \tau_{i_{N}}(t_{N},j+N) \leq {\bar{\tau}}$ and the order at time $(t,j)$ is preserved.
Using these properties of solutions to ${{{\mathcal}H}}_N$, the next section defines the set to which these solutions converge and establishes its stability properties.
Dynamical Properties of ${{{\mathcal}H}}_{N}$ {#sec:analysis}
=============================================
Our goal is to show that the desynchronization configuration [[ of ${{{\mathcal}H}}_N$, which is defined in Section \[sec:AN\]]{}]{}, is asymptotically stable. We recall from [@teel2012hybrid; @Goebel.ea.09.CSM] the following definition of asymptotic stability for general hybrid systems with state $x \in {\mathbb{R}}^n$.
\[def:AS\] A closed set ${\mathcal{A}}\subset {\mathbb{R}}^n$ is said to be
- [*stable*]{} if for each $\varepsilon>0$ there exists $\delta>0$ such that each solution $x$ with $|x(0,0)|_{{\mathcal{A}}}\leq \delta$ satisfies $|x(t,j)|_{{\mathcal{A}}} \leq \varepsilon$ for all $(t,j)\in{\mathop{\rm dom}\nolimits}x$;
- [*attractive*]{} if there exists $\mu > 0$ such that every maximal solution $x$ with $|x(0,0)|_{{\mathcal{A}}}\leq \mu$ is complete and satisfies\
$\lim_{(t,j) \in {\mathop{\rm dom}\nolimits}x, t+j\to\infty} |x(t,j)|_{{\mathcal{A}}}=0$;
- [*asymptotically stable*]{} if stable and attractive;
- [*weakly globally asymptotically stable*]{} if ${\mathcal{A}}$ is stable and if, for every initial condition, there exists a maximal solution that is complete and satisfies $\lim_{(t,j) \in {\mathop{\rm dom}\nolimits}x, t+j\to\infty} |x(t,j)|_{{\mathcal{A}}}=0$.
The set of points from where the attractivity property holds is the basin of attraction and excludes all points where the system trajectories may never converge to ${\mathcal{A}}$. In fact, it will be established in Section \[sec:lyapunov\] that the basin of attraction for asymptotic stability of desynchronization of ${{{\mathcal}H}}_N$ does not include any point $\tau$ such that any two or more timers are equal or become equal after a jump, which is the set ${\mathcal{X}}$ defined in Lemma \[lem:defX\].
Construction of the set ${\mathcal{A}}$ for ${{{\mathcal}H}}_N$ {#sec:AN}
---------------------------------------------------------------
In this section, we identify the set of points corresponding to the impulse-coupled oscillators being desynchronized, namely, we define the [*desynchronization set*]{}. We define desynchronization as the behavior in which the separation between all of the timers’ impulses is equal (and nonzero), see Figure \[fig:example\]. More specifically desynchronization is defined as follows:
\[def:desynch\] A solution $\tau$ to ${{{\mathcal}H}}_N$ is desynchronized if there exists $\Delta > 0$ and a sequence of non-repeated elements $\{i_m\}^N_{m = 1}$ of $I$ (that is, a reordering of the elements of the set $I = \{1,2, \ldots, N \}$) such that $\lim_{j \to \infty} (t_j^{i_m} - t_j^{i_{m+1}} )= \Delta$ for all $m \in \{1,2, \ldots, N-1\}$ and $\lim_{j \to \infty} (t_j^{N} - t_j^{i_{1}}) = \Delta,$ where $\{t_j^{i_m}\}_{j = 0}^\infty$ is the sequence of jump times of the state $\tau_{i_m}$.
In fact, this separation between impulses leads to an ordered sequence of impulse times with equal separation. The desynchronization set ${\mathcal{A}}$ for the hybrid system ${{{\mathcal}H}}_N$ captures such a behavior and is parameterized by $\varepsilon$, the threshold ${\bar{\tau}}$, and the number of impulse-coupled oscillators $N$.
To define this set, first we provide some basic intuition about the dynamics of ${{{\mathcal}H}}_N$ when desynchronized. The set ${\mathcal{A}}$ must be forward invariant and such that trajectories staying in it satisfy the property in Definition \[def:desynch\]. Due to the definition of the flow map $f$, there exist sets in the form of “lines" $\ell_k$, each of them in the direction ${\mathbf 1}$, which is the direction of the flow map, intersecting the jump set at a point which, for the $k$-th line, we denote as ${\widetilde}\tau^{k}$. We define the desynchronization set as the union of sets $\ell_{k}$ collecting points $\tau = {\widetilde}\tau^k + {\mathbf 1}s \in P_N$ parameterized by $s \in {\mathbb{R}}$.
To identify ${\widetilde}\tau^k$, consider a point ${\widetilde}\tau^{k} \in D\setminus {\mathcal{X}}$ with components satisfying ${\widetilde}\tau_{1}^{k} = {\bar{\tau}}> {\widetilde}\tau_{2}^{k} > {\widetilde}\tau_{3}^{k} > ... > {\widetilde}\tau_{N}^{k}$. Due to Definition \[def:desynch\], it must be true that the difference between jump times are constant. This means that there must be some correlation between $\Delta$ and the difference between, in this case, $\tau_{1}^{k}$ and $\tau_{2}^{k}$. Moreover, there must be a correlation between $\tau_{1}^{k}$ and all other states at jumps. It follows that this point belongs to ${\mathcal{A}}$ only if the distance between the expiring timer (${\widetilde}\tau_{1}^{k}$) and each of its other components (${\widetilde}\tau_{i}^{k}$, $i \in I \setminus \{1\}$) is equal to the distance between the value after the jump of the timer expiring next (${\widetilde}\tau_{2}^{k}\null^{+}$) and the value after the jump of its other components (${\widetilde}\tau_{i}^{k}\null^{+}$, $i \in I \setminus \{2\}$), respectively. This property ensures that, when in the desynchronization set, the relative distance between the leading timer and each of the other timers is equal, before and after jumps. More precisely, $$\begin{aligned}
{\widetilde}\tau_{1}^{k} - \widetilde{\tau}_{i}^{k} = {\widetilde}\tau_{2}^{k}\null^{+} - {\widetilde}\tau_{\mbox{\scriptsize next}(i)}^{k}\hspace{-0.21in}\null^{+} \qquad
\qquad \forall \ i \in I \setminus \{1\} \label{eqn:ANcondition},\end{aligned}$$ where ${\widetilde}\tau^{k}\null^{+} = G({\widetilde}\tau^{k})$ and next$(i) = i + 1$ if $i + 1 \leq N$ and $1$ otherwise.[^5] Since ${\mathcal{X}}$ contains all points such that at least two or more timers are the same, we can consider the case when one component of ${\widetilde}\tau^k$ is equal to ${\bar{\tau}}$ at a time. For each such case, we have $(N - 1)!$ possible permutations of the other components and $N$ possible timer components equal to ${\bar{\tau}}$, leading to $N!$ total possible sets $\ell_{k}$.
For the $N$ case, the algorithm above results in the system of equations $\Gamma \tau_s = b$, where and $
b = \bar{\tau} {\bf 1},
$ where $\tau_s$ is the state ${\widetilde}\tau^{k}$ sorted into decreasing order. It can be shown that for any $\varepsilon \in (-1,0)$, a solution $\tau_{s}$ exists (see Lemma \[eqn:tauSsolution\]). Then, $\tau_s$ needs to be unsorted and becomes ${\widetilde}\tau^k$ in the definition of the set $\ell_k$.
The solution to $\Gamma\tau_s = b$ is the result of a single case of $\tau \in D \setminus {\mathcal{X}}$. As indicated above, to get a full definition of the set ${\mathcal{A}}$, the $N!$ sets $\ell_k$ should be computed. For arbitrary $N$, the set ${\mathcal{A}}$ is given as a collection of sets $\ell_{k}$ given by $${\mathcal{A}}= \bigcup_{k =1}^{N!}\ell_{k} \label{eqn:AN},$$ where, for each $k \in \{1,2,\dots,N!\}$, $\ell_k := \{\tau :\tau = {\widetilde}\tau^k + {\mathbf 1}s \in P_N, s \in {\mathbb{R}}\}.$
Lyapunov Stability {#sec:lyapunov}
------------------
Lyapunov theory for hybrid systems is employed to show that the set of points ${\mathcal{A}}$ is asymptotically stable. Our candidate Lyapunov-like function, which is defined below and uses the distance function, is built by observing that there exist points where the distance to ${\mathcal{A}}$ may increase during flows. This is due to the sets $\ell_{k}$ being a subset $P_{N}$. To avoid this issue, we define $${\widetilde}{\mathcal{A}}= \bigcup_{k=1}^{N!} {\widetilde}\ell_k \supset {\mathcal{A}}$$ where ${\widetilde}\ell_k$ is the extension of $\ell_k$ given by $$\begin{aligned}
{\widetilde}\ell_k = \left\{\tau \in {\mathbb{R}}^{N} : \tau = {\widetilde}\tau^k + {\mathbf 1}s, s \in {\mathbb{R}}\right\}. \label{eqn:wtellk}\end{aligned}$$ Then, with this extended version of ${\mathcal{A}}$, the proposed candidate Lyapunov-like function for asymptotic stability of ${\mathcal{A}}$ for ${{{\mathcal}H}}_N$ is given by the locally Lipschitz function $$\label{eqn:lyapunov}
V(\tau) = \min\{|\tau|_{{\widetilde}\ell_1},|\tau|_{{\widetilde}\ell_2}, \ldots ,|\tau|_{{\widetilde}\ell_k}, \ldots ,|\tau|_{{\widetilde}\ell_N!}\} \quad \forall \ \tau \in P_N \setminus {\mathcal{X}}$$ where, for some $k$, $|\tau|_{{\widetilde}\ell_k}$ is the distance between the point $\tau$ and the set ${\widetilde}\ell_k$.[^6] The following theorem establishes asymptotic stability of ${\mathcal{A}}$ for ${{{\mathcal}H}}_N$. We show that the change in $V$ during flows is zero and that at jumps we have a strict decrease of $V$; namely, $V(G(\tau)) - V(\tau) = -|{\varepsilon}| V(\tau)$. A key step in the proof is in using [@teel2012hybrid Theorem 8.2] on a restricted version of ${{{\mathcal}H}}_N$.
\[thm:stability\] For every $N \in {\mathbb{N}}, N > 1$, ${\bar{\tau}}> 0,\omega > 0$, and $\varepsilon \in (-1,0)$, the hybrid system ${{{\mathcal}H}}_N$ is such that the compact set ${\mathcal{A}}$ is
Let the set ${\mathcal{X}}_{v}$ define the $v$-inflation of ${\mathcal{X}}$ (defined in Lemma \[lem:defX\]), that is, the open set[^7] ${\mathcal{X}}_{v} := \{\tau \in {\mathbb{R}}^N : |\tau|_{{\mathcal{X}}} < v\}$, where $v \in (0,v^{*})$ and $v^{*} = \min_{x\in{\mathcal{X}}, y\in{\widetilde}{\mathcal{A}}} |x-y|$. Given any $v \in (0,v^{*})$, we now consider a restricted hybrid system ${\widetilde}{{{\mathcal}H}}_{N} = (f,{\widetilde}{C},G,{\widetilde}{D})$, where ${\widetilde}{C} := C \setminus {\mathcal{X}}_{v}$ and ${\widetilde}{D} := D \setminus {\mathcal{X}}_{v}$, which are closed. We establish that ${\widetilde}{\mathcal{A}}$ is an asymptotically stable set for ${\widetilde}{{{\mathcal}H}}_N$.
Note that the continuous function $V$, given by , is defined as the minimum distance from $\tau$ to ${\widetilde}{\mathcal{A}}$, where ${\widetilde}{\mathcal{A}}$ is the union of $N!$ sets ${\widetilde}\ell_{k}$ in . To determine the change of $V$ during flows[^8], we consider the relationship between the flow map and the sets ${\widetilde}\ell_{k}$. The inner product between a vector pointing in the direction of the set ${\widetilde}\ell_{k}$ and the flow map on ${\widetilde}{C}$ satisfies $${\mathbf 1}^{\top}f(\tau) = {\mathbf 1}^{\top}(\omega{\mathbf 1}) = \omega N=|{\mathbf 1}||\omega {\mathbf 1}| = |{\mathbf 1}||f(\tau)|\cos \theta$$, which is only true if $\theta$ is zero. Therefore, the direction of the flow map and of the vector defining ${\widetilde}\ell_{k}$ are parallel, implying that the distance to the set ${\widetilde}{\mathcal{A}}$ is constant during flows.
The change in $V$ during jumps is given by $V(G(\tau)) - V(\tau)$ for $\tau \in {\widetilde}D \setminus {\widetilde}{\mathcal{A}}$. Due to the fact that we can rearrange the components of $\tau \in P_N \setminus {\mathcal{X}}$, without loss of generality, we consider a single jump condition, namely, we consider $\tau$ such that ${\bar{\tau}}= \tau_{1} > \tau_{2} > \ldots > \tau_{N-1} > \tau_{N}$. Using the formulation in Section \[sec:AN\] and Lemma \[eqn:tauSsolution\], the elements of the vector ${\widetilde}\tau^{k}$ associated with ${\widetilde}\ell_{k}$ for this case of $\tau$ are given by ${\widetilde}\tau_{i}^{k} = \frac{\sum_{p=0}^{N-i}({\varepsilon}+1)^p}{\sum_{p=0}^{N-1}({\varepsilon}+1)^p}{\bar{\tau}}$, which by Lemma \[lem:consum1\] is equal to $\frac{({\varepsilon}+1)^{N - i +1} - 1}{({\varepsilon}+1)^{N} - 1}{\bar{\tau}}$. After the jump, $G(\tau)$ is single valued and is such that its elements are ordered as follows: $g_{2}(\tau) > g_{3}(\tau) > \ldots > g_{N}(\tau) > g_{1}(\tau) = 0.$ Specifically, the jump map is $G(\tau) = [0,(1+{\varepsilon})\tau_{2},\ldots, (1+{\varepsilon})\tau_{N}]^{\top}$. Then, the formulation in Section \[sec:AN\] and [Lemma \[eqn:tauSsolution\]]{} leads to a case of ${\widetilde}\tau^{k}$ denoted as ${\widetilde}\tau^{k'}$. By [Lemma \[lem:consum1\]]{}, the elements of the vector ${\widetilde}\tau^{k'}$ are given by ${\widetilde}\tau^{k'}_{1} = \frac{{\varepsilon}}{({\varepsilon}+1)^{N} - 1}{\bar{\tau}}$ and ${\widetilde}\tau^{k'}_{i} = \frac{({\varepsilon}+1)^{N - i +2} - 1}{({\varepsilon}+1)^{N} - 1}{\bar{\tau}}$ for $i > 1$. Due to the ordering of $\tau$ and $G(\tau)$, ${\widetilde}\tau^{k'}$ is a one-element shifted (to the right) version of ${\widetilde}\tau^{k}$.
From the definition of ${\widetilde}\tau^{k}$ above, $V$ at $\tau$ reduces to $$V(\tau) = |\tau|_{{\widetilde}\ell_{k}} = \left|({\widetilde}\tau^{k} - \tau) - \frac{1}{N}(({\widetilde}\tau^{k} - \tau)^{\top}{\mathbf 1}){\mathbf 1}\right|$$ for some $k$. Note that $$({\widetilde}\tau^{k} - \tau)^{\top}{\mathbf 1}= \sum_{i = 1}^{N} {\widetilde}\tau^{k}_{i} - \sum_{i=1}^{N} \tau_{i}$$ reduces to $\sum_{i = 2}^{N} {\widetilde}\tau^{k}_{i} - \sum_{i=2}^{N} \tau_{i}$ since $\tau_{1} = {\widetilde}\tau_{1}^{k} = {\bar{\tau}}$. Using [Lemmas \[lem:consum1\] and \[lem:consum2\]]{}, it follows that $$\sum_{i = 2}^{N} {\widetilde}\tau^{k}_{i}= \frac{\sum_{i=2}^N \sum_{p=0}^{N-i}({\varepsilon}+1)^p}{\sum_{p=0}^{N-1}({\varepsilon}+1)^p}{\bar{\tau}}= \frac{(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}.$$ Then, the first element of the vector inside the norm in the expression of $V(\tau)$ is given as $$\begin{aligned}
&({\widetilde}\tau^{k}_{1} - \tau_{1}) - \frac{1}{N}\left(\frac{(({\varepsilon}+1)^{N} - 1 )- N{\varepsilon}}{{\varepsilon}(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}- \sum_{i=2}^{N} \tau_{i}\right)
\\ & \qquad\qquad\qquad\qquad= -\frac{(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}N(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}+ \frac{1}{N}\sum_{i=2}^{N} \tau_{i},\end{aligned}$$ while the elements with $m \in \{2,3,\ldots,N\}$ are given by
After the jump at $\tau$, since $G(\tau)$ is single valued, $V(G(\tau))$ is given by $$|G(\tau)|_{{\widetilde}\ell_{k'}} = \left|({\widetilde}\tau^{k'} - G(\tau)) - \frac{1}{N}(({\widetilde}\tau^{k'} - G(\tau))^{\top}{\mathbf 1}){\mathbf 1}\right|.$$ Note that $({\widetilde}\tau^{k'} - G(\tau))^{\top}{\mathbf 1}= \sum_{i = 1}^{N} {\widetilde}\tau^{k'}_{i} - \sum_{i = 1}^{N} g_{i}(\tau)$ reduces to $\sum_{i = 1}^{N} {\widetilde}\tau^{k'}_{i} - \sum_{i = 2}^{N} (1+ {\varepsilon})\tau_i$, since $g_{1}(\tau) = 0$ and $g_{i}(\tau) = (1+{\varepsilon})\tau_{i}$ for $i > 1$. Using [Lemmas \[lem:consum1\] and \[lem:consum2\]]{}, it follows that $$\begin{aligned}
\sum_{i = 1}^{N} {\widetilde}\tau^{k'}_{i} &= \frac{\sum_{i=1}^N \sum_{p=0}^{N-i}({\varepsilon}+1)^p}{\sum_{p=0}^{N-1}({\varepsilon}+1)^p}{\bar{\tau}}\\ &= \frac{({\varepsilon}+1)(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}\end{aligned}$$ which leads to $$({\widetilde}\tau^{k'} - G(\tau))^{\top}{\mathbf 1}= \frac{({\varepsilon}+1)(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}- \sum_{i = 2}^{N} (1+{\varepsilon})\tau_{i}.$$ The first element inside the norm in $V(G(\tau))$ is given by $$\begin{aligned}
&({\widetilde}\tau_{1}^{k'} - g_1(\tau)) - \frac{1}{N}\left(\frac{({\varepsilon}+1)(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}\right. \\ & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \left. - \sum_{i = 2}^{N} (1+{\varepsilon})\tau_{i}\right) \\
& = \frac{{\varepsilon}}{({\varepsilon}+1)^{N} - 1}{\bar{\tau}}- \frac{({\varepsilon}+1)(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}N(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}\\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad + \frac{1}{N}\sum_{i = 2}^{N} (1+{\varepsilon})\tau_{i} \\
&= (1+{\varepsilon})\left(-\frac{(({\varepsilon}+1)^{N} - 1) - N{\varepsilon}}{{\varepsilon}N(({\varepsilon}+1)^{N} - 1)}{\bar{\tau}}+ \frac{1}{N}\sum_{i = 2}^{N} \tau_{i}\right).\end{aligned}$$ For each element $m > 1$, it follows that
Combining the expressions for each of the elements inside the norm of $V(G(\tau))$, it follows that $V(G(\tau)) = (1+{\varepsilon}) V(\tau)$.
Then, the change during jumps is given by $V(G(\tau)) - V(\tau) = {\varepsilon}V(\tau)$ where ${\varepsilon}\in (-1,0)$. With the property of $V$ during flows established above, the change of $V$ along solutions is bounded during flows and jumps by the nonpositive functions $u_{{\widetilde}{C}}$ and $u_{{\widetilde}{D}}$, respectively, defined as follows: $u_{{\widetilde}{C}} (z)= 0$ for each $z \in {\widetilde}{C}$ and $u_{{\widetilde}{C}} (z)= -\infty$ otherwise; $u_{{\widetilde}{D}}(z) = {\varepsilon}V(z)$ for each $z \in {\widetilde}{D}$ and $u_{{\widetilde}{D}}(z) = -\infty$ otherwise. Using Lemma \[lem:BasicConds\], the fact that ${\widetilde}{C}$ and ${\widetilde}{D}$ are closed, and the fact that every maximal solution to ${\widetilde}{{{\mathcal}H}}$ is bounded and complete, by [@teel2012hybrid Theorem 8.2], every maximal solution to ${\widetilde}{{{\mathcal}H}}_N$ approaches the largest weakly invariant subset of $L_{V}(r') \cap {\widetilde}{C} \cap [L_{u_{{\widetilde}{C}}}(0) \cup (L_{u_{{\widetilde}{D}}}(0) \cap G(L_{u_{{\widetilde}{C}}}(0)))] = L_{V}(r') \cap {\widetilde}{C}$ for $r' \in V({\widetilde}{C})$. Since every maximal solution jumps an infinite number of times, the largest invariant set is given for $r' = 0$ due to the fact that $V(G(\tau)) - V(\tau) = {\varepsilon}V(\tau) < 0$ if $r' > 0$. Then, the largest invariant set is given by $L_{V}(0) \cap {\widetilde}{C} = {\widetilde}{\mathcal{A}}\cap {\widetilde}{C}$ which is identically equal to ${\mathcal{A}}$. Hence, the set ${\mathcal{A}}$ is attractive. Stability is guaranteed from the fact that $V$ is nonincreasing during flows and strictly decreasing during jumps. Then, the set ${\widetilde}{\mathcal{A}}$ is asymptotically stable for the hybrid system ${\widetilde}{{{\mathcal}H}}_{N}$. We have that ${\mathcal{A}}$ is (strongly) forward invariant and from Theorem 3.4 we know that ${\mathcal{A}}$ is uniformly attractive from a neighborhood of itself. Then by Proposition 7.5 in [@teel2012hybrid], it follows that ${\mathcal{A}}$ is asymptotically stable.
Note that the set of solutions to ${\widetilde}{{{\mathcal}H}}_{N}$ coincides with the set of solutions to ${{{\mathcal}H}}_{N}$ from $P_{N} \setminus {\mathcal{X}}_{v}$. Therefore, the set ${\mathcal{A}}$ is asymptotically stable for ${{{\mathcal}H}}_N$ with basin of attraction ${{\mathcal}B}_{{\mathcal{A}}} = P_{N} \setminus {\mathcal{X}}_{v}$. [[ Since $v$ is arbitrary, it follows that the basin of attraction is equal to $P_{N} \stopmodif \setminus {\mathcal{X}}$.]{}]{}
Note that the jump map $G$, at points $\tau \in {\mathcal{X}}$, is set valued by definition of $g_i$ in . From these points there exist solutions to ${{{\mathcal}H}}_N$ that jump out of ${\mathcal{X}}$. In fact, consider the case $\tau \in {\mathcal{X}}$. We have that $\tau_{i} = \tau_r$ for some $i,r \in I$. Then, after the jump it follows that $g_{i}(\tau) \in \{0, (1+{\varepsilon}){\bar{\tau}}\}$ and $g_{r}(\tau) \in \{0, (1+{\varepsilon}){\bar{\tau}}\},$ and there exist $g_{i}$ and $g_{r}$ such that $g_{i} = g_{r}$ or $g_{i} \neq g_{r}$. Since for every point in ${\mathcal{X}}$ there exists a solution that converges to ${\mathcal{A}}$ and also a solution that stays in ${\mathcal{X}}$, ${\mathcal{X}}$ is weakly forward invariant.[^9]
Characterization of Time of Convergence {#sec:timetoconverge}
---------------------------------------
In this section, we characterize the time to converge to a neighborhood of ${\mathcal{A}}$. The proposed (upper bound) of the time to converge depends on the initial distance to the set ${\widetilde}{\mathcal{A}}$ and the parameters of the hybrid system $(\varepsilon,{\bar{\tau}})$.
\[thm:timetoconverge\] For every $N \in {\mathbb{N}}$, $N > 1$, and every $c_1, c_2$ such that $\overline{c} > c_2 > c_1 > 0$ with $\overline{c} = \max_{x \in {\mathcal{X}}} |x|_{\widetilde{\mathcal{A}}}$, every maximal solution to ${{{\mathcal}H}}_N$ with initial condition $\tau(0,0) \in (P_{N} \setminus {\mathcal{X}}) \cap {\widetilde}L_{V}(c_2)$ is such that
Let $\tau_0 = \tau(0,0)$ and pick a maximal solution $\tau$ to ${{{\mathcal}H}}_{N}$ from $\tau_{0}$. At every jump time $(t_{j},j) \in {\mathop{\rm dom}\nolimits}\tau$, define $\bar{g}_1 = \tau(t_1,1)$, $\bar{g}_2 = \tau(t_2,2), \ldots, \bar{g}_J = \tau(t_J,J)$, for some $J \in {\mathbb{N}}$. From Theorem \[thm:stability\], we have that there is no change in the Lyapunov function during flows. Furthermore, we have that for each $\tau \in D \setminus {\mathcal{A}}$ the difference $V(G(\tau)) - V(\tau) = \varepsilon V(\tau)$ with $\varepsilon \in (-1,0)$. Since, for every $j,$ $\tau(t_{j},j) \in D$, we have which implies At the next jump, we have Proceeding in this way, after $J$ jumps we have From $V(\bar g_{J}) = (1+\varepsilon)^{J}V(\tau_{0})$, we want to find $J$ so that $V(\bar g_{J}) \leq c_{1}$ when $V(\tau_{0}) \leq c_{2}$. Considering the worst cast for $V(\tau_{0})$, we want $(1+\varepsilon)^{J}c_{2} \leq c_{1}$, which implies $\frac{c_{2}}{c_{1}} \leq \left(\frac{1}{1+\varepsilon} \right)^{J}$, and therefore $J = \left\lceil\frac{\log \frac{c_{2}}{c_{1}}}{ \log \frac{1}{1+\varepsilon}}\right\rceil > 0.$ For each $j$, the time between jumps satisfies $t_1 - t_0 \leq \frac{{\bar{\tau}}}{\omega},
t_2 - t_1 \leq \frac{{\bar{\tau}}}{\omega},
\ldots,
t_j - t_{j-1} \leq \frac{{\bar{\tau}}}{\omega}.$ Then, we have that after $J$ jumps, $
\sum_{j=1}^{J} t_{j} - t_{j-1} \leq J\frac{{\bar{\tau}}}{\omega}.
$ With $t_{0} = 0$, the expression reduces to $t_{J} \leq J\frac{{\bar{\tau}}}{\omega} = \left\lceil\frac{\log \frac{c_{2}}{c_{1}}}{ \log \frac{1}{1+\varepsilon}}\right\rceil\frac{{\bar{\tau}}}{\omega}.$ Then, after $t+j \geq t_{J} +J$, the solution is at least $c_{1}$ close to the set ${\widetilde}{\mathcal{A}}$. Defining $M = t_{J} + J$ we then have
Figure \[fig:timetoconverge\] shows the time to converge (divided by $\frac{{\bar{\tau}}}{\omega}+1$) versus $\varepsilon$ with constant $c_{2} = 0.99{\bar{\tau}}$ and varying values of $c_{1}$. As the figure indicates, the time to converge decreases as $|{\varepsilon}|$ increases, which confirms the intuition that the larger the jump the faster oscillators desynchronize.
Robustness Analysis {#sec:robustness}
-------------------
[[ Lemma \[lem:BasicConds\] establishes that the hybrid model of $N$ impulse-coupled oscillators satisfies the hybrid basic conditions. In light of this property, the asymptotic stability property of ${\mathcal{A}}$ for ${{{\mathcal}H}}_N$ is preserved under certain perturbations; i.e., asymptotic stability is robust [@teel2012hybrid]. In the next sections, we consider a perturbed version of ${{{\mathcal}H}}_N$ and present robust stability results. In particular, we consider generic perturbations to ${{{\mathcal}H}}_N$, and two different cases of perturbations only on the timer rates to allow for heterogeneous timers.]{}]{}
### Robustness to Generic Perturbations
We start by revisiting the definition of perturbed hybrid systems in [@teel2012hybrid].
.1cm
Given a hybrid system ${{{\mathcal}H}}$ and a function $\rho: {\mathbb{R}}^N \to {\mathbb{R}}_{\geq 0}$, the $\rho$-perturbation of ${{{\mathcal}H}}$, denoted ${{{\mathcal}H}}_\rho$, is the hybrid system $$\left\{\begin{array}{cc}x \in C_\rho & \quad \dot{x} \in F_\rho(x) \\
x \in D_\rho & \quad x^+ \in G_\rho(x) \\
\end{array} \right.$$ where
.1cm
Using this definition, we can deduce a generic perturbed hybrid system modeling $N$ impulse-coupled oscillators. Then, for the hybrid system ${{{\mathcal}H}}_N$, we denote ${{{\mathcal}H}}_{N,\rho}$ as the $\rho$-perturbation of ${{{\mathcal}H}}_N$. Given the perturbation function $\rho : {\mathbb{R}}^N \to {\mathbb{R}}_{\geq 0}$, the perturbed flow map is given by where the perturbed flow set $C_\rho$ is given by For example, if $N = 2$ and $\rho(\tau) = \bar\rho > 0$ for all $\tau \in {\mathbb{R}}^N$, which would correspond to constant perturbations on the lower value and threshold, then $C_\rho = C + \rho {\mathbb{B}}$. The perturbed jump map and jump set are defined as where $g_{i,\rho}$ is the $i$-th component of $G_\rho$. The following result establishes that the hybrid system ${{{\mathcal}H}}_N$ is robust to small perturbations.
(robustness of asymptotic stability)\[thm:robustofAS\] If $\rho : {\mathbb{R}}^N \to {\mathbb{R}}_{\geq 0}$ is continuous and positive on ${\mathbb{R}}^{N} \setminus {\mathcal{A}}$, then ${\mathcal{A}}$ is semiglobally practically robustly ${\mathcal}{KL}$ asymptotically stable with basin of attraction $B_{{\mathcal{A}}} = P_N \setminus {\mathcal{X}}$, i.e., for every compact set $K \subset B_{{\mathcal{A}}}$ and every $\alpha > 0$, there exists $\delta \in (0,1)$ such that every maximal solution $\tau$ to ${{{\mathcal}H}}_{N,\delta\rho}$ from $K$ satisfies $|\tau(t,j)|_{{\mathcal{A}}} \leq \beta(|\tau(0,0)|_{{\mathcal{A}}},t+j) + \alpha$ for all $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$.
From Lemma \[lem:BasicConds\], the hybrid system ${{{\mathcal}H}}_N$ satisfies the hybrid basic conditions. Therefore, by [@teel2012hybrid Theorem 6.8] ${{{\mathcal}H}}_N$ is nominally well-posed and, moreover, by [@teel2012hybrid Proposition 6.28] is well-posed. From the proof of Theorem \[thm:stability\], we know that the set ${\mathcal{A}}$ is an asymptotically stable compact set for the hybrid system ${{{\mathcal}H}}_N$ with basin of attraction $B_{{\mathcal{A}}}$. Since by Lemma \[lem:solutions2HS\], every maximal solution is complete, then [@teel2012hybrid Theorem 7.20] implies that ${\mathcal{A}}$ is semiglobally practically robustly ${\mathcal}{KL}$ asymptotically stable.
Section \[sec:jumpperturbs1\] showcases of ${{{\mathcal}H}}_{N}$ with $\rho$-perturbations on the jump map.
### Robustness to Heterogeneous Timer Rates
We consider the case when the continuous dynamic rates are perturbed in the form of for a given solution $\tau$. For example, consider the perturbation of the flow map given by $$\begin{aligned}
f(\tau) = \omega {\mathbf 1}+ \Delta\omega \label{eqn:fdelta}\end{aligned}$$ where $\Delta\omega \in {\mathbb{R}}^n$ is a constant defining a perturbation from the natural frequencies of the impulse-coupled oscillators. Then for some $k$, during flows, along a solution $\tau$ such that over $[t_{j},t_{j+1}]\times\{j\}$ satisfies $V(\tau(t,j)) = |\tau(t,j)|_{{\widetilde}\ell_{k}}$, it follows that $c$ reduces to $c(t,j) = \left(\frac{r_{\ell_{k}}^{\top}(\tau(t,j))(\frac{1}{N}{\bf \underline{1} - \bf I})}{|\tau(t,j)|_{\ell_{k}}}\right) \Delta\omega.$[^10] Furthermore, the norm of the hybrid arc $c$ can be bounded by a constant $\bar{c}$ given by $$\begin{aligned}
\bar{c} = \left|\left(\frac{1}{N}\underline{\bf 1} - {\bf I}\right)\Delta\omega\right| \label{eqn:ctbound}.\end{aligned}$$ Building from this example, the following result provides properties of the distance to ${\widetilde}{\mathcal{A}}$ from solutions $\tau$ to ${{{\mathcal}H}}_N$ under generic perturbations on $f$ (not necessarily as in ).
\[thm:vanishingC\] Suppose that the perturbation on the flow map of ${{{\mathcal}H}}_N$ is such that a perturbed solution $\tau$ satisfies, for each $j$ such that $\{t : (t,j) \in {\mathop{\rm dom}\nolimits}\tau\}$ has more than one point, $\frac{d}{dt} |\tau(t,j)|_{{\widetilde}{\mathcal{A}}} = c(t,j)$ for all $t \in \{t : (t,j) \in {\mathop{\rm dom}\nolimits}\tau\}$ [[ and $\tau(t,j) \in P_N \setminus {\mathcal{X}}$ for all $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$]{}]{}, for some hybrid arc $c$ with ${\mathop{\rm dom}\nolimits}c = {\mathop{\rm dom}\nolimits}\tau$. Then, the following hold:
- The asymptotic value of $|\tau(t,j)|_{{\widetilde}{\mathcal{A}}}$ satisfies $$\begin{aligned}
\lim_{t+j \to \infty}|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq \lim_{t+j \to \infty} \sum_{i = 0}^{j} (1+{\varepsilon})^{j-i} \int_{t_i}^{t_{i+1}} c(t,j) dt\end{aligned}$$
- If there exists $\bar{c} > 0$ such that $|c(t,j)| \leq \bar{c}$ for each $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$ then $$\begin{aligned}
\lim_{t+j \to \infty}|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq \frac{\bar{c}{\bar{\tau}}}{|{\varepsilon}|\omega}. \label{thmeqn:distupperbound}\end{aligned}$$
- If ${\widetilde}{j} : {\mathbb{R}}_{\geq 0} \to {\mathbb{N}}$ is a function that [[ chooses the appropriate minimum $j$ such that $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$]{}]{} for each time $t$ and $t \mapsto c(t,{\widetilde}{j}(t))$ is absolutely integrable, i.e., $\exists B$ such that then $$\begin{aligned}
\lim_{t + j \to \infty} |\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq \frac{B}{{\varepsilon}}. \label{eqn:limboundbeps}\end{aligned}$$
Consider a maximal solution $\tau$ to ${{{\mathcal}H}}_N$ with initial condition $\tau(0,0) \in P_N \setminus {\mathcal{X}}$. This proof uses the function $V$ from the proof of Theorem \[thm:stability\]. With $V$ equal to the distance from $\tau$ to the set ${\widetilde}{\mathcal{A}}$, then, for each $\tau \in D \setminus {\mathcal{X}}$, we have that $V(G(\tau)) - V(\tau) = {\varepsilon}V(\tau)$. Using the fact that $V(\tau) = |\tau|_{{\widetilde}{\mathcal{A}}}$ and the fact that, $G$ along the solution is single valued, it follows that $|\tau|_{{\widetilde}{\mathcal{A}}}$ after a jump can be equivalently written as By assumption, in between jumps, the distance to the set ${\widetilde}{\mathcal{A}}$ is such that $\frac{d}{dt}|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} = c(t,j)$, which implies that at $t_{j+1}$ the distance to the desynchronization set is given by $$|\tau(t_{j+1},j)|_{{\widetilde}{\mathcal{A}}} = \int_{t_{j}}^{t_{j+1}} c(s,j) ds + |\tau(t_{j},j)|_{{\widetilde}{\mathcal{A}}}.$$ It follows that Then, proceeding in this way, we obtain $$\begin{aligned}
|\tau(t_{j},j)|_{{\widetilde}{\mathcal{A}}} &= (1+{\varepsilon})^{j}|\tau(0,0)|_{{\widetilde}{\mathcal{A}}} \\ & \qquad \qquad \qquad +\sum_{i = 0}^{j-1}(1+{\varepsilon})^{j-i}\int_{t_{i}}^{t_{i+1}} c(s,i)ds. \end{aligned}$$ For the case of generic $t_{j+1} \geq t \geq t_j$, we have that $$\begin{aligned}
|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} = (1+{\varepsilon})^{j}|\tau(0,0)|_{{\widetilde}{\mathcal{A}}} + \sum_{i = 0}^{j}(1+{\varepsilon})^{j-i}\int_{t_{i}}^t c(s,i)ds.\end{aligned}$$ Since, we know that as either $t$ or $j$ goes to infinity, $j$ or $t$ go to infinity as well, respectively. The expression reduces to If $c(t,j) \leq \bar{c}$, it follows that
Lastly, since this hybrid system has the property that for any maximal solution $\tau$ with $(t,j) \in {\mathop{\rm dom}\nolimits}\tau$, if $t$ approaches $\infty$ then the parameter $j$ also approaches $\infty$, the expression given by $\lim_{t + j \to \infty} |\tau(t,j)|_{{\widetilde}{\mathcal{A}}}$ can be simplified. To do this, we know that the series $\sum_{i=0}^j (1+{\varepsilon})^{j -i} = \frac{(1+{\varepsilon})^{j+1}-1}{{\varepsilon}}$ approaches $\frac{1}{|{\varepsilon}|}$ as $j \to \infty$. Since $1+{\varepsilon}>0$ for ${\varepsilon}\in (-1,0)$, the series is absolutely convergent and its partial sum $s_j = \sum_{i=0}^j (1+{\varepsilon})^{j -i}$ is such that $\{s_j\}^\infty_{j=m}$ is a nondecreasing sequence (for each $m$). This implies that $s_j \leq 1/|{\varepsilon}|$ for all $j$ and for each $m$. Then, it follows that $(1+{\varepsilon})^{j-i} \leq \frac{1}{|{\varepsilon}|}$ for every $j,i \in {\mathbb{N}}$. Since the expression is a function of $j$ only and, for complete solutions, $t$ is such that as $t \to \infty$, then $j \to \infty$, we obtain
Numerical Analysis {#sec:numerics}
==================
This section presents numerical results obtained from simulating ${{{\mathcal}H}}_N$. First, we present results for the nominal case of ${{{\mathcal}H}}_N$ given by . Then, we present results for ${{{\mathcal}H}}_N$ under different types of perturbations. The Hybrid Equations (HyEQ) Toolbox in [@Sanfelice.ea.13.HSCC] was used to compute the trajectories.
Nominal Case
------------
The possible solutions to the hybrid system ${{{\mathcal}H}}_N$ fall into four categories: always desynchronized, asymptotically desynchronized, never desynchronized, and initially synchronized. The parameters used in these simulations are ${\bar{\tau}}= 1$ and $\varepsilon = -0.2$.
A solution of ${{{\mathcal}H}}_N$ that starts in $P_N \setminus ({\mathcal{X}}\cup {\mathcal{A}})$ asymptotically converges to ${\mathcal{A}}$, as Theorem \[thm:timetoconverge\] indicates. show solutions to both ${{{\mathcal}H}}_{2}$ and ${{{\mathcal}H}}_{3}$ converging to their respective desynchronization sets.
\
\[\]\[\]\[.7\][$\tau_1,\tau_{2}$]{} \[\]\[\]\[.7\][$t$ \[seconds\]]{} \[\]\[\]\[.8\]
For ${{{\mathcal}H}}_2$, if $\tau(0,0) = [0,0.1]^\top$, then the initial sublevel set is ${\widetilde}L_{V}(c_{2})$ with $c_{2} = 0.24$. Using Theorem \[thm:timetoconverge\], the time to converge to the sublevel set ${\widetilde}L_{V}(c_{1})$ with $c_{1} = 0.1$ leads to $M = 7.84$. Figure \[fig:H2notinX2\] shows a solution to the system for 10 seconds of flow time. From the figure, it can be seen that $V(\tau(t,j)) \approx 0.1$ at $(t,j) = (3,4)$. Then, the property guaranteed by Theorem \[thm:timetoconverge\], namely, $V(\tau(t,j)) \leq c_{1}$ for each $(t,j)$ such that $t + j \geq M$, is satisfied. Figure \[fig:H3notinX3\], shows a solution and the distance of this solution to ${\mathcal{A}}$. Notice that the initial sub level set is ${\widetilde}{L}_{V}(c_{2})$ with $c_{2} =0.32$. From Theorem \[thm:timetoconverge\] it follows that the time to converge to ${\widetilde}{L}_{V}(c_{1})$ with $c_{1} = 0.1$ is given by $M =10.14$, which is actually already satisfied at $(t,j) = (2.2,4)$. show solutions to ${{{\mathcal}H}}_N$ that asymptotically desynchronize for $N \in \{7,10\}$.
Perturbed Case
--------------
In this section, we present numerical results to validate the statements in Section \[sec:robustness\].
### Simulations of ${{{\mathcal}H}}_N$ with perturbed jumps {#sec:jumpperturbs1}
$\bullet$ [**Perturbation of the threshold in the jump set:**]{} We replace the jump set $D$ by $D_{\rho} := \{\tau : \exists i \in I \ s.t. \ \tau_i = {\bar{\tau}}+ \rho_i\}$ where $\rho_i \in [0, \bar{\rho}_i]$, $\bar{\rho}_i > 0$ for each $i \in I$. To avoid maximal solutions that are not complete, the flow set $C$ is replaced by $C_\rho := [0,{\bar{\tau}}+\rho_1]\times[0,{\bar{\tau}}+\rho_2] \times \ldots \times [0,{\bar{\tau}}+\rho_N]$. Furthermore, the components of the jump map are also replaced by $$g_{\rho_i} (\tau) = \left\{ \begin{array}{ll} 0 \qquad \qquad \ \ \ \ &\mbox{if } \tau_{i} = {\bar{\tau}}+\rho_i, \tau_r < {\bar{\tau}}+\rho_j \ \ \forall j \in I\setminus \{i\} \\
\{0 , \tau_{i}(1+\varepsilon) \} \ &\mbox{if } \tau_{i} = {\bar{\tau}}+\rho_i \ \exists j \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}+\rho_j \\
(1+\varepsilon)\tau_{i} \qquad \ &\mbox{if } \tau_i < {\bar{\tau}}+\rho_i \ \exists j \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}+\rho_j\end{array} \right. \label{eqn:gi_perturbed} .$$
$\bullet$ [**Perturbations on the reset component of the jump map:**]{} Under the effect of the perturbations considered in this case, instead of reseting $\tau_i$ to zero, the perturbed jump resets $\tau_i$ to a value $\rho_i \in {\mathbb{R}}_{\geq 0}$, for each $i \in I$. The perturbed hybrid system has the following data: and where, for each $i \in I$, the perturbed jump map is given by $$g_i (\tau) = \left\{ \begin{array}{l} \rho_i \qquad \qquad \ \ \ \ \mbox{if } \tau_{i} = {\bar{\tau}}, \tau_r < {\bar{\tau}}\ \ \forall j \in I\setminus \{i\} \\
\{\rho_i , \tau_{i}(1+\varepsilon) \} \ \mbox{if } \tau_{i} = {\bar{\tau}}\ \exists j \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}\\
(1+\varepsilon)\tau_{i} \qquad \ \mbox{if } \tau_i < {\bar{\tau}}\ \exists j \in I \setminus \{i\} \ \mbox{s.t.} \ \tau_r = {\bar{\tau}}\end{array} \right. .\label{eqn:gi_resetperturbation}$$
This case of perturbations exemplifies Theorem \[thm:robustofAS\] with $\rho$ affecting only the jump map of ${{{\mathcal}H}}_N$. Figures \[figs:ResetPerturb\_di\_equal\_di=0.2\] and \[figs:ResetPertub\_di\_notequal\_di\] show several simulations to this perturbation of ${{{\mathcal}H}}_N$. All of the simulations in this section use parameters $\omega = 1$, ${\bar{\tau}}= 3$, ${\varepsilon}= -0.3$, and $N = 2$.
The first case of the perturbed jump map $G_\rho$ considered is for $\rho_1 = \rho_2 = 0.02$. Figure \[ResetPerturb\_di\_equal\_t1t2plot\_di=0.2\] shows a solution to the perturbed ${{{\mathcal}H}}_2$ from the initial condition $\tau(0,0) = [2.4,2.3]^\top$ on the $(\tau_1,\tau_2)$-plane. Notice that for $\tau \in D$ such that $\tau_i = {\bar{\tau}}$ the jump map resets $\tau_i$ to $\rho_i$ (red dashed line) and not to $0$ as in the unperturbed case. The solution for this case approaches a region around ${\widetilde}{\mathcal{A}}$, as Theorem \[thm:robustofAS\] guarantees. Figure \[ResetPerturb\_di\_equal\_timeplot\_di=0.2\] shows the distance to the set ${\widetilde}{\mathcal{A}}$ over time for 10 solutions of the perturbed system ${{{\mathcal}H}}_2$ with initial conditions $\tau(0,0) \in P_2\setminus {\mathcal{X}}_{2}$. This figure shows that solutions approach a distance of about $0.12$ after 25 seconds.
Now, consider the case where $\rho_1 \neq \rho_2$. Figure \[figs:ResetPertub\_di\_notequal\_di\] shows the distance to ${\widetilde}{\mathcal{A}}$ for two sets of solutions with different values for $\rho_1$ and $\rho_2$. More specifically, Figure \[ResetPertub\_di\_NotEqual\_timeplot\_di=\[0.15,0.25\]\] shows the case of $\rho_1 = 0.15$ and $\rho_2 = 0.25$. For this case, it can be seen that the solutions converge after $\approx 28$ seconds of flow time and, after that time, satisfy $|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq 0.25$. Figure \[ResetPertub\_di\_NotEqual\_timeplot\_di=\[0.01,0.02\]\] shows the case of $\rho_1 = 0.02$ and $\rho_2 = 0.01$. For this case, this figure shows that, after $\approx 28$ seconds of flow time, the solutions satisfy $|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq 0.04$. These simulations validate Theorem \[thm:robustofAS\] with $\rho$ affecting only the jump map, verifying that the smaller the size of the perturbation the smaller the steady-state value of the distance to ${\widetilde}{\mathcal{A}}$.
the component $(1+{\varepsilon})\tau_i$ of the jump map is perturbed, namely, we use $\tau_i^+ = (1+{\varepsilon}) \tau_i + \rho_i(\tau_i)$, where $\rho_i : {\mathbb{R}}_{\geq 0} \to P_N \setminus {\mathcal{X}}$ is a continuous function. The perturbed jump map $G_{\rho}$ has components $g_{\rho i}$ that are given as $g_{i}$ in but with $\tau_{i}(1+{\varepsilon}) + \rho_{i}(\tau_{i})$ replacing $\tau_{i}(1+{\varepsilon})$.
Consider the case $\rho_i(\tau_{i}) = {\widetilde}\rho_i \tau_i$ with ${\widetilde}\rho_i \in (0,|{\varepsilon}|)$ and let ${\widetilde}{\varepsilon}_i = {\varepsilon}+ {\widetilde}\rho_i \in (-1,0)$. Then $\tau_i^+$ reduces to $\tau_i^+ = (1+{\widetilde}{\varepsilon}_i)\tau_i$ and the jump map $g_{\rho i}$ is given by with ${\widetilde}{\varepsilon}_{i}$ in place of ${\varepsilon}$. This type of perturbation is used to verify Theorem \[thm:robustofAS\] with $\rho$ affecting only the “bump” portion of the jump map. and \[figs:ResetBumpPerturb\_di\_notequal\_di\] show simulations to ${{{\mathcal}H}}_N$ with the parameters $\omega = 1$, ${\bar{\tau}}= 3$, ${\varepsilon}= -0.3$, and $N = 2$.
Consider the case of ${{{\mathcal}H}}_2$ with $G_\rho$ when ${\widetilde}\rho_1 = {\widetilde}\rho_2 = 0.1$, leading to ${\widetilde}{\varepsilon}_1 = {\widetilde}{\varepsilon}_2$ = 0.2. shows a solution on the $(\tau_1,\tau_2)$-plane for this case with initial condition $\tau(0,0) = [0.1,0.2]^\top$. Notice that the solution approaches a region around ${\mathcal{A}}$ (green line), as Theorem \[thm:robustofAS\] guarantees. Figure \[ResetBumpPertub\_di\_equal\_timeplot\_di=\[0.1,0.1\]\] shows the distance to the set ${\widetilde}{\mathcal{A}}$ over time for 10 solutions with initial conditions $\tau(0,0) \in C$. It shows that solutions approach a distance to ${\widetilde}{\mathcal{A}}$ of $\approx0.09$ after $\approx40$ seconds of flow time.
Next, we consider the case of $G_{\rho}$ with ${\widetilde}{\varepsilon}_1 \neq {\widetilde}{\varepsilon}_2$. Figure \[ResetBumpPertub\_di\_NotEqual\_timeplot\_di=\[0.15,0.1\]\] shows the distance to ${\widetilde}{\mathcal{A}}$ for 10 solutions with perturbations given by ${\widetilde}\rho_1 = 0.15$ and ${\widetilde}\rho_2 = 0.1$. For this case, the distance to ${\widetilde}{\mathcal{A}}$ satisfies $|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq 0.3$ after $\approx40$ seconds of flow time. Figure \[fig:ResetBumpPertub\_di\_NotEqual\_timeplot\_di=\[0.02,0.01\]\] shows simulation results with ${\widetilde}\rho_1 = 0.02$ and ${\widetilde}\rho_2 = 0.01$. Notice that the smaller the value of the perturbation is, the closer the solutions get to the set ${\widetilde}{\mathcal{A}}$. For this case, after $\approx30$ seconds of flow time, the distance to ${\widetilde}{\mathcal{A}}$ satisfies $|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq 0.06$. These simulations validate Theorem \[thm:robustofAS\] with $\rho$ affecting only the jump map, verifying that the smaller the size of the perturbation the smaller the steady-state value of the distance to ${\widetilde}{\mathcal{A}}$ would be.
### Perturbations on the Flow Map {#sec:flowpertrubs}
In this section, we consider a class of perturbations on the flow map. More precisely, consider the case when there exists a function $(t,j) \mapsto c(t,j)$ such that $c(t,j) \leq \bar{c}$ with $\bar{c}$ as in . Then, from Theorem \[thm:vanishingC\] with , we know that $$\begin{aligned}
\lim_{t+j \to \infty} |\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq \left|\frac{\bar{c}{\bar{\tau}}}{{\varepsilon}\omega}\right| \leq \left|\frac{\left|(\frac{1}{N}\underline{\bf 1} - {\bf I})\Delta\omega\right|{\bar{\tau}}}{{\varepsilon}\omega}\right|. \label{eqn:flowpert2}\end{aligned}$$
Figure shows a simulation so as to verify this property. The parameters of this simulation are $N = 2$, $\omega = 1$, ${\varepsilon}= -0.3$, ${\bar{\tau}}= 4$, and $\Delta\omega = [0.120,0.134]^\top$. It follows from that $\overline{c} = 0.0105$. Then, from , it follows that $\lim_{t+j \to \infty}|\tau(t,j)|_{{\widetilde}{\mathcal{A}}} \leq 0.1047$. Specifically, Figure \[fig:c=constant\_tau1tau2plot\] shows a solution on the $(\tau_1,\tau_2)$-plane of the perturbed hybrid system ${{{\mathcal}H}}_2$ with initial condition $\tau(0,0) = [0,0.01]^\top$. This figure shows the solution (blue line) converging to a region around ${\widetilde}{\mathcal{A}}$ (between dash-dotted lines about ${\mathcal{A}}$ in green). Figure \[fig:c=constant\_disttimetraj\] shows the distance to the set ${\widetilde}{\mathcal{A}}$ of 10 solutions with initial conditions $\tau(0,0) \in C$ with a dashed line denoting the upper bound on the distance in . Notice that all solutions are within this bound after approximately 15 seconds of flow time and stay within this region afterwards.
Conclusion {#sec:conclusion}
==========
We have shown that desynchronization in a class of impulse-coupled oscillators is an asymptotically stable and robust property. These properties are established within a solid framework for modeling and analysis of hybrid systems, which is amenable for the study of synchronization and desynchronization in other impulse-coupled oscillators in the literature. The main difficulty in applying these tools lies on the construction of a Lyapunov-like quantity certifying asymptotic stability. As we show here, invariance principles can be exploited to relax the conditions that those functions have to satisfy, so as to characterize convergence, stability, and robustness in the class of systems under study. Future directions of research include the study of nonlinear reset maps, such as those capturing the phase-response curve of spiking neurons, as well as impulse-coupled oscillators connected via general graphs.
Appendix {#sec:appendix}
========
The following result derives the solution to $\Gamma\tau_{s} = b$ with $\Gamma$ given in and $b = {\bar{\tau}}{\mathbf 1}$ via Gaussian elimination.
\[eqn:tauSsolution\] For each ${\varepsilon}\in (-1,0)$, the solution $\tau_s$ to $\Gamma\tau_{s} = b$ with $\Gamma$ given in and $b = {\bar{\tau}}{\mathbf 1}$ is such that its elements, denoted as $\tau_s^k$ for each $k \in \{1,2,\ldots,N\}$, are given by $\tau_{s}^{k} = \frac{\sum_{i=0}^{N-k}({\varepsilon}+1)^i}{\sum_{i=0}^{N-1}({\varepsilon}+1)^i}{\bar{\tau}}$.
The $N \times N$ matrix in and the $N \times 1$ matrix $b = {\bar{\tau}}{\mathbf 1}$ leads to the augmented matrix $[\Gamma|b]$ given by To solve for $\tau_{s}^{k}$, we apply the Gauss-Jordan elimination technique to to remove the elements $-({\varepsilon}+1)$ above the diagonal. Starting from the $N$-th row to remove the $-({\varepsilon}+1)$ component in the $N-1$ row, and continuing up to the second row, gives Denoting the augmented matrix in as $[\Gamma'|b']$, with $\tau_{s}^{1} = {\bar{\tau}}$ and $\tau_{s}^{2} = \frac{\sum_{i=0}^{N-2}({\varepsilon}+1)^i}{\sum_{i=0}^{N-1}({\varepsilon}+1)^i}{\bar{\tau}}$, the solution for each element of $\tau_{s}^{k}$ with $k > 2$ can be derived from as $\Gamma'_{k,2}\tau^{2}_{s} + \tau^{k}_{s} = b'_{k}$ where $\Gamma'_{k,2}$ denotes the $(k,2)$ entry of $\Gamma'$. Noting that $\tau_{s}^{1}$ can be rewritten as $\tau_{s}^{1} = \frac{\sum_{i=0}^{N-1}({\varepsilon}+1)^i}{\sum_{i=0}^{N-1}({\varepsilon}+1)^i}{\bar{\tau}}$ leads to $\tau_{s}^{k} = \frac{\sum_{i=0}^{N-k}({\varepsilon}+1)^i}{\sum_{i=0}^{N-1}({\varepsilon}+1)^i}{\bar{\tau}}$.
\[lem:consum1\] For each $x \neq 1$, and $m,n \in {\mathbb{N}}$ such that $n-1 \geq m$, the finite sum $\sum_{i=m}^{n-1} x^i$ satisfies $
\sum_{i=m}^{n-1} x^i = \frac{x^n - x^m}{x-1}.
$
\[lem:consum2\] For each $x \neq 1$, [[ and each $m,N \in {\mathbb{N}}$ such that $N \geq m$]{}]{}, the finite sum $\sum_{n=m}^N\sum_{i=0}^{N-n}x^i$ satisfies
[^1]: Department of Computer Engineering, University of California, Santa Cruz, CA 95064. Email: [seaphill,[email protected]]{}. This research has been partially supported by the National Science Foundation under CAREER Grant no. ECS-1150306 and by the Air Force Office of Scientific Research under Grant no. FA9550-12-1-0366.
[^2]: Cf. the model for synchronization in [@Mirollo.90.SIAMJAM.BiologicalOscillators] where $\varepsilon > 0$.
[^3]: In [@mauroy:037122], a more general flow map and a jump map incrementing $\tau_i$ by ${\varepsilon}> 0$ are considered.
[^4]: A set-valued mapping $G : {\mathbb{R}}^N{\rightrightarrows}{\mathbb{R}}^N$ is [*outer semicontinuous*]{} if its graph $\{(x,y): x \in {\mathbb{R}}^N, y \in G(x)\}$ is closed, see [@teel2012hybrid Lemma 5.10] and [@RockafellarWets98].
[^5]: Note that $G$ is single valued at each ${\widetilde}\tau^k \notin {\mathcal{X}}$.
[^6]: The set ${\widetilde}\ell_k$ can be described as a straight line in ${\mathbb{R}}^n$ passing through a point ${\widetilde}\tau^k$ and with slope ${\mathbf 1}$. Then, $|\tau|_{{\widetilde}\ell_k}$ can be written as the general point-to-line distance $|({\widetilde}\tau^k - \tau) - 1/N (({\widetilde}\tau^k - \tau)^\top{\mathbf 1}){\mathbf 1}|$.
[^7]: The set ${\mathcal{X}}_{v}$ is open since every point $\tau \in {\mathcal{X}}_{v}$ is an interior point of ${\mathcal{X}}_{v}$.
[^8]: Its derivative can be computed using Clarke’s generalized gradient [@Clarke90].
[^9]: For example, consider the case $N = 2$. If $\tau(0,0) = [{\bar{\tau}},{\bar{\tau}}]^\top \in D$, then there are nonunique solutions due to the jump map begin set valued. It follows that after the jump, each $\tau_i$ can be mapped to any point in $\{0,\tau_i(1+{\varepsilon})\}$, which leads to any of the following four options of the states $(\tau_1,\tau_2)$ after such a jump: $(0,0),(0,{\bar{\tau}}(1+{\varepsilon})),({\bar{\tau}}(1+{\varepsilon}),0)$ or $({\bar{\tau}}(1+{\varepsilon}),{\bar{\tau}}(1+{\varepsilon}))$. If the state is mapped to either $(0,0)$ or $({\bar{\tau}}(1+{\varepsilon}),{\bar{\tau}}(1+{\varepsilon}))$, then it remains in ${\mathcal{X}}_2$. Conversely, if any of the other options are chosen, then $({\tau_1},{\tau_2})$ leaves ${\mathcal{X}}_2$ and converges to ${\mathcal{A}}$ asymptotically.
[^10]: Let $r_{\ell_{k}}(\tau)$ be the vector defined by the minimum distance from $\tau$ to the line $\ell_{k}$. Then, it follows that $V(\tau) = (r_{\ell_{k}}^{\top}(\tau)r_{\ell_{k}}(\tau))^{\frac{1}{2}}$. To determine its change during flows, note that on $C\setminus ({\mathcal{X}}\cup {\mathcal{A}})$ the gradient is given by $
\nabla V(\tau) = \frac{\partial}{\partial \tau} \left( r_{\ell_{k}}^{\top}(\tau)r_{\ell_{k}}(\tau) \right)^{\frac{1}{2}}
= \frac{\left(r_{\ell_{k}}^{\top}(\tau) \frac{\partial}{\partial \tau} r_{\ell_{k}}(\tau) \right)}{|\tau|_{\ell_{k}}}
$ where each $j$-th entry of $\frac{\partial}{\partial \tau} r_{\ell_{k}}(\tau)$ is given by $
\frac{\partial}{\partial \tau} r^{j}_{\ell_{k}}(\tau) = \frac{\partial}{\partial \tau}\left(({\widetilde}\tau_j\null\hspace{-.08cm}^{k} - \tau_{j}) - \frac{1}{N}\sum_{i=1}^N({\widetilde}\tau_i\null\hspace{-.08cm}^k - \tau_i)^\top \right)
= \left[\frac{1}{N}, \frac{1}{N},\ldots, \frac{1}{N}, -1 + \frac{1}{N}, \frac{1}{N}, \ldots, \frac{1}{N} \right]
$ – the term $-1 + \frac{1}{N}$ corresponds to the $j$-th element of the vector. It follows that $\frac{\partial}{\partial \tau} r_{\ell_{k}}(\tau) = \frac{1}{N}{\bf \underline{1}} - {\bf I}$. Then, for each $\tau \in C \setminus {\mathcal{X}}$, $
\langle \nabla V(\tau), f(\tau) \rangle = \left(\frac{r_{\ell_{k}}^{\top}(\tau)(\frac{1}{N}{\bf \underline{1} - \bf I})}{|\tau|_{\ell_{k}}}\right)f(\tau)
$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present [[WHFast]{}]{}, a fast and accurate implementation of a Wisdom-Holman symplectic integrator for long-term orbit integrations of planetary systems. [[WHFast]{}]{}is significantly faster and conserves energy better than all other Wisdom-Holman integrators tested.
We achieve this by significantly improving the Kepler-solver and ensuring numerical stability of coordinate transformations to and from Jacobi coordinates. These refinements allow us to remove the linear secular trend in the energy error that is present in other implementations. For small enough timesteps we achieve Brouwer’s law, i.e. the energy error is dominated by an unbiased random walk due to floating-point round-off errors.
We implement symplectic correctors up to order eleven that significantly reduce the energy error. We also implement a symplectic tangent map for the variational equations. This allows us to efficiently calculate two widely used chaos indicators the Lyapunov characteristic number (LCN) and the Mean Exponential Growth factor of Nearby Orbits (MEGNO).
[[WHFast]{}]{}is freely available as a flexible C package, as a shared library, and as an easy-to-use python module.
author:
- |
Hanno Rein$^{1,2}$ and Daniel Tamayo$^{1,3,4}$\
$^1$ Department of Physical and Environmental Sciences, University of Toronto at Scarborough, Toronto, Ontario M1C 1A4, Canada\
$^2$ Department of Astronomy and Astrophysics, University of Toronto, Toronto, Ontario, M5S 3H4, Canada\
$^3$ Canadian Institute for Theoretical Astrophysics, 60 St. George St, University of Toronto, Toronto, Ontario M5S 3H8, Canada\
$^4$ Centre for Planetary Sciences Fellow
bibliography:
- 'full.bib'
date: 'Submitted: 8th May 2015. Accepted: 2nd June 2015.'
title: '[[WHFast]{}]{}: A fast and unbiased implementation of a symplectic Wisdom-Holman integrator for long term gravitational simulations'
---
methods: numerical — gravitation — planets and satellites: dynamical evolution and stability
Introduction {#sec:intro}
============
Celestial mechanics, the field that deals with the motion of celestial objects, has been an active field of research since the days of Newton and Kepler. Analytic solutions only exist for a few special cases. Historically, the main driver for the development of perturbation theory has been the problem of planets orbiting the Sun. Because the central body is so much more massive than the planets, it is profitable to ask how the small mutual tugs between the planets modify the Keplerian orbits they would each individually follow around the Sun in the absence of the other bodies. This analytical approach has been, and continues to be, successful in explaining many important features of planetary orbits. However, the Solar System is chaotic, and the rise of computing power has yielded many important insights. There is therefore considerable interest in developing fast and accurate numerical integrators.
A large number of such integrators have been developed over the years to perform this task. For many long term integrations, symplectic integrators have proven to be a favourable choice. Symplectic schemes incorporate the symmetries of Hamiltonian systems, and therefore typically conserve quantities like the energy and angular momentum better than non-symplectic integrators.
For integrations of planetary systems, [@WisdomHolman1991], and independently [@Kinoshita1991], developed a widely used class of symplectic integrators. The ideas of [@WisdomHolman1991] developed from the original ideas of the mapping method of [@Wisdom1981]. We refer to these as a Wisdom-Holman mapping or a Wisdom-Holman integrator. Since then, many authors have modified and built upon this method, and several have made their integrators publicly available to the astrophysics community [e.g. @Chambers1997; @Duncan1998].
The Wisdom-Holman integrator exploits the intuition from perturbation theory that one can separate the problem into a system of Keplerian orbits about the Sun, modified by small perturbations among the planets. The nuisance is that while Newton provided us the solution to the two-body problem, Poincar[é]{} showed that the remaining superimposed perturbations are not integrable. Analytically, the traditional way forward is to average over the short-period oscillations in the problem to yield approximate solutions. The great insight of Wisdom and Holman was that, at the same level of approximation, one can [*add*]{} high frequency terms. By judicious choice of these additional frequencies, the perturbations among the planets can be transformed into trivially integrated delta functions. The result is an exceedingly efficient integrator that has proven an indispensable tool for modern studies in celestial mechanics.
In this paper, we present results from a complete reimplementation of the Wisdom-Holman integrator. We show how to speed up the algorithm in several ways and dramatically increase its accuracy. Many of the improvements are related to finite double floating-point precision on modern computers [IEEE754, @IEEE754]. The fact that almost all real numbers cannot be represented exactly in floating-point precision leads to important consequences for the numerical stability of any algorithm and the growth of numerical round-off error.
To our knowledge, we present the first publicly available Wisdom-Holman integrator that is unbiased, i.e. the errors are random and uncorrelated. This leads to a very slow error growth. For sufficiently small timesteps, we achieve Brouwer’s law, i.e., the energy error grows as time to the power of one half.
We have also sped up the integrator through various improvements to the integrator’s Kepler-solver. Our implementation allows for the evolution of variational equations (to determine whether orbits are chaotic) at almost no additional cost. Additionally, we implement so-called symplectic correctors up to order eleven to increase the accurary [@Wisdom1996], allow for arbitrary unit choices, and do not tie the integration to a particular frame of reference.
We make our integrator, which we call [[WHFast]{}]{}, publicly available in its native C99 implementation and as an easy-to-use python module.
The remainder of this paper is structured as follows. We first summarize the concepts and algorithms used in this paper, including Jacobi Coordinates, our choice of Hamiltonian splitting, the symplectic Wisdom-Holman map, symplectic correctors and the variational equations in Sect. \[sec:background\]. We then go into detail discussing the improvements we have made to these algorithms in Sect. \[sec:improvements\]. Numerical tests are presented in Sect. \[sec:numericalresults\] before we conclude in Sect. \[sec:conclusions\].
Background {#sec:background}
==========
The Hamiltonian $\mathcal{H}$ of the gravitational $N$-body system can be written as the sum of kinetic and potential terms in Cartesian coordinates $$\begin{aligned}
\mathcal{H} &=& \sum_{i=0}^{N-1} \frac{\mathbf{p}_i^2}{2m_i} - \sum_{i=0}^{N-1} \sum_{j=i+1}^{N-1} \frac{Gm_im_j}{|\mathbf{r}_i-\mathbf{r}_j|}.\label{eq:H}\end{aligned}$$ One way forward toward separating out the two-body Keplerian Hamiltonians is to transform to heliocentric coordinates involving the centre-of-mass and the $\mathbf{r}_i-\mathbf{r}_0$. However, rewriting the Cartesian momenta in terms of heliocentric momenta (which have an additional component along the centre-of-mass momentum), leads to several cross-terms. Alternatively, Jacobi worked out a coordinate system in which the kinetic terms are particularly clean, and the kinetic energy remains a sum of squares. For readers that may not be familiar, and because our improved accuracy is largely due to modifications of the manner in which we transform between Cartesian and Jacobi coordinates, we briefly review them [see also @Plummer1918; @SussmanWisdom2001; @solarsystemdynamics].
Jacobi Coordinates
------------------
Rather than reference planet positions to the central star, a planet’s Jacobi coordinates are measured relative to the centre-of-mass of all bodies with lower indices. For concreteness, consider a system of $N$ particles with masses $m_i$, $i=0,\ldots, N-1$. Let $\mathbf{r}_i$ be the position vector of the $i$-th particle with respect to an arbitrary origin that is fixed in an inertial frame. Here we assume that the particles are ordered such that $i=0$ corresponds to the central object, $i=1$ to the innermost object orbiting the central object and so on. The existence of such an ordering does not restrict the architecture of the system. For example, the coordinates of an equal-mass binary with a circumbinary particle can be expressed in Jacobi coordinates. But note that the ordering might in general be non-unique and that it can change during an integration. This can have important implications for a numerical scheme using Jacobi coordinates.
The Jacobi coordinate $\mathbf{r}'_i$ of the $i$-th particles is the position relative to $\mathbf{R}_{i-1}$, the centre-of-mass of all the particles interior to the $i$-th particle: $$\begin{aligned}
\mathbf{r}'_i &=& \mathbf{r}_i - \mathbf{R}_{i-1},\quad\quad\quad\quad\text{for } \;i=1,\ldots,N-1\\
\text{where} \quad \mathbf{R}_{i} &=& \frac{1}{M_i}\sum_{j=0}^i m_j \mathbf{r}_j \quad \text{and } \quad M_{i} \;=\; \sum_{j=0}^i m_j.\end{aligned}$$ Other quantities such as the velocity and acceleration (also the coordinates in the variational equations, see below) transform in the same way. This is because the Jacobi coordinates are a linear function of the Cartesian coordinates, and the velocity is the time derivative of the position in both coordinates systems.
The momenta, however, transform differently[^1]. The momentum conjugate to $\mathbf{r}'_i$ and the corresponding Jacobi mass are given by $$\begin{aligned}
\mathbf{p}'_i = m'_i \dot{\mathbf{r}'}_i = m'_i \mathbf{v'}_i \quad\quad\text{ and } \quad\quad m'_i = m_i \frac{M_{i-1}}{M_{i}}= \frac{m_i\,M_{i-1}}{m_i + M_{i-1}}.\label{eq:p}\end{aligned}$$ Note that the Jacobi mass $m'_i$ is the reduced mass of $m_i$ and $M_{i-1}$. Explicit expressions for the momenta can be found by evaluating the time derivative r[Eq. \[eq:p\].]{}
The Jacobi coordinates above are relative coordinates for $i=1,\ldots, N-1$. For the $0$-th coordinate, a different convention is used, $$\begin{aligned}
\mathbf{r}'_0 = \mathbf{R}_{N-1},\quad\quad\quad m'_0 = M_{N-1},\quad\quad\quad \mathbf{p}'_0 = \sum_{j=0}^{N-1} \mathbf{p}_j.\end{aligned}$$ Thus, $\mathbf{r}'_0$ points towards the centre-of-mass of the entire system, $\mathbf{p}'_0$ is the total momentum and $m'_0$ is the total mass.
Hamiltonian Splitting {#sec:splitting}
---------------------
After some algebra, we can rewrite the Hamiltonian in Eq. \[eq:H\] in terms of the conjugate momenta of the Jacobi coordinates [e.g. @solarsystemdynamics; @SussmanWisdom2001]. We only rewrite the kinetic term and keep the potential term expressed as a function of the Cartesian coordinates: $$\begin{aligned}
\mathcal{H} &=& \sum_{i=0}^{N-1} \frac{{\mathbf{p}'_i}^2}{2m'_i} - \sum_{i=0}^{N-1} \sum_{j=i+1}^{N-1} \frac{Gm_im_j}{|\mathbf{r}_i-\mathbf{r}_j|}.\end{aligned}$$ Note that the kinetic term is still diagonal, i.e. there are no cross terms involving $\mathbf{p}_i\mathbf{p}_j$ with $i\neq j$. Next, we add and subtract the term $$\begin{aligned}
\operatorname{\mathcal{H}}_\pm = \sum_{i=1}^{N-1}\frac{G m'_i M_{i} }{|\mathbf{r}'_i|}. \label{eq:hpm}\end{aligned}$$ After grouping terms in the Hamiltonian, we arrive at $$\begin{aligned}
\mathcal{H} &=&
\underbrace{\frac{{\mathbf{p}'_0}^2}{2m'_0}\vphantom{\sum_{i=1}^{N-1} \frac{{\mathbf{p}'_i}^2}{2m'_i} }}_{\mathcal{H}_{0}}
+\underbrace{\sum_{i=1}^{N-1} \frac{{\mathbf{p}'_i}^2}{2m'_i} -\sum_{i=1}^{N-1}\frac{G m'_i M_{i} }{|\mathbf{r}'_i|}}_{\mathcal{H}_{\rm Kepler}} \nonumber \\
&& + \underbrace{\sum_{i=1}^{N-1}\frac{G m'_i M_{i} }{|\mathbf{r}'_i|}- \sum_{i=0}^{N-1} \sum_{j=i+1}^{N-1} \frac{Gm_im_j}{|\mathbf{r}_i-\mathbf{r}_j|}}_{\mathcal{H}_{\rm Interaction}}.\label{eq:hamsplit}\end{aligned}$$ The first term, $\mathcal{H}_0$, simply describes the motion of the centre-of-mass $\mathbf{r}'_0$ along a straight line. For that reason this term is often ignored. However, we keep it which will allow us to integrate particles without any restriction to a particular frame of reference.
The terms $\mathcal{H}_{\rm Kepler}$ can be split up further into a sum of $$\begin{aligned}
\left(\mathcal{H}_{\rm Kepler}\right)_i &=& \frac{{\mathbf{p}'_i}^2}{2m'_i} - \frac{G m'_i M_{i} }{|\mathbf{r}'_i|}. \label{eq:HKepler}\end{aligned}$$ Each of the Hamiltonians $\left(\mathcal{H}_{\rm Kepler}\right)_i$ describes the Keplerian motion of the $i$-th particle with mass $m_i$ around the centre-of-mass of all interior particles with total mass $M_{i-1}$.
After some more algebra, the interaction term can be simplified and split into two parts, one of which can be easily computed in Jacobi coordinates and the other in Cartesian coordinates of the inertial frame $$\begin{aligned}
\operatorname{\mathcal{H}}_{\rm Interaction} =
\sum_{i=2}^{N-1} \frac{ G{m'}_i M_i }{|\mathbf{r}'_i|}
- \sum_{i=0}^{N-1} \sum_{\substack{j=i+1\\j\neq 1}}^{N-1} \frac{Gm_im_j}{|\mathbf{r}_i-\mathbf{r}_j|}.\end{aligned}$$
One important point to note is that our choice of $\operatorname{\mathcal{H}}_\pm$ is slightly different from that used by [@solarsystemdynamics] and [@WisdomHolman1991]. These authors use $$\begin{aligned}
\left(\operatorname{\mathcal{H}}_\pm\right)_{\rm WH1991} &=& \sum_{i=1}^{N-1}\frac{G m'_i \, \mathcal{M}_i }{|\mathbf{r}'_i|}.\end{aligned}$$ where $\mathcal{M}_i = m_0 \frac{M_i}{M_{i-1}}$. Their choice leads to the usual disturbing function in perturbation theory. We conducted various tests but found no significant difference between these mass choices. We therefore chose our prescription, Eq.\[eq:hpm\], which has a simpler physical interpretation: the mass entering Kepler’s third law is simply the interior mass.
Wisdom-Holman Mapping
---------------------
Our goal is to find a solution to the equations of motion for particles governed by the Hamiltonian in Eq. \[eq:H\]. No analytic solution exists to the full Hamiltonian and we thus need to find an approximate solution. There are many different ways to do that. Here, we describe the idea of constructing a symplectic integrator by means of splitting the Hamiltonian into smaller parts, each of which can be easily integrated.
The introduction of Jacobi coordinates led us to the Hamiltonian splitting described in Sect. \[sec:splitting\]. Analytic solutions can be found for the evolution of the system under each of the individual Hamiltonians $\operatorname{\mathcal{H}}_0$ and $\operatorname{\mathcal{H}}_{\rm Interaction}$. The solution to $\operatorname{\mathcal{H}}_0$ simply corresponds to motion along a straight line. The solution to $\operatorname{\mathcal{H}}_{\rm Interaction}$ is a kick step where the velocities change due the inter-particle accelerations but the position remain constant. The solution to $\operatorname{\mathcal{H}}_{\rm Kepler}$ is a set of two-body Kepler orbits, which can also be easily solved with an iterative algorithm. We discuss the details related to the Kepler problem in Sect. \[sec:kepler\].
Now that we have broken down the full Hamiltonian into individual Hamiltonians, to all of which we know the solution (or can easily calculate it), we can construct a symplectic integrator for the total Hamiltonian using an operator split method [e.g. @SahaTremaine1992]. Let us describe the evolution of particles under a Hamiltonian $\operatorname{\mathcal{H}}$ for a time $\mathit{dt}$ using the operator notation $\hat{\operatorname{\mathcal{H}}}(\mathit{dt})$. The notation $\hat{\operatorname{\mathcal{H}}}_2(\mathit{dt})\circ\hat{\operatorname{\mathcal{H}}}_1(\mathit{dt})$ means applying operator $\hat{\operatorname{\mathcal{H}}}_1$ first, then applying operator $\hat{\operatorname{\mathcal{H}}}_2$. It is easy to see that many of the operators commute with each other, i.e. $$\begin{aligned}
\left[\hat \operatorname{\mathcal{H}}_0, \hat \operatorname{\mathcal{H}}_{\rm Kepler} \right] &=& 0 \\
\left[\hat \operatorname{\mathcal{H}}_0, \hat \operatorname{\mathcal{H}}_{\rm Interaction} \right] &=& 0\\
\left[\left(\hat \operatorname{\mathcal{H}}_{\rm Kepler}\right)_i , \left(\hat \operatorname{\mathcal{H}}_{\rm Kepler}\right)_j \right] &=& 0\quad \forall i,j,\end{aligned}$$ where $[\hat{\operatorname{\mathcal{H}}}_1,\hat{\operatorname{\mathcal{H}}}_2] = \hat{\operatorname{\mathcal{H}}}_1\circ \hat{\operatorname{\mathcal{H}}}_2- \hat{\operatorname{\mathcal{H}}}_2\circ \hat{\operatorname{\mathcal{H}}}_1$. This leads to the following Drift-Kick-Drift (DKD) operator splitting scheme, which we refer to as the Wisdom-Holman map:
1. Evolve the system under $\hat \operatorname{\mathcal{H}}_{\rm Kepler}(\mathit{dt}/2) \circ \hat\operatorname{\mathcal{H}}_0(\mathit{dt}/2)$.
2. Evolve the system under $\hat \operatorname{\mathcal{H}}_{\rm Interaction}(\mathit{dt})$.
3. Evolve the system under $\hat \operatorname{\mathcal{H}}_{\rm Kepler}(\mathit{dt}/2) \circ \hat\operatorname{\mathcal{H}}_0(\mathit{dt}/2)$.
The ordering of $\hat \operatorname{\mathcal{H}}_{\rm Kepler}$ and $\operatorname{\mathcal{H}}_0$ in the first and last step doesn’t matter as they commute. The first and last steps can be combined if the system is evolved for multiple timesteps.
Note that the evolution of $\operatorname{\mathcal{H}}_{\rm Kepler}$ and $\operatorname{\mathcal{H}}_0$ is most easily accomplished in Jacobi coordinates. The interaction Hamiltonian $\operatorname{\mathcal{H}}_{\rm Interaction}$, however, contains terms that depend on both the Cartesian and Jacobi coordinates. The simplest way to calculate these terms is to convert to Cartesian coordinates, evaluate the $\mathbf{r}_i-\mathbf{r}_j$ term, convert the accelerations back to Jacobi accelerations, and calculate the remaining terms.
Symplectic Correctors {#sec:correctors}
---------------------
The operator splitting method used in the symplectic integrator discussed above effectively adds high frequency terms to the Hamiltonian. An argument often used in favour of symplectic integrators is that, although these high-frequency terms alter the Hamiltonian, they do not change the long term evolution as they average out. However, they do lead to relatively large short term oscillations, for example in the energy error.
The idea of a symplectic corrector, first used by [@TittemoreWisdom1989] and fully developed by [@Wisdom1996], is to remove some of these high frequency terms using perturbation theory. The basic procedure is as follows. Before the start of an integration, we convert from real coordinates to so-called mapping coordinates. Then we perform the integration using our standard symplectic map. After the simulation has finished (or whenever we need an output) we convert back from mapping to real coordinates. The symplectic corrector operator that we use is a combination of several $\hat \operatorname{\mathcal{H}}_{\rm Interaction}(\mathit{dt})$ and $\hat \operatorname{\mathcal{H}}_{\rm Kepler}(\mathit{dt}) \circ \hat\operatorname{\mathcal{H}}_0(\mathit{dt})$ operators applied for different (positive and negative) intervals $\mathit{dt}$. If $\epsilon$ is the order of the perturbations, i.e. the mass ratio and therefore the relative magnitude of $\hat \operatorname{\mathcal{H}}_{\rm Interaction}$ compared to $\hat \operatorname{\mathcal{H}}_{\rm Kepler}$, then one can show that the use of symplectic correctors can lead to a scheme of order $O(\epsilon \mathit{dt}^{K} ) + O(\epsilon^2 \mathit{dt}^2)$ where $K$ is the order of the symplectic corrector [@MikkolaPalmer2000]. A second order Wisdom-Holman map without symplectic correctors has an energy error of order $O(\epsilon^2 \mathit{dt}^2)$. Because this coordinate transformation for the symplectic corrector is only performed for outputs and at the beginning and end of the simulation, its effect on the speed of the algorithm is negligible for sparse output.
A full derivation of the symplectic correctors would go beyond the scope of this paper and we refer the reader to [@Wisdom1996] and [@MikkolaPalmer2000]. The corrector coefficients are listed in a compact form in [@Wisdom2006].
We implement a third, fifth, seventh and eleventh order symplectic corrector for [[WHFast]{}]{}. Whether the high order-symplectic correctors provide any improvement over the low-order ones depends on the mass ratios in the system. For Jupiter-mass planets, a symplectic corrector of fifth order is no less accurate than a higher order one. If in doubt, there is no harm done in using a higher-order corrector as the speed implications are minimal. Thus, we implement the eleventh-order symplectic corrector by default.
Chaos Indicators
----------------
A powerful tool for studying the long term evolution of Hamiltonian systems is the Lyapunov characteristic number (LCN). The inverse of the LCN is the Lyapunov timescale and gives an estimate of how fast two nearby particle trajectories diverge. If the system is chaotic, the divergence is exponential in time and the Lyapunov timescale is finite. Thus, measuring the LCN gives us an estimate of whether the system is chaotic and, if so, on what timescale.
A more recent approach with similar informative value is the Mean Exponential Growth factor of Nearby Orbits, or MEGNO for short [@Cincotta2003]. The MEGNO, $Y(t)$, is a scalar function of time, and provides a clear picture of resonant structures and of the locations of stable and unstable periodic orbits.
There are two ways to calculate the LCN or the MEGNO. Conceptually the simplest is to integrate an additional shadow particle for each body in the simulation, i.e. a particle with slightly perturbed initial conditions. One can then directly measure the divergence of each particle’s path from its shadow. The second approach is to consider each body’s six-dimensional displacement vector $\pmb{\delta_i}$ from its shadow (in both position and velocity) as a dynamical variable. One can then obtain differential equations for each $\pmb{\delta_i}$ vector by applying a variational principle to the trajectories of the original bodies. We choose to follow the latter approach, as it is both faster and numerically more robust [@Tancredi2001]. In this scheme, one can imagine shadow particles with phase-space coordinates $\pmb \xi^s_i = \pmb\xi_i + \pmb\delta_i$, where $\pmb\xi_i=(\mathbf{r_i}, \mathbf{v_i})$ is the phase-space coordinate of the $i$-th original particle. Initially, we set each component of $\pmb \delta_i$ to a small value.
We follow the work of [@MikkolaInnanen1999] who describe how to efficiently couple the variational equations to the original equations of motion. This allows us to construct a symplectic integrator for the variational equations (a symplectic tangent map). An important advantage of this method is that we only solve Kepler’s equation once for each particle/shadow-particle pair (one of the most time-consuming steps in a Wisdom-Holman integrator for small particle numbers).
The MEGNO is then straightforwardly computed from the variations as [@Cincotta2003] $$\begin{aligned}
Y(t) = \frac{2}{t} \int_0^t t' \frac{\sum_{i=0}^{N-1} \dot{\pmb{\delta}_i}(t') \cdot \pmb{\delta}_i(t')}{\sum_{i=0}^{N-1} \pmb{\delta}_i^2(t')} dt'.\end{aligned}$$ If $Y(t)\rightarrow \infty$, then the system is chaotic. For quasi-periodic orbits, the MEGNO converges to a finite value, $Y(t)\rightarrow 2$ [e.g. @Hinse2010].
One can obtain the Lyapunov characteristic number (LCN), the inverse of the Lyapunov timescale, from the time evolution of the MEGNO via a linear least square fit to $Y(t)$.
Kepler Problem with Variations {#sec:kepler}
------------------------------
In this section, we summarize how to solve the two-body Kepler problem numerically, including the variational equations. Although the solution has been known since the days of Newton, the transcendental nature of Kepler’s equation does not admit a closed-form mathematical expression.
We closely follow the work of [@MikkolaInnanen1999] where the reader can find additional information that we have left out. The equivalent one-body Hamiltonian for the Kepler problem is $$\begin{aligned}
H_{\rm Kepler} &=& \frac12 \mathbf{v}^2 - \frac{M}{|\mathbf{r}|},\end{aligned}$$ where $M$ is the total mass of the two bodies. For consistency with [@MikkolaInnanen1999], we have dropped the primes, have scaled out $m'_i$ from $p'_i$, and rewritten Eq. \[eq:HKepler\] in non-dimensional form, i.e. the gravitational constant $G=1$ for the remainder of this paper. However, we have taken care to remove any dependence on the choice of units from our implementation, so $G$ can be freely set by the user in our implementation of the algorithm.
Our task is to find the final positions and velocities $\mathbf{r}$ and $\mathbf{v}$ of a particle evolving under this Hamiltonian for some time $\mathit{dt}$, given the initial conditions $\mathbf{r}_0$ and $\mathbf{v}_0$. Thus, we seek the effect of the operator $\hat H_{\rm Kepler}(\mathit{dt})$.
It is advantageous to solve the Kepler problem numerically using the Gauss f and g functions, which express the relevant quantities in terms of $\mathbf{r}_0$ and $\mathbf{v}_0$ [@WisdomHolman1991]. This avoids the computationally expensive conversion between Cartesian and classical orbital elements, and avoids coordinate singularities associated with circular orbits. We find that it is advantageous to use universal variables in this solution [@Stumpff1962]. This approach provides greater speed and numerical stability compared to a solution using elliptic elements. It also avoids the singularity associated with the transition from elliptic to hyperbolic motion.
To solve the analogue of Kepler’s equation for the particle’s position in time, we make use of several special functions. Let us begin by defining the $c$-functions [@Stumpff1962] as a series expansion: $$\begin{aligned}
c_n(z) \equiv \sum_{j=0}^\infty \frac{(-z)^j}{(n+2j)!} \label{eq:c},\end{aligned}$$ which satisfy the recursion relation $$\begin{aligned}
c_n(z) &=& \frac{1}{n!} - z \,c_{n+2}. \label{eq:recur}\end{aligned}$$ The $c$-functions are related to trigonometric functions, for example $$\begin{aligned}
c_0(z) = \cos \sqrt{z} \quad\quad \text{and}\quad\quad c_1(z) = \frac{\sin\sqrt{z}}{\sqrt{z}},\end{aligned}$$ and thus satisfy the following relationships [@Mikkola1997], which are related to the half-angle formula for trigonometric functions: $$\begin{aligned}
c_5(z) &=& \frac1{16} \left[ c_5(z/4) + c_4(z/4) + c_3(z/4)c_2(z/4)\label{eq:crel5}\right]\\
c_4(z) &=& \frac1{8} c_3(z/4) \left[1 + c_1(z/4).\label{eq:crel4}\right]\end{aligned}$$ Values for $c_0$ through $c_3$ are then readily computed from Eq. \[eq:recur\]. Next, we introduce the so called $G$-functions [@StiefelScheifele1971] which in turn depend on the $c$-functions: $$\begin{aligned}
G_n(\beta,X) \equiv X^n c_n(\beta X^2). \label{eq:G}\end{aligned}$$ The $G$-functions also satisfy recursion relationships similar to those mentioned above for the $c$-functions. We can easily calculate derivatives of $G_n$ by looking at the series expansion of $c_n$ [see @MikkolaInnanen1999 for details]. With this framework, we can now write down the steps needed to find the solution to the Kepler Hamiltonian in compact form.
First, we need to calculate the following three quantities from the initial conditions $\mathbf{r}_0, \mathbf{v}_0$: $$\begin{aligned}
\beta &=& \frac{2M}{r_0}-v_0^2\\
\eta_0 &=& \mathbf{r}_0 \cdot \mathbf{v}_0 \\
\zeta_0 &=& M-\beta r_0\end{aligned}$$ where $r_0 = |\mathbf{r}_0|$ and $v_0 = |\mathbf{v}_0|$. Note that the semi-major axis $a$ can be written as $a=M/\beta$.
Second, we need to solve Kepler’s equation which, using the above notation, takes the form $$\begin{aligned}
r_0 X + \eta_0 G_2(\beta,X) + \zeta_0 G_3(\beta,X) - \mathit{dt} = 0. \label{eq:kepler} \end{aligned}$$ We solve this equation for $X$. This is a non-algebraic (i.e. transcendental) equation that we need to solve iteratively, for example using Newton’s method. In Sect. \[sec:newton\], we describe our algorithm in detail.
Third, having solved Kepler’s Equation, we can calculate the so called Gau[ß]{} $f$ and $g$-functions as well as their time derivatives via $$\begin{aligned}
f &= 1 - M \frac{G_2}{r_0}\quad & \dot f &= -\frac{M\,G_1}{r_0r} \label{eq:fg}&\\
g &= dt - M G_3 & \dot g &= 1-\frac{M\,G_2}{r},\label{eq:fg2}\end{aligned}$$ where $r = r_0 + \eta_0 G_1 + \zeta_0 G_2$. Note that all the $G$-functions depend on $\beta$ and the $X$ value found in the second step.
Fourth, we write the final positions and velocities as a linear transformation of the initial conditions using the Gau[ß]{} $f$ and $g$-functions: $$\begin{aligned}
\mathbf{r} &= f \mathbf{r}_0 + g \mathbf{v}_0
&\mathbf{v} &= \dot f \mathbf{r}_0 + \dot g \mathbf{v}_0.\label{eq:fgupdate}&&\end{aligned}$$ This completes the solution of the Kepler problem.
To solve for the variational equations, we also make use of the $G$-functions. Fortunately, we only need to solve Kepler’s equation once (to solve for $X$). We then get the solution for the variational equations without solving another transcendental equation and thus have only one iteration loop per timestep for both the particle and its variational counterpart. The position and velocity components of $\pmb \delta$ at the end of the timestep, $\delta \mathbf{r}$ and $\delta \mathbf{v}$, can be written as $$\begin{aligned}
\delta \mathbf{r} &=& f \;\delta \mathbf{r}_0 + g \;\delta\mathbf{v}_0 + \mathbf{r}_0 \;\delta f+ \mathbf{v}_0 \;\delta g\label{eq:var1}\\
\delta \mathbf{v} &=& \dot f \;\delta \mathbf{r}_0 + \dot g \;\delta\mathbf{v}_0 + \mathbf{r}_0 \;\delta \dot f+ \mathbf{v}_0 \;\delta \dot g,\label{eq:var2}\end{aligned}$$ where the variations $\delta f$, $\delta g$, $\delta \dot f$ and $\delta \dot g$ can be derived from Eqs. \[eq:fg\]-\[eq:fg2\] (see @MikkolaInnanen1999 for the explicit expressions).
Types of Numerical Errors {#sec:errors}
-------------------------
There are three distinct effects contributing to the energy error of a symplectic integrator [see e.g., @QuinnTremaine1990]. See also [@ReinSpiegel2015] for a similar discussion for non-symplectic integrators.
First, there is an error term associated with the integrator itself because we are not solving the equations of motion for the Hamiltonian $\mathcal{H}$ exactly. For symplectic integrators such as those discussed here, this error term is bound and we call it $E_{\rm bound}$. If the mass ratio of the planets to the star is $\epsilon$, then the order of this error term is roughly $O(\epsilon\, \mathit{dt}^{2})$ for integrators without symplectic corectors and $O(\epsilon \,\mathit{dt}^{K} ) + O(\epsilon^2 \,\mathit{dt}^2)$ for those with symplectic correctors (see Sect. \[sec:correctors\]). Note that $E_{\rm bound}$ is independent of time $t$.
Second, there is an error term associated with the finite precision of numbers represented on a computer. We can only represent a small subset of all real numbers exactly in floating-point precision. Thus after every operation such as an addition or multiplication, the computer rounds to a nearby floating-point number. For CPUs and compilers that follow the IEEE754 standard [@IEEE754], we are guaranteed to round to the nearest floating-point number. Thus, if all operations follow the IEEE754 standard, then as long as the algorithm itself is unbiased, we expect the error to grow as the square root of the number of operations, i.e. $E_{\rm rand} \sim \sqrt{N}\sim \sqrt{t}$, where $N$ is the number of timesteps. This is the best behaviour achievable; to do better we would have to move to extended precision or use fewer operations. This fundamental limit is known as Brouwer’s law [@Newcomb1899; @Brouwer1937].
Third, if any parts of the integration algorithm are biased, the errors will be correlated. This leads to a faster long-term energy-error growth than if errors are uncorrelated; it grows linearly with time, i.e. $E_{\rm bias}\sim N \sim t$.
For a given integrator, which of these three error terms dominates depends on the nature of the simulation, the timestep, and the total integration time (number of timesteps).
Improvements {#sec:improvements}
============
The algorithms we describe in Sect. \[sec:background\] have been used successfully for many years. In the following, we show how to significantly improve the speed and accuracy of the algorithms by taking special care in the implementation of several details, many of which are related to finite floating-point precision on modern computers.
For the remainder of this paper, we will assume that we work with a CPU that follows the IEEE 754 standard for floating-point arithmetic. Most importantly, we assume that all floating-point operations follow the *rounding to nearest, ties to even* rule [@IEEE754]. What follows is in principle applicable to any precision. However, we work exclusively in double floating-point precision (64 bit) which is used on almost all modern CPUs.
Jacobi Coordinate Transformations {#sec:jacobi}
---------------------------------
The evolution under the effect of the interaction Hamiltonian is most efficiently done in Cartesian coordinates. On the other hand, the evolution of the Kepler Hamiltonian is easier in Jacobi coordinates. We thus need an efficient way to convert to and from Jacobi coordinates.
Luckily, the conversion from Cartesian to Jacobi coordinates and back can be done efficiently in $\mathcal{O}(N)$. We construct the algorithms from the definitions above and list them here in pseudo code. As before, primes denote Jacobi coordinates, Note that these algorithms work even if some of the bodies are test particles with $m_i=0$ (for $i\neq 0$). To convert from Cartesian to Jacobi coordinates:
$\mathbf{R} \gets m_0 \cdot \mathbf{r}_0$ $\mathbf{r}'_i \gets \mathbf{r}_i - \mathbf{R}/M_{i-1}$ $\mathbf{R} \gets \mathbf{R} \cdot (1 +m_i/M_{i-1}) + m_i\cdot\mathbf{r}'_i$ $\mathbf{r}'_0 \gets \mathbf{R} / M_{N-1}$
Similarly, we construct the algorithm to convert back from Jacobi to Cartesian coordinates as follows:
$\mathbf{R} \gets \mathbf{r}'_0\cdot M_{N-1}$ $\mathbf{R} \gets (\mathbf{R}- m_i \cdot \mathbf{r}'_i) / M_i$ $\mathbf{r}_i \gets \mathbf{r}'_i + \mathbf{R}$ $\mathbf{R} \gets \mathbf{R} \cdot M_{i-1}$ $\mathbf{r}_0 \gets \mathbf{R}/m_0$
We thoroughly tested the conversions to and from Jacobi coordinates to ensure they are unbiased. This task turns out to be much harder than we na[ï]{}vely expected. As an example, consider the following algorithm which is formally equivalent to the above but numerically much less stable.
$\mathbf{R} \gets 0$ $\mathbf{r}_i \gets \mathbf{r}'_0 + M_{i-1}/M_i \cdot \mathbf{r}'_i -\mathbf{R}$ $\mathbf{R} \gets \mathbf{R} + m_i/M_i \cdot \mathbf{r}'_i$ $\mathbf{r}_0 \gets \mathbf{r}'_0 - \mathbf{R}$
In the above algorithm, we access $\mathbf{r}_0'$ multiple times and have to do a subtraction in the last step. This significantly promotes error propagation and leads to floating-point errors that can be orders of magnitudes higher than in the other implementation. After many timesteps, this leads to a linear secular growth in the energy error.
Implementation of Newton’s Method {#sec:newton}
---------------------------------
To solve Kepler’s equation (Eq. \[eq:kepler\]) for $X$, we need to use an iterative scheme. We now describe our implementation of Newton’s method in floating-point arithmetic. The straightforward implementation is an iteration loop that terminates when the change to $X$ is small, e.g.,
$X \gets initial\; guess$ $dX \gets -f(X)/f'(X)$ $X \gets X + dX$ .
Here, ${\ensuremath{\mathit{eps}}}$ is a small number just above machine precision, typically ${\ensuremath{\mathit{eps}}}\sim10^{-15}$. We use a different implementation of Newton’s method that is both faster and more accurate, despite the fact that it is algebraically equivalent to the above implementation.
$X \gets initial\; guess$ $X_{\rm prev 1} \gets \mathit{NaN}$ $X_{\rm prev 2} \gets X_{\rm prev 1}$ $X_{\rm prev 1} \gets X$ $X \gets (X\cdot f'(X)-f(X))/f'(X)$
Note that the equal sign in the above breakout condition is evaluated in floating-point precision. In comparison to the first algorithm, at each iteration step we test whether the iteration has converged by a simple comparison rather than by a slow division and absolute-value operation.
We keep track of two previous values instead of just one because for certain initial conditions, the iteration can cycle indefinitely between two nearby floating-point numbers and not converge to a single floating-point number.
Our implementation thus ensures that the value of $X$ is more accurately calculated than in the straightforward implementation using a heuristic value of ${\ensuremath{\mathit{eps}}}$. A further advantage of rewriting Newton’s method in the above form is that the term on the right-hand-side of the last line can be simplified significantly for the Kepler problem, giving: $$\begin{aligned}
X \gets \frac{X (\eta_0 G_1 + \zeta_0 G_2) -\eta_0 G_2 -\zeta_0 G_3+dt }{r_0+\eta_0 G_1 +\zeta_0 G_2} \label{eq:kepeq}\end{aligned}$$ where the $G$’s on the right-hand-side all depend on $X$ and $\beta$ (see Eq. \[eq:G\]).
We also experimented with higher-order generalizations of Newton’s method (Householder’s methods). For typical cases where the orbits are not extremely elliptical ($e\lesssim 0.99$) and the timestep is much smaller than the shortest orbital period, we found Newton’s method to always be fastest. This is because when the value and derivatives of the function are easily evaluated, the precision gain from these higher-order methods does not compensate for the increased computation cost of each iteration. In other words, while higher-order methods will converge in fewer iterations than Newton’s method, the overall computation time is longer. At large eccentricities and long timesteps, the $G$-function evaluations become expensive (one must recursively apply the quarter-angle formulas described in Sect. \[sec:cs\]), and higher-order methods are helpful. For large eccentricities we use a higher order method described in detail in Sect. \[sec:largeE\]. To safeguard against rare cases where Newton’s method might fail, we also implemented a failsafe bisection method. We find that the bisection method is only triggered when the timestep is comparable to the orbital period.
The Initial Guess for Kepler’s Equation: Short Timesteps
--------------------------------------------------------
The quantity X in Eq. \[eq:kepeq\] can also be expressed as $$\begin{aligned}
X = \int_{t_0}^{t_0 + \mathit{dt}} \frac{\mathit{dt}'}{r} = \mathit{dt} \cdot{\langle r^{-1}\rangle} \label{eq:xint}\end{aligned}$$ where $t_0$ is the time at the beginning of the timestep, and $\langle r^{-1}\rangle$ is the time-averaged value of $r^{-1}$ over the interval $[t_0, t_0+\mathit{dt}]$. Thus, if the orbit’s eccentricity $e$ is low, or more generally if the timestep is short enough that the orbital radius does not vary much, then $X \approx \mathit{dt}/r_0$. The troublesome cases are highly eccentric orbits near pericentre where the radius changes rapidly. For such cases, the radius varies by a factor of $1+e \approx 2$ from pericentre to a true anomaly of $90^\circ$. We can therefore estimate the timescale over which the orbital radius varies near pericentre as $$\begin{aligned}
T_{char} = \frac{q}{v_q} = \frac{a(1-e)}{na}\Bigg(\frac{1-e}{1+e}\Bigg)^{1/2} \sim \frac{(1-e)^{3/2}}{n}, \label{eq:tchar}\end{aligned}$$ where $q$ is the pericentre distance, $v_q$ is the speed at pericentre and $n$ is the mean motion. Thus, if one does not resolve pericentre passages (i.e., $n\,\mathit{dt} = \Delta M \gtrsim (1-e)^{3/2}$), $X$ will differ from $\mathit{dt}/r_0$ near pericentre (but may nevertheless conform to the simple approximation at apocentre where the body moves slowly).
More quantitatively, one can non-dimensionalize Eq. \[eq:kepeq\], setting $\tilde{X} = r_0 X / \mathit{dt}$. One can then solve the equation perturbatively, assuming the deviations from $\tilde{X} = 1$ are small. This procedure requires that the following three non-dimensional parameters in the equation also be much smaller than unity, $$\begin{aligned}
\chi \equiv \frac{\beta \mathit{dt}^2}{r_0^2}\quad \eta \equiv \frac{\eta_0\mathit{dt}}{r_0^2} \quad \zeta \equiv \frac{\zeta_0\mathit{dt}^2}{r_0^3}.\end{aligned}$$ One can show that when our heuristic estimate $\Delta M \ll (1-e)^{3/2}$ is satisfied, $\chi$, $\eta$, $\zeta \ll 1$. In this case, one can extend the solution of Eq. \[eq:kepeq\] to higher order. For the initial guess in our algorithm, we go up to second order $$\begin{aligned}
X = \frac{\mathit{dt}}{r_0} \cdot \left(1 - \frac12 \eta \right). \label{eq:xinit}\end{aligned}$$ We experimented with higher-order initial guesses (see @Danby1987 for explicit expressions), but found these to be slower, even for small eccentricities and timesteps. This can again be attributed to the computational efficiency of each iteration of Newton’s method.
Large Eccentricities and Timesteps {#sec:largeE}
----------------------------------
The previous two sections describe an optimized algorithm for solving Kepler’s equation when the timestep and eccentricities are low. We have also developed an improved handling of high-eccentricity/long-timestep cases. In this regime, both the solver and initial guess should be modified.
Like previous authors [@Conway1986; @Danby1987], we found the root-finding method of Laguerre-Conway to be most stable. However, unlike [@Danby1987], who finds the method to always converge (presumably using comparatively small timesteps), we often have to resort to bisection when the timestep is comparable to the orbital period. Of course, such long timesteps should not be chosen anyway, since they poorly sample inter-planet interactions, and are more susceptible to timestep resonances [@WisdomHolman1992; @ToumaWisdom1993; @Rauch1999].
We also had to modify the breakout condition used for Newton’s method. While the Laguerre-Conway algorithm sometimes also bounces between two floating-point values once it has converged, in this regime the method often executes larger-period cycles (e.g., it will periodically repeat the last eight floating-point numbers). We therefore chose to store the values from each iteration and exit the loop whenever a result was repeated.
One way to determine which solver should be used is to check whether $\mathit{dt}$ is smaller than $T_{char}$ (Eq. \[eq:tchar\]). However, because $T_{char}$ is expensive to compute from $\mathbf{r}_0$ and $\mathbf{v}_0$, we instead check how much the first iteration of Newton’s method deviates from the initial guess, as a fraction of $2\pi \beta^{-1/2}$. The latter is a natural quantity to compare against since it is the value of $X$ when the timestep is equal to the orbital period. We found a threshold of $\sim 1\%$ to strike a good balance over a wide parameter range in timestep/eccentricity space, though the algorithm’s speed is not particularly sensitive to the exact value adopted.
Finally, the method can be sped up in this regime with an improved initial guess for $X$, since $\mathit{dt}/r_0$ in Eq. \[eq:xinit\] blows up near pericentre as the eccentricity gets large. [@Danby1983] provide a widely used initial guess using classical orbital elements but, to our knowledge, no comparably simple initial guess has been found for universal variables.
In this high-eccentricity / long timestep regime, most existing methods using universal variables choose to make the expensive conversion to orbital elements and use Danby’s guess. We instead observed that because $\langle r^{-1} \rangle = a^{-1}$ over one orbital period, $X = \mathit{dt}/a$ for a timestep of one orbit. We find that over a relevant parameter range with timesteps logarithmically spaced between 0.03 and 1 orbital periods, and eccentricities between 0.999 and 0.9999, our improved guess is faster than converting to orbital elements and using Danby’s by $\approx 30\%$. In a manner analogous to that described in the previous section, we also solved Eq. \[eq:kepeq\] perturbatively around $X = \mathit{dt}/a = \beta\; \mathit{dt} / M$ in the regime $\chi \gg 1$, but we found the second-order solution to be a slower initial guess than the simple $X = \beta \; \mathit{dt} / M$.
Implementation of $c$-functions {#sec:cs}
-------------------------------
Finding a solution to Kepler’s equation is done iteratively and is thus the most expensive step in solving the Kepler problem. The iteration itself involves the calculation of multiple $G$-functions, which in turn require the calculation of $c$-functions. Thus, it is particularly important to optimize these functions for both speed and accuracy. When calculating chaos indicators, we need $c_0$, $c_1$, $c_2$, $c_3$, $c_4$ and $c_5$. If we are not integrating the variational equations, we only need $c_0$, $c_1$, $c_2$ and $c_3$.
We first ensure that $z$ is smaller than $0.1$ to guarantee that the series expansion of $c$ in Eq. \[eq:c\] converges. We do this by dividing $z$ repeatedly by 4. Note that divisions by powers of 2 are fast and exact in floating-point arithmetic. To calculate the series expansion, we need an inverse factorial for every term. Calculating this inverse factorial by multiplying floating-point numbers and then implementing a floating-point division would be very slow. We found that the fastest way to calculate the inverse factorial is to use a simple lookup table. We checked that the series expansions of the $c$-functions converge very quickly for small $z$ and thus we only store inverse factorials up to $1/34!$ in the lookup table. Any larger factorial would contribute less than one part in $10^{16}$ to the sum and can thus be neglected (as we work in double floating-point precision).
We always calculate the first two terms in the series expansion. We then enter a loop and add more terms until the result no longer changes. Because $z$ is small and the inverse factorials decrease quickly, we are assured that the series will converge to a single floating-point number. This allows us to simply check whether the value changes from one iteration to the next, which is much faster than evaluating relative changes (cf. Sect. \[sec:newton\]).
Once the $c$-functions are calculated for the small $z$ value, we use the relations in Eqs. \[eq:crel5\]-\[eq:crel4\] with Eq. \[eq:recur\] to calculate the $c$-functions for the original $z$ value.
Because this algorithm is an integral part of the integrator, we list the function to calculate $c(z)$ in pseudo code:
$n \gets 0$ $z \gets z/4$ $n \gets n+1$ $c_4 \gets \frac{1}{4!} - z \cdot \frac{1}{6!}$ $c_5 \gets \frac{1}{5!} - z \cdot \frac{1}{7!}$ $\bar z \gets -z$ $p \gets \bar z$ $k \gets 8$
$p \gets p \cdot \bar z$ $c_4 \gets c_4 + p \cdot \frac{1}{k!}$
$k \gets k+1$ $c_5 \gets c_5 + p \cdot \frac{1}{k!}$ $k \gets k+1$ $c_3 \gets \frac16 - z \cdot c_5$ $c_2 \gets \frac12 - z \cdot c_4$ $c_1 \gets 1 - z \cdot c_3$ $z \gets 4 \cdot z$ $c_5 \gets \frac1{16}\cdot (c_5+c_4+c_3+c_2) $ $c_4 \gets \frac18\cdot c_3 \cdot (1-c_1)$ $c_3 \gets \frac16 -z \cdot c_5$ $c_2 \gets \frac12 -z \cdot c_4$ $c_1 \gets 1 -z \cdot c_3$ $n\gets n-1$ $c_0 \gets 1 -z \cdot c_2$
Implementation of Gau[ß]{} $f$ and $g$-functions {#sec:gauss}
------------------------------------------------
The precise implementation of Gauss $f$ and $g$ functions matters for long term integrations. The straightforward implementation following [@MikkolaInnanen1999] leads to the $f$ and $g$-functions in Eq. \[eq:fg\]. Note that for timesteps smaller than half an orbital period, the term $MG_2/r_0$ in $f$ is small compared to the first term (which is just 1). The same argument holds true for $\dot g$. We can define new $\hat f$ and $\hat{\dot g}$-functions $$\begin{aligned}
&\hat f = - M \frac{G_2}{r_0}\quad & \dot f = -\frac{M\,G_1}{r_0r} \label{eq:fghat}\\
&g = dt - M G_3 & \hat {\dot g} = -\frac{M\,G_2}{r}.\end{aligned}$$ This allows us to rewrite the last step in solving the Kepler problem as $$\begin{aligned}
\mathbf{r} = \left( \hat f \mathbf{r}_0 + g \mathbf{v}_0 \right) + \mathbf{r}_0 \quad\quad\quad\quad
\mathbf{v} = \left( \dot f \mathbf{r}_0 + \hat{\dot g} \mathbf{v}_0 \right) + \mathbf{v}_0.\end{aligned}$$ Although this step is algebraically equivalent to the original Eq. \[eq:fgupdate\], we achieve higher precision. The reason is that we can now ensure that the small quantities in brackets are summed before they are added to the larger quantity (the initial value). We implement the same trick for the variational equations, Eqs. \[eq:var1\] and \[eq:var2\]
A full integration in Jacobi coordinates {#sec:intinjac}
----------------------------------------
The algorithms to convert to and from Jacobi coordinates that we describe in Sect. \[sec:jacobi\] are unbiased and fast. Nevertheless, we aim to avoid as many conversion as possible.
As it turns out, we can reduce the number of conversions per timestep to two, one for the positions from Jacobi coordinates to the inertial frame, and one for the accelerations from the inertial frame to Jacobi accelerations. But note that this is only possible under the following assumptions: 1) the particle position and velocities are not changed in-between timesteps, e.g. manually by the user or by collisions, 2) outputs are not required at every timestep, 3) variational equations are not integrated, 4) no additional velocity-dependent forces are present. In such a case, an integration starting from an arbitrary inertial frame is achieved as follows:
calculate Jacobi coordinates drift all particles under $H_{\rm Kepler}$ for half a timestep, $\mathit{dt}/2$ calculate 1st part of $H_{\rm Interaction}$ in Jacobi coordinates update positions in the inertial frame calculate 2nd part of $H_{\rm Interaction}$ in inertial frame convert accelerations from 2nd part to Jacobi accelerations apply kick from Jacobi accelerations to Jacobi velocities drift all particles under $H_{\rm Kepler}$ for a full timestep $\mathit{dt}$ drift all particles under $H_{\rm Kepler}$ for half a timestep $\mathit{dt}/2$ update both positions and velocities in the inertial frame.
Note that we never update the velocities in the inertial frame until the end of the simulation (or when an output is needed). We only convert the positions and velocities to Jacobi coordinates at the very beginning and not at every timestep. Besides the obvious speed-up, avoiding to go back and fourth between different coordinate systems reduces the build-up of round-off errors and thus makes the integrator more robust.
LCN calculation {#sec:megnocalc}
---------------
To calculate the Lyapunov characteristic number and the Lyapunov timescale we need to perform a linear least square fit to the function $Y(t)$. Thus we need the mean and the covariance of $Y(t)$. Storing all previous values of $Y(t)$ just to calculate its mean and covariance is inefficient. We therefore implement an efficient one-pass method described by [@Pebay2008]. This method lets us calculate the LCN at every timestep in $\mathcal{O}(1)$ and has the further advantage of being numerically more robust than the standard implementation.
Numerical Results {#sec:numericalresults}
=================
In this section, we test the speed, accuracy and numerical stability of [[WHFast]{}]{}and compare it to other publicly available and widely used integrators. We begin by briefly defining our nomenclature for these other integrators and summarizing their properties.
[[MERCURY]{}]{}is a mixed-variable symplectic integrator implemented in fortran and provided by the [[MERCURY]{}]{}package [@Chambers1997]. This Wisdom-Holman style integrator uses high-order symplectic correctors. We directly call the fortran code without any modifications.
[[SWIFTER-WHM]{}]{}is again a classical 2nd-order Wisdom-Holman integrator without symplectic correctors [@WisdomHolman1991]. We use the integrator provided by the [[SWIFTER]{}]{}package. It is implemented in fortran and we directly call the [[SWIFTER]{}]{}executable without any modifications.
[[SWIFTER-HELIO]{}]{}is also 2nd-order symplectic integrator without symplectic correctors [@Duncan1998]. It uses democratic heliocentric coordinates. We again use the integrator provided by the [[SWIFTER]{}]{}package. It is implemented in fortran and we directly call the [[SWIFTER]{}]{}executable without any modifications.
[[SWIFTER-TU4]{}]{}is a 4th-order symplectic integrator. It is *not* a Wisdom-Holman integrator but splits the Hamiltonian in kinetic and potential terms [@Gladman1991]. We also use the integrator provided by the [[SWIFTER]{}]{}package. It is implemented in fortran and we directly call the [[SWIFTER]{}]{}executable without any modifications.
For a more direct comparison, we also make use of an integrator that we simply refer to as [[WH]{}]{}. It is based on the [[SWIFTER-WHM]{}]{}integrator in [[SWIFTER]{}]{}but ported to C and available in the [[REBOUND]{}]{}[@ReinLiu2012] package. Like the [[SWIFTER-WHM]{}]{}integrator, it is a symplectic integrator that works in the heliocentric frame, and does not implement any symplectic correctors. Note that this is not the original integrator used by [@WisdomHolman1991], which is not publicly available.
[[WHFast]{}]{}is C99 compliant. The C99 standard guarantees that floating point operations are not re-ordered by the compiler (unless one of the fast-math options is turned on). Because of that, the final positions and velocities of particles agree down to the last bit across different platforms. This makes [[WHFast]{}]{}platform independent and the simulation results reproducible. We verified this on different architectures (Linux, MacOSX), different CPUs (Intel Core i5-3427U, Intel Xeon E5-2697 v2, Intel Xeon E5-2620 v3) and different compilers (Apple LLVM 6.1.0, gcc 4.4.7).
Two-body Kepler Solver
----------------------
The kernel of every Wisdom-Holman integrator is the Kepler solver. We describe our implementation in detail in Sections \[sec:newton\]-\[sec:gauss\]. Here, we test the Kepler solver using a two-body problem. The two body problem is invariant with respect to rescaling of the total mass, the mass ratio, the value of the gravitational constant and the orbital period. What does matter is the eccentricity of the orbit and the ratio of the timestep to the orbital period. We thus scan the parameter space in those two dimensions by integrating two bodies for 100 orbital periods. We explore an extremely wide parameter space. The eccentricities range from zero to $0.999\,999\,99 = 1-10^{-8}$. The range of timesteps goes from 0.1% of the orbital period all the way up to one orbital period.
Fig. \[fig:2body\] shows the performance of [[WHFast]{}]{}(right column) compared to [[WH]{}]{}(left column). The top row shows the absolute value of the relative energy error at the end of the simulation. The middle row shows the sign of the energy error. The bottom row shows the average runtime for a single timestep. The vertical lines visible in the top row correspond to timestep resonances [@WisdomHolman1992; @ToumaWisdom1993; @Rauch1999].
One can see that [[WHFast]{}]{}is significantly more accurate than the standard [[WH]{}]{}integrator for the most important parts of parameter space (eccentricities less than $\sim 0.99$). The relative energy is conserved better by two to three orders of magnitude. Most importantly, note that the energy error in the standard [[WH]{}]{}integrator is biased over large regions of the parameter space (there are large blue and red areas in the second row). On the other hand, [[WHFast]{}]{}has a random energy error throughout the parameter space. Having a biased energy error will lead to a long-term linear growth of the energy error (see below).
In the entire parameter space explored, [[WHFast]{}]{}requires less time to complete a timestep than [[WH]{}]{}. The speed-up is typically between 20% and 100%. For the integrations performed in this section, we convert to and from Jacobi coordinates at every timestep to provide a fair comparison. Thus, the speed-up and the energy-conservation properties of [[WHFast]{}]{}are in fact even better than shown here in any actual production run (see Sect. \[sec:intinjac\]).
Short Term Energy Conservation {#sec:shorttermenergy}
------------------------------
To compare the accuracy of the different integrators in a realistic test case, we run simulations of the outer Solar System for one thousand Jupiter orbits ($12\,000$ years). We include the Sun and four massive bodies with approximate initial conditions corresponding to those of Jupiter, Saturn, Uranus and Neptune. In each simulation the initial conditions and masses are randomly perturbed by 0.1%. In Fig. \[fig:shorttermenergy\], we plot the relative energy errors at the end of the simulation as a function of the number of timesteps imposed per Jupiter orbit.
One can see that all the integrators except [[SWIFTER-TU4]{}]{}are second-order schemes. For timesteps between 20% and 0.1% of the orbital period of Jupiter (50 to 1000 timesteps per orbit), their error decreases quadratically with decreasing timesteps. This is the error term $E_{\rm bound}$ introduced in Sect. \[sec:errors\].
However, decreasing the timestep also increases the number of floating point operations. There will therefore be a timestep value at which the numerical round-off error dominates over the error associated with the symplectic method itself $E_{\rm bound}$. For that reason we find that for small timesteps, less than 0.1% of the shortest orbital period, the errors of all integrators rise instead of decreasing further. Thus there is an optimum timestep $\mathit{dt}_{\rm opt}$ that yields the minimum energy error. This optimum timestep depends on the length of the integration and will be larger for longer simulations.
In Fig. \[fig:shorttermenergy\] one can see that the errors of [[WH]{}]{}, [[SWIFTER-WHM]{}]{}, [[SWIFTER-HELIO]{}]{}and [[MERCURY]{}]{}rise very rapidly after reaching $\mathit{dt}_{\rm opt}$, scaling as at least $\mathit{dt}^{-2}$ for the first decade.
The optimum timestep for [[WHFast]{}]{}is roughly 0.1% of the shortest orbital period. However, [[WHFast]{}]{}’s error grows much more slowly with decreasing timestep than that of the other second-order integrators. In fact, the error is dominated by $E_{\rm rand}$ and thus follows $\mathit{dt}^{-1/2}$ as the number of timesteps $N_{\rm steps}$ increases as $\sim\mathit{dt}^{-1}$ if we keep the total integration time constant. Thus the behaviour of [[WHFast]{}]{}in Fig. \[fig:shorttermenergy\] for small timesteps can be seen as the first indication that [[WHFast]{}]{}follows Brouwer’s law (see Sects. \[sec:errors\] and \[sec:longtermtest\]).
The [[SWIFTER-TU4]{}]{}integrator is the only other integrator we tested that seems to follow Brouwer’s law, but it performs poorly at large timesteps. This is expected, since unlike the other integrators, [[SWIFTER-TU4]{}]{}does not assume a Keplerian splitting and must therefore take smaller timesteps to accurately reproduce the orbital motions.
Integrators with symplectic correctors, [[MERCURY]{}]{}and [[WHFast]{}]{}, perform significantly better for long timesteps. Their energy conservation is three orders of magnitude better ($E_{\rm bound}$ is three orders of magnitude smaller) compared to integrators without symplectic correctors. This is due to the mass ratio of Jupiter and the Sun being roughly $10^{-3}$. The order of the symplectic corrector is not very important for relatively high mass ratios such as these, i.e. a fifth-order symplectic corrector performs as well as an 11th-order one. For much smaller mass ratios (when the mass ratio is less than the timestep ratio), higher-order symplectic correctors are advantageous.
Note that $\mathit{dt}_{\rm opt}$ for almost all of the integrators is $10^{-3}$ orbital periods of Jupiter, i.e. 4 days. This is significant because Mercury’s orbital period is 88 days. Thus if we included Mercury in our simulation, we would be very restricted in our timestep choice. We need more than 20 timesteps ($\mathit{dt}\approx4$ days) to resolve Mercury’s orbit accurately. However, if we choose choose a timestep smaller than 4 days, we start to accumulate errors in the outer Solar System. It is worth reiterating that the simulations shown in Fig. \[fig:shorttermenergy\] all ran for only 1000 orbits. If we ran a longer simulation with the same timestep, we would have more timesteps and thus accumulate more round-off errors by the end of the simulation. One can therefore reach better energy conservation with a longer timestep. In other words, $\mathit{dt}_{\rm opt}$ is larger for longer integration times.
Speed Comparison {#sec:speedcomparison}
----------------
We run the same simulations as in Sect. \[sec:shorttermenergy\] to compare the speed of the different integrators. Fig. \[fig:speedcomparison\] shows the relative energy error as a function of runtime. The results show that no matter what the desired energy error is, [[WHFast]{}]{}is the fastest integrator. In the large timestep limit, the speed-up compared to [[MERCURY]{}]{}is roughly a factor of 5.
In the small timestep limit, $\mathit{dt} < \mathit{dt}_{\rm opt}$, we can only compare [[WHFast]{}]{}to [[SWIFTER-TU4]{}]{}, as all the other integrators’ errors are significantly larger (by 4 to 5 orders of magnitude) due to numerical roundoff errors (see below). [[SWIFTER-TU4]{}]{}is as fast for small timesteps as [[WHFast]{}]{}but, as noted above, is unsuitable for large timesteps since it is not a Wisdom-Holman integrator. It is only shown here as a comparison.
Long Term Energy Conservation {#sec:longtermtest}
-----------------------------
Let us finally address the most important benchmark, the long term energy conservation properties of [[WHFast]{}]{}compared to other integrators in a real world test case. In this section we only study the energy error, but other conserved properties like the angular momentum behave the same way. In Fig. \[fig:longtermtest\], we show the time evolution of the relative energy error in a simulation of the outer Solar System. As in Sect. \[sec:shorttermenergy\], we include the Sun and four massive bodies with approximate initial conditions corresponding to those of Jupiter, Saturn, Uranus and Neptune. The timestep for all simulations is 1.5 days. Note that this timestep is smaller than what one would typically choose for this kind of integration. However, with a 1.5 day timestep we reach machine precision for integrators that use symplectic correctors, allowing us to better quantify the long-term behaviour of [[WHFast]{}]{}. All the effects we discuss here are also present in simulations with longer timesteps, but they would manifest themselves in the relative energy error at a later time. We run four simulations for each integrator and randomly perturb the initial conditions and masses by 0.1% in each simulation. We plot the individual simulations as thin lines, and the average error as a bold line.
The lower relative energy bound is set by machine precision for all integrators, roughly $10^{-16}$. [[WHFast]{}]{}and [[MERCURY]{}]{}, the integrators in our sample that have symplectic correctors, almost reach this limit early on in the simulation. The bound energy error $E_{\rm bound}$ is approximately $10^{-14}$. The integrators without symplectic correctors, [[SWIFTER-WHM]{}]{}and [[WH]{}]{}have an energy error roughly three order of magnitudes higher $E_{\rm bound}\approx 10^{-10.5}$.
From Fig. \[fig:longtermtest\] it is clear that the integrators [[MERCURY]{}]{}, [[WH]{}]{}and [[SWIFTER-WHM]{}]{}show a linear behaviour in the energy error at late times. This is due to the term $E_{\rm bias}$. The $E_{\rm bias}$ term already dominates at early times (after 100 Jupiter orbits) for [[MERCURY]{}]{}because the symplectic correctors lower the value of $E_{\rm bound}$. For [[WH]{}]{}and [[SWIFTER-WHM]{}]{}the $E_{\rm bias}$ term dominates after $10\,000$ Jupiter orbits. This result shows that one or more steps in these integration algorithms are biased. We found that the two main contributions were the inaccurate implementation of the rootfinder for Kepler’s equation and the conversions to and from Jacobi coordinates. In [[WHFast]{}]{}, $E_{\rm bias}$ is absent, showing that its implementation is completely unbiased.
Since all integrators are implemented in double floating-point precision and use the same timestep, they all have roughly the same error term $E_{\rm rand}$. However, it is only visible in Fig. \[fig:longtermtest\] for the [[WHFast]{}]{}integrator. For all other integrators the linearly growing term $E_{\rm bias}$ dominates over $E_{\rm rand}$.
If we increase the timestep, the linear error growth will show up at a later time because $E_{\rm bound}$ will be larger. However, it is still present at all times. Let us think of a symplectic integrator as an exact integrator for a perturbed Hamiltonian $\tilde{\mathcal{H}}$ with high frequency terms added compared to $\mathcal{H}$ in Eq. \[eq:H\], see e.g. [@Wisdom1996]. Then the quantity related to the energy error for $\tilde{\mathcal{H}}$, let us call this $\tilde{E}$, should be conserved exactly at all times (that is the idea of a symplectic integrator). However, if the implementation is biased, $\tilde{E}$ will undergo a linear growth at all times. With [[WHFast]{}]{}, we improve the conservation of $\tilde{E}$ by many orders of magnitude in any integration, regardless of timestep.
This difference could have important implication for the dynamical evolution of the system and could for example push it from a stable to an unstable region of parameter space. We plan to study the effect of different integrators on systems near a chaotic/non-chaotic separatrix in a follow up paper.
Conclusions {#sec:conclusions}
===========
In this paper, we presented [[WHFast]{}]{}, a new implementation of a symplectic Wisdom-Holman integrator. Key advantages and improvements over other publicly available implementations of symplectic integrators are:
*[[WHFast]{}]{}is faster by a factor of $1.5$ to $5$.* Of that, a 50% speedup comes from the improved Kepler solver, where we use a fast convergence criteria for Newton’s method and an efficient implementation of $c$ and $G$-functions. The remainder of the speedup is due to combining drift steps at the end and beginning of each timestep and to only converting to and from Jacobi coordinates when needed.
*The Kepler solver is more accurate and unbiased.* We achieve this thanks to improvements to the convergence criteria in Newton’s method, a Laguerre-Conway solver for highly eccentric orbits with long timesteps, the high accuracy implementations of the $c$ and $G$-functions and a careful ordering of floating-point operations.
*We remove the secular energy error that grows linearly with integration time.* This is due to two improvements. First, the unbiased Kepler solver. Second, the improved and also unbiased coordinate transformations to and from Jacobi coordinates. To our knowledge, [[WHFast]{}]{}is the first publicly available implementation of a Wisdom-Holman integrator that follows Brouwer’s law over long timescales for small enough timesteps and does not show a linear growth in the energy error.
*We implement variational equations that allow us to compute the Lyapunov timescale and the MEGNO.* Our algorithm to calculate the Lyapunov timescale uses a numerically stable algorithm that is based on a one-pass covariance filter. The variational equations do not require us to solve Kepler’s equation and are thus very inexpensive to calculate.
*Symplectic correctors of order 3, 5, 7, and 11 are implemented.* These symplectic corrector allow for high-accuracy simulations of systems with small mass ratios. Even for relatively massive planets like those in the Solar System, symplectic correctors achieve an improvement of three orders of magnitude. For long integrations, the performance cost of symplectic correctors is negligible and so our default setting uses an 11th-order corrector.
*[[WHFast]{}]{}lets the centre-of-mass move freely during an integration.* We integrate an additional degree of freedom in order for our integrator to work in any inertial frame, i.e. one is not restricted to the heliocentric or barycentric frame. Additionally, we do not tie our implementation to a specific choice of units.
*The integrator is available as an easy to use python module.* The module works on both python 2 and 3. It can be installed on most Unix and MacOS systems with a single command:
pip install rebound
The following python script imports the rebound module, adds particles to the simulation, selects an integrator and timestep and runs the integration.
import rebound
rebound.add(m=1)
rebound.add(m=0.001, a=1.)
rebound.add(m=0.001, a=2., e=0.1)
rebound.integrator = 'whfast'
rebound.dt = 0.01
rebound.integrate(6.2831)
More complicated examples and the source code of [[WHFast]{}]{}(written in C, compliant with the C99 standard) can be found in the [[REBOUND]{}]{}package. [[REBOUND]{}]{}includes several other integrators, collision detection algorithms, a gravity tree code and much more. The [[REBOUND]{}]{}git repository is hosted at <https://github.com/hannorein/rebound>.
We also provide an experimental hybrid integrator for simulations in which close encounters occur. The hybrid integrator switches over to a high-order non-symplectic integrator [IAS15, @ReinSpiegel2015] during a close encounter. A detailed discussion of this integrator and its properties will be given in a follow-up paper.
We hope that with the speed and accuracy improvements, [[WHFast]{}]{}will become the go-to integrator package for short and long-term orbit simulations of planetary systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research has been supported by the NSERC Discovery Grant RGPIN-2014-04553. We thank Wayne Enright, Philip Sharp and Scott Tremaine for stimulating discussions and Jack Wisdom for a helpful referee report.
[^1]: But note that we do not need to calculate the momenta explicitly in our algorithm.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the problem of testing if a function depends on a small number of linear directions of its input data. We call a function $f$ a *linear $k$-junta* if it is completely determined by some $k$-dimensional subspace of the input space. In this paper, we study the problem of testing whether a given $n$ variable function $f : {{\mathbb{R}}}^n \to \{0,1\}$, is a linear $k$-junta or $\epsilon$-far from all linear $k$-juntas, where the closeness is measured with respect to the Gaussian measure on ${{\mathbb{R}}}^n$. Linear $k$-juntas are a common generalization of two fundamental classes from Boolean function analysis (both of which have been studied in property testing) **1.** $k$- juntas which are functions on the Boolean cube which depend on at most k of the variables and **2.** intersection of $k$ halfspaces, a fundamental geometric concept class.
We show that the class of linear $k$-juntas is not testable, but adding a surface area constraint makes it testable: we give a $\mathsf{poly}(k \cdot s/\epsilon)$-query non-adaptive tester for linear $k$-juntas with surface area at most $s$. We show that the polynomial dependence on $s$ is necessary. Moreover, we show that if the function is a linear $k$-junta with surface area at most $s$, we give a $(s \cdot k)^{O(k)}$-query non-adaptive algorithm to learn the function *up to a rotation of the basis*. In particular, this implies that we can test the class of intersections of $k$ halfspaces in ${{\mathbb{R}}}^n$ with query complexity independent of $n$.
author:
- 'Anindya De[^1]'
- 'Elchanan Mossel [^2]'
- 'Joe Neeman[^3]'
bibliography:
- 'allrefs.bib'
title: 'Is your function low-dimensional?'
---
Introduction
============
Preliminaries
=============
[^1]: Northwestern. Email: ` [email protected]` Supported by NSF grant CCF 1814706
[^2]: MIT. Email: `[email protected]`. Partially supported by NSF award DMS-1737944 and ONR award N00014-17-1-2598
[^3]: UT Austin. Email: `[email protected]`.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The global pandemic of COVID-19 has infected millions of people since its first outbreak in last December, and the daily new cases are still climbing by hundreds of thousands as of May 2020. A key challenge for preventing and controlling COVID-19 is how to quickly, widely, and effectively implement the test for the disease, because testing is the first step to break the chains of transmission. To assist and speed up the diagnosis of the disease, radiology imaging is used to complement the screening process and triage patients into different risk levels. Deep learning methods have been considered as very powerful tools and have taken a more active role in automatically detecting COVID-19 disease in chest x-ray images, as witnessed in many recent works in the past few weeks. Most of these works first train a Convolutional Neural Network (CNN) on an existing large-scale chest x-ray image dataset and then fine-tune it with a COVID-19 dataset at a much smaller scale. However, direct transfer across datasets from different domains may lead to poor performance due to domain shift, especially on the biomedical datasets which can be collected and preprocessed quite differently from different hospitals. Also, the small scale of the COVID-19 dataset on the target domain can make the training fall into the overfitting trap. To solve all these crucial problems and fully exploit the available large-scale chest x-ray image dataset[@wang2017chestx], we formulate the problem of COVID-19 chest x-ray image classification in a semi-supervised open set domain adaptation setting, through which we are motivated to reduce the domain shift and avoid overfitting when training on a very small dataset of COVID-19. In addressing this formulated problem, we propose a novel emi-supervised pen set omain dversarial network (SODA), which is able to align the data distributions across different domains in a general domain space and also in a common subspace of source and target data. In our experiments, SODA achieves a leading classification performance compared with recent state-of-the-art models, as well as effectively separating COVID-19 with common pneumonia.'
author:
- Jieli Zhou
- Baoyu Jing
- Zeya Wang
bibliography:
- 'sample-base.bib'
title: 'SODA: Detecting Covid-19 in Chest X-rays with Semi-supervised Open Set Domain Adaptation'
---
Introduction
============
Since the Coronavirus disease 2019 (COVID-19) was first declared as a Public Emergency of International Concern (PHEIC) on January 30, 2020[^1], it has quickly evolved from a local outbreak in Wuhan, China to a global pandemic, costing millions of lives and dire economic loss worldwide. In the US, the total COVID-19 cases grew from just one confirmed on Jan 21, 2020 to over 1 million on April 28, 2020 in a span of 3 months. Despite drastic actions like shelter-in-place and contact tracing, the total cases in US kept increasing at an alarming daily rate of 20,000 - 30,000 throughout April, 2020. A key challenge for preventing and controlling COVID-19 right now is the ability to quickly, widely and effectively test for the disease, since testing is usually the first in a series of actions to break the chains of transmission and curb the spread of the disease.
COVID-19 is caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [^2] By far, it is the most reliably diagnosed through Reverse Transcription Polymerase Chain Reaction (RT-PCR) [^3] in which a sample is taken from the back of throat or nose of the patients and tested for viral RNA. While taking samples from the patients, aerosol pathogens could be released and would put the healthcare workers at risk. Furthermore, once the sample is collected, the testing process usually takes several hours and recent study reports that the sensitivity of PT-PCR is around 60-70% [@ai2020correlation], which suggests that many people tested negative for the virus may actually carry it thus could infect more people without knowing it. On the other hand, the sensitivity of chest radiology imaging for COVID-19 was much higher at 97% as reported by [@ai2020correlation; @fang2020sensitivity].
Due to the shortage of viral testing kits, the long period of waiting for results, and low sensitivity rate of RT-PCR, radiology imaging has been used as a complementary screening process to assist the diagnosis of COVID-19 and triage patients into different risk levels. Unlike PT-PCR, imaging is readily available in most healthcare facilities around the world, and the whole process can be done rapidly. In recent years, with the rapid advancement in deep learning and computer vision, many breakthroughs have been developed in using Artificial Intelligence (AI) for medical imaging analysis, especially disease detection [@wang2017chestx; @irvin2019chexpert; @wang2018tienet] and report generation [@jing2017automatic; @li2018hybrid; @jing2019show; @biswal2020clinical], and some AI models achieve expert radiologist-level performance [@lakhani2017deep]. Right now, with most healthcare workers busy at front lines saving lives, the scalability advantage of AI-based medical imaging systems stand out more than ever. Some AI-based chest imaging systems have already been deployed in hospitals to quickly inform healthcare workers to take corresponding actions[^4].
Annotated datasets are required for training AI-based methods, and a small chest x-ray dataset with COVID-19 is collected recently: COVID-ChestXray [@cohen2020covid]. In the last few weeks, several works [@wang2020covidnet; @li2020artificial; @apostolopoulos2020covid; @minaee2020deep] apply Convolutional Neural Networks (CNN) and transfer learning to detect COVID-19 cases from chest x-ray images. They first train a CNN on a large dataset like Chexpert [@irvin2019chexpert] and ChestXray14 [@wang2017chestx], and then fine-tune the model on the small COVID-19 dataset. By far, due to the lack of large-scale open COVID-19 chest x-ray imaging datasets, most works only used a very small amount of positive COVID-19 imaging samples [@cohen2020covid]. While the reported metrics like accuracy and AUC are high, it is likely that these models overfit on this small dataset and may not achieve the reported performance on a larger COVID-19 x-ray dataset. Besides, these methods suffer a lot from label domain shift: these newly trained models lose the ability to detect common thoracic diseases like “Effusion” and “Nodule” since these labels do not appear in the new dataset. Moreover, they also ignored the visual domain shift between the two datasets. On the one hand, the large-scale datasets like ChestXray14 [@wang2017chestx] and Chexpert [@irvin2019chexpert] are collected from top U.S. health institutes like National Institutes of Health (NIH) clinical center and Stanford University, which are well-annotated and carefully processed. On the other hand, COVID-ChestXray [@cohen2020covid] is collected from a very diverse set of hospitals around the world and they are of very different qualities and follow different standards, such as the viewpoints, aspect ratios and lighting, etc. In addition, COVID-ChestXray contains not only chest x-ray images but also CT scan images.
In order to fully exploit the limited but valuable annotated COVID-19 chest x-ray images and the large-scale chest x-ray image dataset at hand, as well as prevent the above-mentioned drawbacks of those fine-tuning based methods, we define the problem of learning a classifier for COVID-19 from the perspective of open set domain adaptation (Definition \[def:uoda\]) [@panareda2017open]. Different from traditional unsupervised domain adaptation which requires the label set of both source and target domain to be the same, the open set domain adaptation allows different domains to have different label sets. This is more suitable for our problem because COVID-19 is a new disease which is not included in the ChestXray14 or Chexpert dataset. However, since our task is to train a new classifier for COVID-19 dataset, we have to use some annotated samples. Therefore, we further propose to view the problem as a Semi-Supervised Open Set Domain Adaptation problem (Definition \[def:soda\]).
Under the given problem setting, we propose a novel emi-supervised pen set omain dversarial network (SODA) comprised of four major components: a feature extractor $G_f$, a multi-label classifier $G_y$, domain discriminators $D_g$ and $D_c$, as well as common label recognizer $R$. SODA learns the domain-invariant features by a two-level alignment, namely, domain level and common label level. The general domain discriminator $D_g$ is responsible for guiding the feature extractor $G_f$ to extract domain-invariant features. However, it has been argued that the general domain discriminator $D_g$ might lead to false alignment and even negative transfer [@pei2018multi; @wang2019adversarial]. For example, it is possible that the feature extractor $G_f$ maps images with “Pneumonia” in the target domain and images with “Cardiomegaly” in the source domain into similar positions, which might result in the miss-classification of $G_y$. In order to solve this problem, we propose a novel common label discriminator $D_c$ to guide the model to align images with common labels across domains. For labeled images, $D_c$ only activates when the input image is associated with a common label. For unlabeled images, we propose a common label recognizer $R$ to predict their probabilities of having a common label.
The main contributions of the paper are summarized as follows:
- To the best of our knowledge, we are the first to tackle the problem of COVID-19 chest x-ray image classification from the perspective of domain adaptation.
- We formulate the problem in a novel semi-supervised open set domain adaptation setting.
- We propose a novel two-level alignment model: emi-supervised pen set omain dversarial network (SODA).
- We present a comprehensive evaluation to demonstrate the effectiveness of the proposed SODA.
Preliminary
===========
Problem Definition
------------------
\[def:uoda\] *Unsupervised Open Set Domain Adaptation*
Let $\mathcal{D}^{s}=\{(\mathbf{x}_n^{s}, y^{s}_n)\}_{n=1}^{N^s}$ be a source domain with $N^s$ labeled samples, and $\mathcal{D}^{t}=\{(\mathbf{x}_n^{t})\}_{n=1}^{N^t}$ be a target domain with $N^t$ unlabeled samples, where the underlying label set $\mathcal{L}^{t}$ of the target domain might be different from the label set $\mathcal{L}^{s}$ of the source domain. Define $\mathcal{L}^c = \mathcal{L}^s\cap\mathcal{L}^t$ as the *set of common labels* shared across different domains, $\bar{\mathcal{L}}^s=\mathcal{L}^s\textbackslash \mathcal{L}^c$ and $\bar{\mathcal{L}}^t=\mathcal{L}^t\textbackslash \mathcal{L}^c$ be *sets of domain-specific labels* which only appear in the source and the target domain respectively.
The task of *Unsupervised Open Set Domain Adaptation* is to build a model which could accurately assign common labels in $\mathcal{L}^c$ to samples $\mathbf{x}^t_n$ in the target domain, as well as distinguish those $\mathbf{x}^t_n$ belonging to $\bar{\mathcal{L}}^t$.
\[def:soda\] *Semi-supervised Universal Domain Adaptation*
Given a source domain $\mathcal{D}^{s}=\{(\mathbf{x}_n^{s}, y^{s}_n)\}_{n=1}^{N^s}$ with $N^s$ labeled samples, and a target domain $\mathcal{D}^{t}\cup\mathcal{D}^{t'}$ consisting of $\mathcal{D}^{t}=\{(\mathbf{x}_n^{t})\}_{n=1}^{N^t}$ with $N^t$ unlabeled samples and $\mathcal{D}^{t'}=\{(\mathbf{x}_n^{t'}, y_n^{t'})\}_{n=1}^{N^{t'}}$ with $N^{t'}$ labeled samples.
The task of *Semi-supervised Open Set Domain Adaptation* is to build a model to assign labels from $\mathcal{L}^t$ to unlabeled samples in $\mathcal{D}^{t}$.
Notations
---------
We summarize the symbols used in the paper and their descriptions in Table \[tab:notations\].
Symbols Description
----------------------- ------------------------------------------------------------------
$\mathcal{D}^{s}$ set of labeled samples in the source domain
$\mathcal{D}^{t}$ set of unlabeled samples in the target domain
$\mathcal{D}^{t'}$ set of labeled samples in the target domain
$\mathcal{L}^s$ set of labels for the source domain
$\mathcal{L}^t$ set of labels for the target domain
$\mathcal{L}^c$ set of common labels across domains
$\bar{\mathcal{L}}^s$ set of domain-specific labels in the source domain
$\bar{\mathcal{L}}^t$ set of domain-specific labels in the target domain
$\mathcal{L}$ set of all labels from all domains
$N^s$ number of labeled samples in the source domain
$N^t$ number of unlabeled samples in the target domain
$N^{t'}$ number of labeled samples in the target domain
$G_f$ feature extractor
$G_y$ multi-label classifier for $\mathcal{L}$
$G_{y_l}$ binary classifier for label $l$ (part of $G_y$)
$R$ common label recognizer
$D_{c}$ domain discriminator for common labels $\mathcal{L}^c$
$D_{g}$ general domain discriminator
$L_{G_y}$ loss of multi-label classification over the entire dataset
$L_R$ loss of $R$ over the entire dataset
$L_{D_g}$ loss of $D_g$ over the entire dataset
$L_{D_c}$ loss of $D_c$ over the entire dataset
$\lambda$ the coefficient of losses
$\mathbf{x}$ input image
$\mathbf{h}$ hidden features
$y$ ground-truth label
$\hat{y}$ predicted probability
$\hat{d}$ predicted probability that $\mathbf{x}$ belongs to source domain
$\hat{r}$ predicted probability that $\mathbf{x}$ has common labels
: Notations
\[tab:notations\]
Methodology
===========
Overview
--------
An overview of the proposed emi-supervised pen Set omain dversarial network (SODA) is shown in Fig. \[fig:model\]. Given an input image $\mathbf{x}$, it will be first fed into a feature extractor $G_f$, which is a Convolutional Neural Network (CNN), to obtain its hidden feature $\mathbf{h}$ (green part). The binary classifier $G_{y_l}$ (part of the multi-label classifier $G_y$) takes $\mathbf{h}$ as input, and will predict the probability $\hat{y}_l$ for the label $l\in\mathcal{L}$ (blue part).
We propose a novel two-level alignment strategy for extracting the domain invariant features across the source and target domain. On the one hand, we perform *domain alignment* (Section \[sec:align\_domain\]), which leverages a general domain discriminator $D_g$ to minimize the domain-level feature discrepancy. On the other hand, we emphasize the *alignment of common labels* $\mathcal{L}^c$ (Section \[sec:align\_common\]) by introducing another domain discriminator $D_c$ for images associated with common labels. For labeled images in $\mathcal{D}^s$ and $\mathcal{D}^{t'}$, we compute loss for $D_c$ and conduct back-propagation only if the input image $\mathbf{x}$ is associated with a common label $l\in\mathcal{L}^c$. As for unlabeled data in $\mathcal{D}^{t}$, we propose a common label recognizer $R$ to predict the probability $\hat{r}$ that an image $\mathbf{x}$ has a common label, and use $\hat{r}$ as a weight in the losses of $D_c$ and $D_g$.
Domain Alignment {#sec:align_domain}
----------------
Domain adversarial training [@ganin2016domain] is the most popular method for helping feature extractor $G_f$ learn domain-invariant features such that the model trained on the source domain can be easily applied to the target domain. The objective function of the domain discriminator $D_g$ can be written as: $$\label{eq:loss_D_g}
\begin{split}
L_{D_g} =& -\mathbb{E}_{(\mathbf{x}^s\in\mathcal{D}^s)}[\log\hat{d}_g] \\
&-\mathbb{E}_{(\mathbf{x}^t\in\mathcal{D}^t\cup\mathcal{D}^{t'})}[\log(1 -\hat{{d}}_g)]
\end{split}$$ where $\hat{{d}}_g$ denotes the predicted probability that the input image belongs to the source domain.
In SODA, we use a Multi-Layer Perceptron (MLP) as the general domain discriminator $D_g$.
{width=".95\textwidth"}
Common Label Alignment {#sec:align_common}
----------------------
In the field of adversarial domain adaptation, most of the existing methods only leverage a general domain discriminator $D_g$ to minimize the discrepancy between the source and target domain. Such a practice ignores the label structure across domains, which will result in false alignment and even negative transfer [@pei2018multi; @wang2019adversarial]. If we only use a general domain discriminator $D_g$ in the open set domain adaptation setting (Definition \[def:uoda\] and Definition \[def:soda\]), it is possible that the feature extractor $G_f$ will map the target domain images with a common label $l\in\mathcal{L}^c$, say “Pneumonia”, and the source domain images with a specific label $l\in\bar{\mathcal{L}}^s$, “Cardiomegaly”, to similar positions in the hidden space, which might lead to the classifier miss-classifying a “Pneumonia” image in the target domain as “Cardiomegaly”.
To address the problem of the miss-matching between the common and specific label sets, we propose a domain discriminator $D_c$ to distinguish the domains for the images with a common label. For the labeled data from the source domain $\mathcal{D}^s$ and the target domain $\mathcal{D}^t$, we know whether an image $\mathbf{x}$ has a common label or not, and we only calculate the loss $L_{d_c}$ for $D_c$ on the samples with common labels: $$\label{eq:loss_D_c_label}
\begin{split}
L_{D_c}^{label} =& -\mathbb{E}_{(\mathbf{x}^s\in\mathcal{D}^s, y^s\in\mathcal{L}^c)}[\log\hat{d}_c] \\
&-\mathbb{E}_{(\mathbf{x}^t\in\mathcal{D}^{t'}, y^{t'}\in\mathcal{L}^c)}[\log(1 -\hat{{d}}_c)]
\end{split}$$ where $\hat{d}_c$ denotes the predicted probability that the input images is associated with a common label.
However, a large number of images in the target domain are unlabeled, and thus extra effort is required for determining whether an unlabeled image is associated with a common label. To address this problem, we propose a novel common label recognizer $R$ to predict the probability $\hat{r}$ whether an unlabeled image has at least one common label. The probability $\hat{r}$ will be used as a weight in the loss function of $D_c$[^5]:
$$\label{eq:loss_D_c_un}
L_{D_c}^{un} = -\mathbb{E}_{(\mathbf{x}^t\in\mathcal{D}^{t}, y^{t}\in\mathcal{L}^c)}[\hat{r}\log(1 -\hat{{d}}_c)]$$
In addition, we also use $\hat{r}$ to re-weigh unlabeled samples in $D_g$ (Equation \[eq:loss\_D\_g\]) to further emphasize the alignment of common labels:
$$\label{eq:loss_D_g_2}
\begin{split}
L_{D_g} =& -\mathbb{E}_{(\mathbf{x}^s\in\mathcal{D}^s)}[\log\hat{d}_g] \\
& -\mathbb{E}_{(\mathbf{x}^t\in\mathcal{D}^{t'})}[\log(1 -\hat{{d}}_g)] \\
& -\mathbb{E}_{(\mathbf{x}^t\in\mathcal{D}^t)}[\hat{r}\log(1 -\hat{{d}}_g)]
\end{split}$$
Finally, the recognizer $R$ is trained on the labeled set $\mathcal{D}^s\cup\mathcal{L}^{t'}$ via cross-entropy loss: $$\label{eq:loss_R}
\begin{split}
L_{R} =& -\mathbb{E}_{(\mathbf{x}\in\mathcal{D}^s\cup\mathcal{D}^{t'}, y\in\mathcal{L}^c)}[\log\hat{r}] \\
&-\mathbb{E}_{(\mathbf{x}\in\mathcal{D}^s\cup\mathcal{D}^{t'}, y\notin\mathcal{L}^c)}[\log(1 -\hat{{r}})]
\end{split}$$
Overall Objective Function
--------------------------
The overall objective function of SODA can be written as a min-max game between classifiers $G_y$, $R$ and discriminators $D_g$, $D_c$:
$$\min_{G_y, R}\max_{D_g, D_c} L_{G_y} + \lambda_R L_{R} - \lambda_{D_g}L_{D_g} - \lambda_{D_c}^{label}L_{D_c}^{label} - \lambda_{D_c}^{un}L_{D_c}^{un}$$
where $L_R$, $L_{D_g}$, $L_{D_c}^{label}$ and $L_{D_c}^{un}$ are respectively defined in Equation \[eq:loss\_R\], \[eq:loss\_D\_g\_2\], \[eq:loss\_D\_c\_label\] and \[eq:loss\_D\_c\_un\]; $L_{G_y}$ denotes the cross-entropy loss for multi-label classification; $\lambda$ denotes the coefficient for each loss function.
Experiments
===========
Experiment Setup
----------------
### Datasets
#### Source Domain
We use ChestXray-14 [@wang2017chestx] as the source domain dataset. This dataset is comprised of 112,120 anonymized chest x-ray images from the National Institutes of Health (NIH) clinical center. The dataset contains 14 common thoracic disease labels: “Atelectasis”, “Consolidation”, “Infiltration”, “Pneumothorax”, “Edema”, “Emphysema”, “Fibrosis”, “Effusion”, “Pneumonia”, “Pleural thickening”, “Cardiomegaly”, “Nodule”, “Mass” and “Hernia”.
#### Target Domain
The newly collected COVID-ChestXray [@cohen2020covid] is adopted as the target domain dataset. This dataset contains images collected from various public sources and different hospitals around the world. This dataset (by the time of this writing) contains 328 chest x-ray images in which 253 are labeled positive as the new disease “COVID-19”, whereas 61 are labeled as other well-studied “Pneumonia”.
### Evaluation Metrics
We evaluate our model from four different perspectives. First, to test the classification performance, following the semi-supervised protocol, we randomly split the 328 x-ray images in COVID-ChestXray into 40% labeled set, and 60% unlabeled set. We run each model 3 times and report the average AUC-ROC score. Second, we compute the Proxy-$\mathcal{A}$ Distance (PAD) [@ben2007analysis] to evaluate models’ ability for minimizing the feature discrepancy across domains. Thirdly, we use t-SNE to visualize the feature distributions of the target domain. Finally, we also qualitatively evaluate the models by visualizing their saliency maps.
### Baseline Methods
We compare the proposed SODA with two types of baselines methods: fine-tuning based transfer learning models and domain adaptation models. For fine-tuning based models, we select the two most popular CNN models DenseNet121 [@huang2017densely] and ResNet50 [@he2016deep] as our baselines. These models are first trained on the ChestXray-14 dataset and then fine-tuned on the COVID-ChestXray dataset. As for the domain adaptation models, we compare our model with two classic models, Domain Adversarial Neural Networks (DANN) [@ganin2016domain] and Partial Adversarial Domain Adaptation (PADA) [@cao2018partial]. Note that DANN and PADA were designed for unsupervised domain adaptation, and we implement a semi-supervised version of them.
### Implementation Details
We use DenseNet121 [@huang2017densely], which is pretrained on the ChestXray-14 dataset [@wang2017chestx], as the feature extractor $G_f$ for SODA. The multi-label classifier $G_y$ is a one layer neural network and its activation is the sigmoid function. We use the same architecture for $D_g$, $D_c$ and $R$: a MLP containing two hidden layers with ReLU [@nair2010rectified] activation and an output layer. The hidden dimension for all of the modules: $G_y$, $D_g$, $D_c$ and $R$ is 1024. For fair comparison, we use the same setting of $G_f$, $G_y$ and $D_g$ for DANN [@ganin2016domain] and PADA [@cao2018partial]. All of the models are trained by Adam optimizer [@kingma2014adam], and the learning rate is $10^{-4}$.
Classification Results
----------------------
To investigate the effects of domain adaptation and demonstrate the performance improvement of the proposed SODA, we present the average AUC-ROC scores for all models in Table \[tab:results\]. Comparing the results for ResNet50 and DenseNet121, we observe that deeper and more complex models achieve better classification performance. For the effects of domain adaptation, it is obvious that the domain adaptation methods (DANN, PADA, and SODA) outperform those fine-tuning based transfer learning methods (ResNet50 and DenseNet121). Furthermore, the proposed SODA achieves higher AUC scores on both COVID-19 and Pneumonia than DANN and PADA, demonstrating the effectiveness of the proposed two-level alignment.
Model COVID-19 Pneumonia
--------------------------------- ------------ ------------
ResNet50 [@he2016deep] 0.8143 0.8342
DenseNet121 [@huang2017densely] 0.8202 0.8414
DANN [@ganin2016domain] 0.8785 0.8961
PADA [@cao2018partial] 0.8822 0.9038
SODA **0.9006** **0.9082**
: Target Domain Average AUC-ROC Score
\[tab:results\]
Proxy $\mathcal{A}$-Distance
----------------------------
Proxy $\mathcal{A}$-Distance [@ben2007analysis] has been widely used in domain adaptation for measuring the feature distribution discrepancy between the source and target domains. PAD is defined by $$d_{\mathcal{A}}=2(1-2\min(\epsilon))$$ where $\epsilon$ is the domain classification error (e.g. mean absolute error) of a classifier (e.g. linear SVM [@cortes1995support]).
Following [@ganin2016domain], we train SVM models with different $C$ and use the minimum error to calculate PAD. In general, a lower $d_\mathcal{A}$ means a better ability for extracting domain invariant features. As shown in Fig. \[fig:pad\], SODA has a lower PAD compared with the baseline methods, which indicates the effectiveness of the proposed two-level alignment strategy.
![Proxy $\mathcal{A}$-Distance[]{data-label="fig:pad"}](pad.png){width="40.00000%"}
{width=".9\textwidth"}
Feature Visualization
---------------------
We use t-SNE to project the high dimensional hidden features $\mathbf{h}$ extracted by DANN, PADA, and SODA to low dimensional space. The 2-dimensional visualization of the features in the target domain is presented in Fig. \[fig:tsne\], where the red data points are image features of “Pneumonia” and the blue data points are image features of “COVID-19”. It can be observed from Fig. \[fig:tsne\] that SODA performs the best for separating “COVID-19” from “Pneumonia”, which demonstrates the effectiveness of the proposed common label recognizer $R$ as well as the domain discriminator for common labels $D_c$.
Grad-CAM
--------
Grad-CAM [@selvaraju2017grad] is used to visualize the features extracted from all compared models. Fig. \[fig:gradcam\] shows the Grad-CAM results on seven different COVID-19 positive chest x-rays. These seven images have annotations (small arrows and box) indicating the pathology locations. We observe that ResNet50 and DenseNet121 can focus wrongly on irrelevant locations like the dark corners and edges. In contrast, domain adaptation models have better localization in general, and our SODA model gives more focused and accurate pathological locations than other models compared. In addition, we consult a professional radiologist with over 15 years of clinical experience from Wuxi People’s Hospital and received positive feedback on the pathological locations as indicated by the Grad-CAM of SODA. In the future, we plan to do a more rigorous evaluation study with more inputs from radiologists. We believe the features extracted from SODA can assist radiologists to pinpoint the suspect COVID-19 pathological locations faster and more accurately.
Related Work
============
Domain Adaptation
-----------------
Domain adaptation is an important application of transfer learning that attempts to generalize the models from source domains to unseen target domains [@ganin2015unsupervised; @ganin2016domain; @tzeng2017adversarial; @tzeng2014deep; @you2019universal; @jing2018cross; @wang2020coarse]. Deep domain adaptation approaches are usually implemented through discrepancy minimization [@tzeng2014deep] or adversarial training [@ganin2015unsupervised; @ganin2016domain; @tzeng2017adversarial]. Adversarial training, inspired by the success of generative adversarial modeling [@goodfellow2014generative], has been widely applied for promoting the learning of transfer features in image classification. It takes advantage of a domain discriminator to classify whether an image is from the source or target domains. On top of these methods, a couple of works have been presented for exploring the high-level structure in the label space, which aim at further improving the domain adaptation performance for multi-class image classification [@wang2019adversarial] or fundamentally solving the application problem when the label sets from source domains and target domains are different [@you2019universal]. In order to meet the latter target, more and more researchers have started to study the open set domain adaptation problem, in which case the target domain has images that do not come from the classes in the source domain [@you2019universal; @panareda2017open]. Universal domain adaptation is the latest method that is proposed through using an adversarial domain discriminator and a non-adversarial domain discriminator to successfully solve this problem. [@you2019universal]. Although domain adaptation has been well explored, its application in medical imaging analysis, such as domain adaptation for chest x-ray images, is still under-explored.
Semi-supervised Learning
------------------------
Semi-supervised learning is a very important task for image classification, which can make use of both labeled and unlabeled data at the same time [@saito2019semi]. Recently it has been used to solve image classification problems on a very large (1 billion) set of unlabelled images [@yalniz2019billion]. In spite of many progresses that have been made with unsupervised domain adaptation methods, the domain adaptation with semi-supervised learning has not yet been fully explored.
Chest X-Ray Image Analysis
--------------------------
There has been substantial progress in constructing publicly available databases for chest x-ray images as well as a related line of works to identify lung diseases using these images. The largest public datasets of chest x-ray images are Chexpert [@irvin2019chexpert] and ChestXray14 [@wang2017chestx], which respectively include more than 200,000 and 100,000 chest x-ray images collected by Stanford University and National Institute of Healthcare. The creation of these datasets have also motivated and promoted the multi-label chest x-ray classification for helping the screening and diagnosis of various lung diseases. The problems of disease detection [@wang2017chestx; @irvin2019chexpert; @wang2018tienet] and report generation using chest x-rays [@jing2017automatic; @li2018hybrid; @jing2019show; @biswal2020clinical] are fully investigated and have achieved much-improved results upon recently. However, there have been very few attempts for studying the domain adaptation problems with the multi-label image classification problem using chest x-rays.
Conclusion
==========
In this paper, in order to assist and complement the screening and diagnosing of COVID-19, we formulate the problem of COVID-19 chest x-ray image classification in a semi-supervised open set domain adaptation framework. Accordingly, we propose a novel deep domain adversarial neural network, emi-supervised pen set omain dversarial network (SODA), which is able to align the data distributions across different domains at both domain level and common label level. Through evaluations of the classification accuracy, we show that SODA achieves better AUC-ROC scores than the recent state-of-the-art models. We further demonstrate that the features extracted by SODA is more tightly related to the lung pathology locations, and get initial positive feedback from an experienced radiologist. In practice, SODA can be generalized to any semi-supervised open set domain adaptation settings where there are a large well-annotated dataset and a small newly available dataset. In conclusion, SODA can serve as a pilot study in using techniques and methods from domain adaptation to radiology imaging classification problems.
[^1]: <https://www.statnews.com/2020/01/30/who-declares-coronavirus-outbreak-a-global-health-emergency/>
[^2]: <https://www.who.int/emergencies/diseases/novel-coronavirus-2019/technical-guidance/naming-the-coronavirus-disease-(covid-2019)-and-the-virus-that-causes-it>
[^3]: <https://spectrum.ieee.org/the-human-os/biomedical/diagnostics/how-do-coronavirus-tests-work>
[^4]: <https://spectrum.ieee.org/the-human-os/biomedical/imaging/hospitals-deploy-ai-tools-detect-covid19-chest-scans>
[^5]: Note that gradients stop at $\hat{r}$ in the training period.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we propose a generalization of deep neural networks called deep function machines (DFMs). DFMs act on vector spaces of arbitrary (possibly infinite) dimension and we show that a family of DFMs are invariant to the dimension of input data; that is, the parameterization of the model does not directly hinge on the quality of the input (eg. high resolution images). Using this generalization we provide a new theory of universal approximation of bounded non-linear operators between function spaces. We then suggest that DFMs provide an expressive framework for designing new neural network layer types with topological considerations in mind. Finally, we introduce a novel architecture, RippLeNet, for resolution invariant computer vision, which empirically achieves state of the art invariance.'
author:
- |
William H. Guss\
Machine Learning at Berkeley\
University of California, Berkeley\
`[email protected]`
bibliography:
- 'dfm.bib'
title: 'Deep Function Machines: Generalized Neural Networks for Topological Layer Expression'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We continue our study of the knot Floer homology invariants of cable knots. For large $|n|$, we prove that many of the filtered subcomplexes in the knot Floer homology filtration associated to the $(p,pn+1)$ cable of a knot, $K$, are isomorphic to those of $K$. This result allows us to obtain information about the behavior of the concordance invariant under cabling, which has geometric consequences for the cabling operation. Applications considered include quasipositivity in the braid group, the knot theory of complex curves, smooth concordance, and lens space (or, more generally, L-space) surgeries.'
address: 'Department of Mathematics, Massachusetts Institute of Technology, MA '
author:
- Matthew Hedden
bibliography:
- 'mybib.bib'
---
Introduction
============
A powerful knot invariant was introduced by in [@Knots] and independently by Rasmussen in his thesis, [@RasThesis]. The invariant takes the form of the filtered chain homotopy type of a filtered chain complex. The chain complex is the so-called “infinity" chain complex associated to a $\SpinC$ three-manifold, $\CFinf(Y,\spinc)$, and the filtration of this chain complex is induced by a knot $K\subset Y$. Definitions of the chain complex can be found in [@HolDisk] and the filtration induced by the knot is defined in [@Knots; @RasThesis]. Throughout, we will work with null-homologous knots equipped with a fixed Seifert surface, $F$, though more general constructions are possible [@RationalSurgeries]. This paper, and its predecessor [@Cabling], study the knot Floer filtration induced by a class of knots called cable knots.
Let $K$ be a knot. Recall that the $(p,q)$ cable of $K$, denoted $K_{p,q}$, is a satellite knot with pattern the $(p,q)$ torus knot, $T_{p,q}$. More precisely, $K_{p,q}$ is the image of a torus knot living on the boundary of a tubular neighborhood of $K$.[^1] Thus $p$ is the number of times $K_{p,q}$ traverses the longitudinal direction of $K$, and $q$ the meridional number. Throughout, we will assume $p>0$.[^2] Our original motivation for studying cable knots and, more generally, satellite knots, lay in the fact that their complements decompose as the union of two three-manifolds joined along a torus, and hence provide a testing ground for the topological quantum field theoretic (TQFT) behavior of the Heegaard Floer invariants in $(2+1)$ dimensions.
The knot Floer filtration is, in fact, a $\Z\oplus\Z$ filtration of $\CFinf(Y,\spinc)$, and it is the filtered chain homotopy type of this $\Z\oplus\Z$ filtration which is the primary knot invariant coming from Heegaard Floer homology. The existence of two independent $\Z$ filtrations allows one to define many auxilary knot invariants and in this paper we deal with a less robust invariant - the filtered chain homotopy type of the $\Z$-filtration of $\CFa(Y,\spinc)$ obtained by setting one of the $\Z$ filtrations equal $0$. We denote this filtration by $\FiltY(K)$ so that we have the sequence of inclusions: $$0=\FiltY(K,-i) \subseteq \FiltY(K,-i+1)\subseteq \ldots \subseteq
\FiltY(K,n)=\CFa(Y,\spinc).$$
The associated graded complexes of this filtration, $\frac{\FiltY(K,j)}{\FiltY(K,j-1)}$, will be denoted by $\CFKa_\spinc(Y,K, j)$, and their homology by $\HFKa_\spinc(Y,K, j)$. The homology groups $\HFKa_\spinc(Y,K,j)$ are commonly referred to as the [*knot Floer homology groups of $K\subset Y$*]{}. These groups were studied for the $(p,pn\pm1)$ cables of an arbitrary knot, $K\subset S^3$, in [@Cabling; @MyThesis]. In that paper a stabilization theorem was proved which provided a formula for the groups $\HFKa(S^3,K_{p,pn\pm1},i)$ in the case when the parameter $n$ was sufficiently large. The formula required $H_*(\Filt(K))$ as input, while the output was merely the associated graded object; hence, there was a loss of information. Despite this loss of information the formulas and techniques of [@Cabling] have proved to be quite useful and, in particular, were implemented by [@OSThurston] and Ni [@NiThurston] in the proof that link Floer homology detects the Thurston norm.
The purpose of this paper is to extend our knowledge of the filtered chain homotopy type of $\Filt(K_{p,pn\pm1})$ beyond the level of its associated graded object. A primary motivation for this extension comes from the relationship between the filtered chain homotopy type of $\Filt(K)$ and the smooth four-ball genus of $K$, $g_4(K)$. [@FourBall] and Rasmussen [@RasThesis] define the following numerical invariant of a knot, $K\subset S^3$: $$\tau(K)=\mathrm{min}\{i\in\Z|H_*(\Filt(K,i))\longrightarrow \HFa(S^3)\ \mathrm{is \ non\-trivial}\}.$$
It is shown that this invariant provides a lower bound for the four-ball genus, $$|\tau(K)|\le g_4(K),$$ and that $\tau$ provides a homomorphism from the the smooth concordance group of knots, $\mathcal{C}$, to $\Z$. Moreover, the above inequality is sharp for torus knots, providing a new proof of Milnor’s famous conjecture [@Milnor1968] on the four-genera and unknotting numbers of this family.
The main result of this paper is the following theorem which, for simplicity, we state for knots in the three-sphere.
\[thm:filt\] Let $K\subset S^3$ be a knot. Pick any $M\in \Z$. Then there exists a constant $N>0$ so that $\forall \ n>N$, the following holds for each $j> M$: $$H_*(\Filt(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-1))\cong H_*(\Filt(K,j-1)).$$ Furthermore, $$\begin{array}{ll}
H_*(\Filt(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-i))\cong & \\
H_*(\Filt(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-i-1)) & \forall \ i=2,\ldots,p-1. \\
\end{array}$$ In particular, $$\tau(K_{p,pn+1})=\left\{\begin{array}{ll}
p \tau(K)+\frac{(pn)(p-1)}{2}+p-1 & {\text{or}} \\
p \tau(K)+\frac{(pn)(p-1)}{2}. & \\
\end{array}
\right.$$
The theorem has an analogue for $n<0$ (stated in Section \[sec:proof\]) which we use with the above to prove:
\[thm:tau\]
Let $K\subset S^3$ be a non-trivial knot, then the following inequality holds for all $n$, $$p \tau(K)+ \frac{(pn)(p-1)}{2}\ \ \le\tau(K_{p,pn+1})\le \ \
p\tau(K)+\frac{(pn)(p-1)}{2} +p-1.$$
In the special case when $K$ satisfies $\tau(K)=g(K)$ we have the equality, $$\tau(K_{p,pn+1}) = p \tau(K)+\frac{(pn)(p-1)}{2},$$ whereas when $\tau(K)=-g(K)$ we have $$\tau(K_{p,pn+1}) = p \tau(K)+\frac{(pn)(p-1)}{2} + p-1.$$
Here, and throughout, $g(K)$ denotes the Seifert genus of $K$. We emphasize that while Theorem \[thm:filt\] requires the cabling parameter $n$ to be sufficiently large, there is no restriction on $n$ in the statement of Theorem \[thm:tau\].
Recently, we generalized $\tau(K)$ to a sequence of invariants of a knot $K\subset Y$ in an arbitrary $3$-manifold [@tbbounds]. For each non-vanishing Floer homology class, $\alpha\in \HFa(Y,\spinc)$ we obtain an integer, $\tau_\alpha(Y,K)$. Theorem \[thm:filt\] is a special case of Theorem \[thm:filtY\], found in the next section. This latter theorem holds for knots in arbitrary manifolds and we use this more general result to obtain information for each $\tau_\alpha(Y,K)$.
Of particular interest is the case when $\alpha$ is the contact invariant, $c(\xi)\in \HFa(-Y)$. In this case $\tau_\alpha(Y,K)$ provides upper bounds for the classical framing invariants of Legendrian and transverse representatives of $K$ in the contact structure $\xi$. Indeed, using our present results, Theorem $1.4$ of [@tbbounds] was able to provide the first systematic construction of prime knot types in many tight contact structures whose classical framing invariants are constrained to be arbitrarily negative.
Geometric Consequences
----------------------
In addition to the theory of Legendrian knots and the connection with the four-ball genus, the Floer invariants of cable knots can be used in several other contexts. We take some time to discuss these results.
### Concordance information
It is straightforward to see that cabling induces a well-defined operation on the smooth concordance group, $\mathcal{C}$. Indeed, if $K$ and $J$ are concordant, their $(p,q)$ cables will be concordant via a concordance which “follows along" the concordance between $K$ and $J$ (see, for instance, [@Kawauchi1980] for more details). Thus cabling defines a sequence of maps: $$\phi_{p,q}: \mathcal{C}\rightarrow \mathcal{C},$$ where $\phi_{p,q}([K])=[K_{p,q}]$ (here $[K]$ denotes the smooth concordance class of $K$). By pre-composing with $\phi_{p,q}$, it follows that any smooth concordance invariant provides a sequence of smooth concordance invariants. In the present context, we obtain functions $$\ctau{n}:= \tau \circ \phi_{p,pn+1}([K]) = \tau(K_{p,pn+1}).$$
Theorem \[thm:tau\] shows that $\ctau{n}$ is a piecewise linear function of $n$. Indeed, the graph of $\ctau{n}$ lies entirely on the lines of slope $\frac{(p-1)(p)}{2}$ whose $y$ intercepts range between $p\tau(K)$ and $p\tau(K)+p-1$. Moreover, a crossing change inequality for $\tau$ (Equation ) shows that $\ctau{n}$ is monotonically decreasing. This indicates that $\ctau{n}$ has a finite set of discontinuities: $$J_\tau^p(K)= \{ n\in \Z | \ \ctau{n} \ne \ctau{n+1}-\frac{(p-1)(p)}{2} \}.$$ The cardinality of this “jumping locus" is at most $p-1$ and may be zero. For instance, Theorem \[thm:tau\] indicates $J_p(K)=\emptyset$ if $\tau(K)=\pm g(K)\ne 0$. On the other hand, for the unknot we have $J_p(\text{unknot})=\{-1\}$.
One should compare this with recent work of Van Cott [@VanCott2008], which uses formal properties of $\tau$ to reprove the fact that $\ctau{n}$ is bounded between two lines of slope $\frac{(p-1)p}{2}$. Her results extend to $(p,q)$ cables, but are unable to recover the possible $y$ intercepts of the lines which bound the graph of $\ctau{n}$, showing only that they differ by $p-1$.
In light of this, one might hope that formal properties of $\tau$ could be pushed further to reprove Theorem \[thm:tau\] without having to understand the Floer chain complexes. This seems unlikely, due to the fact that the techniques of [@VanCott2008] can also be employed in the study of the Rasmussen concordance invariant, $s(K)$. This latter invariant is defined using Khovanov homology [@RasSlice], and while it shares several important properties of $\tau$ the two invariants are independent [@Stau]. Indeed, we expect the behavior of the corresponding functions $\cs{n}$ to be quite different from that of $\ctau{n}$. Our intuition comes from the fact that the Alexander polynomial of cable knots is determined by the formula $$\label{eq:satpoly}
\Delta_{K_{p,q}}(t)=\Delta_{T_{p,q}}(t)\cm\Delta_K(t^p).$$ On the other hand, there cannot exist a formula which computes the Jones polynomial of cables of $K$ in terms of the Jones polynomial of $K$. Since knot Floer homology and Khovanov homology categorify the Alexander and Jones polynomial, respectively, we expect any invariant derived from these theories, e.g. $\tau$ and $s$, to have quite different behavior under cabling. For this reason, we expect invariants derived from $s$ of cables to be very interesting new concordance invariants, and a pursuit of effective means of computation is well-motivated. In fact, one can ask:
Does the Rasmussen concordance invariant, applied to all iterated cables of $K$, determine if $K$ is smoothly slice?
This question has many variants obtained by using other satellite operations or asking for more refined concordance information (see, for instance, [@Doubling; @LN2008]). One can also ask the question where $s(K)$ is replaced by $\tau(K)$ (or any other smooth concordance invariant). The results of the present paper, together with further expected TQFT properties of Floer homology lead us to conjecture that in this case the answer is “no".
In a related direction, we point out that the $(p,1)$ cabling operation $\phi_{p,1}$ was studied in detail by Kawauchi in [@Kawauchi1980]. He showed that $\phi_{p,1}$ is a homomorphism from the algebraic concordance group to itself, and remarked that it does not appear to be so on the level of the (smooth) concordance group in general (see the parenthetical remark at the end of the proof of Proposition $4.1$ of [@Kawauchi1980]). Using Theorem \[thm:tau\] we are able to prove that this is so.
\[cor:concordance\] Let $\mathcal{C}$ denote the smooth concordance group, and let $$\phi_{p,1}: \mathcal{C}\rightarrow \mathcal{C}$$ denote the map defined by $\phi_{p,1}([K])=[K_{p,1}]$. Then $\phi_{p,1}$ is not a homomorphism for any $p$
### Cable knots and complex curves
Let $V_f$ be a complex curve $$V_f=\{ (z,w)\in \C^2| 0=f(z,w)\in \mathbb{C}[z,w]\},$$ and let $$S^3=\{(z,w)\subset \mathbb{C}^2| |z|^2+|w|^2=1\}$$ be the three-sphere. Further suppose that $$K=V_f \cap S^3$$ is a transverse intersection. In this case, $K\subset S^3$ is a knot or link, and we call knots that arise in this way (transverse) [*$\C$-knots*]{} (see [@Rudolph2005] for a thorough introduction to these knots).
It is well-known that some iterated cables of the unknot are $\C$-knots. Indeed, the class of cabled $\C$-knots contains the so-called [*links of singularities*]{}, which come from complex curves with a single isolated singularity at the origin. In fact, the links of singularities are precisely the iterated cables of the unknot satisfying a positivity condition, see [@EN1985] for a discussion. A notable feature of the link of a singularity is that its Milnor fiber (a smoothing of the singular complex curve contained in the four-ball [@Milnor1968]) can be isotoped into the three-sphere to be a Seifert surface for the knot.
In light of these classical results, a natural question to ask is to what extent cabling can be performed in the complex category. That is, when can a cable knot be a $\C$-knot? If it is, when is the piece of the complex curve contained in the four-ball isotopic to a Seifert surface? To this end, our results provide the following obstructions:
\[cor:cor1\] Suppose that $K_{p,pn+1}$ is a $\C$-knot. Then $n\ge -2(\frac{\tau(K)}{p-1}+ \frac{1}{p})$.
\[cor:cknot\] Suppose $K_{p,pn+1}$ is a $\C$-knot with defining complex curve, $V_f$. Further, suppose the genus of the piece of $V_f$ contained in the four-ball is equal to the Seifert genus, $g(K)$. Then $n\ge 0$ and $\tau(K)=g(K)$.
Thus, for instance, no negative cable (i.e. $n<0$) of a knot with $\tau(K)<0$ (e.g. the left-handed trefoil) will ever be a $\C$-knot. Note, too, that since $\tau(K)\le g_4(K)$, Corollary \[cor:cor1\] could also be phrased as a (weaker) obstruction which depends solely on the smooth four-genus of $K$.
Another feature of links of singularities is that they are fibered and, as alluded to above, there is an isotopy taking their fiber surface to their Milnor fiber. Restricting to the category of fibered knots whose fiber surface is isotopic to a piece of a complex curve, we have the following characterization theorem:
Corollary $1.4$ of [@ComplexCable]\[thm:ComplexCable\] Let $K$ be a fibered knot. Then $K_{p,q}$ has a Seifert surface which is isotopic to a piece of a complex curve $V_f \cap B^4$ if and only if
- $K$ has a Seifert surface which is isotopic to a piece of a complex curve [*and*]{}
- $q>0$
In particular, the fiber surface of an iterated cable of the unknot is isotopic to a piece of a complex curve if and only if all the cabling coefficients are positive.
Note the lack of restriction on $q$. We also remark that the “if" direction of the theorem holds for the class of knots which bound [*quasipositive Seifert surfaces*]{}. Quasipositive Seifert surfaces are those which can be obtained from parallel disks by attaching bands with a positive half twist (see Figure $1$ of [@ComplexCable]). We postpone the proof of Theorem \[thm:ComplexCable\] until [@ComplexCable]. There, we determines the relationship between the contact structure associated to a fibered knot and those associated to its cables. Theorem \[thm:ComplexCable\] is a corollary of this relationship and of a connection between contact geometry and the knot theory of complex curves established in [@SQPfiber; @Rudolph2005].
### Cable knots and lens space (L-space) surgeries
One area of low-dimensional topology where the invariants have had a significant impact is in the study of Dehn surgery [@BGH; @Ghiggini2007; @LensMe; @NiFibered; @Lens; @AbsGrad; @Figure8; @RasGoda; @Ras2007; @Wang2006]. Many of these results draw on the close relationship between the knot Floer homology invariants of a knot, $K$, and the invariants of the closed three-manifolds obtained by Dehn surgery on $K$ (see [@Knots; @IntegerSurgeries; @RationalSurgeries]). In the case that surgery on $K$ yields a manifold with particularly simple Floer homology, this relationship tightly constrains the knot Floer homology invariants. The knot Floer homology, in turn, determine various geometric and topological properties of the knot e.g. the genus, fiberedness.
The three-manifolds with simplest Floer homology are the rational homology spheres, $Y$, for which the rank of the Floer homology is equal to the order of the first (singular) homology, i.e. rk$\ \HFa(Y) = |H_1(Y;\Z)|$. These manifolds are called [*L-spaces*]{}, and the name stems from the fact that lens spaces are L-spaces. In the case that positive surgery on $K$ yields an L-space, we call $K$ an [*L-space knot*]{}. show that the knot Floer homology invariants of L-space knots are determined by the Alexander polynomial [@Lens]. In particular, a corollary of their theorem is that the coefficients of the Alexander polynomial of an L-space knot are all $\pm 1$, and that the knot must be fibered [@Ghiggini2007; @NiFibered] with four-genus equal to the Seifert genus [@FourBall]. Combining Theorem \[thm:tau\] with their result yields the following obstruction to a cable knot admitting an L-space surgery.
\[cor:cor2\] Suppose that positive surgery on $K_{p,pn+1}$ yields a lens space or, more generally, an L-space. Then $n\ge 0$ and $\tau(K)=g(K)$.
As counterpoint to this obstruction, we also have the following existence theorem:
\[thm:lspace\] Let $K\subset S^3$ be a non-trivial knot which admits a positive $L$-space space surgery. Then $K_{p,q}$ admits positive L-space surgeries whenever $q\ge p(2g(K)-1)$,
Note the above result does not require $q$ of the form $q=pn+1$. Indeed, it is proved somewhat differently from Theorem \[thm:filt\], using a standard cut-and-paste topological argument together with known properties of the invariants.
Theorem \[thm:lspace\] is interesting in light of a paucity of examples. The theorem is the first general construction of L-space knots outside of the double primitive knots [@Berge]. Indeed, any sufficiently positive iterated cable of a knot which has an actual lens space surgery will itself have L-space surgeries.
As noted above, L-space knots have the property that the $\Z\oplus\Z$ filtered chain homotopy type of the knot’s filtration of $\CF^\infty(S^3)$ is determined by the Alexander polynomial [@Lens]. Since the Alexander polynomial of a cable knot is determined by Formula , Theorem \[thm:lspace\] provides an efficient method for calculating the Floer homology of a large class of cable knots. For instance, $+5$ surgery on the trefoil is the lens space, $L(5,4)$, and thus the $(p,q)$ cable of the trefoil admits L-space surgeries whenever $q\ge p+1$. In particular, the Floer homology of the $(2,3)$ cable of the trefoil is determined by its Alexander polynomial which, from Equation is $$(t-1+t^{-1})(t^2-1+t^{-2})= t^{3}-t^{2}+1-t^{-2}+t^{-4}.$$ On the other hand, this is the same Alexander polynomial as that of the $(3,4)$ torus knot. As this knot also admits lens space surgeries, we find that the two distinct knots have identical Floer invariants. This example was obtained by a rather lengthy calculation in [@MyThesis].
We find this example noteworthy as it appears difficult to produce families of L-space knots with the same Floer invariants. On the other hand, we will show in an upcoming paper that infinite families of knots with identical Floer invariants are rather abundant. These families, however, do not admit L-space surgeries. It would be interesting to probe Theorem \[thm:lspace\] for further examples of L-space knots with identical Floer homology.
[**[Remarks:]{}**]{} Versions of Theorems \[thm:filt\] and \[thm:tau\] appear in the author’s dissertation, [@MyThesis].
[**[Acknowledgment:]{}**]{} I wish to thank Peter Ozsv[á]{}th for his advice and encouragement throughout my time as a graduate student, in which the heart of this work was done. I also thank Chuck Livingston for his interest, and Cornelia Van Cott for informing me of Kawauchi’s results and suggesting Corollary \[cor:concordance\].
Proof of Theorems {#sec:proof}
=================
In this section we prove the theorems stated in the introduction. Several of the proofs rely heavily on the results of [@Cabling].
Notational Background
---------------------
Before beginning, we recall a few facts about knot Floer homology. Our purpose here is to establish notation and is not intended as an introduction to theory or knot Floer homology. First, recall that to a knot $K\subset Y$ we can associate a doubly-pointed Heegaard diagram $$(\Sigma,\{\alpha_1,\ldots,\alpha_g\},\{\beta_1,\ldots,\beta_{g} \},w,z).$$ (see Definition $2.4$ of [@Knots]). The chain complex $\CFa(Y)$ is generated (as a $\Z/2\Z$ vector space) by $g$-tuples $\x=x_1\times ... \times x_g$ of intersection points, where $x_i\in \alpha_i\cap \beta_{\sigma(i)}$ (here, $\sigma$ is a permutation in the symmetric group on $g$ letters). The chain complex is equipped with a differential $\partial$ which counts points in moduli spaces of pseudo-holomorphic disks [@HolDisk].
Let $\mathcal{G}$ be the set of generators. Using the basepoint, $w$, define a map: $$\spinc_w: \mathcal{G}\rightarrow \SpinC(Y),$$ where $\SpinC(Y)$ is the set of $\SpinC$ structures on $Y$. The chain complex splits as direct sum $$\CFa(Y)\cong \underset{\spinc\in \SpinC(Y)}\bigoplus \CFa(Y,\spinc),$$ with the summand $\CFa(Y,\spinc)$ generated by those $\x\in\mathcal{G}$ with $\spinc_w(\x)=\spinc$. Using both basepoints, define a map: $$\underline{\spinc}:=\spinc_{w,z}: \mathcal{G}\rightarrow \underline{\SpinC}(Y,K),$$ where $ \underline{\SpinC}(Y,K)$ is the set of relative $\SpinC$ structures on $Y-K$. Picking a Seifert surface, $F$, for $K$, we obtain a map: $$A : \mathcal{G}\rightarrow \Z,$$ by associating the quantity $A(\x)=\OneHalf\langle c_1(\spincrel(\x)),[F,\partial F]\rangle $ to a generator. Here $c_1(\spincrel(\x))\in H^2(Y,K;\Z)$ is the relative first Chern class of the relative $\SpinC$ structure. We will refer to $A$ as the [*Alexander grading*]{}. $A$ defines a filtration on $\CFa(Y,\spinc)$, in the sense that $A(\partial (\x))\le A(\x)$ for each $\x\in \mathcal{G}$. Note that $A$ depends on the Seifert surface, but only through its homology class.
Thus, once we pick a homology class of Seifert surface, it makes sense to define $\FiltY(K,i)$ to be the subcomplex of $\CFa(Y,\spinc)$ generated by $\x\in \mathcal{G}$ satisfying $A(\x)\le i$. We then have the finite length filtration mentioned in the introduction: $$0=\FiltY(K,-i) \subseteq \FiltY(K,-i+1)\subseteq \ldots \subseteq
\FiltY(K,n)=\CFa(Y,\spinc).$$ We will habitually omit the Seifert surface from the notation and assume throughout that we have chosen a fixed Seifert surface, drawing attention to its role only when there may be some ambiguity.
Finally, we recall the following definition from [@tbbounds]. To state it, let $\alpha\in \HFa(Y,\spinc)$ be a non-vanishing Floer homology class, and let $\iota_m: \FiltY(K,m)\rightarrow \CFa(Y,\spinc)$ be the inclusion map.
\[defn:tau\_alpha\] $$\tau_{\alpha}(Y,K)=\mathrm{min}\{m\in\Z|\ \alpha \subset \mathrm{Im} \ (\iota_m)_*\},$$
where $(\iota_m)_*$ is the map induced on homology.
Proof of Theorem \[thm:filt\]
-----------------------------
With the above notation, we will prove the following more general version of Theorem \[thm:filt\]:
\[thm:filtY\] Let $K\subset Y$ be a knot and let $\spinc\in\SpinC(Y)$. Pick any $M\in \Z$. Then there exists a constant $N(M)>0$ so that $\forall \ n>N(M)$, the following holds for each $j> M$: $$H_*(\FiltY(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-1))\cong H_*(\FiltY(K,j-1)).$$ Furthermore, $$\begin{array}{ll}
H_*(\FiltY(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-i))\cong & \\
H_*(\FiltY(K_{p,pn+1},pj+\frac{(pn)(p-1)}{2}-i-1)) & \forall \ i=2,\ldots,p-1. \\
\end{array}$$ In particular, $$\tau_\alpha(Y,K_{p,pn+1})=\left\{\begin{array}{ll}
p \tau_\alpha(Y,K)+\frac{(pn)(p-1)}{2}+p-1 & {\text{or}} \\
p \tau_\alpha(Y,K)+\frac{(pn)(p-1)}{2}. & \\
\end{array}
\right.$$
In the above, the filtration for $K_{p,pn+1}$ is defined using a Seifert surface $F_{p,pn+1}$ satisfying $[F_{p,pn+1}]=p[F]$, where $F$ is the Seifert surface used to define the filtration for $K$.
The key tool in proving the theorem is
\[lemma:hd\]Lemma 2.2 of [@Cabling]Let $$H=(\Sigma,\{\alpha_1,\ldots,\alpha_g\},\{\beta_1,\ldots,\beta_{g-1},\mu\},z,w),$$ be a Heegaard diagram for a knot K, where $\mu$ is a meridian for $K$. Then $$H(p,n)= (\Sigma,\{\alpha_1,\ldots,\alpha_g\},\{\beta_1,\ldots,\beta_{g-1},\tilde{\beta}\},z',w),$$
is a Heegaard diagram for $K_{p,pn+1}$. $H$ and $H(p,n)$ differ only in the final curve $\tilde{\beta}$. Here, $\tilde{\beta}$ is obtained by winding $\mu$ around an $n$-framed longitude for the knot $(p-1)$ times. $w$ is to remain fixed under this operation. $z$ is replaced by a basepoint $z'$ so that the arc connecting $z'$ and $w$ has intersection number $p$ with $\tilde{\beta}$. See Figure \[fig:wind\].
![\[fig:wind\] Shown are parts of the diagrams $H$ and $H(p,n)$ (here, $p=3$ and $n=2$). On $H(p,n)$, we have slid the meridian, $\mu$, $2$ times along a $2$-framed longitude $\lambda$. This figure establishes the ordering on exterior intersection points in terms of the ordering on $x_i\in \alpha_g\cap \tilde{\beta}$. ](winding.eps)
Lemma 3.1 of [@Cabling] shows that with the addition of a third basepoint (which, by an abuse of notation, we also denote by $z$) $H(p,n)$ specifies $K$ by using the pair $(w,z)$ while it specifies $K_{p,pn+1}$ using the pair $(w,z^\prime)$. See Figure \[fig:wind\]. Increasing the parameter $n$ for the cable has the effect of adding many generators to the knot Floer chain complexes derived from $H(p,n)$. These generators, $\x'$, were called [*exterior intersection points*]{} in the predecessor, and are characterized by the property of having one component of $\x'$ of the form $x_i\in \alpha_g\cap \tilde{\beta}$, with $x_i$ lying in a small neighborhood of the meridian, $\mu$. The exterior points are in $(2n(p-1)+1)$-to-one correspondence with the generators of the chain complex for $K$ which came from $H$, i.e. there is $(2n(p-1)+1)$-to-one map: $$\pi:\{\mathrm{exterior \ intersection \ points \ of \ H(p,n)} \}\longrightarrow \{\mathrm{intersection \ points \ of\ H}\}.$$
More precisely, any generator for the Heegaard diagram $H$ is of the from $\{p,\y\}$, where $p\in \alpha_g\cap \mu$ and $\y$ is a $(g-1)$-tuple of intersection points. The fibers of $\pi$ are given by $$\pi^{-1}(\{p,\y\}) = \{ \ \{x_i,\y\} \ | \ x_i\in \alpha_g\cap \tilde{\beta}, \ i=0,..., 2n(p-1)\}.$$ Figure \[fig:wind\] establishes an ordering for the fiber $\pi^{-1}(\{p,\y\})$, in terms of an ordering on the $x_i$. In terms of this ordering, we call intersection points of the form $\{x_0,\y\}$ [*outermost intersection points*]{}. Let us establish some notation: $$C(i):= \{ (g-1)\sd \text{tuples},\ \y \ | \ \{p,\y\} \text{\ has\ Alexander\ grading\ } i \text{\ for\ }H \}$$ $$A:= \text{\ Alexander\ grading\ for\ generators\ from\ } H(p,n),\text{\ with\ respect\ to\ } (w,z)$$ $$A':= \text{\ Alexander\ grading\ for\ generators\ from\ } H(p,n),\text{\ with\ respect\ to\ } (w,z')$$
We will determine the Alexander gradings, $A$ and $A'$, of exterior intersection through a sequence of lemmas. The first is an adaptation of Lemma $3.6$ of [@OSThurston] to the present notation. It determines $A$ and $A'$ for the outermost intersection points.
\[lemma:pindown\] Let $\x=\{x_0,\y\}$ be an outermost intersection point with $\y\in C(i)$. Then
1. The $\SpinC$ structure $\spinc\in \SpinC(Y)$ associated to $\x$ is independent of $z$ or $z'$, and agrees with the $\SpinC$ structure associated to the corresponding generator for $H$, $\{p,\y\}$.
2. $A(\{x_0,\y\})=i$. That is, the Alexander grading of $\{x_0,\y\}$ with respect to $(w,z)$ agrees with that of $\{p,\y\}$.
3. $A'(\{x_0,\y\})= pA(\{x_0,\y\})+\frac{(pn)(p-1)}{2} $
The first two parts follow from the fact that $(1)$ the map $\spinc_w$ does not depend on $z$ or $z'$ and $(2)$ there is an isotopy taking $\tilde{\beta}$ to $\mu$ in the complement of $z$, under which the generator $\{x_0,\y\}$ becomes identified with $\{p,\y\}$. The third is the content of Lemma $3.6$ of [@OSThurston] which establishes the result in the context of multi-pointed Heegaard diagrams for links. One can pass from their result to the present case by setting $l=1$, $p_1=p$, and $q_1=pn+1$ and then considering Chern classes of relative $\SpinC$ structures to determine the Alexander grading.
Next, we have Lemmas $3.3$ and $3.4$ of [@Cabling]
Let $H(p,n)$ be as above. Then for odd integers $i < 2n$, we have $$\begin{gathered}
A(x_{i-1} ,\y)-A(x_{i},\y)=A'(x_{i-1},\y)-A'(x_{i},\y)=1\\
A(x_i,\y)-A(x_{i+1},\y)=0\\
A'(x_i,\y)-A'(x_{i+1},\y)=p-1.\end{gathered}$$
\[lemma:varytuple\] Suppose $\y \in C(j), \z \in C(k)$. Then, $$\begin{gathered}
A(x_i,\y)-A(x_i,\z)=j-k\\
A'(x_i,\y)-A'(x_i,\z)=p(j-k).\end{gathered}$$
Lemma 3.5 of [@Cabling] shows that, by making the cabling parameter $n$ sufficiently large, we can ensure that exterior intersection points generate the highest $A$ and $A'$ gradings. More precisely, we have the following restatement of Lemma 3.5 of [@Cabling]:
\[lemma:isolate\] Pick $l\in \Z$, and let $$g=\mathrm{max}\{ A(\x) \ | \ \x \text{\ is\ any\ intersection\ point,\ exterior\ or\ not} \}.$$ Then there exists a constant $N > 0$ such that for all $n$ with $n> N$, the only intersection points with $A(\x)\ge g-l$ are exterior.
The lemmas are summarized by the table and caption in Figure \[fig:chaincomplex\].
![\[fig:chaincomplex\] Table of $A$ and $A'$ gradings of exterior points for the Heegaard diagram, $H(p,n)$. For each ordered pair, the number on the left is the $A$ grading (i.e. the companion knot’s grading). The number on the right is $A'$ (the cabled knot’s grading). According to Lemma \[lemma:pindown\], we have $g'=pg+\frac{(pn)(p-1)}{2}$. The area below the top (solid) line represents the chain complex $\FiltY(K_{p,pn+1},g'-p-1)=\FiltY(K,g-2)$. The area below the bottom (dashed) line represents the chain complex $\FiltY(K_{p,pn+1},g'-2p-1)=\FiltY(K,g-3)$. ](chaincomplextau.eps)
Theorem \[thm:filt\] follows quickly from the table. The key point is that $H(p,n)$ is a diagram for $Y$, regardless of whether we use the basepoint $z$ or $z^\prime$. The only restriction on the differential for the filtered subcomplexes is that $n_w(\phi)=0$, which is independent of $z$ and $z^\prime$. In light of these remarks and the table of filtrations, we have the following isomorphisms of [*chain complexes*]{}, for each $\spinc\in \SpinC(Y)$ and $k<l$ $$\begin{array}{ll} \FiltY(K_{p,pn+1},pg+\frac{(pn)(p-1)}{2}-p(k-1)-1))& = \FiltY(K,g-k) \\
\FiltY(K_{p,pn+1},pg+\frac{(pn)(p-1)}{2}-p(k-1)-i)& = \FiltY(K_{p,pn+1},pg+\frac{(pn)(p-1)}{2}-p(k-1)-i-1)) \\
& \forall \ i=2,\ldots,p-1 \\ \end{array}$$
Note the appearance of the $\SpinC$ structure, $\spinc$. Up to this point we had not distinguished between intersection points corresponding to different $\SpinC$ structures; indeed, the table represents all intersection points. However, it is straightforward to see that the table splits as a direct sum of complexes according to $\SpinC$ structures on $Y$, yielding the above. This follows from the fact that part $(1)$ of Lemma \[lemma:pindown\] actually applies to any exterior point in the table which, in turn, follows from the fact that $\{x_j,\y\}$ is connected to $\{x_0,\y\}$ by a Whitney disk (c.f the proof of Lemmas $3.3$ and $3.4$ of [@Cabling]).
Taking $l>g-M$ in Lemma \[lemma:isolate\] and changing variables $j-1=g-k$ yields the first part of the theorem. For the second part we let $M<-g(K)$, where $g(K)$ is the genus of $K$. It follows from the adjunction inequality for knot Floer homology that
$$\begin{array}{ll} H_*(\FiltY(K,j))=0 & \forall j<-g(K). \end{array}$$
The second part of the theorem now follows from the definition of $\tau_\alpha$. $\square$
Examining the Heegaard diagram $H(p,n)$ when $n<0$, we are lead to:
\[thm:filtYneg\] Let $K\subset Y$ be a knot and let $\spinc\in\SpinC(Y)$. Pick any $M\in \Z$. Then there exists a constant $N(M)>0$ so that $\forall \ n>N$, the following holds for each $j< M$: $$H_*(\FiltY(K_{p,-pn+1},pj-\frac{(pn)(p-1)}{2}-1))\cong H_*(\FiltY(K,j-1)).$$ Furthermore, $$\begin{array}{ll}
H_*(\FiltY(K_{p,-pn+1},pj-\frac{(pn)(p-1)}{2}+i))\cong & \\
H_*(\FiltY(K_{p,-pn+1},pj-\frac{(pn)(p-1)}{2}+i+1)) & \forall \ i=1,\ldots,p-2. \\
\end{array}$$ In particular, $$\tau_\alpha(Y,K_{p,-pn+1})=\left\{\begin{array}{ll}
p \tau_\alpha(Y,K)-\frac{(pn)(p-1)}{2}+p-1 & {\text{or}} \\
p \tau_\alpha(Y,K)-\frac{(pn)(p-1)}{2}. & \\
\end{array}
\right.$$
Proof of Theorem \[thm:tau\]
----------------------------
We now turn to the proof of Theorem \[thm:tau\]. We focus first on proving the inequality: $$\label{eq:tauinequality} p \tau(K)+ \frac{(pn)(p-1)}{2}\ \ \le\tau(K_{p,pn+1})\le \ \
p \tau(K)+\frac{(pn)(p-1)}{2} +p-1.$$
This will follow rather quickly from Theorems \[thm:filtY\] and \[thm:filtYneg\], together with a crossing change inequality satisfied by $\tau$ to interpolate between the cases $n>0$ and $n<0$. Recall that two knots $K_+, K_- \subset S^3$ are said to differ by a crossing change if there exists an embedded three-ball $B^3\subset S^3$, outside of which the knots agree,$$(S^3,K_+)\backslash B^3 \cong (S^3,K_-)\backslash B^3,$$ and such that the local pictures of $(B^3, B^3\cap K_+)$ (resp. $(B^3, B^3\cap K_-)$) are given by Figure \[fig:crossingchange\].
![\[fig:crossingchange\] A crossing change. $K_+,K_-\subset S^3$ agree, except in the three-ball pictured here.](crossingchange.eps)
Corollary $1.5$ of [@FourBall] states that if $K_+$ and $K_-$ differ by a crossing change, then we have the following inequality: $$\label{eq:crossing}
\tau(K_+)-1\le \tau(K_-) \le \tau(K_+).$$ Now it is straightforward to see that $K_{p,pl+1}$ can be changed into $K_{p,p(l-1)+1}$ by a sequence of $\frac{p(p-1)}{2}$ crossing changes, each of which changes a positive crossing to a negative (to see this induct on $p$ and, for the induction step, change the first $p-1$ crossings of a full twist on $p$ strands). Thus $$\label{eq:ninequality} \tau(K_{p,pl+1}) -\frac{p(p-1)}{2} \le \tau(K_{p,p(l-1)+1})\le \tau(K_{p,pl+1}).$$ Theorems \[thm:filtY\] and \[thm:filtYneg\] tell us that Inequality is satisfied for $K_{p,pn+1}$ provided $|n|>N$. Combining this with Inequality yields for all $n$.
The proof of Inequality will easily extend to each of the invariants $\tau_\alpha(Y,K)$, once a generalization of the crossing change inequality is established for knots in $Y$. We defer the proof of this latter inequality to [@Slice], where it will follow from the fact that $|\tau_\alpha(Y,K)|$ bounds the genus of any smoothly embedded surface $$i: (F,\partial F) \hookrightarrow Y\times [0,1],$$ with $i|_{\partial F} = K\subset Y\times \{1\}$.
We turn now to the case when $\tau(K)=\pm g(K)$. Consider the function defined in the introduction: $$\ctauprime{n}:=\tau(K_{p,pn+1})$$
Inequality says that the graph of $\ctauprime{n}$ is bounded between the parallel lines $y^{\pm}(n)=\frac{(pn)(p-1)}{2} + c^{\pm},$ where $$\begin{array}{ll} c^+= & p \tau(K) + p-1 \\ c^-= & p \tau(K) \end{array}$$
On the other hand, Theorem \[thm:filtY\] tells us that for $n>N$, $\ctauprime{n}$ is either $y^+(n)$ or $y^-(n)$. Similarly, Theorem \[thm:filtYneg\] says that $\ctauprime{n}$ agrees with either $y^+(n)$ or $y^-(n)$, provided $n<-N$.
Next, we have a lemma:
Suppose $\ctauprime{n}=y^+(n)$ for all $n>N$. Then $\ctauprime{n}=y^+(n)$ for all $n$.
Likewise, suppose $\ctauprime{n}=y^-(n)$ for all $n<-N$. Then $\ctauprime{n}=y^-(n)$ for all $n$.
The lemma follows easily from the fact $\ctauprime{n}$ is bounded between $y^\pm(n)$ for all $n$, together with Inequality . More precisely, Inequality says that $$\ctauprime{n}-\ctauprime{n-1}\le \frac{p(p-1)}{2},$$ for all $n$. However, if $\ctauprime{n}=y^+(n)$ for all $n>N$, then the only way for $\tau^p$ to be bounded by $y^+$ (for all $n$) is if $\tau^p$ decreases by exactly $\frac{p(p-1)}{2}$ each time we decrease $n$ by one. Similar considerations hold if $\ctauprime{n}=y^-(n)$ for all $n<-N$.
We will show that if $\tau(K)=-g(K)$, then $\ctauprime{n}=y^+(n)$ for all $n>N$. Similarly, if $\tau(K)=g(K)$ then $\ctauprime{n}=y^-(n)$ for all $n<-N$. To this end, we have:
\[thm:largen\] Theorem $1.2$ of [@Cabling] Let $K\subset S^3$ be a knot of genus $g$. Pick any $M\in \Z$. Then there exists a constant $N(M)>0$ so that $\forall \ n>N$, the following holds for each $j< M$: $$\HFKa_*(K_{p,pn+1},i)\cong
\left\{\begin{array}{ll}
H_{*+2(j-g)}(\Filt(K,j-g)) & {\text{for
$i=pg+\frac{(p-1)(pn)}{2}-pj$}} \\
H_{*+2(j-g)+1}(\Filt(K,j-g)) & {\text{for $i=pg+\frac{(p-1)(pn)}{2}-pj-1$}} \\
0 & {\text{otherwise.}}\\
\end{array}
\right.$$
Considering $j=2g$ in the above theorem shows that $$\HFKa(K_{p,pn+1},-pg+\frac{(p-1)(pn)}{2})\cong \Z_{(-2g)}$$ This shows that $\tau(K_{pn+1})\ne -pg+\frac{(p-1)(pn)}{2}$; there is no homology in Alexander grading $-pg+\frac{(p-1)(pn)}{2}$ of the appropriate Maslov grading (provided $K$ is a non-trivial knot). If $\tau(K)=-g$, then $y^-(n)= -pg+\frac{(p-1)(pn)}{2}$, and hence $\ctauprime{n}=y^+(n)$ for all $n>N$. Using Theorem $3.8$ of [@Cabling] in place of Theorem \[thm:largen\] above, the same argument shows that if $\tau(K)=g(K)$ then $\ctauprime{n}=y^-(n)$ for all $n<-N$. This completes the proof.
Proof of Theorem \[thm:lspace\]
-------------------------------
The strategy here is to show that if $K$ is an L-space knot, then surgery on $K_{p,q}$ will be an L-space for $q$ large enough. We will achieve this through a standard topological argument, together with formulas for the Floer homology of manifolds obtained by Dehn surgery on knots and connected sums, respectively. More precisely, Theorem \[thm:lspace\] is an immediate consequence of the following facts:
1. $pq$ surgery on $K_{p,q}$ is the manifold $S^3_{q/p}(K)\# L(p,q)$, where $S^3_{q/p}(K)$ is the manifold obtained by $q/p$ Dehn surgery on $K$.
2. The lens space, $L(p,q)$, is an L-space
3. If any positive surgery on $K$ yields an L-space, then $q/p$ surgery on $K$ is an L-space for any $q/p\ge 2g(K)-1$, where $g(K)$ is the genus of $K$.
4. If $Y_1$ and $Y_2$ are L-spaces, then $Y_1\#Y_2$ is an L-space.
The first fact is well-known to those working with Dehn surgery. For completeness, we include a proof below. The Floer homology of lens spaces can easily be computed from their genus one Heegaard diagram, verifying $(2$). The third fact follows from a general formula which computes the Floer homology of manifolds obtained by surgery on $K$ in terms of the knot Floer homology invariants [@RationalSurgeries; @Lens]. Specifically, we have
\[lemma:CalcRanks\] Let $K\subset S^3$ be an L-space knot, and fix a pair of relatively prime integers $p$ and $q$. Then $$\Rk \HFa(S^3_{q/p}(K)) =
|q| + 2\max(0,(2g(K)-1)|p|-|q|).$$
The lemma is a particular case of Proposition $9.5$ of [@RationalSurgeries]. Specifically, if $K$ is an L-space knot, results of [@Lens] show that $\tau(K)=g(K)$. This implies that the term $\nu(K)$ appearing in Proposition $9.5$ of [@RationalSurgeries] is equal to the genus, since $\nu(K)$ is equal to $\tau(K)$ or $\tau(K)+1$ by definition, and $\nu(K)\le g(K)$ by the adjunction inequality (Theorem $5.1$ of [@Knots]). The term in Proposition $9.5$ involving rk $H_*(\widehat{A}_s)$ vanishes under the assumption that $K$ is an L-space knot, since in this case rk $H_*(\widehat{A}_s)=1$.
With the lemma in hand, $(3)$ follows immediately: if $q/p\ge 2g(K)-1$, the second term in the proposition vanishes and $$\Rk\ \HFa(S^3_{q/p}(K)) =
|q| = |H_1(S^3_{q/p}(K);\Z)|.$$
The last fact follows from a K[ü]{}nneth type formula for the Floer homology of manifolds obtained by connected sum, Theorem $1.5$ of [@HolDiskTwo]. This theorem says that the Floer homology of the connected sum, $Y_1\#Y_2$, can be computed from a chain complex quasi-isomorphic to the tensor product of Floer chain complexes for $Y_1$ and $Y_2$. In particular, it implies that $$\mathrm{rk}\ \HFa(Y_1\#Y_2)=\mathrm{rk}\ \HFa(Y_1)\cm \mathrm{rk} \ \HFa(Y_2).$$ Thus $(4)$ follows from the definition of an L-space and the elementary observation that $$|H_1(Y_1\#Y_2;\Z)|=|H_1(Y_1;\Z)|\cm |H_1(Y_2;\Z)|.$$
We conclude by showing that $pq$ surgery on $K_{p,q}$ is $S^3_{q/p}(K)\# L(p,q)$. To see this, decompose $S^3$ as $$S^3=E(K) \underset{T_K} \amalg N(K),$$ where $N(K)$ is a tubular neighborhood of $K$, $E(K)=S^3-N(K)$, and $T_K=\partial N(K)$ (see Figure \[fig:cablesurgery\]). $K_{p,q}$ is embedded in $T_K$ as a curve of slope $p/q$. Here, the meridian of $K$ has slope $0/1$, while the longitude has slope $1/0$. meridional Consider next the tubular neighborhood of the cable. Denote this by $N(K_{p,q})$. The intersection $A=N(K_{p,q})\cap T_K$ is an annular neighborhood of $K_{p,q}$ in $T_K$. The boundary of this annulus consists of two parallel copies of $K_{p,q}$, which we denote by $\lambda$ and $\lambda'$, each of which have linking number $pq$ with $K_{p,q}$. Let us examine the result of surgery on $K_{p,q}$ with framing given by $\lambda$ (or equivalently, $\lambda'$). The fact that lk$(K_{p,q},\lambda)=pq$ is equivalent to the slope of the surgery being $pq/1$. Now the exterior of $K_{p,q}$ can be decomposed as $$E(K_{p,q})= E(K) \underset{T_K-A} \amalg N(K).$$ (Figure \[fig:cablesurgery2\]) Since $K_{p,q}$ is an essential curve on $T_K$, we see that $T_K-A$ is an annulus. Now the surgery is performed by gluing a solid torus $D^2\times S^1$ to $E(K_{p,q})$ in such a way that the boundary of each meridional disk is identified with a curve on $\partial N(K_{p,q})$ isotopic to $\lambda$. This gluing can, equivalently, be thought of as attaching a pair of two-handles (Figure \[fig:cablesurgery3\]) $H_1=D^2\times [0,\pi]$, $H_2= D^2\times [\pi, 2\pi]$ to $E(K_{p,q})$, so that $$S^3_{pq}(K_{p,q})= [ E(K)\underset{\partial D^2\times [0,\pi] }\sqcup H_1 ] \underset {D^2 \times \{0\} \sqcup (T_K-A) \sqcup D^2\times \{\pi\} } \amalg [H_2 \underset{\partial D^2 \times [\pi,2\pi] }\sqcup N(K)].$$ Since $K_{p,q}$ is a $p/q$ curve on $T_K=\partial N(K)$, the handle $H_2$ is attached to the solid torus $N(K)$ along a curve of slope $p/q$. From the perspective of $E(K)$, however, $K_{p,q}$ is a curve of slope $q/p$. It follows that term on the left is $S^3_{q/p}(K)-D^3$, while the term on the right is $L(p,q)-D^3$. These two three-manifolds are joined along their common ($2$-sphere) boundary, $D^2 \times \{0\} \sqcup (T_K-A) \sqcup D^2\times \{\pi\}$, completing the proof. $\square$
![\[fig:cablesurgery\] ](cablesurgery.eps)
![\[fig:cablesurgery2\] ](cablesurgery2.eps)
![\[fig:cablesurgery3\] ](cablesurgery3.eps)
Proof of Corollaries {#sec:cor}
====================
In this section we prove the corollaries stated in the introduction. The heart of the corollaries is that $\tau(K)$ carries a great deal of geometric information, and thus can be used in conjunction with Theorem \[thm:tau\] to obstruct cables of $K$ from having certain geometric or braid theoretic properties.
For instance, we can derive several consequences from the second half of Theorem \[thm:tau\]. To make this precise, define $$\Pos := \{ K\subset S^3 \ | \ \tau(K)=g(K)\}.$$ The following is an immediate corollary of Theorem \[thm:tau\]:
\[cor:pos\]
- If $K\in \Pos$, then $K_{p,pn+1}\in \Pos$ if and only if $n\ge 0$.
- If $K\notin \Pos$, then $K_{p,pn+1}\notin \Pos$ for any $n$.
This corollary derives its power from the fact that there are several classes of knots which we know to be contained in $\Pos$. For instance, the following classes of knots are contained in $\Pos$:
1. Knots which bound a complex curve, $V_f\subset B^4$, satisfying $g(V_f)=g(K)$ [@SQPfiber].
2. Positive knots i.e. those knots which admit a diagram containing only positive crossings [@Livingston2004].
3. L-space knots i.e. knots for which positive slope Dehn surgery on $K$ yields an L-space (in particular, lens space knots) [@Lens].
4. Any non-negatively twisted, positive-clasped Whitehead double of a knot satisfying $\tau(K)>0$ [@Doubling].
5. Strongly quasipositive knots i.e. those knots bounding a Seifert surface obtained from parallel disks by attaching bands with a positive half twist [@Livingston2004] (see also [@SQPfiber]).
6. Fibered knots whose associated contact structure is tight [@SQPfiber].
These classes overlap highly. For instance, $(2)\subset (5)\subset (1)$, and $(3)\subset (6) \subset (5)$ (see [@SQPfiber] for a discussion of these inclusions.) Combining Corollary \[cor:pos\] with $(1)$ yields Corollary \[cor:cknot\] of the introduction. Combining with $(3)$ yields Corollary \[cor:cor2\], and $(6)$ is instrumental in the results of [@ComplexCable]. Item $(2$) yields an obstruction for cabling to produce positive knots:
Suppose $K\notin \Pos$. Then $K_{p,pn+1}$ is not a positive knot for any $n\in \Z$
Item $(5)$ produces the most precise information to date on the smooth four-genera of knots obtained by iterated doubling and cabling:
Suppose $\tau(K)>0$. Then any knot, $S$, obtained by an arbitrary sequence of non-negative cabling and Whitehead double operations is in $\Pos$. In particular, $g_4(S)=g(S)$.
We have seen that understanding when $K_{p,pn+1}\in \Pos$ has geometric consequences for cabling. Likewise, understanding when $\tau(K_{p,pn+1})\ge 0$ is also tied to geometry. In this case, Theorem \[thm:tau\] yields:
Suppose $\tau(K_{p,pn+1})\ge 0$. Then $n\ge -2\left(\frac{\tau(K)}{p-1}+ \frac{1}{p}\right)$.
This is relevant in light of the connection between $\tau(K)$ and complex curves. Suppose that $K$ is a $\C$-knot. Then results of [@Olga2004; @SQPfiber; @Rudolph1983] show that $\tau(K)=g_4(K)$. In particular, $\tau(K)\ge 0$. Combining this fact with the corollary yields Corollary \[cor:cor1\] of the introduction.
Corollary \[cor:cor1\] could alternatively be stated in terms of the braid group. Let $B_n$ denoted the braid group on $n$ strands, with generators $\sigma_1,\ldots,\sigma_{n-1}$. A [*quasipositive*]{} knot is any knot which can be realized as the closure of a braid of the form: $$\beta = \Pi_{k=1}^m w_k \sigma_{i_k} w_k^{-1}.$$
Results of [@Rudolph1983] and [@BO2001] indicate that $\C$-knots are exactly the class of knots obtained as the closures of quasipositive braids. Thus, we obtain
Suppose $K_{p,pn+1}$ is the closure of a quasipositive braid. Then $$n\ge -2(\frac{\tau(K)}{p-1}+ \frac{1}{p}).$$
We conclude with a proof of Corollary \[cor:concordance\].
[**Proof of Corollary \[cor:concordance\].**]{} We wish to show that $\phi_{p,1}$ is not a homomorphism, so we must find knots $K_1$ and $K_2$ for which $$\phi_{p,1}([K_1\#K_2])\ne [\phi_{p,1}(K_1)\#\phi_{p,1}(K_2)].$$ To do this, it suffices to show that $$\tau^p(K_1\#K_2,1)\ne \tau^p(K_1,1) + \tau^p(K_2,1),$$ This is accomplished with $$K_1=\mathrm{right\sd handed \ trefoil}$$ $$K_2=\mathrm{left\sd handed \ trefoil}$$ $K_1\#K_2$ is slice, and so $\tau^p(K_1\#K_2,1)=\tau^p(U,1)=0,$ where $U$ is the unknot. Now $\tau(K_1)=g(K_1)=1$, so Theorem \[thm:tau\] implies $\tau^p(K_1,1)=p\tau(K_1)=p$. For $K_2$, however, we have $\tau(K_2)=-g(K_2)=-1$. Thus $\tau^p(K_2,1)=p\tau(K_2)+p-1=-1$ (again by Theorem \[thm:tau\]). This completes the proof. $\square$
[^1]: The depends on an identification of the tubular neighborhood with a solid torus which, in turn, comes from the longitude specified by a Seifert surface.
[^2]: There will be no loss of generality in doing this, since $K_{-p,-q}\simeq -K_{p,q}$ where $-K_{p,q}$ is $K_{p,q}$ with reversed string orientation. Our invariants are not sensitive to this orientation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider stochastic resonance for a diffusion with drift given by a potential, which has two metastable states and two pathways between them. Depending on the direction of the forcing, the height of the two barriers, one for each path, will either oscillate alternating or in synchronisation. We consider a simplified model given by a continuous time Markov Chains with two states. This was done for alternating and synchronised wells. The invariant measures are derived for both cases and shown to be constant for the synchronised case. A PDF for the escape time from an oscillatory potential is studied. Methods of detecting stochastic resonance are presented, which are linear response, signal-noise ratio, energy, out-of-phase measures, relative entropy and entropy. A new statistical test called the conditional Kolmogorov-Smirnov test is developed, which can be used to analyse stochastic resonance. An explicit two dimensional potential is introduced, the critical point structure derived and the dynamics, the invariant state and escape time studied numerically. The six measures are unable to detect the stochastic resonance in the case of synchronised saddles. The distribution of escape times however not only shows a clear sign of stochastic resonance, but changing the direction of the forcing from alternating to synchronised saddles an additional resonance at double the forcing frequency starts to appear. The conditional KS test reliably detects the stochastic resonance. This paper is mainly based on the thesis [@tommy_thesis].'
author:
- 'Tommy Liu[^1]'
bibliography:
- 'paper\_reportreferences.bib'
title: 'A comparative study of stochastic resonance for a model with two pathways by escape times, linear response, invariant measures and the conditional Kolmogorov-Smirnov Test'
---
Introduction
============
Outline of Problem
------------------
Consider the following problem. Let $X_t^\epsilon$ be the random variable describing the trajectory of a diffusion process in $\mathbb{R}^r$ where $t$ is the time and $\epsilon^2$ is the variance level. More precisely we consider processes described by the following type of stochastic differential equation $$\begin{aligned}
dX^\epsilon_t=b\left(X^\epsilon_t,t\right)dt+\epsilon\,dW_t\end{aligned}$$ where $b:\mathbb{R}^r\times \mathbb{R}\longrightarrow \mathbb{R}^r$ and $W_t$ is a Wiener process in $\mathbb{R}^r$. We suppose that the drift term $b$ has the form $$\begin{aligned}
b(x,t)=-\nabla V_0 (x) + F\cos \Omega t \end{aligned}$$ where $F,x\in\mathbb{R}^r$ and $V_0:\mathbb{R}^r\longrightarrow \mathbb{R}$ is called the unperturbed potential. We consider unperturbed potentials with two or more minimas (wells). Most importantly, we consider potentials where there are multiple pathways between the wells. To our knowledge systems with two pathways have not been studied in the context of stochastic resonance.
Consider the case $\Omega=0$ and where the noise $\epsilon$ is very small. The particle will stay very close to one of the wells of the potential and will occasionally escape to the other well. The time of the actual transition from one well to the other is very short compare to the time it stays in any particular well.
Now consider the case where $\Omega>0$. For particular choices of $\Omega>0$ and $\epsilon>0$, these transitions between the two wells will become synchronised with the driving frequency $\Omega$. This is called stochastic resonance. Thus the term *noise induced synchronisation* was used for systems where the amplitude of the forcing $F$ was not large [@tk_14; @tk_15] (see also the discussions in [@tk_16]). New insights into the exact manner of these synchronised transitions will be studied in this paper, which may be more appropriate in light of the results obtained in this paper.
For small noise $\epsilon$, one would expect that stochastic resonance depends only on the essential properties of the system, such as the height difference between the wells and the pathways for escape. We investigate what effects these multiple pathways have on the appearance of stochastic resonance. Varying $F$, $\Omega$ and $\epsilon$ should thus reveal the qualitative structure of the unperturbed potential $V_0$. In this paper we test this paradigm by studying a two dimensional example with two wells and two independent pathways between them, see Section \[sum\_chap\_mexican\_hat\_toy\_model\].
Historical Background
---------------------
Stochastic resonance has attracted interest among mathematicians and physicist. An overview of the studies that have occurred in both physics and mathematics are given here.
### Physical Background
Stochastic resonance was first observed in 1981 [@benzi81; @tk_17c; @tk_17b]. The first example [@benzi81] considered transitions between two metastable states to model the cyclic occurrences of ice ages. Since then many examples of stochastic resonance were found in optics [@opt_1_PhysRevLett.60.2626; @opt_2_Guidoni1995; @opt_3_PhysRevA.49.2199; @opt_4_PhysRevLett.68.3375], electronics [@elec_1_FAUVE19835; @elect_2_PhysRevE.49.R1792; @elect_3_Mantegna1995; @elect_4_PhysRevLett.76.563; @elect_5_PhysRevLett.74.3161; @elect_6_:/content/aip/journal/jap/76/10/10.1063/1.358258; @elect_7_PhysRevA.39.4323; @elect_8_PhysRevE.49.4878; @elect_9_PhysRevLett.67.1799], neuronal systems [@neuron_1_PhysRevLett.67.656], quantum systems [@quant_1_:/content/aip/journal/jap/77/6/10.1063/1.358720; @quant_2_:/content/aip/journal/apl/66/1/10.1063/1.114161] and paddlefish [@fish_1_PhysRevLett.84.4773; @fish_2_FREUND200271]. Stochastic resonance could be thought of as quasi-deterministically periodic transition between two metastable states. For example, the climate of the Earth could be modelled by two states. There is a state corresponding to an Ice Age and another corresponding to the opposite of an Ice Age, a so-called “Hot Age”. As the Earth’s climate cyclically changes many times between Cold Ages and Hot Ages, its behaviour could be modelled by stochastic resonance.
A range of techniques for example linear response [@lin_1_TUS:TUS1787; @lin_2_doi:10.1137/0143037], signal-to-noise ratio [@sign_1_PhysRevA.39.4668; @sign_2_escape_2_PhysRevA.41.4255] and distribution of escape times [@escape_1_PhysRevA.42.3161; @sign_2_escape_2_PhysRevA.41.4255; @escape_3_PhysRevE.49.4821] were used to define, analyse and study stochastic resonance. These techniques along with other examples of stochastic resonance are reviewed in the long overview paper by Gammaitoni, Hänggi, Jung and Marchesoni [@RevModPhys.70.223]. We will evaluate the usefulness of some of these techniques for our problem, see Section \[sum\_chap\_results\].
### Mathematical Background
There are various mathematical studies of stochastic resonance. These often involve different orders of approximations for small noise levels. The first and second order of approximations are discussed below. Adiabatic large deviation is also presented.
In the first leading order of approximation, a key element of study is to control the escape times from the wells as given by the so called large deviation theory, see the monograph of Freidlin and Wentzell [@freidlin98]. The distribution of the exit time was derived by Day in [@tk_19] and by Galves, Kifer, Olivieri and Vares [@freid_2_doi:10.1137/1119057; @oliveri_expoen; @freid_3]. To go beyond leading order has been much more difficult for the transition problem between two wells as WKB theory could up to now not be rigorously applied.
The next order of approximation was rigorously derived by Bovier, Eckhoff, Gayrard, Klein [@Bovier02metastabilityin] and Berglund and Gentz [@kram_2_2008arXiv0807.1681B] using techniques from potential theory. Berglund and Gentz in a series of papers studied the situation of low, non-quadratic barriers and drifts not given by autonomous potentials [@kram_2_2008arXiv0807.1681B; @tk_16]. A review of different techniques used to derive Kramers’ formula can be found in the review paper [@krammer_review_2011arXiv1106.5799B].
In [@ld_1_comp_1_Freidlin2000333] Friedlin considered stochastic resonance in the adiabatic regime. This means the diffusion can effectively be described by a Markov process which describes the jumps between wells. This problem was revisited by Hermann, Imkeller and Pavlyukevich, see Chapter 4 in [@tran2014] and references therein, to derive results uniformly for varying time scale to identify the optimal resonance point asymptotically for small noise even outside the adiabatic regime leading to different logarithmic corrections including the famous cycling effect discovered by Day [@tk_20], see also [@tk_18] for the connection with stochastic resonance. Escape time outside of adiabatic regime is studied in [@tk_00_doi:10.1137/120887965].
As mentioned above in leading order the transitions of the diffusion process $X^\epsilon_t$ between the wells can be approximated by a two state Markov Chain $Y^\epsilon_t=\pm1$ which have been studied [@pav_thesis; @pav_imkell; @pav03; @tran2014]. Further comparative studies of the stochastic resonance for the diffusion case $X^\epsilon_t$ versus the Markov Chain $Y^\epsilon_t$ case were done by Hermann, Imkeller, Pavlyukevich and Peithmann in [@ld_3_comp_4_imkeller2004stochastic; @comp_2_Herrmann2005; @ld_2_comp_3_herrmann2005large; @herrmann2005]. A collection of papers on comparative studies between stochastic resonance in diffusion and Markov Chains can be found in the monograph [@tran2014]. One of the main conclusions in [@ld_2_comp_3_herrmann2005large; @herrmann2005; @comp_2_Herrmann2005; @tran2014] is rigorously showing that using linear response and signal-to-noise ratio to analyse stochastic resonance in the diffusion case $X^\epsilon_t$ gives a different result to analysing the Markov Chain case $Y^\epsilon_t=\pm1$ with the same techniques even asymptotically in the small noise limit. Other common methods used to study stochastic resonance include invariant measures and Fourier transforms. We consider six measures of stochastic resonance frequently used and considered by Pavlyukevich in his thesis [@pav_thesis; @tran2014] which are linear response, signal-to-noise ratio, energy, out-of-phase measure, relative entropy and entropy.
In this paper we will study stochastic resonance on a two dimensional toy model, in both the diffusion and Markov Chain cases, and where there are two independent pathways between the wells going through two different saddles. The escape times and the six measures of stochastic resonance introduced above are studied.
Static Potential
================
We remind ourselves of the theory of escape times and escape rate from a well of a static potential. These results follow from large deviation and potential theory.
Large Deviation, Potential Theory and Kramers Formula {#chap_sect_kram}
-----------------------------------------------------
Let $V:\mathbb{R}^r:\longrightarrow\mathbb{R}$. Let $x\in\mathbb{R}^r$ be a well and $z_i\in\mathbb{R}^r$ be saddles labelled by $i=1\ldots n$. The saddles would be gateways providing a passage for escape from the well. Define $$\begin{aligned}
\Delta V_i=V(z_i)-V(x)\end{aligned}$$ which is the height difference between the well and the $i$th saddle. For small noise $\epsilon$, an approximate expression can be estimated for the escape time of the particle going through the $i$th saddle. It is well known from the theory of large deviation [@freidlin98], that in the lowest order of the noise $\epsilon$ the mean exit time is given by $$\begin{aligned}
\tau_i=e^{+2\Delta V_i /\epsilon^2}\end{aligned}$$ Inverting this gives the escape rate $$\begin{aligned}
R_i=e^{-2\Delta V_i /\epsilon^2}\end{aligned}$$ and the total escape rate would be to sum over all the saddles $$\begin{aligned}
R=\sum_{i=1}^n R_i
=\sum_{i=1}^n e^{-2\Delta V_i /\epsilon^2}.\end{aligned}$$ The order correction is done by adding a coefficient called Kramers coefficient and the resulting corrected rate is called Kramers rate $$\begin{aligned}
R_i=k_ie^{-2\Delta V_i /\epsilon^2}
\quad
\text{where}
\quad
k_i=\frac{\sqrt{\left|\nabla^2V(x)\right|}}{2\pi}
\frac{\left|\lambda(z_i)\right|}{\sqrt{\left\Vert \nabla^2V(z_i) \right\Vert}}\end{aligned}$$ where $\left|\nabla^2(x)\right|$ denotes the determinant of the Hessian of the potential at the well $x$, $\left\Vert \nabla^2V(z_i) \right\Vert$ denotes the modulus of the determinant of the potential at the saddle $z_i$ and $\left|\lambda(z_i)\right|$ denotes the minimum eigenvalue of the Hessian of the potential at the saddle $z_i$. This gives the escape rate in the next order of approximation to be $$\begin{aligned}
R&=\sum_{i=1}^nR_i
=\sum_{i=1}^n k_ie^{-2\Delta V_i /\epsilon^2}\end{aligned}$$ which is rewritten as $$\begin{aligned}
R=
\frac{\sqrt{\left|\nabla^2V(x)\right|}}{2\pi}
\sum_{i=1}^n
\frac{\left|\lambda(z_i)\right|}{\sqrt{\left\Vert \nabla^2V(z_i) \right\Vert}}
\exp\left\{\frac{-2\left(V(z_i)-V(x)\right)}{\epsilon^2}\right\}\end{aligned}$$ The last order of approximation for higher noise $\epsilon$ is done by bounding the error on Kramers coefficient. This is $$\begin{aligned}
k_i=\frac{\sqrt{\left|\nabla^2V(x)\right|}}{2\pi}
\frac{\left|\lambda(z_i)\right|}{\sqrt{\left\Vert \nabla^2V(z_i) \right\Vert}}
\left(
\frac{1}{1+\mathcal{O}\left(\frac{\epsilon^2}{2}\ln\frac{\epsilon^2}{2}\right)}
\right)\end{aligned}$$ which is rewritten as $$\begin{aligned}
\label{chap_2_krammers}
\frac{1}{k_i}
=
\frac{2\pi}{\sqrt{\left|\nabla^2V(x)\right|}}
\frac{\sqrt{\left\Vert \nabla^2V(z_i) \right\Vert}}{\left|\lambda(z_i)\right|}
\left[
1+\mathcal{O}\left(\frac{\epsilon^2}{2}\ln\frac{\epsilon^2}{2}\right)
\right]\end{aligned}$$ The derivation of Kramers formula was done in [@Bovier02metastabilityin].
Oscillatory Potential {#chapter_oscil_times}
=====================
Let $V:\mathbb{R}^r\rightarrow\mathbb{R}$ be a potential with two wells. This potential is subjected to a periodic forcing $F\in\mathbb{R}^r$ with frequency $\Omega$ and perturbed by noise $\epsilon$, which is described by the SDE. $$\begin{aligned}
\dot{X^\epsilon_t}=-\nabla V+F\cos(\Omega t)+\epsilon \dot{W_t}
\label{chap_4:markov}\end{aligned}$$ where $W_t$ is a Wiener process in $\mathbb{R}^r$ and $F\in\mathbb{R}^r$. We call $X^\epsilon_t$ the diffusion case. In this Section we study Equation \[chap\_4:markov\] by considering the escape times between the two wells, modelling Equation \[chap\_4:markov\] by a continuous time Markov Chain, considering six measures of stochastic resonance and a new statistical test called the conditional Kolmogorov-Smirnov Test.
Markov Chain Reduction
----------------------
Stochastic resonance usually involves studying transitions between two stable states. If one solely concentrate on the transitions times one can reduce the model to a continuous time Markov chain $Y_t^\epsilon$ with state space $Y^\epsilon_t=\pm 1$ symbolizing the two stable states. $$\begin{aligned}
X^\epsilon_t
\longrightarrow
Y^\epsilon_t\end{aligned}$$ Let $w_{l}(t)$ denote the position of the left well at time $t$ and $w_{r}(t)$ the position of the right well at time $t$. Note that $w_l(t)$ and $w_r(t)$ are also continuous in time. Let $R\in\mathbb{R}$ be constant. The reduction from the $X^\epsilon_t$ to the Markov Chain $Y^\epsilon_t$ is $$\begin{aligned}
Y^\epsilon_t&=
\left\{
\begin{array}{lll}
-1 & \text{if} &
\left|X^\epsilon_t-w_l(t)\right|\leq R\\[0.5em]
+1 & \text{if} &
\left|X^\epsilon_t-w_r(t)\right|\leq R\\[0.5em]
Y^\epsilon_s & \text{if} & \left|X^\epsilon_t-w_l(t)\right|> R \quad \text{and} \quad \left|X^\epsilon_t-w_r(t)\right|> R
\end{array}
\right.\end{aligned}$$ where $s$ is given by $$\begin{aligned}
s&=\max\left\{s_1,s_2\right\}\\
\text{where}\quad s_1&=\max_{s<t}\left\{s:\left|X^\epsilon_s-w_l(s)\right|\leq R\right\}\\
\text{where}\quad s_2&=\max_{s<t}\left\{s:\left|X^\epsilon_s-w_r(s)\right|\leq R\right\}\end{aligned}$$ When $Y^\epsilon_t=-1$ we say the particle is in the left well and when $Y^\epsilon_t=+1$ we say the particle is in the right well. Hence we keep $Y^\epsilon_t$ constant even when the particle is in neither well. Only when it enters the other well would $Y^\epsilon_t$ change sign.
The escape time from the left to right well $\tau_{-1+1}$ and from the right to left well $\tau_{+1-1}$ are defined in the following way $$\begin{aligned}
\tau_{-1+1}&=l\left\{t:Y^\epsilon_t=-1\right\}\quad \text{where} \quad \left\{t:Y^\epsilon_t=-1\right\} \quad \text{is an interval}\\
\tau_{+1-1}&=l\left\{t:Y^\epsilon_t=+1\right\}\quad \text{where} \quad \left\{t:Y^\epsilon_t=+1\right\} \quad \text{is an interval}\end{aligned}$$ where $l$ denotes the Lebesgue measure. In other words the time spent being in the state $Y^\epsilon_t=-1$ is $\tau_{-1+1}$ and the time spent being in the state $Y^\epsilon_t=+1$ is $\tau_{+1-1}$. These intervals will always be closed intervals. The process $Y^\epsilon_t$ has two states, hence each sample is a piecewise constant function. The length of each piece is the escape time $\tau_{-1+1}$ or $\tau_{+1-1}$. At every point in time it is possible to define a state probability, that is the probability of the trajectory being $-1$ or $+1$ $$\begin{aligned}
P\left(Y_t=-1\right)=\nu_-(t)
\quad \text{and} \quad
P\left(Y_t=+1\right)=\nu_+(t)\end{aligned}$$ A continuous time Markov Chain model for Equation \[chap\_4:markov\] is studied next by studying its state probabilities.
Continuous Time Markov Chain
----------------------------
Consider a two state continuous time Markov Chain given by $Y_t=\pm1$. The probability of transiting from $Y_t=-1$ to $Y_t=+1$ in a small time interval $[t,t+d t]$ is $p_{-1+1}\left([t,t+d t]\right)\sim p(t)dt$. Similarly the probability of transiting from $Y_t=+1$ to $Y_t=-1$ in a small time interval $[t,t+d t]$ is $p_{+1-1}\left([t,t+d t]\right)\sim q(t)dt$. The probability of $Y_t$ staying at $-1$ in the small time interval $[t,t+d t]$ is $p_{-1-1}\left([t,t+d t]\right)$. Similarly the probability of $Y_t$ staying at $+1$ in the small time interval $[t,t+d t]$ is $p_{+1+1}\left([t,t+d t]\right)$. These probabilities satisfy $$\begin{aligned}
p_{-1-1}+p_{-1+1}=1
\quad \text{and} \quad
p_{+1-1}+p_{+1+1}=1\end{aligned}$$ Let $p:\mathbb{R}\rightarrow \mathbb{R}$ and $q:\mathbb{R}\rightarrow \mathbb{R}$ be cyclic functions on the interval $[0,T]$ where $T$ is the period. The behaviour of $\nu(t)=
\left(
\begin{array}{cc}
\nu_-(t)&\nu_+(t)
\end{array}
\right)^\dagger$ is described by $$\begin{aligned}
\frac{d\nu}{dt}&=\label{markov_chain_main_eqn}
Q^\dagger
\nu
\quad
\text{where}
\quad
Q=
\left(
\begin{array}{cc}
-p(t) & p(t)\\[0.5em]
q(t) & -q(t)
\end{array}
\right)\end{aligned}$$ and $Q$ is the infinitesimal generator. The aim now is to derive the state probability by solving this differential equation for various forms of $p$ and $q$. The extra conditions we use are $$\begin{aligned}
\nu_-(t)+\nu_+(t)=1\label{markov_chain_main_condition_1}
\quad \text{and} \quad
\nu'_-(t)+\nu'_+(t)=0\end{aligned}$$ for all times $t$ and the initial conditions at $t=0$ are $\nu_-(0)$ and $\nu_+(0)$.
After a very long time the state probabilities $\nu_\pm(\cdot)$ should not depend on the initial state probabilities $\nu_\pm(0)$. At time infinity $\nu_\pm(\cdot)$ should also be cyclic on $[0,T]$. Let the time be given by $t=NT+n$ where $N$ is a discrete number of periods. This leads us to define the invariant measure as the state probabilities in the limit as $N\longrightarrow\infty$ $$\begin{aligned}
\overline{\nu}(n):=\lim_{N\longrightarrow\infty}\nu(NT+n)\end{aligned}$$ The rate of convergence to the invariant measure would depend on the value of $p$ and $q$ themselves. Define the relaxation time $T_{relax}$ as the first time $t=T_{relax}$ such that $$\begin{aligned}
\left|
\overline{\nu}\left(T_{relax}\right)
-\nu\left(T_{relax}\right)
\right|
\leq
e^{-1}\end{aligned}$$ which is a measure of the rate of convergence to the invariant measure.
### Continuous Time Markov Chain - Alternating Saddles $p\neq q$
Notice that $p$ may be interpreted as the probability of escape from the left well and $q$ as the probability of escape from the right well. If $p$ and $q$ are cyclic over $[0,T]$, then this can be interpreted as modelling a potential with periodic forcing in continuous time.
\[chap\_4\_thm:equal\_contin\] Let $p\neq q$ and $t\geq0$. The state probabilities are given by $$\begin{aligned}
\nu_-(t)&=\frac{\nu_-(0)+\int_0^tq(s)\exp\left\{\int_0^sp(u)+q(u)\,du\right\}\,ds}{\exp\left\{\int_0^tp(u)+q(u)\,du\right\}}\\[0.5em]
\nu_+(t)&=\frac{\nu_+(0)+\int_0^tp(s)\exp\left\{\int_0^sp(u)+q(u)\,du\right\}\,ds}{\exp\left\{\int_0^tp(u)+q(u)\,du\right\}}\end{aligned}$$
\[corollary\_invariant\_measure\_contin\] For the state probabilities in Theorem \[chap\_4\_thm:equal\_contin\] the invariant measures are $$\begin{aligned}
\overline{\nu}_-(t)
&=
\frac{\int^t_0p(s)g(s)\,ds}{g(t)}
+\frac{\int^T_0p(s)g(s)\,ds}{g(t)\left(g(T)-1\right)}\\[0.5em]
\overline{\nu}_+(t)
&=
\frac{\int^t_0q(s)g(s)\,ds}{g(t)}
+\frac{\int^T_0q(s)g(s)\,ds}{g(t)\left(g(T)-1\right)}\\[0.5em]
\text{where}\quad
g(t)&=\exp\left\{\int^t_0p(u)+q(u)\,du\right\}\end{aligned}$$
We derive the invariant measure for $\overline{\nu}_-(t)$. The case for $\overline{\nu}_+(t)$ is similar. Consider the fact that $p(\cdot)$ and $q(\cdot)$ are cyclic on $[0,T]$ and let $i$ be an integer, then the following integral can be rewritten as $$\begin{aligned}
\int^{(i+1)T}_{iT}
p(s)g(s)\,ds
=
g(T)^i\int^T_0p(s)g(s)\,ds\end{aligned}$$ Let the time be given by $NT+t$ where $N$ is an integer number of periods. This means the following integral can be written as $$\begin{aligned}
\int^{NT+t}_0
p(s)g(s)\,ds
=
g(T)^N
\int^t_0
p(s)g(s)\,ds
+
\int^T_0p(s)g(s)\,ds
\sum_{i=0}^{N-1}
g(T)^i\end{aligned}$$ So the state probability is equal to $$\begin{aligned}
\nu_-(NT+t)
=
\frac{\nu_-(0)}{g(T)^Ng(t)}
+
\frac{\int^t_0p(s)g(s)\,ds}{g(t)}
+
\frac{\int^T_0p(s)g(s)\,ds}{g(t)}
\frac{1}{g(T)-1}
\left(
1-\frac{1}{g(T)^N}
\right)\end{aligned}$$ Letting $N\longrightarrow\infty$ gives the required result.
### Continuous Time Markov Chain - Synchronised Saddles $p=q$
If the forcing is such that the height of the barrier stays the same for both pathways the same then this corresponds to the case $p=q$
\[chap\_4\_thm:not:equal\_contin\] Let $p=q$ and $t\geq0$. The state probabilities are given by $$\begin{aligned}
\nu_-(t)&=\frac{1}{2}-\frac{\nu_+(0)-\nu_-(0)}{2}\exp\left\{-2\int_0^tp(s)\,ds\right\}\\[0.5em]
\nu_+(t)&=\frac{1}{2}+\frac{\nu_+(0)-\nu_-(0)}{2}\exp\left\{-2\int_0^tp(s)\,ds\right\}\end{aligned}$$
\[cor\_half\_continuous\] For the state probabilities in Theorem \[chap\_4\_thm:not:equal\_contin\] the invariant measures are $$\begin{aligned}
\overline{\nu}_-(t)=\overline{\nu}_+(t)=\frac{1}{2}\end{aligned}$$
Probability Density Function of Escape Times
--------------------------------------------
The escape rates from the left to right are denoted by $R_{-1+1}(\cdot)$ and right to left escape rates are denoted by $R_{+1-1}(\cdot)$. The PDFs for the escape times are given by the Theorem below.
\[chap\_4\_pdf\_thm\] Let $u$ be the time coordinate of entry into a well, then the PDFs for the escape occurring at time coordinate $t>u$ are $$\begin{aligned}
p_{-}(t,u)&=R_{-1+1}(t)\exp\left\{-\int^t_uR_{-1+1}(s)\,ds\right\}\\[0.5em]
p_{+}(t,u)&=R_{+1-1}(t)\exp\left\{-\int^t_uR_{+1-1}(s)\,ds\right\}\end{aligned}$$ where $p_{-}(t,u)$ is for left to right and $p_{+}(t,u)$ is for right to left.
We consider escaping from the left well. The right well is similar. Divide the time interval $[u,t]$ into many small time intervals $$\begin{aligned}
\delta t =\frac{t-u}{N}\end{aligned}$$ Similar to how we derived the invariant measures we want to derive the probability of escape in a very small time interval $[t,t+\delta t]$. This is given by $$\begin{aligned}
p_{-1+1}([t,t+\delta t])
&=p(t)\delta t \\
&=1-e^{-R_{-1+1}(t)\delta t}\\
&\approx R_{-1+1}(t)\delta t \end{aligned}$$ which is valid for small $\delta t$. Large deviations allow us to say even more about the escape time $\tau_{-1+1}$ and $\tau_{+1-1}$. Theorem 1 in [@oliveri_expoen] shows that it is an exponentially distributed random variable. The probability of staying in the left well is given by $$\begin{aligned}
p_{-1-1}([t,t+\delta t])
&=1-p_{-1+1}([t,t+\delta t])\\
&=1-p(t)\delta t \\
&=1-\left(1-e^{-R_{-1+1}(t)\delta t}\right)\\
&=e^{-R_{-1+1}(t)\delta t}\end{aligned}$$ We want to know the probability of escaping in the time interval $[t,t+\delta t]$ given that the particle has entered at $u$ and stayed up to time $t$. This is given by $$\begin{aligned}
p_{-1-1}([u,t])p_{-1+1}([t,t+\delta t])
&=\prod_{i=1}^{N}p_{-1-1}\left([u+(i-1)\delta t,u+i\delta t]\right)
p_{-1+1}([t,t+\delta t])\\
&=\prod_{i=1}^{N}\exp\left\{-R_{-1+1}\left(u+(i-1)\delta t\right)\delta t\right\}
p_{-1+1}([t,t+\delta t])\\
&=\exp\left\{\sum_{i=1}^{N}-R_{-1+1}\left(u+(i-1)\delta t\right)\delta t\right\}
p_{-1+1}([t,t+\delta t])\\
&=\exp\left\{-\int^t_uR_{-1+1}(s)\,ds\right\}
R_{-1+1}(t)\delta t\end{aligned}$$ This completes the proof.
### Perfect Phase Approximation of Probability Density Function of Escape Times {#chap_approx_pdf}
The PDF for the escape times derived in Theorem \[chap\_4\_pdf\_thm\] had to differentiate between left and right escapes and are conditioned on the time $u$ of entrance into the well. Suppose now that $t$ is the escape time from any well, which does not differentiate between left and right escape. Note that $t$ is the actual time it takes to escape from a well and is not a time coordinate. The PDF for $t$ is given by $$\begin{aligned}
p_{tot}(t)=
\frac{1}{2}
\int_{0}^{T}p_-(t+u,u)m_-(u)+p_+(t+u,u)m_+(u)\,du\end{aligned}$$ This is because after a long time has elapsed we would expect that many transitions would have occurred between left and right. The number of transitions escaping from the left and right should be roughly the same. The $m_-(u)$ is a PDF for the time of entrance into the left well and the $m_+(u)$ is a PDF for the time of entrance into the right well. We may not have explicit expressions for $m_-(u)$ and $m_+(u)$. We derive an approximate expression for $p_{tot}$ without an explicit expressions for $m_-(u)$ and $m_+(u)$. Let $m_-(u)$ and $m_+(u)$ be approximated by $$\begin{aligned}
m_-(u)\approx\delta\left(u-T/2\right)
\quad \text{where} \quad
m_+(u)\approx
\frac{1}{2}\delta\left(u\right)
+
\frac{1}{2}\delta\left(u-T\right)\end{aligned}$$ where $\delta(\cdot)$ is the Dirac delta function. This approximation is used because in the SDEs which we will simulate, the times when transition into the left well is greatest is at half the period $u=\frac{T}{2}$ and the times when transition into the right well is greatest is at $u=0$ and $u=T$. Due to the fact that $m_-(u)$ and $m_+(u)$ are probabilities a factor of $\frac{1}{2}$ is used in $m_+(u)$. Progressing we have $$\begin{aligned}
p_{tot}(t)&=
\frac{1}{2}
\int_{0}^{T}p_-(t+u,u)m_-(u)+p_+(t+u,u)m_+(u)\,du\\
&\approx
\frac{1}{2}
\int_{0}^{T}
p_-(t+u,u)\delta\left(u-T/2\right)
+p_+(t+u,u)
\left(
\frac{1}{2}\delta\left(u\right)
+
\frac{1}{2}\delta\left(u-T\right)
\right)
\,du\\[0.5em]
&=
\frac{1}{2}
\left\{
p_-(t+T/2,T/2)
+\frac{1}{2}
p_+(t,0)
+\frac{1}{2}
p_+(t+T,T)
\right\}\\[0.5em]
&=
\frac{1}{2}
\left\{
p_-(t+T/2,T/2)
+\frac{1}{2}
p_+(t,0)
+\frac{1}{2}
p_+(t+0,0)
\right\}\\[0.5em]
&=
\frac{1}{2}
\left\{
p_-(t+T/2,T/2)
+
p_+(t,0)
\right\}\\[0.5em]
&=p_+(t,0)\end{aligned}$$ This is because for the simulations which we are going to do, the Kramers rate satisfy $R_{-1+1}(t)=R_{+1-1}(t+T/2)$ (see later in Section \[chap\_mexican\_hat\_toy\_model\] and \[sum\_chap\_results\] for the geometry of the Mexican Hat Toy Model which justifies this). Thus the following approximation $$\begin{aligned}
p_{tot}\approx p_+(t,0)\end{aligned}$$ is only valid for the simulations we do, and not for a general potential. We call this way of approximating $m_-(u)$ and $m_+(u)$ the perfect phase approximation.
Six Measures of Stochastic Resonance {#chap_analysis_theory}
------------------------------------
\[sum\_chap\_analysis\_theory\]
We introduce six possible criteria of measuring how close a process is to exhibiting stochastic resonance. We call them the six measures denoted by $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$. In what follows we will consider so large times, that the relaxation time has effectively elapsed for both the diffusion $X^\epsilon_t$ and Markov Chain $Y^\epsilon_t$, in other words the state probability would have effectively converged to the invariant measure $\overline{\nu}$. This means that over one period $T=2\pi/\Omega$ of the forcing, the invariant measures will have the properties $$\begin{aligned}
\overline{\nu}_\pm(t)=\overline{\nu}_\pm(t+T)
\quad \text{and} \quad
\overline{\nu}_\pm(t)=\overline{\nu}_\mp(t+T/2)\end{aligned}$$ We obtain the averaged trajectories given by $$\begin{aligned}
\left\langle X^\epsilon_t \right\rangle
=E\left(X^\epsilon_t\right)
\quad \text{and} \quad
\left\langle Y^\epsilon_t \right\rangle
=E\left(Y^\epsilon_t\right)\end{aligned}$$ which are the trajectories obtained after averaging over many realisations. Note that $\left\langle X^\epsilon_t \right\rangle$ is calculated by averaging over many realisation over many periods. The $\left\langle X^\epsilon_t \right\rangle$ is cyclic over one period. Notice that $\left\langle Y^\epsilon_t \right\rangle$ is related to the invariant measures by $$\begin{aligned}
\left\langle Y^\epsilon_t \right\rangle=\overline{\nu}_+(t)-\overline{\nu}_-(t)\end{aligned}$$ We introduce the Out-of-Phase Markov Chain defined by $$\begin{aligned}
\overline{Y}^\epsilon_t=
\left\{
\begin{array}{l}
0 \quad \text{if} \quad Y^\epsilon_t=-1 \quad \text{and} \quad mod(t,T)\leq T/2\\
1 \quad \text{if} \quad Y^\epsilon_t=-1 \quad \text{and} \quad mod(t,T)> T/2\\
1 \quad \text{if} \quad Y^\epsilon_t=+1 \quad \text{and} \quad mod(t,T)\leq T/2\\
0 \quad \text{if} \quad Y^\epsilon_t=+1 \quad \text{and} \quad mod(t,T)> T/2
\end{array}
\right.\end{aligned}$$ and similarly the averaged Out-of-Phase Markov Chain is defined by $$\begin{aligned}
\left\langle \overline{Y}^\epsilon_t \right\rangle=E\left(\overline{Y}^\epsilon_t\right)\end{aligned}$$ Define two new functions by $$\begin{aligned}
\phi^-(t)=
\left\{
\begin{array}{c}
1 \quad \text{if} \quad mod(t,T)\leq T/2\\
0 \quad \text{if} \quad mod(t,T)> T/2
\end{array}
\right.
\quad \text{and} \quad
\phi^+(t)=
\left\{
\begin{array}{c}
0 \quad \text{if} \quad mod(t,T)\leq T/2\\
1 \quad \text{if} \quad mod(t,T)> T/2
\end{array}
\right.\end{aligned}$$ The following trajectories are Fourier transformed $$\begin{aligned}
\tilde{x}(\omega)&=\mathcal{F}\left(
\langle x_t \rangle
\right)
=
\langle \mathcal{F} \left(x_t\right) \rangle
\\[0.5em]
\tilde{Y}(\omega)&=\mathcal{F}\left(
\left\langle Y^\epsilon_t \right\rangle
\right)
=
\langle \mathcal{F} \left(x_t\right) \rangle\end{aligned}$$ The linear response is defined as the intensity of the Fourier Transform at the driving frequency $\Omega$. $$\begin{aligned}
X_{lin}=\left|\tilde{x}\left(\frac{\Omega}{2\pi}\right)\right|
\quad \text{and} \quad
Y_{lin}=\left|\tilde{Y}\left(\frac{\Omega}{2\pi}\right)\right|\end{aligned}$$ Note that the forcing is periodic and monochromatic. Now we can define the six measures. For the diffusion case only $M_1$ and $M_2$ are defined $$\begin{aligned}
M_1=\frac{1}{F}X_{lin}
\quad \text{and} \quad
M_2=\frac{1}{\epsilon F}X_{lin}\end{aligned}$$ where $F$ is the magnitude of the forcing. For the Markov Chain $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$ are all defined as $$\begin{aligned}
M_1&=\frac{1}{F}Y_{lin}\\
M_2&=\frac{1}{\epsilon F}Y_{lin}\\
M_3&=\int_0^T \left\langle Y^\epsilon_t \right\rangle^2 dt\\
M_4&=\int_0^T \left\langle \overline{Y}^\epsilon_t \right\rangle dt\\
M_5&=\int_0^T
\phi^-(t)\ln\left(\frac{\phi^-(t)}{\overline{\nu}_-(t)}\right)+
\phi^+(t)\ln\left(\frac{\phi^+(t)}{\overline{\nu}_+(t)}\right)
dt\\
M_6&=\int^T_0
-\overline{\nu}_-(t)\ln\overline{\nu}_-(t)
-\overline{\nu}_+(t)\ln\overline{\nu}_+(t)\,
dt\end{aligned}$$ Note that in definition of the six measures it is assumed that the process has relaxed to equilibrium. We give a few physical interpretation of the six measures $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$. The $M_1$ is the intensity of the driving frequency $\Omega$ in the spectrum of the Fourier transform. The $M_2$ is sometimes called signal-to-noise ratio as it compares this intensity to the noise level $\epsilon$. The $M_3$ is sometimes called the energy. The $M_4$ is sometimes called the out-of-phase measure since it measures the amount of time the Markov Chain spends in the “wrong” well. The $M_5$ and $M_6$ are sometimes called relative entropy and entropy respectively, since they measure how far away the invariant measures are from being constant. If the invariant measures are constant then these six measures will also be constant. Thus it can be understood that these six measures is a measure of how far away the invariant measures are from being constant.
Escape Time Statistics
----------------------
We will measure the escape time for many consecutive transitions. This will result in a collection of measurements of escape times $$\begin{aligned}
\tau_1,\tau_2,\ldots,\tau_n\end{aligned}$$ A new method for analysing such a collection of measurements is presented. One may be led to think that in the case of synchronized saddle no stochastic resonance is possible. To study this in detail we need to have a more careful look at the escape time statistics. The problem we are facing is that the distribution of the escape times strictly speaking depends on the entrance time phase $u$. As we have no theoretical result for the distribution of $u$, we are developing here a technique to study the distribution of the escape conditioned on the entrance phase $u$, that is $p(t ,u)$.
### Kolmogorov-Smirnov Test
First we recall well known results about the Kolmogorov-Smirnov statistic and the Kolmogorov-Smirnov test [@ks_test_1]. Let $
\xi_1, \xi_2, \ldots, \xi_n
$ be $n$ independently and identically distributed real random variables, where each $\xi_i$ is distributed with CDF $F(\cdot)$. The empirical CDF $F_n(\cdot)$ is defined. This gives $$\begin{aligned}
P(\xi_i\leq x)=F(x)
\quad \text{and} \quad
F_n(x)=
\frac{1}{n}
\sum_{i=1}^n
\mathbf{1}_{(-\infty,x]}
(\xi_i)\end{aligned}$$ where $\mathbf{1}_{A}$ is the indicator function for a set $A$. Consider the following $$\begin{aligned}
D_n=
\left\Vert
F_n-F
\right\Vert_\infty
=\sup_{x\in\mathbb{R}}
\left|
\frac{1}{n}
\sum_{i=1}^n
\mathbf{1}_{(-\infty,x]}
(\xi_i)
-F(x)
\right|\end{aligned}$$ where $D_n$ is called the Kolmogorov-Smirnov statistic or KS statistic. We define what we mean by the null hypothesis.
Let $\xi_1,\xi_2,\ldots,\xi_n$ be $n$ real random variables. The null hypothesis is that each $\xi_i$ is independently distributed with CDF $F(x)$.
We want to know how large or small $D_n$ needs to be before deciding whether to reject the null hypothesis. The following Theorem offers a remarkable answer to this problem.
\[chap\_4\_thm\_ks\_test\] Suppose the null hypothesis is true, then the distribution of $D_n$ depends only on $n$.
This distribution is called the KS distribution. There is a Theorem which describes the asymptotic behaviour of the KS distribution.
In the limit $n \longrightarrow \infty$, $\sqrt{n}D_n$ is asymptotically Kolmogorov distributed with the CDF $$Q(x)=1-2\sum_{k=1}^\infty (-1)^{k-1}e^{-2k^2x^2}$$ that is to say $$\lim_{n\longrightarrow \infty} P(\sqrt{n} D_n \leq x)=Q(x).$$
### Conditional Kolmogorov-Smirnov Test
This conditional Kolmogorov-Smirnov Test is developed in the paper[@tommy_paper_A] and the thesis [@tommy_thesis]. Let $\zeta_1,\zeta_2,\ldots,\zeta_n$ be $n$ iid real random variables. They are $n$ empirical observations of a random variable $\zeta$. Now suppose that each of the $\xi_1,\xi_2,\ldots,\xi_n$ is conditioned and dependent on the corresponding $\zeta_1,\zeta_2,\ldots,\zeta_n$. The conditional CDF $F(\cdot,\cdot)$ is $$\begin{aligned}
P(\xi_i\leq x\,|\,\zeta_i)=F_{\zeta_i}(x)=\int^x_{-\infty}f(s,\zeta_i)\,ds.\end{aligned}$$ But $\xi_1,\xi_2,\ldots,\xi_n$ are empirical measurements of the same random variable $\xi$. The CDF for $\xi$ is $$\begin{aligned}
P(\xi\leq x)=F(x)=\int^x_{-\infty}\int^{u=+\infty}_{u=-\infty}f(s,u)m(u)\,du\,ds\end{aligned}$$ where $m(\cdot)$ is the PDF for $\zeta$. In our context we have the problem that the random variables are not identically distributed under the null hypothesis. The $\xi_1,\xi_2,\ldots,\xi_n$ and $\zeta_1,\zeta_2,\ldots,\zeta_n$ are obtained experimentally and $F_{\zeta_i}(\xi_i)$ can be calculated but a PDF for $\zeta_i$, that is $m(\cdot)$, has no easy expression. We still want to perform a statistics test that is similar to the KS test even in such situations where the distribution $m(\cdot)$ of $\zeta$ is unknown. First we define what we call the total null hypothesis and the conditional null hypothesis.
Let $\xi_1, \xi_2, \ldots, \xi_n$ be $n$ empirical observations of a random variable $\xi$. The total null hypothesis is that $\xi$ is distributed with the CDF $F(\cdot)$. The conditional null is that each $\xi_i$ is distributed with the conditional CDF $F_{\zeta_i}(\cdot)$.
A new statistical test is developed, which is similar to the KS test.
Suppose the conditional null hypothesis is true. Let $F_{\zeta_i}(\cdot)$ be continuous. Let $S_n$ be the statistic given by $$\begin{aligned}
S_n=\sup_{x\in[0,1]}
\left|
\frac{1}{n}
\sum^n_{i=1}
\mathbf{1}_{[0,x]}
\left(
F_{\zeta_i}(\xi_i)
\right)
-x
\right|\end{aligned}$$ then $S_n$ is KS distributed.
Adiabatic Large Deviation
-------------------------
We have to stress that this paper is built on three approximations, which form the backbone of all the research presented. These are small noise approximation, adiabatic approximation and perfect phase approximation.
Perfect phase approximation only works for small noise. This is because the noise is so small the particle will only escape when the maximum probability to escape has arrived. When the minimum probability to escape is present it will almost never escape. This is the idea behind the perfect phase approximation.
Notice one subtlety behind all the theory presented in this Section. The derivations involved probabilities of escape $p$ and $q$ and the escape rates $R_{-1+1}$ and $R_{+1-1}$. But it was assumed that $p$, $q$, $R_{-1+1}$ and $R_{+1-1}$ are accurately known no matter how large or small the noise level $\epsilon$ is and no matter how fast or slow the driving frequency $\Omega$ is. But such ideal expressions for $p$, $q$, $R_{-1+1}$ and $R_{+1-1}$ are not known.
When we come to do the analysis in Section \[sum\_chap\_results\], the $p_{tot}$ is calculated with the approximation $p_{tot}\approx p_+(t,0)$. When the rates $R_{-1+1}$ and $R_{+1-1}$ are needed they are calculated using Kramers formula as though it is escape from a static potential in the small noise limit. This means an oscillatory potential is being approximated by a static potential which is the adiabatic approximation.
In the paper [@adiabatic_large_deviation_herrmann2006] the adiabatic approximation was justified in the small noise, slow forcing limit using time dependent large deviation theory, that is, it was shown asymptotically the escape times are given by the adiabatic approximation. This result is only for the leading term, whether the analogue result holds for the Kramers rate is unknown.
Mexican Hat Toy Model {#sum_chap_mexican_hat_toy_model}
=====================
\[chap\_mexican\_hat\_toy\_model\]
The main object of consideration of this paper, which is called the Mexican Hat Toy Model, is now introduced. Let $a>0$, $b>0$ and $V_0:\mathbb{R}^2\longrightarrow \mathbb{R}$ be a real function from the plane to the line. The unperturbed potential is defined as $$\begin{aligned}
V_0(x,y)=\frac{1}{4}r^4-\frac{1}{2}r^2-ax^2+by^2
\quad \text{where} \quad r=\sqrt{x^2+y^2}\end{aligned}$$ Let $F_x, F_y \in \mathbb{R}$ be the forcing. The potential with forcing $V_F$ is defined as $$\begin{aligned}
V_F(x,y)
&=\frac{1}{4}r^4-\frac{1}{2}r^2-ax^2+by^2+F_xx+F_yy\\
&=\frac{1}{4}r^4-\frac{1}{2}r^2-ax^2+by^2+\mathbf{F}\cdot\mathbf{x}\\
&=V_0+\mathbf{F}\cdot\mathbf{x}\end{aligned}$$ written more compactly in vector notation. The forcing will clearly have a magnitude and direction given by $$\begin{aligned}
F=\sqrt{F_x^2+F_y^2}
\quad \text{and} \quad
\phi=\tan^{-1}\left(\frac{F_y}{F_x}\right)\end{aligned}$$ We will study the critical points which are solutions to the simultaneous equations $$\begin{aligned}
\frac{\partial V_F}{\partial x}=0
\quad \text{and} \quad
\frac{\partial V_F}{\partial y}=0
\label{solv_equation_hat}\end{aligned}$$ It is easy to show that when there is no forcing $F=0$, there are five critical points, that is two wells, two saddles and one hill. Stochastic resonance is studied when the forcing is small enough such that the topology of the potential does not change significantly. This is because if the forcing is too large (beyond criticality) then transitions are almost certain, and there is little point to consider stochastic resonance.
In our context, what we mean by the topology of the potential not changing significantly is when the forcing $F$ is small enough such that there are still five critical points and none of them have changed their nature. We have the following Theorem.
\[x\_crit\_thm\] Let $F_x>0$, $F_y=0$ and $b<\frac{1}{2}$. Let $F_x$ be bounded by $F_x^{sad}$ and $F_x^{crit}$ where $$\begin{aligned}
F_x^{sad}=2(a+b)\sqrt{1-2b}
\quad \text{and} \quad
F_x^{crit}=\sqrt{\frac{4(1+2a)^3}{27}}\end{aligned}$$ then $V_f$ has five critical points. Let $F_x=0$, $F_y>0$ and $b<\frac{1}{2}$. Let $F_y$ be bounded by $F_y^{sad}$ and $F_y^{crit}$ where $$\begin{aligned}
F_y^{sad}=2(a+b)\sqrt{1+2a}
\quad \text{and} \quad
F_y^{crit}=\sqrt{\frac{4(1-2b)^3}{27}}\end{aligned}$$ then $V_f$ has five critical points.
We prove the case for $F_x>0$, $F_y=0$ and $b<\frac{1}{2}$. The case for $F_x=0$, $F_y>0$ and $b<\frac{1}{2}$ is similar. Solving Equation \[solv\_equation\_hat\] leads to $
\partial V_F/\partial y
=y(x^2+y^2)-(1-2b)y=0
$ which holds if either $(x^2+y^2)-(1-2b)=0$ or $y=0$. The first case is $(x^2+y^2)-(1-2b)=0$ which gives two solutions in $y$. Having $F_x<F_x^{sad}$ gives two real solutions. The second case is $y=0$ which would yield a cubic equation with three unknowns. Having $F_x<F_x^{crit}$ gives three real solutions.
The exact properties and behaviour of the critical points are most easily studied for the cases $F_x\neq0$, $F_0$ and $F_x=0$, $F_y\neq0$. For $F_x\neq0$, $F_y\neq0$ a quintic equation with five unknowns is involved. This means that although an explicit value for the critical forcing when $F_x\neq0$ and $F_y\neq0$ cannot be given analytically, an educated guess can be made $$\begin{aligned}
F^{crit}=\min\left\{F_x^{sad}, F_x^{crit}, F_y^{sad}, F_y^{crit}\right\}
\label{chap_5_critical_forcing}\end{aligned}$$ that is because a critical force in a general direction must encompass all the other directions. More details of the properties of the critical points can be found in the thesis [@tommy_thesis]. We give an example of how the Mexican Hat Toy Model look like
Experimental Method {#sum_chap_results}
===================
We simulate a series of stochastic trajectories for the Mexican Hate Toy Model and analyse them. The SDE we want to simulate is $$\begin{aligned}
dx&=\left[-\frac{\partial V_0}{\partial x}+F_x\cos \Omega t \ \right]dt+\epsilon \ dw_x\\
dy&=\left[-\frac{\partial V_0}{\partial y}+F_y\cos \Omega t \ \right]dt+\epsilon \ dw_y\end{aligned}$$ where $V_0:\mathbb{R}^2\longrightarrow\mathbb{R}$ is the unperturbed potential of the Mexican Hat Toy Model, $F_x$ and $F_y$ are the $x$ and $y$ components of the forcing, $\Omega$ is the forcing frequency, $\epsilon$ is the noise level and $w_x$ and $w_y$ are two independent Wiener processes. This SDE can be written alternatively as $$\begin{aligned}
dx&=\left[-\frac{\partial V_0}{\partial x}+F\cos\phi\cos \Omega t \ \right]dt+\epsilon \ dw_x\\
dy&=\left[-\frac{\partial V_0}{\partial y}+F\sin\phi\cos \Omega t \ \right]dt+\epsilon \ dw_y\end{aligned}$$ The Euler method was used to simulate this SDE with the following parameters being fixed at the following values (see Equation \[chap\_5\_critical\_forcing\] for $F^{crit}$) $$\begin{aligned}
a=0.15 \quad
b=0.1 \quad
F=0.7F^{crit} \quad
\Omega=0.001\end{aligned}$$ The angle of the forcing $\phi$ and the noise level $\epsilon$ were varied. The values used were $$\begin{aligned}
\epsilon=0.15, 0.16, \ldots, 0.30
\quad \text{and} \quad
\phi=0^\circ,75^\circ,78^\circ,81^\circ,84^\circ,87^\circ,90^\circ\end{aligned}$$ The averaged diffusion trajectories $\langle x_t \rangle$ and $\langle y_t \rangle$ were collected. The averaged Markov Chain $\langle Y_t^\epsilon \rangle$ and the averaged Out-of-Phase Markov Chain $\langle \overline{Y}_t^\epsilon \rangle$ were collected as well. This would allow for the calculation of the invariant measures $\overline{\nu}_-(\cdot)$ and $\overline{\nu}_+(\cdot)$. The time coordinates of the entrance and exit to and from the left and right wells were also collected. This would allow for the calculation of the escape times. We use the following values for the time step and the radius around the wells. $$\begin{aligned}
t_{step}=0.014
\quad \text{and} \quad
R=0.19\end{aligned}$$ Note that the period of the forcing is denoted by $$\begin{aligned}
T=\frac{2\pi}{\Omega}\end{aligned}$$ The averaged trajectories were simulated by taking the averaged of 200 realisations. Each realisation was 30 periods long, that is a trajectory over the interval $[0,30T]$. The initial value of the state probabilities were set at $$\begin{aligned}
\nu_-(0)=\nu_+(0)=\frac{1}{2}\end{aligned}$$ which assists in giving a faster convergence to the invariant measures (see Theorems \[chap\_4\_thm:equal\_contin\] and \[chap\_4\_thm:not:equal\_contin\]). We should also stress that a lot of the data and results presented in this Section is just a selection of out a much wider range of results. All 112 combinations of the parameters were simulated and analysed. Details as to why these range of parameters are chosen for the experiment are given in [@tommy_thesis].
Results {#sum_chap_results}
=======
Six Measures Analysis {#conclusion_six_measures}
---------------------
The six measures are calculated for the diffusion and Markov Chain case for all angles of the forcing $\phi$ and all noise levels $\epsilon$ used in the simulations. When $\phi=90^\circ$ the wells were moving up and down but they were always at the same height as each other. The distance from either wells to the saddles, which is a gateway for escape, is the same in both wells at all times. This means the $\phi=90^\circ$ can be modelled by a synchronised Markov Chain with $p=q$. The invariant measures for the $\phi=90^\circ$ case as predicted by Corollary \[cor\_half\_continuous\] is $\overline{\nu}_-=\overline{\nu}_+=\frac{1}{2}$, which means the Fourier Transform of the averaged Markov Chain is predicted to be zero. This predicts the six measures at $\phi=90^\circ$ to be $$\begin{aligned}
M_1=M_2=M_3=0\quad
M_4=\frac{1}{2}T\quad
M_5=M_6=+T\ln(2)\end{aligned}$$ Note that $\ln(2)=0.6931\approx0.7$. Notice that for very low noise level $\epsilon\approx0$ the probabilities of escape from either well is so small it may be approximately modelled by a synchronised Markov Chain with $p\approx q$. The results below confirm the predictions for the case of $\phi=90^\circ$. Since $M_1$ and $M_2$ differ by a factor, only $M_2$ is shown.
![The measure $M_2$ for the diffusion case for various angles and noise levels.[]{data-label="chap_8_g75_diff_m_2_x_single_measures"}](chap_8_g75_diff_m_2_x_single_measures.jpg)
![The measure $M_2$ for the Markov Chain for various angles and noise levels.[]{data-label="chap_8_g75_markov_m_2_single_measures"}](chap_8_g75_markov_m_2_single_measures.jpg)
![The measure $M_3$ for the Markov Chain for various angles and noise levels.[]{data-label="chap_8_g75_markov_m_3_measures"}](chap_8_g75_markov_m_3_measures.jpg)
![The measure $M_4$ for the Markov Chain for various angles and noise levels.[]{data-label="chap_8_g75_markov_m_4_measures"}](chap_8_g75_markov_m_4_measures.jpg)
![The measure $M_5$ for the Markov Chain for various angles and noise levels.[]{data-label="chap_8_g79_markov_m_5_measures"}](chap_8_g79_markov_m_5_measures.jpg)
![The measure $M_6$ for the Markov Chain for various angles and noise levels.[]{data-label="chap_8_g75_markov_m_6_measures"}](chap_8_g75_markov_m_6_measures.jpg)
### Interpretation of the Six Measures Analysis
The six measures $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$ were plotted as a function of the noise level $\epsilon$. The six measures show a regular systematic behaviour in the angle $\phi$. The shape of the graphs of the six measures were very similar for all the angles. As the angle increased from $\phi=0^\circ$ to $\phi=90^\circ$ the six measures gradually tended to being nearly constant in $\epsilon$.
This effect can be explained with the invariant measures. When $\phi=0^\circ$ the probabilities for escaping from left to right $p_{-1+1}$ was different to to the probabilities for escaping from right to left $p_{+1-1}$. But in the $\phi=90^\circ$ case they are the same, that is $$\begin{aligned}
\phi=0^\circ\,\,\, \quad p_{-1+1}&\neq p_{+1-1}\\
\phi=90^\circ \quad p_{-1+1}&= p_{+1-1}\end{aligned}$$ The is can be understood geometrically. For $\phi=0^\circ$ we have $F_x\neq0$ and $F_y=0$. The two wells in the Mexican Hat potential move up and down and are alternating with each other. When one well is high the other is low. For $\phi=90^\circ$ we have $F_x=0$ and $F_y\neq0$. The two wells are always at the same height as each other and the distance to the saddles (which is a gateway to escape) is also the same in both wells.
Recall our discussions on the Markov Chain in Section \[chapter\_oscil\_times\]. The $p$ is related to left to right escape $p_{-1+1}$ and $q$ was related to right to left escape $p_{+1-1}$. For $\phi=0^\circ$ the Markov Chain can be modelled with $p\neq q$ and for $\phi=90^\circ$ the Markov Chain can be modelled with $p=q$. In the case of $\phi=0^\circ$ the invariant measure was cyclically changing in time. In the case of $\phi=90^\circ$ the invariant measure was constant at $\overline{\nu}_-(\cdot)=\overline{\nu}_+(\cdot)=\frac{1}{2}$. This explains why the six measures $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$ were nearly constant for angle $\phi=90^\circ$. As $\phi$ changed from $\phi=0^\circ$ to $\phi=90^\circ$, the Markov Chain changed from being modelled by $p\neq q $ to being modelled by $p=q$. This explains the change in the six measures tending to being constant in $\epsilon$ as $\phi$ was varied. The six measures can be thought of as a way of measuring how far away the invariant measures are from being constant. If the invariant measures are constant then the six measures will also be constant.[^2]
For fixed $\phi$ near $\phi=90^\circ$ there is no pronounced maximum of any measure for varying $\epsilon$. Hence the six measures indicate the absence of a pronounce stochastic resonance near $\phi=90^\circ$. But consider the trajectories at a range of angles.
![The blue trajectory is $x(t)$ and the green trajectory is $y(t)$.[]{data-label="chap_8_path_p0"}](chap_8_path_p0.jpg)
![The blue trajectory is $x(t)$ and the green trajectory is $y(t)$.[]{data-label="chap_8_path_p84"}](chap_8_path_p84.jpg)
![The blue trajectory is $x(t)$ and the green trajectory is $y(t)$.[]{data-label="chap_8_path_p90"}](chap_8_path_p90.jpg)
When $\phi=0^\circ$ the $x(t)$ show quasi-deterministic behaviour. The transitions are very regular and $y(t)$ fluctuates around zero. As the angle varies the transitions become less regular and $y(t)$ starts to oscillate. This suggest that there is some regularity in the behaviour of the trajectories but the six measures are not detecting it. Further studies with the escape times would tell us more.
Escape Time and Conditional KS Test Analysis
--------------------------------------------
We remind ourselves of the PDF of escape times and the way the conditional KS test can be applied in our context. The conditional PDF of the escape times are $$\begin{aligned}
p_{-}(t,u)&=R_{-1+1}(t)\exp\left\{-\int^t_uR_{-1+1}(s)\,ds\right\}\\[0.5em]
p_{+}(t,u)&=R_{+1-1}(t)\exp\left\{-\int^t_uR_{+1-1}(s)\,ds\right\}\end{aligned}$$ where $R_{-1+1}$ and $R_{+1-1}$ are the Kramers escape rate from left to right and right to left. In the case of $p_-(t,u)$, $t$ is the time coordinate of escape from the left well and $u$ is the time coordinate of entrance into the left well. In the case of $p_+(t,u)$, $t$ is the time coordinate of escape from the right well and $u$ is the time coordinate of entrance into the right well. If we do not differentiate between escaping from the left or right then the PDF for an escape time $t$ is (note that $t$ here is an escape time as it is and not a time coordinate) $$\begin{aligned}
p_{tot}(t)=
\frac{1}{2}
\int_{0}^{T}p_-(t+u,u)m_-(u)+p_+(t+u,u)m_+(u)\,du\end{aligned}$$ where $m_-(\cdot)$ and $m_+(\cdot)$ are PDFs of the time of entrance into the left and right well respectively. We do not have explicit expressions for $m_-(\cdot)$ and $m_+(\cdot)$. The $p_{tot}(t)$ is approximated by $$\begin{aligned}
p_{tot}(t)\approx p_+(t,0)\end{aligned}$$ The times it took to escape from both the left or right wells are plotted in histograms. This is an empirical approximation to the PDF $p_{tot}(t)\approx p_+(t,0)$. A selection of some of the results are given below for various angles of the forcing $\phi$ and noise level $\epsilon$. They are examples of the Singles, Intermediate and Double Frequencies which we will explain later. Note that the escape times are given in units of normalised time, which is in the number of periods $T$.
![This is an example of the Single Frequency. []{data-label="chap_8_g75_p0_e18_pdf"}](chap_8_g75_p0_e18_pdf.jpg)
![ This is an example of the Intermediate Frequency. []{data-label="chap_8_g75_p84_e20_pdf"}](chap_8_g75_p84_e20_pdf.jpg)
![This is an example of the Double frequency.[]{data-label="chap_8_g75_p90_e21_pdf"}](chap_8_g75_p90_e21_pdf.jpg)
It is important to note that Figures \[chap\_8\_g75\_p0\_e18\_pdf\], \[chap\_8\_g75\_p84\_e20\_pdf\] and \[chap\_8\_g75\_p90\_e21\_pdf\] are histograms of the actual times it took to escape from either wells without differentiation between wells on the left or right. The times of entrance into the wells are not shown. The PDF used is $p_{tot}(\cdot)$ which is being approximated by $p_{tot}(t)\approx p_+(t,0)$.
These escape times can be analysed in a different way. Let $u$ be the time of entrance into a well and $t$ the time of exit from a well. Figures \[chap\_8\_g75\_p0\_e18\_pdf\], \[chap\_8\_g75\_p84\_e20\_pdf\] and \[chap\_8\_g75\_p90\_e21\_pdf\] are therefore histograms of the $(t-u)$ for both left and right escapes combined. Thus $0\leq mod(u,T)\leq1$ is the phase of entrance into a well and $mod(t-u,T)$ is the escape time itself in normalised time. Such an analysis is done for the times in Figure \[chap\_8\_g75\_p0\_e18\_pdf\] for both the left and right wells respectively.
![The $u$ is the time of entrance into the well and $t$ is the time of exit from the well.[]{data-label="chap_8_scatter_g75_p0_e18_pmaT"}](chap_8_scatter_g75_p0_e18_pmaT.jpg)
![The $u$ is the time of entrance into the well and $t$ is the time of exit from the well.[]{data-label="chap_8_scatter_g75_p0_e18_ppaT"}](chap_8_scatter_g75_p0_e18_ppaT.jpg)
Notice the general behaviour of the data for $mod(u,T)$ and $mod(t-u,T)$. For the $\phi=0^\circ$ case the wells are alternating and one well is higher than the other. Entrance into the left well tend to occur near $u=0.5$ and entrance into the right well tend to occur near $u=0$ and $u=1$. For $\phi=90^\circ$ the wells are synchronised and are always at the same height as each other. Entrance and exit to and from either well tend to occur at $u=0$, $u=0.5$ and $u=1$.
Notice also in Figure \[chap\_8\_scatter\_g75\_p0\_e18\_pmaT\] the data points are tiled near $0.5$. This seems to suggest that the use of the Dirac delta function to approximate $p_{tot}\approx p_+(t,0)$ (see Section \[chap\_approx\_pdf\]) may not be very good. The main problem here is the fact that we do not have an explicit formula for a probability measure of the time of entrance into a well, that is we do not have expressions for $m_-(u)$ and $m_+(u)$. This motivates us into developing the conditional KS test.
We want to test whether the escape times we have measured are really distributed by the conditional PDFs $p_-(t,u)$ and $p_+(t,u)$. This is testing the conditional null hypothesis. Define the conditional CDFs by $$\begin{aligned}
F^-_u(t)&=\int_u^tp_-(s,u)\,ds=1-\exp\left\{-\int_u^tR_{-1+1}(s)\,ds\right\}\\[0.5em]
F^+_u(t)&=\int_u^tp_-(s,u)\,ds=1-\exp\left\{-\int_u^tR_{-1+1}(s)\,ds\right\}\end{aligned}$$ The time coordinates of the entrance and exit from the wells are collected. These are $$\begin{aligned}
\left(
\begin{array}{cccc}
u_1&u_2&\ldots&u_n\\
t_1&t_2&\ldots&t_n
\end{array}
\right)\end{aligned}$$ where $u_i$ is the time coordinate of the $i$th entrance into a well and $t_i$ is the time coordinate of the $i$th exit from a well. The conditional KS statistic is calculated by $$\begin{aligned}
S_n^-=\sup_{x\in[0,1]}
\left\Vert
\frac{1}{n}
\sum_{i=1}^n\mathbf{1}_{[0,x]}
\left(
F^-_{u_i}(t_i)
-x
\right)
\right\Vert
\quad \text{and} \quad
S_n^+=\sup_{x\in[0,1]}
\left\Vert
\frac{1}{n}
\sum_{i=1}^n\mathbf{1}_{[0,x]}
\left(
F^+_{u_i}(t_i)
-x
\right)
\right\Vert\end{aligned}$$ where in $S_n^-$ we sum over the time coordinates of entrance and exit to and from the left well and in $S_n^+$ we sum over the time coordinates of entrance and exit to and from the right well. Recall that if the conditional null hypothesis is true then $S_n^-$ and $S_n^+$ are asymptotically distributed by $$\begin{aligned}
\lim_{n\longrightarrow \infty} P(\sqrt{n} S_n \leq x)=Q(x)
\quad \text{where} \quad
Q(x)=1-2\sum_{k=1}^\infty (-1)^{k-1}e^{-2k^2x^2}\end{aligned}$$ We want 99% confidence. Note that $$\begin{aligned}
P\left(
\sqrt{n}S_n\leq 1.6920
\right)
=Q(1.6920)
=0.99\end{aligned}$$ The $Q\left(\sqrt{n}S_n\right)$ is also calculated. The smaller $Q\left(\sqrt{n}S_n\right)$ is the more certain we are in accepting the null hypothesis. A selection of some of the data being implemented with the conditional KS test are given below for various angles of the forcing $\phi$ and noise level $\epsilon$. These are examples of the KS test being implemented for the histograms of escape times just given in Figures \[chap\_8\_g75\_p0\_e18\_pdf\] and \[chap\_8\_g75\_p90\_e21\_pdf\]
![ This is an example of the conditional KS test being implemented for the data in Figure \[chap\_8\_g75\_p0\_e18\_pdf\]. Note that $\epsilon=0.18$, $\phi=0^\circ$, $n=200$, $\sqrt{n}S^-_n=0.5233$ and $Q\left(\sqrt{n}S^-_n\right)=0.0529$. ](chap_8_g80_p0_e18_m1.jpg)
![ This is an example of the conditional KS test being implemented for the data in Figure \[chap\_8\_g75\_p90\_e21\_pdf\]. Note that $\epsilon=0.21$, $\phi=90^\circ$, $n=200$, $\sqrt{n}S^-_n=1.0465$ and $Q\left(\sqrt{n}S^-_n\right)=0.7766$. []{data-label="chap_8_g80_p90_e21_m1"}](chap_8_g80_p90_e21_m1.jpg)
### Interpretation of the Escape Time and Conditional KS Test Analysis
When $\phi=0^\circ$ there were peaks in the empirical PDF of the escape times. These occurred at times $\frac{1}{2}T$, $\frac{3}{2}T$, $\frac{5}{2}T$, …. This effect we call the Single frequency. When $\phi=90^\circ$ the peaks occurred at $\frac{1}{2}T$, $\frac{3}{2}T$, $\frac{5}{2}T$, …and $0$, $T$, $2T$, $3T$, $4T$, …. This effect we call the Double Frequency. When $0^\circ<\phi<90^\circ$ an intermediate effect is seen. There were major peaks at $\frac{1}{2}T$, $\frac{3}{2}T$, $\frac{5}{2}T$, …and minor peaks at $0$, $T$, $2T$, $3T$, $4T$.
The behaviour of the Single, Intermediate and Double Frequencies can be explained geometrically. When the height between a well and a saddle is minimum, the optimal probability of escape has occurred. When $\phi=0^\circ$ the frequency of the return of the optimal probability of escape is the same as the driving frequency $\Omega$. This optimal probability comes back very $T$ which is once in a period. When $\phi=90^\circ$ the frequency of the return of the optimal probability of escape is the double the driving frequency at $2\Omega$. This optimal probability comes back very $\frac{T}{2}$ which is twice in a period. This explains why the peaks in the Single and Double Frequencies are seen where they are.
As the angle changed from $\phi=0^\circ$ to $\phi=90^\circ$ the Single Frequency gradually changes into the Double Frequency with the Intermediate Frequency seen in between. Thus the angle of the forcing is leaving a mark in the PDFs of escape times.
When the conditional KS test was implemented, the functions $$\begin{aligned}
y_0(x)=x,
\quad
y_-(x)=\sum_{i=1}^n\mathbf{1}_{[0,x]}
\left(
F^-_{u_i}(t_i)
\right)
\quad \text{and} \quad
y_+(x)=\sum_{i=1}^n\mathbf{1}_{[0,x]}
\left(
F^+_{u_i}(t_i)
\right)\end{aligned}$$ were used to calculate the following distances which are the conditional KS statistics $$\begin{aligned}
S^-_n=\left\Vert y_0-y_-\right\Vert_\infty
\quad \text{and} \quad
S^+_n=\left\Vert y_0-y_+\right\Vert_\infty\end{aligned}$$ It is reasonable to say that $y_-(\cdot)$ and $y_+(\cdot)$ were close enough to $y_0(\cdot)$ that we can accept the conditional null hypothesis. This can be seen and judged graphically with $S^-_n$ and $S^+_n$ calculated as well. This is an example of the conditional KS test giving a reasonable result.
Remarks on Analysis of Stochastic Resonance
-------------------------------------------
There are a few subtleties, heavily dependent approximations and setbacks to the analysis which is worth mentioning here.
### Remarks on Implementing the Conditional KS Test
Notice that all the theories developed about the KS Test were based on the assumption that the null hypothesis is true. This means strictly speaking a small KS statistic, that is a small $S_n^-$ or $S_n^+$ does not immediately allow us to accept the null hypothesis but good reasons not to reject it. Also when there were many transitions, that is for large $n$, the terms $Q(\sqrt{n}S_n^-)$ and $Q(\sqrt{n}S_n^+)$ were also calculated. The smaller $Q(\sqrt{n}S_n^-)$ and $Q(\sqrt{n}S_n^+)$ are the more confidence we have in not rejecting the null hypothesis. This is because for very large $n$, we would expect $$\begin{aligned}
\lim_{n\rightarrow\infty}\sqrt{n}S_n^-=0
\quad \text{and} \quad
\lim_{n\rightarrow\infty}\sqrt{n}S_n^+=0\end{aligned}$$ so the smaller $Q(\sqrt{n}S_n^-)$ and $Q(\sqrt{n}S_n^+)$ are the more certain we are in not rejecting the null hypothesis. Note that we have used $n=200$ transitions for implementing the KS test. This still works with sparse data. For examples with few data say $n=20$ see the thesis [@tommy_thesis].
### Remarks on Adiabatic Approximation
Notice that in the PDFs $p_-(t,u)$, $p_+(t,u)$ and $p_{tot}(t)$ expressions for the escape rates $R_{-1+1}(t)$ and $R_{+1-1}(t)$ were required. These rates were also required for the conditional KS test. Strictly speaking these rates are dependent on the driving frequency $\Omega$, but we stress that these rates were calculated using Kramers formula as though the particle is escaping from a static potential. This is the adiabatic approximation where an oscillatory potential is approximated by a static potential.
It is worth summarising all the approximations which the analysis of the data have been based. There is the small noise approximation and slow forcing approximation from Kramers formula, the adiabatic approximation and the perfect phase approximation where $p_{tot}$ is approximated by $p_{tot}\approx p_{tot}(t,0)$.
Conclusion
==========
In this paper we have considered the following problem. Let $X^\epsilon_t$ be a stochastic process in $\mathbb{R}^2$ which is described by the the SDE $$\begin{aligned}
dX^\epsilon_t=b\left(X^\epsilon_t,t\right)dt+\epsilon\,dW_t\end{aligned}$$ and the drift term $b(\cdot,\cdot)$ is expressed by $$\begin{aligned}
b(x,t)=-\nabla V_0 (x) + F\cos \Omega t \end{aligned}$$ where $V_0:\mathbb{R}^2\longrightarrow \mathbb{R}$ is a time independent function, the unperturbed potential, with two metastable states, and two pathways between these states. The $F\in\mathbb{R}^r$ is the magnitude of the forcing and $\Omega$ is the driving frequency. Our aim was to see characteristics of the trajectory $X^\epsilon_t$ which only depends on the qualitative structure of $V_0$, that is the existence of two metastable states and two pathways.
For concreteness we considered a model, which we call the Mexican Hat Toy Model $$\begin{aligned}
V_0(x,y)=\frac{1}{4}r^4-\frac{1}{2}r^2-ax^2+by^2
\quad \text{where} \quad r=\sqrt{x^2+y^2}\end{aligned}$$ The magnitude and angle of the forcing are given by $$\begin{aligned}
F=\sqrt{F_x^2+F_y^2}
\quad \text{and} \quad
\phi=\tan^{-1}\left(\frac{F_y}{F_x}\right)\end{aligned}$$ The angle $\phi$ and noise level $\epsilon$ were varied. At $\phi=0$ the wells were alternating, that is one well is higher than the other, in the sense that it is easer to jump from one well to the other than vice versa. At $\phi=90^\circ$ the wells are synchronised, that is both wells are always at the same height but the heights of the barrier for the two paths is alternating.
A potential with two pathways has never been considered before in the context of stochastic resonance. We studied it using approximation techniques and direct simulations. In an adiabatic regime the Freidlin-Wentzell theory allows one to give analytical solutions of the jump type distributions asymptotically in this regime. This theory predicted the appearance of additional resonance peaks at half the frequency when the angle approaches $\phi=90^\circ$.
We simulated $X_t^\epsilon$ for different values of $\phi$ and $\epsilon$ and computed for the values of angle increasing from $\phi=0$ to $\phi=90^\circ$ the six measures $M_1$, $M_2$, $M_3$, $M_4$, $M_5$ and $M_6$ as function of the noise level. The first major surprise was that the graphs showed less and less pronounced minima (or maxima) and hence suggests that the phenomena of stochastic resonance gets less and less pronounced, see Section \[conclusion\_six\_measures\]. The effect of resonance seems to disappear overall.
However, considering the path $X_t^\epsilon$ itself, one sees that there may be nevertheless some synchronisation, see Figure \[chap\_8\_path\_p0\], \[chap\_8\_path\_p84\] and \[chap\_8\_path\_p90\]. To properly quantify synchronisation we considered the histograms of the escape times, which to our knowledge has not been considered thoroughly before. The histograms showed a clear periodicity and also the emergence of peaks at the Double Frequency for increasing angle. For a quantitative consideration we assume that the entrance time is in perfect phase (this is when $m_-(u)$ and $m_+(u)$ *can* be approximated by Dirac delta functions). This gives for several cases good quantitative and in general good qualitative agreement with the combined adiabatic and small noise approximation. Summarizing, the theoretical and the simulation results are in very good agreement. We want to stress that in the comparison no free parameters were present and so no fitting took place.
The fact that the six measures are blind can be explained using Markov chain models approximating the SDE. As one expects from large deviation theory, for small noise and in an adiabatic regime the SDE can be approximated by a continuous time Markov chain. In this Markov chain model we showed that the invariant measures are constant when $\phi=90^\circ$. Hence we expect that the invariant measure gives in the diffusion case equal weights to the left and the right well. Together, this gives us the following qualitative picture of the dynamics for any angle. At a fixed time the probability that one sees a jump from the left to the right well or vice versa has the same probability. However, conditioned on the phase and the direction of the last jump, for concreteness assume that it was at phase $u$ and from the left to the right (that is to say the particle entered the well at time $u$) the next jump will be at phase which is near to a multiple of $T/2$ (that is to say the particle will leave the well near the times $t=nT/2$ where $n$ is an integer). The jump rates will be given by the height of the potential barriers.
At $\phi=90^\circ$, the path $X_t^\varepsilon$ and $-X_t^\varepsilon$ will appear with the same probability if one starts in the invariant measure. This explains why the six measures are all insensitive in this case. The equilibration happens because the process will skip some of the jump opportunities and in this way the left-right synchronization will get lost quickly.
This new phenomena we discovered has added an additional motivation to the observation of Hermann, Imkeller, Pavlyukevich, Berglund and Gentz that the appropriate consideration has to be on the path level. Averaged quantities like the six measures can be very misleading and masking the real behaviour of the system. The escape time distribution shows a clear signal of stochastic resonance in accordance with the theoretical consideration. The presence of a two pathways manifests itself in an appearance of peaks at the Double Frequency. We showed that adiabatic small noise approximation gives a good statistical model. We demonstrated that this appearance can be detected also when only a limited number of transitions is available. Our analysis provides us with a clear footprint indicating the existence of a second pathway. The angle dependence of our result should also allow us to predict the orientation of the saddles with respect to the wells.
Numerical Methods for calculating $M_5$ and $M_6$ {#appendix_six_measures}
=================================================
Here we present how we computed $M_5$ and $M_6$ numerically. This is how $M_5$ and $M_6$ are calculated in theory $$\begin{aligned}
M_5&=\int_0^T
\phi^-(t)\ln\left(\frac{\phi^-(t)}{\overline{\nu}_-(t)}\right)+
\phi^+(t)\ln\left(\frac{\phi^+(t)}{\overline{\nu}_+(t)}\right)
dt\\
M_6&=\int^T_0
-\overline{\nu}_-(t)\ln\overline{\nu}_-(t)
-\overline{\nu}_+(t)\ln\overline{\nu}_+(t)\,
dt\end{aligned}$$ where $$\begin{aligned}
\phi^-(t)&=
\left\{
\begin{array}{c}
1 \quad \text{if} \quad mod(t,T)\leq T/2\\
0 \quad \text{if} \quad mod(t,T)> T/2
\end{array}
\right.\\[0.5em]
\phi^+(t)&=
\left\{
\begin{array}{c}
0 \quad \text{if} \quad mod(t,T)\leq T/2\\
1 \quad \text{if} \quad mod(t,T)> T/2
\end{array}
\right.\end{aligned}$$ When the invariant measures are generated numerically they are finite discrete objects described by $$\begin{aligned}
\nu_-&=\left\{\nu^-_1, \nu^-_2, \cdots, \nu^-_N\right\}\\
\nu_+&=\left\{\nu^+_1, \nu^+_2, \cdots, \nu^+_N\right\}\end{aligned}$$ The real invariant measure were close to zero sometimes and in the numerical approximation they became actually zero or even negative which lead to numerical artefacts. Note that $$\begin{aligned}
\lim_{x\longrightarrow0}\ln\left(\frac{1}{x}\right)=\infty
\quad \text{and} \quad
\lim_{x\longrightarrow0}x\ln\left(x\right)=0\end{aligned}$$ Define $$\begin{aligned}
\nu_-^{lim}&=\min_{\substack{i=1,2,\ldots,N\\\nu^-_i>0}}\left\{\nu^-_1, \nu^-_2, \cdots, \nu^-_N\right\}\\
\nu_+^{lim}&=\min_{\substack{i=1,2,\ldots,N\\\nu^+_i>0}}\left\{\nu^+_1, \nu^+_2, \cdots, \nu^+_N\right\}\end{aligned}$$ The quantities $M_5$ and $M_6$ are computed numerically in the following way $$\begin{aligned}
M_5&=\sum_{\substack{i\leq\frac{N}{2}\\\nu^-_i>0}}t_{step}\ln\left(\frac{1}{\nu^-_i}\right)
+
\sum_{\substack{i\leq\frac{N}{2}\\\nu^-_i\leq0}}t_{step}\ln\left(\frac{1}{\nu_-^{lim}}\right)
+
\sum_{\substack{i>\frac{N}{2}\\\nu^+_i>0}}t_{step}\ln\left(\frac{1}{\nu^+_i}\right)
+
\sum_{\substack{i>\frac{N}{2}\\\nu^+_i\leq0}}t_{step}\ln\left(\frac{1}{\nu^{lim}_+}\right)\\[0.5em]
M_6&=\sum_{\substack{i=1,2,\cdots,N\\\nu^-_i>0}}\nu^-_i\ln(\nu^-_i)(-t_{step})\quad+\quad
\sum_{\substack{i=1,2,\cdots,N\\\nu^+_i>0}}\nu^+_i\ln(\nu^+_i)(-t_{step})\end{aligned}$$
[^1]: University of Reading [email protected]
[^2]: See Appendix \[appendix\_six\_measures\] for how $M_5$ and $M_6$ were numerically calculated. The ideas were not that trivial.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The Duru-Kleinert (DK) method of solving unknown path integrals of quantum mechanical systems by relating them to known ones does not apply to Markov processes since the DK-transform of a Fokker-Planck equation is in general not a Fokker-Planck equation. In this note, we present a significant modification of the method, based again on a combination of path-dependent time and coordinate transformations, to obtain such relations after all. As an application we express unknown Green functions for a one-parameter family of Markov processes in terms of the known one for the Schenzle-Brand process.\
address: 'Institut für Theoretische Physik, Freie Universität Berlin, Arnimallee 14, 14195 Berlin, Germany'
author:
- 'A. Pelster and H. Kleinert[^1]'
title: Relations Between Markov Processes Via Local Time and Coordinate Transformations
---
\#1
[**1.**]{} The stochastic theory of Markov processes [@Stratonovich; @Kampen] explains many phenomena where fluctuations play a significant role. Prominent examples are provided by the emergence of self-organization [@Haken1], the occurrence of quantum dissipation [@Weiss] and the appearance of stochastic resonance [@Haengi]. In accordance with their wide range of applicability there exist various powerful solution methods for Markov processes based either on the global characterization of the probability evolution by the Onsager-Machlup path integral or on its local equivalent, the Fokker-Planck equation. Examples for the former are the small noise expansion and the adiabatic elimination procedure of fast random variables, for the latter the eigenfunction expansion and the continued-fraction method applied to periodically driven systems [@Risken].
In this note we demonstrate the use of path-dependent time transformations, which have become a powerful tool for solving quantum mechanical problems since Duru and Kleinert’s original work on the path integral of the hydrogen atom [@Duru1; @Kleinert]. In one dimension with coordinate $q$, the crucial time transformation has the form $$\begin{aligned}
\label{TT}
\frac{d t}{d s} = f ( q ) \, ,\end{aligned}$$ where $f ( q )$ is some positive but otherwise arbitrary function. Such a transformation does not change the standard formulations of quantum mechanics, a form invariance which has recently been emphasized in [@Pelster1; @Pelster2]. Due to this form invariance different quantum mechanical systems can be related to each other.
Some years ago, the DK-method was also applied to the stochastic theory of Markov processes [@Blanchard; @Morette], thereby relating the Fokker-Planck equation of Markov processes to other stochastic differential equations. The latter had, however, an important disadvantage: they were no longer Fokker-Planck equations so that the DK-transformations did not link different Markov processes. This defect will be eliminated in the sequel by a significant modification of the DK-method.
\
[**2.**]{} Consider a one-dimensional Markov process of a single random variable $x$, whose conditional probability density $P(x,x_0;t)$ possesses the initial condition $$\begin{aligned}
P ( x , x_0 ; 0 ) = \delta ( x - x_0 )\end{aligned}$$ and obeys the Fokker-Planck equation for $t > 0$ $$\partial_t P(x,x_0;t)= \hat{H} ( x ) P ( x , x_0 ; t ) \, .
\label{}$$ Here $\hat{H} ( x )$ denotes the infinitesimal time evolution operator $$\begin{aligned}
\hat{H} ( x ) \bullet =
- \frac{\partial}{\partial x}
\left[ K ( x ) \bullet \right]
+ \frac{1}{2} \frac{\partial^2}{\partial x^2}
\left[ D ( x ) \bullet \right],
\label{EQ1B}\end{aligned}$$ containing the drift coefficient $K ( x )$ and the diffusion coefficient $D ( x )$. Then the Laplace transform $$\begin{aligned}
G(x,x_0;E) = \int\limits_0^{\infty} dt e^{-Et}P(x,x_0;t)\end{aligned}$$ represents a fixed-energy Green function solving the time-independent equation $$\begin{aligned}
&& \left[ \hat{H}^{({\rm i})} ( x ) - E^{({\rm i})} \right]
G^{({\rm i})} ( x , x_0 ; E^{({\rm i})} )
+ \delta ( x - x_0 ) = 0 \, . \label{EQ1A}\end{aligned}$$ We have added a superscript $({\rm i})$ to emphasize that this equation describes the initial stochastic system from which we depart.
It is well-known [@Risken] that the Fokker-Planck equation remains form invariant under arbitrary invertible coordinate transformations $$\begin{aligned}
\label{RT}
x = x ( q ) \, ,\end{aligned}$$ where the Green function transforms as $$\begin{aligned}
\label{P1}
G^{({\rm 1})} ( q , q_0 ; E^{({\rm i})} ) = \pm \, x' ( q )
G^{({\rm i})} ( x ( q ) , x ( q_0 ) ; E^{({\rm i})} ) \, .\end{aligned}$$ The different signs take into account whether (\[RT\]) is monotonously in- or decreasing. In fact, we conclude from (\[EQ1B\]) and (\[EQ1A\]) that $G^{({\rm 1})} ( q , q_0 ; E^{({\rm i})} )$ satisfies the Fokker-Planck equation $$\begin{aligned}
&& \left[ \hat{H}^{({\rm 1})} ( q ) - E^{({\rm i})}
\right] G^{({\rm 1})} ( q , q_0 ; E^{({\rm i})} ) + \delta ( q - q_0 )
= 0 \label{EQ2A} ,\end{aligned}$$ where the new infinitesimal time evolution operator $$\begin{aligned}
\hat{H}^{({\rm 1})} ( q ) \bullet = -
\frac{\partial}{\partial q}
\left[ K^{({\rm 1})} ( q ) \bullet \right]
+ \frac{1}{2} \frac{\partial^2}{\partial q^2}
\left[ D^{({\rm 1})} ( q ) \bullet \right]
\label{EQ2B}\end{aligned}$$ contains the drift and diffusion coefficients $$\begin{aligned}
K^{({\rm 1})} ( q ) & = & \frac{1}{x' ( q )} K^{({\rm i})} ( x ( q ) ) -
\frac{x'' ( q )}{2 x'{}^3 ( q )} D^{({\rm i})} ( x ( q ) ) \, , \label{DR1}\\
D^{({\rm 1})} ( q ) & = & \frac{1}{x' {}^2( q )} D^{({\rm i})} ( x ( q ) ) \, .
\label{DI1}\end{aligned}$$ These coordinate transformations are of standard use in finding unknown solutions from known ones [@Haken2; @Schenzle].
Let us now supplement these transformations by the path-dependent time transformation (\[TT\]). First we proceed in analogy with [@Kleinert Chap. 12], and change the Green function according to $$\begin{aligned}
G^{({\rm 1})} ( q , q_0 ; E^{({\rm i})} ) & =& f ( q ) \,
\frac{F ( q_0 ;
E^{({\rm i})} )}{F ( q ; E^{({\rm i})} )} \nonumber\\
& & \times
G^{({\rm f})} ( q , q_0 ; E^{({\rm f})} ( E^{({\rm i})} ) ) ,
\label{P2}\end{aligned}$$ where $F(q ; E^{({\rm i})} )$ and $E^{({\rm f})} ( E^{({\rm i})} )$ are as yet unknown trial functions. Applying (\[EQ2A\]) with (\[EQ2B\]), we find the equation for the final Green function $$\begin{aligned}
&& \left[ \hat{H}^{({\rm f})} ( q ) - E^{({\rm f})} ( E^{({\rm i})} ) +
X ( q ; E^{({\rm i})} ) \right]
G^{({\rm f})} ( q , q_0 ; E^{({\rm f})}
( E^{({\rm i})} ) ) \nonumber\\
&&+ \, \delta ( q - q_0 )
=0 ,
\label{EQ3A}\end{aligned}$$ with the infinitesimal time evolution operator $$\begin{aligned}
\hat{H}^{({\rm f})} ( q ) \bullet = -
\frac{\partial}{\partial q}
\left[ K^{({\rm f})} ( q ) \bullet \right]
+ \frac{1}{2} \frac{\partial^2}{\partial q^2}
\left[ D^{({\rm f})} ( q ) \bullet \right] \, ,
\label{EQ3B}\end{aligned}$$ containing the transformed drift and diffusion coefficients $$\begin{aligned}
K^{({\rm f})} ( q ) & = & f ( q ) \left[ K^{({\rm 1})} ( q ) +
\frac{F' ( q ; E^{({\rm i})} )}{F ( q ; E^{({\rm i})} )}
D^{({\rm 1})} ( q ) \right] \, , \label{DR2}\\
D^{({\rm f})} ( q ) & = & f ( q ) D^{({\rm 1})} ( q ) \label{DI2}\end{aligned}$$ and the additional term $$\begin{aligned}
&&X ( q ; E^{({\rm i})} ) = f ( q ) \left[ \frac{1}{2}
D^{({\rm 1})} ( q ) \frac{F'' ( q ; E^{({\rm i})} )}{F ( q ; E^{({\rm i})} )}
\right. \nonumber \\
&& \left. + K^{({\rm 1})} ( q )
\frac{F' ( q ; E^{({\rm i})} )}{F ( q ; E^{({\rm i})} )}
+ {\cal E} ( q ;E^{({\rm i})}) - E^{({\rm i})}
\right] \, ,\label{EQQ}\end{aligned}$$ where $$\begin{aligned}
{\cal E} ( q ;E^{({\rm i})}) = \frac{
E^{({\rm f})}(E^{({\rm i})})}{f(q)} \, .
\label{EQQ2}\end{aligned}$$ The equation (\[EQ3A\]) has the above-mentioned defect of not being a Fokker-Planck equation, due to the presence of the additional term $X ( q ; E^{({\rm i})} )$. This term can, however, be removed by choosing any functions $E^{({\rm f})} ( E^{({\rm i})} )$ and $F ( q ; E^{({\rm i})})$ which solve the differential equation $$X ( q ;
E^{({\rm i})} ) \equiv 0.
\label{EQ}$$ Note that although this equation is of the same complexity as the initial Fokker-Planck equation, only a [*particular solution*]{} is required, so that labor will definitely be saved by our method.
An alternative procedure which avoids solving the differential equation (\[EQ\]) is by leaving the time transformation function $f(q)$ open, choosing some trial function $F ( q ; E^{({\rm i})})$, and calculating ${\cal E}(q ;
E^{({\rm i})} )$ from (\[EQQ\]) with (\[EQ\]). If this happens to be factorizable as in (\[EQQ2\]), the $q$-dependent prefactor may be chosen as the transformation function $f(q)$ and the $E^{({\rm i})}$-dependent one as the energy function $E^{({\rm f})} ( E^{({\rm i})} )$.
In either procedure, the function $F ( q ; E^{({\rm i})})$ is subject to an important restriction. In the limit $E^{({\rm i})}
\rightarrow 0$ it has to satisfy $$\begin{aligned}
\label{LI0}
\lim_{E^{({\rm i})} \rightarrow 0}
F ( q ; E^{({\rm i})} ) = 1 \end{aligned}$$ identically in $q$, so that the energy function $E^{({\rm f})} ( E^{({\rm i})} )$ obeys $$\begin{aligned}
\label{LI}
\lim_{E^{({\rm i})} \rightarrow 0} E^{({\rm f})} ( E^{({\rm i})} )
= 0 \,.\end{aligned}$$ Only under this condition do initial and final Green functions possess proper stationary limits $$\begin{aligned}
p^{({\rm i})}_{\rm st} ( x ) & = & \lim_{E^{({\rm i})} \rightarrow 0}
E^{({\rm i})} G^{({\rm i})} ( x , x_0 ; E^{({\rm i})} ) \, ,
\label{ST1}\\
p^{({\rm f})}_{\rm st} ( q ) & = & \lim_{E^{({\rm f})} \rightarrow 0}
E^{({\rm f})} G^{({\rm f})} ( q , q_0 ; E^{({\rm f})} ) \, .
\label{ST2}\end{aligned}$$ >From (\[P1\]), (\[P2\]) and (\[LI0\])–(\[ST2\]) we read off a relation between them $$\begin{aligned}
\label{STR}
p^{({\rm i})}_{\rm st} ( x ) = \pm \,
\left[ \frac{d E^{({\rm f})} (
E^{({\rm i})} )}{d E^{({\rm i})}} \right]^{-1}_{E^{({\rm i})} = 0} \,
\frac{f ( q ( x ) )}{
x' ( q ( x ) )} \,
p^{({\rm f})}_{\rm st} ( q ( x ) ) \, ,\end{aligned}$$ which guarantees the normalization of the probability: $$\int \, p^{({\rm i})}_{\rm st} ( x ) \, d x
=\int \, p_{\rm st}^{({\rm f})} ( q )\, d q =1
\, .$$ An interesting feature of the present method is that it permits us in the stationary limit to relate the probability distributions of two [*arbitrary*]{} Markov processes to each other. Given initial and final drift and diffusion coefficients $K ( x )$ and $D ( x )$, we satisfy (\[DI1\]) and (\[DI2\]) by choosing the time transformation function as $$\begin{aligned}
\label{STR1}
f ( q ) =x'{}^2 ( q ) \frac{D^{({\rm f})} ( q )}{D^{({\rm i})}
( x ( q ) )} \, .\end{aligned}$$ Using this together with (\[DR1\]), (\[DI1\]), (\[DR2\]), (\[DI2\]) and (\[LI0\]), we obtain the desired coordinate transformation from the differential equation $$\begin{aligned}
\label{STR2}
x' ( q ) = C \exp \left[ \int\limits^{x ( q )} d \tilde{x}
\frac{2 K^{({\rm i})} ( \tilde{x} )}{D^{({\rm i})} ( \tilde{x} ) }
- \int\limits^q d \tilde{q} \frac{2 K^{({\rm f})} ( \tilde{q} )}{D^{({\rm f})}
( \tilde{q} )} \right] \, ,\end{aligned}$$ where $C$ is an integration constant. The final stationary solution $$\begin{aligned}
p^{({\rm f})}_{\rm st} ( q ) = \frac{N^{({\rm f})} }{D^{({\rm f})} ( q )}
\exp \left[ \int\limits^q d \tilde{q}
\frac{2 K^{({\rm f})} ( \tilde{q} )}{D^{({\rm f})} ( \tilde{q} )}
\right]\end{aligned}$$ is then related to the initial one $$\begin{aligned}
p^{({\rm i})}_{\rm st} ( x ) = \frac{N^{({\rm i})} }{D^{({\rm i})} ( x )}
\exp \left[ \int\limits^x d \tilde{x}
\frac{2 K^{({\rm i})} ( \tilde{x} )}{D^{({\rm i})} ( \tilde{x} )}
\right] \,\end{aligned}$$ by (\[STR\]), (\[STR1\]) and (\[STR2\]), if the normalization constants satisfy $$\begin{aligned}
C = \frac{N^{({\rm i})} }{N^{({\rm f})} }
\, \left. \frac{d E^{({\rm f})} ( E^{({\rm i})} )}{d
E^{({\rm i})} }\right|_{E^{({\rm i})} = 0} \, .\end{aligned}$$
\
[**3.**]{} In order to demonstrate the applicability of the new transformation method we consider a Markov process with a multiplicative noise for a random variable $x \in ( 0 , \infty )$ where the drift and the diffusion coefficient depend on an arbitrary parameter $\alpha > 0$ as follows: $$\begin{aligned}
\label{IM}
K^{({\rm i})} ( x ) = a^{({\rm i})} x - b^{({\rm i})} x^{2 \alpha + 1} \, ,
\hspace*{0.3cm} D^{({\rm i})} ( x ) = Q^{({\rm i})} x^{2 \alpha + 2} \, .\end{aligned}$$ Performing a transformation of the random variable (\[RT\]) and the time (\[TT\]) with $$\begin{aligned}
\label{XF1}
x ( q ) = q^{\beta} \, , \hspace*{0.5cm} f ( q ) = q^{\gamma} \, ,\end{aligned}$$ the associated differential equation (\[EQ\]) with (\[EQQ\]) is solved by a function $$\begin{aligned}
\label{AN}
F ( q ; E^{({\rm i})} ) = q^{\delta ( E^{({\rm i})} )} \, ,\end{aligned}$$ if the parameters $\beta$ and $\gamma$ are related according to $$\begin{aligned}
\label{XF2}
2 \alpha \beta + \gamma = 0 \,.\end{aligned}$$ The function $\delta ( E^{({\rm i})} )$ and the energy relation $E^{({\rm f})} = E^{(\rm f)} ( E^{({\rm i})} )$ are determined by $$\begin{aligned}
\delta ( E^{({\rm i})} ) & = & \frac{\beta}{a^{({\rm i})} }
E^{({\rm i})} \, , \label{I1} \\
E^{({\rm f})} ( E^{({\rm i})} ) & = & -
\frac{ Q^{({\rm i})} }{2 a^{({\rm i})}{}^2 } E^{({\rm i})}{}^2 +
\left[ \frac{ b^{({\rm i})} }{a^{({\rm i})} } + \frac{
Q^{({\rm i})} }{2 a^{({\rm i})} } \right] E^{({\rm i})} \, .
\label{I2}\end{aligned}$$ Note that (\[AN\]), (\[I1\]) and (\[I2\]) satisfy the correct limits (\[LI0\]) and (\[LI\]).
The transformed drift and diffusion coefficients then follow from (\[DR1\]), (\[DI1\]), (\[DR2\]) and (\[DI2\]) as $$\begin{aligned}
\label{BS}
K^{({\rm f})} ( q ) = a^{({\rm f})} q - b^{({\rm f})} q^{- 2 \alpha \beta
+ 1} \, , \hspace*{0.3cm} D^{({\rm f})} ( q ) = Q^{({\rm f})} q^2,\end{aligned}$$ where $$\begin{aligned}
a^{({\rm f})}& = & - \frac{b^{({\rm i})} }{\beta} +
\frac{Q^{({\rm i})} }{\beta} \left[
\frac{E^{({\rm i})} }{a^{({\rm i})} } - \frac{\beta - 1}{2 \beta}
\right] \, , \label{I3}\\
b^{({\rm f})} & = & - \frac{a^{({\rm i})} }{\beta}
\, , \\
Q^{({\rm f})} & = & \frac{Q^{({\rm i})} }{\beta^2} \, .\label{I4}\end{aligned}$$ These relations supply us with solutions of the Fokker-Planck equation for the one-parameter family of Markov processes (\[IM\]) if we specialize $$\begin{aligned}
\label{SP}
\beta = - \frac{1}{\alpha} \, .\end{aligned}$$ Then the final Markov process (\[BS\]) for a random variable $q \in ( 0 , \infty )$ coincides with the well-understood Schenzle-Brand process [@Schenzle], which is a standard model in nonlinear optics and chemical reaction dynamics. It can be derived as an approximation to a number of different processes by adiabatically eliminating fast random variables in the limit of large external fluctuations. For instance, the Schenzle-Brand process describes the electrical field near a laser threshold, the multiplicative noise being due to inversion fluctuations.
With standard methods [@Risken], the Fokker-Planck equation of the Schenzle-Brand process with (\[BS\]) and (\[SP\]) can be transformed to the Schrödinger equation of the Morse oscillator. As the quantum mechanical Green function of this system has been explicitly calculated from path integrals [@Inomata; @Duru], the Green function of the Schenzle-Brand process is known: $$\begin{aligned}
&& G^{({\rm f})} ( q , q_0 ; E^{({\rm f})} ) =
\frac{\Gamma \left(
\frac{k^{({\rm f})} }{2} + \frac{1}{4} -
\frac{ a^{({\rm f})} }{2 Q^{({\rm f})} } \right) }{b^{({\rm f})}
\Gamma \left( 1 + k^{({\rm f})} \right) }\,
q^{ \frac{ a^{({\rm f})} }{Q^{({\rm f})} } - \frac{5}{2} }
\nonumber \\
&& \times q_0^{- \frac{ a^{({\rm f})} }{Q^{({\rm f})} } - \frac{1}{2} }
\exp \left[ \frac{ b^{({\rm f})} ( q_0^2 - q^2 )}{ 2 Q^{({\rm f})} } \right]
\Bigg\{ W_{\frac{1}{4} +
\frac{ a^{({\rm f})} }{2 Q^{({\rm f})} } ,
\frac{k^{({\rm f})} }{2} }
\left( \frac{b^{({\rm f})} q^2}{Q^{({\rm f})} } \right) \nonumber\\
\label{G1}
&& \times M_{\frac{1}{4} +
\frac{ a^{({\rm f})} }{2 Q^{({\rm f})} } ,
\frac{k^{({\rm f})} }{2} }
\left( \frac{b^{({\rm f})} q_0^2}{Q^{({\rm f})} } \right)
\Theta ( q - q_0 ) +
\left( q \leftrightarrow q_0 \right) \Bigg\} \, , \label{BBSS}\end{aligned}$$ where $k^{({\rm f})} $ denotes the abbreviation $$\begin{aligned}
\label{K}
k^{({\rm f})} =
\sqrt{ \left( \frac{ a^{({\rm f})} }{Q^{({\rm f})}} - \frac{1}{2}
\right)^2
+ \frac{2 E^{({\rm f})} }{Q^{({\rm f})} } }\, .\end{aligned}$$ With the help of the transformation formulas (\[P1\]), (\[P2\]), and taking into account (\[XF1\])–(\[SP\]), we find from this the unknown Green function of the initial Markov processes (\[IM\]): $$\begin{aligned}
&& G^{({\rm i})} ( x , x_0 ; E^{({\rm i})} ) =
\frac{\Gamma \left(
\frac{ E^{({\rm i})} }{ 2 \alpha a^{({\rm i})} } \right) }{a^{({\rm i})}
\Gamma \left( 1 + k^{({\rm i})} \right) }
\,x^{ - \alpha k^{({\rm i})} - \alpha - 1 }
\nonumber \\
&& \times x_0^{\alpha k^{({\rm i})} + \alpha }
\exp \left[ \frac{ a^{({\rm i})} }{ 2 \alpha Q^{({\rm i})} } \left(
\frac{1}{x_0^{2\alpha}} - \frac{1}{x^{2\alpha}} \right) \right]
\nonumber \\
&& \times
\Bigg\{ \Theta ( x_0 - x )
W_{\frac{k^{({\rm i})} }{2} +
\frac{1}{2}
- \frac{E^{({\rm i})} }{2 \alpha a^{({\rm i})} } ,
\frac{ k^{({\rm i})} }{2 } }
\left( \frac{a^{({\rm i})} }{\alpha Q^{({\rm i})} x^{2\alpha} } \right)
\nonumber \\
&& \times
M_{\frac{k^{({\rm i})} }{2}
+ \frac{1}{2}
- \frac{E^{({\rm i})} }{2 \alpha a^{({\rm i})} } ,
\frac{ k^{({\rm i})} }{2} }
\left( \frac{a^{({\rm i})} }{\alpha Q^{({\rm i})} x_0^{2\alpha}} \right)
\label{G2}
+ \left(
x \leftrightarrow x_0 \right) \Bigg\},\end{aligned}$$ with $$\begin{aligned}
k^{({\rm i})} =
\frac{b^{({\rm i})} }{\alpha Q^{({\rm i})} } + \frac{1}{2 \alpha}.\end{aligned}$$ Furthermore we obtain from (\[ST2\]), (\[BBSS\]), (\[K\]) the stationary solution of the Schenzle-Brand process (\[BS\]), (\[SP\]) $$\begin{aligned}
p^{({\rm f})}_{\rm st} ( q ) = \frac{ 2 \left(
\frac{b^{({\rm f})} }{Q^{({\rm f})} }
\right)^{\frac{a^{({\rm f})} }{Q^{({\rm f})} }- \frac{1}{2} } }{
\Gamma \left( \frac{a^{({\rm f})} }{Q^{({\rm f})} }- \frac{1}{2}
\right) }
q^{ 2 \frac{a^{({\rm f})} }{Q^{({\rm f})} } - 2 }
\exp \left[ - \frac{b^{({\rm f})} }{Q^{({\rm f})} } q^2 \right] \, ,\end{aligned}$$ which is mapped via (\[STR\]), (\[XF1\])–(\[SP\]) to the initial Markov process (\[IM\]), yielding $$\begin{aligned}
&& \!\!\!\!\!\!\!\!\!\!p^{({\rm i})}_{\rm st} ( x ) = \frac{2 \alpha \left(
\frac{a^{({\rm i})} }{\alpha Q^{({\rm i})} } \right)^{
\frac{b^{({\rm i})} }{ \alpha Q^{({\rm i})} }+ \frac{1}{2 \alpha} + 1 } }{
\Gamma \left(
\frac{b^{({\rm i})} }{ \alpha Q^{({\rm i})} } + \frac{1}{2 \alpha} + 1
\right) } \nonumber \\
&& \times
x^{- 2 \frac{b^{({\rm i})} }{Q^{({\rm i})} } - 2 \alpha - 2}
\exp \left[ - \frac{a^{({\rm i})} }{\alpha Q^{({\rm i})} x^{2 \alpha} }
\right] \, .\end{aligned}$$ The analytic properties of the Green functions $G^{({\rm i})} ( x , x_0 ; E^{({\rm i})} )$ and $G^{({\rm f})} ( q , q_0 ; E^{({\rm f})} )$ in the energies $E^{({\rm i})}$ and $E^{({\rm f})}$ determine the spectra of the infinitesimal time evolution operators $\hat{H}^{({\rm i})} ( x )$ and $\hat{H}^{({\rm f})} ( q )$ [@Kleinert Chap. 9]. From (\[G1\])–(\[G2\]) we deduce that the initial multiplicative process (\[IM\]) has only a discrete spectrum, whereas the final one (\[BS\]), (\[SP\]) contains both a discrete and a continuous branch. Such differences between spectral types were encountered before in quantum mechanical DK-transformations: the hydrogen atom has discrete and continuous states whereas the DK-equivalent oscillator has only discrete states [@Duru1; @Kleinert]. The spectra are usually related by a Sommerfeld-Watson transformation of the Green functions [@Kleinert Chap. 14], [@Mustapic].
The discrete levels closest to zero are in a one-to-one correspondence [@Pelster1; @Zeile]. For stochastic systems, these levels rule the approach of the conditional probability density to its stationary limit. In our example the associated poles of the initial Green function (\[G2\]) $$\begin{aligned}
E^{({\rm i})}_n = - 2 \alpha a^{({\rm i})} n \, , \hspace*{0.5cm}
n = 0 , 1 , \ldots\end{aligned}$$ are mapped to the corresponding final ones of (\[G1\]), (\[K\]) $$\begin{aligned}
E^{({\rm f})}_n = 2 Q^{({\rm f})} n^2 + ( Q^{({\rm f})} - 2 a^{({\rm f})}
) n\end{aligned}$$ by $$\begin{aligned}
\label{MAP}
E^{({\rm f})}_n = E^{({\rm f})} ( E^{({\rm i})}_n ) \, ,\end{aligned}$$ as long as $n$ is bounded by $$\begin{aligned}
n\le \frac{a^{({\rm f})} }{2 Q^{({\rm f})} } - \frac{1}{4} .\end{aligned}$$ The evaluation of (\[MAP\]) requires the use of relations (\[I2\]) and (\[I3\])–(\[SP\]).
\
[**4.**]{} In summary, we have shown that a combination of local time and coordinate transformations opens new possibilities of relating different Markov processes. By extending the method to several random variables, we expect many useful applications. Furthermore we hope that this method might help to solve non-Markov processes [@Haenggi2].
Let us finally mention that the local time transformation (\[TT\]) is nonholonomic in spacetime, i.e., it carries a flat spacetime into a spacetime with nonzero torsion and curvature [@Pelster1; @Pelster2]. If we allow for purely spatial nonholonomic changes of coordinates, we can reach also spacetime geometries with spatial curvature and torsion. This will enable us to describe technically relevant diffusion processes in crystals with defects [@Kleinert2; @Bausch1; @Shabanov; @Bausch2].
[199]{} R.L. Stratonovich, [*Topics in the Theory of Random Noise, Volume 1 – General Theory of Random Processes, Nonlinear Transformations of Signals and Noise,*]{} Gordon and Breach, Second Printing, 1967 N.G. van Kampen, [*Stochastic Processes in Physics and Chemistry,*]{} North-Holland Publishing Company, 1981 H. Haken, [*Synergetics – An Introduction, Nonequilibrium Phase Transitions and Self-Organization in Physics, Chemistry and Biology,*]{} Third Revised and Enlarged Edition, Springer, 1983 U. Weiss, [*Quantum Dissipative Systems,*]{} Series in Modern Condensed Matter Physics, Vol. 2, World Scientific, 1992 L. Gammaitoni, P. Hänggi, P. Jung, F. Marchesoni, to appear in Rev. Mod. Phys. (1996) H. Risken, [*The Fokker-Planck Equation – Methods of Solution and Applications,*]{} Springer, Second Edition, 1988 I.H. Duru, H. Kleinert, Phys. Lett. [**B 84**]{}, 185 (1979); Fortschr. Phys. [**30**]{}, 401 (1982) H. Kleinert, [*Path Integrals in Quantum Mechanics, Statistics and Polymer Physics*]{}, World Scientific, Second Edition, 1995 A. Pelster, [*Zur Theorie und Anwendung nichtintegrabler Raum-Zeit-Transformationen in der klassischen Mechanik und in der Quantenmechanik,*]{} Dissertation, Universität Stuttgart, Shaker-Verlag (1996) A. Pelster, H. Kleinert, [*Transformation Properties of Classical and Quantum Laws under Some Nonholonomic Spacetime Transformations,*]{} Proceedings of the [*5th International Conference on Path Integrals from meV to MeV, Dubna, May 27–31, 1996*]{} Ph. Blanchard, M. Sirugue, J. Math. Phys. [**22**]{}, 1372 (1981) A. Young, C. DeWitt-Morette, Ann. Phys. [**169**]{}, 140 (1984) H. Haken, Z. Phys. [**B 24**]{}, 321 (1976) A. Schenzle, H. Brand, Phys. Rev. [**A 20**]{}, 1628 (1979) P.Y. Cai, A. Inomata, R. Wilson, Phys. Lett. [**A 96**]{}, 117 (1983) I.H. Duru, Phys. Rev. [**D 28**]{}, 2689 (1983) H. Kleinert, I. Mustapic, J. Math. Phys. [**33**]{}, 643 (1992) K. Zeile, A. Pelster, A. Wunderlin, Phys. Lett. [**A 179**]{}, 161 (1993) P. Hänggi, Z. Phys. [**B 30**]{}, 85 (1978) H. Kleinert, [*Gauge Fields in Condensed Matter, Vol. II, Part IV, Differential Geometry of Defects and Gravity,*]{} World Scientific, 1989 R. Bausch, R. Schmitz, [Ł]{}.A. Turski, Phys. Rev. Lett. [**73**]{}, 2382 (1994) H. Kleinert, S.V. Shabanov, [*Theory of Brownian Motion of Massive Particle in a Space with Curvature and Torsion and Crystals with Defects,*]{} Berlin preprint (1995), http://www.physik.fu-berlin.de/\~kleinert/kleiner\_re234/ brcurtor.html, cond-mat/9504121 R. Bausch, R. Schmitz, [Ł]{}.A. Turski, Z. Phys. [**B 97**]{}, 171 (1995)
[^1]: Email: [email protected], [email protected]; URL: http://www.physik.fu-berlin.de/\~kleinert; Phone/Fax: 0049/30/8383034.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a sequence of high resolution (R$\sim$20,000 or 15 km s$^{-1}$) infrared spectra of stars and brown dwarfs spanning spectral types M2.5 to T6. Observations of 16 objects were obtained using eight echelle orders to cover part of the $J$-band from 1.165-1.323 $\mu$m with NIRSPEC on the Keck II telescope. By comparing opacity plots and line lists, over 200 weak features in the $J$-band are identified with either FeH or H$_{2}$O transitions. Absorption by FeH attains maximum strength in the mid-L dwarfs, while H$_{2}$O absorption becomes systematically stronger towards later spectral types. Narrow resolved features broaden markedly after the M to L transition. Our high resolution spectra also reveal that the disappearance of neutral Al lines at the boundary between M and L dwarfs is remarkably abrupt, presumably because of the formation of grains. Neutral Fe lines can be traced to mid-L dwarfs before Fe is removed by condensation. The neutral potassium (K I) doublets that dominate the $J$-band have pressure broadened wings that continue to broaden from $\sim$ 50 km s$^{-1}$ (FWHM) at mid-M to $\sim$ 500 km s$^{-1}$ at mid-T. In contrast however, the measured pseudo-equivalent widths of these same lines reach a maximum in the mid-L dwarfs. The young L2 dwarf, G196-3B, exhibits narrow potassium lines without extensive pressure-broadened wings, indicative of a lower gravity atmosphere. Kelu-1AB, another L2, has exceptionally broad infrared lines, including FeH and H$_{2}$O features, confirming its status as a rapid rotator. In contrast to other late T objects, the peculiar T6 dwarf 2MASS 0937+29 displays a complete absence of potassium even at high resolution, which may be a metallicity effect or a result of a cooler, higher-gravity atmosphere.'
author:
- '[IAN S. MCLEAN, L. PRATO, MARK R. MCGOVERN, ADAM J. BURGASSER, J. DAVY KIRKPATRICK, EMILY L. RICE AND SUNGSOO S. KIM]{}'
title: 'THE NIRSPEC BROWN DWARF SPECTROSCOPIC SURVEY II: HIGH-RESOLUTION J-BAND SPECTRA OF M, L and T DWARFS[^1]'
---
Introduction
============
With effective temperatures $\la$ 2200 K, the cool atmospheres of L and T dwarfs generate complex spectra that are rich in molecular features, especially at near-infrared (NIR) wavelengths where ro-vibrational transitions of many molecules dominate. Fortunately, these cool, very low luminosity objects are also brightest in the NIR. Until recently, most infrared spectroscopic investigations of L and T dwarfs have concentrated on the identification of strong, broad spectral features, useful for the establishment of spectral classification, and have employed resolving powers of $R = \lambda/\Delta\lambda \la 2000$ (Burgasser et al. 2002, 2004, 2006; Cushing et al. 2003, 2005; Geballe et al. 1996, 2002; Jones et al. 1994; Leggett et al. 2000, 2001; McLean et al. 2000a, 2001, 2003a; Reid et al. 2000; Testi et al. 2001). Observations with significantly higher spectral resolution are potentially very important because line-blending from molecular transitions is reduced and weak features are resolved. Higher resolution spectra are more useful for constraining models of the complex molecular chemistry of brown dwarf atmospheres and for characterizing properties such as gravity and metallicity (Mohanty et al. 2004). For example, less massive brown dwarfs and younger brown dwarfs have smaller surface gravities which results in less pressure broadening and a different line shape. Furthermore, spectra with $R\ga$ 20,000 ($\la$15 km s$^{-1}$) are required for the measurement of radial and rotational velocities, and to search for radial velocity variability associated with brown dwarf spectroscopic binaries.
Obtaining high signal-to-noise observations with an increase in spectral resolution of a factor of ten is difficult because brown dwarfs are so faint. Basri et al. (2000) successfully resolved the resonance absorption lines of Cs and Rb in the far-red visible regime for a sample of M and L dwarfs using the HIRES echelle spectrograph on the Keck 10-m telescope and derived effective temperatures through comparison with available model atmospheres. Reid et al. (2002) also used high-resolution optical echelle spectroscopy to study 39 dwarfs with spectral types between M6.5 and L0.5. However, because brown dwarf fluxes are significantly less in visible light, high-resolution observations of fainter L dwarfs and of the even dimmer T dwarfs are not tenable in the optical and require infrared observations. With the advent and development of sensitive large-format IR array detectors, IR spectroscopy with the requisite spectral resolution is now possible (McLean et al. 1998, 2000b).
In this paper we present the first well-sampled spectral sequence of late M, L and T dwarfs observed at high resolution (R $\sim$ 20,000) in the NIR. This work is part of the NIRSPEC Brown Dwarf Spectroscopic Survey (BDSS) being carried out at the Keck Observatory; preliminary results were presented in McLean et al. (2003b). The goals of the BDSS are to obtain a significant sample of NIR spectra of low-mass stars and brown dwarfs of differing ages, surface gravities, and metallicities at both medium (R $\sim$ 2,000) and high spectral resolution for spectral classification studies and comparisons with model atmospheres. McLean et al. (2003a), hereafter M03, describes the lower resolution part of the survey; spectra from that study are available online.[^2] Here, we investigate the $J$-band using ten times higher spectral resolution than in M03. The $J$-band (defined as 1.15-1.36 $\mu$m in this paper) is important because this region contains four strong lines of neutral potassium (K I) that are both temperature and gravity-sensitive, and which persist throughout the M, L, and T dwarf sequence. In §2 we describe our observations and data reduction procedures. §3 provides a discussion of the rich, spectral morphology. In addition to atomic K I, there are lines of Al I, Fe I, Mn I, Na I and Ti I, and transitions of molecular species such as CrH, FeH, and H$_{2}$O that can provide a unique resource for improving model atmospheres at these low temperatures. We show that the sudden disappearance of the Al I lines critically defines the M-L boundary at these resolutions. Concentrating on the strongly pressure-broadened K I lines, we look for correlations between spectral type and equivalent widths, velocity widths (FWHM), and residual intensity. The relation between molecular line strengths and spectral type is also investigated. The effects of rotation, surface gravity and metallicity are explored in §4. A summary of the overall results and concluding remarks is given in §5.
Observations and Data Reduction
===============================
Targets and Instrumentation
---------------------------
Targets for the initial survey, the BDSS (M03), were selected primarily from well-known M dwarfs and from L and T dwarfs identified in the Two Micron All Sky Survey (2MASS; Kirkpatrick et al. 1997, 1999, 2000, 2001; Burgasser et al. 2000, 2002; Reid et al. 2000; Wilson et al. 2001), augmented with discoveries from the Deep Near-Infrared Survey of the Southern Sky (DENIS; Delfosse et al. 1997, 1999), the Sloan Digital Sky Survey (SDSS; Leggett et al. 2000; Geballe et al. 2002), and other investigations (Becklin & Zuckerman 1988; Ruiz, Leggett and Allard 1997). To ensure high signal-to-noise spectra for the high-resolution part of the survey, a subset of 12 of the brightest objects ($J=7-15$), spanning the spectral type range from M6 to T6, was selected. Of these, only 2MASS 0140+27 was not part of the initial survey. The M2.5 star G196-3A was also observed along with its L2 companion G196-3B, both examples of objects substantially younger than 1 Gyr (Rebolo et al. 1998) and therefore most likely to exhibit gravity effects (McGovern et al. 2004). In addition, the peculiar T6 dwarf 2MASS 0937+29 (Burgasser et al. 2002) was added to the list because of its apparent lack of K I features in lower resolution spectra. Another late T dwarf (2MASS 2356$-$15, T5.5) was observed after completion of the initial set for comparison to the 2MASS 0937+29. Although the signal-to-noise ratio was sufficient to establish the presence of K I absorption in this T5.5 dwarf, the fainter magnitude ($J=15.8$) and stronger 2o absorption made quantitative analysis too difficult so the spectrum is not shown.
Table 1 provides the complete list of 16 targets and the observing log. Shorthand names such as 2MASS 1507$-$16 are used in the text for simplicity, but the full designations are listed in Table 1. Two targets were known visual doubles at the time of observing, 2MASS 0746+20 (L0.5) and DENIS 0205$-$11 (L7), but in neither case did we have sufficient angular resolution to separate the components. Subsequent to making our observations, DENIS 0205-11 was reported as a possible triple brown dwarf system (Bouy et al. 2005) based on Hubble Telescope images. Burgasser et al. (2005) subsequently found SDSS 0423$-$04 to be double, the average spectral type of T0 being due to an L6 and T2 combination. Even more recently, the binary nature of Kelu-1, a 0$\farcs$29 pair, was revealed using Laser Guide Star adaptive optics on the Keck telescope (Liu & Leggett 2005; Gelino et al. 2006). Again, in neither case were these targets resolved in our NIRSPEC observations.
All of the observations were made using the NIRSPEC cryogenic spectrometer on the Keck II 10-m telescope on Mauna Kea, Hawaii. Detailed descriptions of the design and performance of this UCLA-built instrument are given elsewhere (McLean et al. 1998; 2000b). For this study, NIRSPEC was used in its cross-dispersed echelle mode. High resolution spectra are dispersed across the 1024$\times$1024 InSb detector at 0$\farcs$143 per pixel while the spatial scale in the cross-dispersed direction is 0$\farcs$19 per pixel. An independent slit-viewing camera with a scale of 0$\farcs$18 per pixel is available for centering and guiding. With the gratings used in NIRSPEC, the relationship between the blaze wavelength ($\lambda_{b}$) and echelle order number (m) is $m\lambda_{b}$ = 76.56 $\mu$m; together with the free spectral range (see below), this equation gives the order location of a given wavelength. The spectrometer was set up with the NIRSPEC-3 order-sorting filter and specific echelle and cross-dispersion grating angles to record 11 echelle orders ($m$ = 66 to $m$ = 56) covering the wavelength range from 1.15–1.36$\mu$m, corresponding approximately to the standard $J$-band. The free spectral range ($\lambda_{b}/m$) at 1.255 $\mu$m (order 61) is $\sim$206 Å, but the effective dispersion is 0.179 Å/pixel, allowing for only 183 Å (89%) of this order to be captured by the detector. In fact, the captured wavelength range varies from 171 Å (94%) in order 65 to 192 Å (84%) in order 58. Thus, because the spectral interval captured by the detector is slightly smaller than the free spectral range in each order, there are small gaps, increasing with wavelength, in the total spectral coverage. Table 2 summarizes the spectral range for each order used in the subsequent analysis. In practice, for an entrance slit 0$\farcs$43 (3 pixels) wide, the final spectral resolution in the reduced data is $R \sim 20,000$, (or 15 km s$^{-1}$), compared to the theoretical value of $R = 24,000$. The average value of one spectral resolution element is $\sim$0.625Å (equivalent to 3.6 pixels) over most of the $J$-band region.
Spectroscopic observations were made as nodded pairs. Typically, integrations of 600 s each were taken with the object placed at two positions, designated A and B, separated by $\sim$7$''$ on the $\sim$12$''$ long entrance slit of NIRSPEC. Shorter exposure times were used for brighter objects. Exposures of 300 s per nod position were used for 2MASS 0746+2000AB, Kelu-1AB and 2MASS 1507-1627, 120 s for Wolf 359 and 60 s per nod position for G196-3A. Total integration times per object ranged from a few minutes to 1.5 hours depending on the apparent $J$ magnitude. Signal-to-noise ratios were typically greater than 20 (5%) per resolution element over most orders, and sometimes greater than 100 (1% noise). Seeing conditions were $\sim0\farcs5-0\farcs6$ and therefore a slit width of 0$\farcs$43 (3 pixels) was used for all observations, except in the case of 2MASS 1507$-$16, for which we used a $0\farcs576$ (4 pixels) slit because of poorer seeing. A0 V stars were observed at an airmass very close to that of the target object to calibrate for absorption features caused by the Earth’s atmosphere. Arc lamp spectra, taken immediately after each observation, and OH night sky lines in the observed spectra, were used for wavelength and dispersion calibration. A white-light spectrum and a corresponding dark frame were obtained for flat-fielding.
Data Reduction Methods
----------------------
For the data reduction we used REDSPEC, an IDL-based software package developed at UCLA for NIRSPEC by S. Kim, L. Prato, and I. McLean[^3]. For each echelle order, REDSPEC uses the position of the two-dimensional spectra on the NIRSPEC array and the calibration line spectra to construct spatial and spectral maps necessary to transform the raw data onto a uniform grid. If the target spectrum itself is too faint to provide the spatial rectification, then the A0V star observed with the same set up was used instead. Although four arc lamps are available, it is often the case that there are too few well-distributed lines per echelle order for good spectral rectification. Consequently, OH night sky lines were also used. The dispersion was more than adequately fit by a second order polynomial of the form $\lambda = c_{0} + c_{1} x + c_{2} x^{2}$ where $c_{1} \sim$ 0.17$\pm$0.01 Å/pixel and $c_{2} \sim$ 7 x 10$^{-6}$ Å/pixel$^2$. To extract spectra free from atmospheric background and uneven detector response, the difference of an A/B image pair was formed and flat-fielded. The flat-fielded difference frame was then rectified using the spatial and spectral maps and the raw spectrum produced by summing 5$-$10 rows from each trace in the rectified image. The extracted traces (one positive, one negative) are subtracted again to produce a positive spectrum with residual night sky emission line features removed, unless a line was saturated. In the $J$-band, none of the night sky emission lines are saturated. A0 V star spectra were reduced in the same way, interpolating over the intrinsic Pa$\beta$ hydrogen absorption line at 1.28$\mu$m in the $J$-band spectra. The raw target spectrum was then divided by the raw A0 V star spectrum to remove telluric features. The true slope of the target spectrum was restored by multiplication with a blackbody spectrum of T$_{eff}=9500$ K for an A0 V star (Tokunaga 2000). Finally, the spectra reduced from multiple A/B pairs were averaged together to improve the signal-to-noise ratio.
J-band Spectral morphology at R$\sim$20,000
===========================================
Overview
--------
For each of the 16 targets we have extracted 8 echelle orders (see Table 2) yielding a total of 128 spectra. Before examining and interpreting the new spectra in detail, it is very useful to have a broad overview of the basic spectral features present and an awareness of the general trends that occur in the high-resolution $J$-band data as a function of spectral type. A convenient way of doing this is to select a representative source for a few spectral types and present all eight echelle orders on the same plot, thus enabling the entire $J$-band to be viewed at a glance. Figures 1$-$6 show the reduced spectra of Wolf 359 (M6), 2MASS 0140+27 (M9), 2MASS 0345+25 (L0), 2MASS 1507$-$16 (L5), SDSS 0423$-$04AB (T0) and 2MASS 0559$-$14 (T4.5). The double nature of SDSS 0423$-$04AB means that we do not have a true T0 spectrum, but lower resolution studies (M03) show that J-band spectral variations are relatively weak from L6 to T2 and therefore this binary remains a useful proxy for a T0 dwarf. In these plots, echelle orders 58 through 65 are shown together; the remaining orders at the edges of the $J$-band are too contaminated by strong atmospheric absorption to be useful. For ease of comparison, all spectra are shown in the laboratory reference frame and vacuum wavelengths are used throughout; radial velocities and searches for radial velocity variations will be reported and discussed in a separate forthcoming paper (Prato et al. 2006 in prep.). Each order is normalized to unity at the same wavelength. Comparison of the spectra in these six figures, all of which have excellent signal to noise ratios (at least 20:1 per pixel), shows that the region is densely populated with numerous weak absorption features and a few stronger lines. We will show that the fine-scale spectral structure is real and repeatable, and that it is mainly attributable to FeH or H$_{2}$O. The strongest atomic features are the doublets of K I that occur in orders 61 and 65. These lines persist from M6-T4.5 but clearly change their character with spectral type. In later sections the K I lines will be singled out for closer inspection. For reference, Table 3 summarizes the main spectral transitions observed in the $J$-band over the spectral type range from M6-T4.5, including the energy levels of the atomic transitions.
As shown in Figure 1, the M6 dwarf Wolf 359 has at least one distinguishing feature in each order. Atomic lines of Al I at 1.31270 and 1.31543 $\mu$m appear in order 58. There is a moderately strong line of Mn I at 1.29033 $\mu$m in order 59, plus some weaker lines of Ti I at 1.28349 and 1.28505 $\mu$m. A weak unresolved Na I doublet is seen at 1.26826 $\mu$m in order 60. The first pair of strong K I lines at 1.24357 and 1.25256 $\mu$m appears in order 61. Multiple weak absorption features occur in both orders 62 and 63, the most notable grouping being the set of lines around 1.222 $\mu$m. Several of the stronger features have been identified with FeH from lower resolution studies (Jones et al. 1996; Cushing et al. 2003; M03). Note however, that a major FeH band head at 1.24 $\mu$m is just off the detector at the short wavelength edge of this order. In order 64 there is a pair of strong Fe I lines at 1.18861 and 1.18873 $\mu$m and another Fe I line at 1.19763 $\mu$m. Order 65 contains the second set of strong K I lines, one at 1.169342 $\mu$m and the close pair at 1.177286, 1.177606 $\mu$m. In general, the lines are relatively sharp and well-resolved. Wolf 359 is a bright source and therefore the signal-to-noise ratio in this spectrum is at least 100:1.
Following these spectral features order by order through Figures 1-6 reveals certain general trends as a function of spectral type. Comparing the M9 object (Figure 2) with the M6 source (Figure 1) we see that the Al I lines at 1.3127 and 1.3154 $\mu$m in order 58 are somewhat weaker at M9 and then suddenly they are no longer present at L0 (Figure 3), or in any later spectral types (Figures 4-6). This is an important observation that relates to the M-L transition and we will discuss the Al I lines in the next section. Throughout order 58 there are other weaker spectral features, the so-called fine-scale spectral structure. This spectral structure becomes more pronounced at M9 (Figure 2), seems broader in the L0 and L5 objects (Figures 3 and 4), weakens at T0 or more accurately from L6 to T2 (Figure 5) and then completely changes character by T4.5 (Figure 6). The most likely interpretation of this spectral sequence is that it represents changes in the physical structure of these cool atmospheres (temperature, pressure, chemistry). In subsequent sections we compare opacity data for different molecular species to identify the primary absorbers at each spectral class.
In order 59 the sharp Mn I line at 1.2903 $\mu$m seen at M6 and M9 (Figures 1 and 2) broadens and disappears after L0 (Figure 3). The fine-scale spectral structure in this order is dominant until T0 composite type (Figure 5) when the spectrum becomes remarkably smooth. Here again the high resolution data reveal a striking effect, this time at the transition from L to T dwarfs. New spectral structure develops in this order between types T0 at T4.5 (Figure 6) but, as was the case for order 58, the pattern is different, indicating different atmospheric conditions.
The weak Na I line at 1.2683 $\mu$m detected in order 60 in the M6 object (Figure 1) is already absent in the M9 object (Figure 2). Otherwise, the behavior of the fine-scale structure follows a pattern similar to order 59 becoming remarkably weak at T0 (Figure 5) and leading to a smoother appearance for these spectra near the L-T transition.
Order 61 contains one of the pairs of strong K I doublets located at 1.2436 and 1.2525 $\mu$m. These lines deepen and widen slightly from M6 to M9, and then become increasingly broader and shallower from L0 to T4.5. NIRSPEC spectral order 61 also has many fine-scale features attributable to molecular transitions. Two features, one at 1.24637 and the other at 1.24825 $\mu$m have been identified previously with FeH (Cushing et al. 2003). These FeH lines strengthen slightly from M6 to M9 (Figures 1 and 2), become much broader in the L dwarfs (Figures 3 and 4) and then vanish completely in the T dwarfs (Figures 5 and 6) to leave, once again, a remarkably smooth continuum between the K I lines.
Comparing orders 62 and 63 in Figures 1-6, the known FeH features in these spectral bands strengthen from M6 to M9, broaden markedly at L0 and remain strong and broad through L5 before becoming weaker in the T0 and T4.5. As with the atomic lines, the individual FeH lines seem to broaden significantly at the transition from spectral types M9 to L0. Evidence of weak FeH absorption is still present around 1.222 $\mu$m at spectral type T0, and possibly even at T4.5, as shown in Figures 5 and 6, but this molecular species is clearly not dominant in T dwarfs.
For the M6 dwarf (Figure 1), order 64 is characterized by a pair of strong lines of Fe I at 1.18861 and 1.18873 $\mu$m that are easily resolved, and another Fe I line at 1.19763 $\mu$m that is blended with FeH. These features remain strong at M9 (Figure 2) and persist into the L-dwarf range, becoming broader at L0 (Figure 3), and then undetectable by L5 (Figure 4). From T0 to T4.5 (Figures 5 and 6), order 64 becomes increasingly chopped up by new spectral features, some of which are quite sharp and deep. In section §3.3 we show that these features are caused by absorption by 2o.
Finally, there is order 65, which contains the second K I doublet and exhibits some of the largest changes with spectral type. The slightly weaker K I companion line at 1.17728 $\mu$m, only 3.3Å from the longer wavelength member of the doublet is easily resolved in the M9 object (Figure 2), already blended from line broadening in the L0 (Figure 3), barely discernable at L5 (Figure 4), and completely washed out by line broadening and numerous molecular features at T4.5 (Figure 6). Order 65, being close to the short wavelength edge of the $J$-band where terrestrial water vapor absorption is expected, also contains many strong intrinsic transitions of hot H$_{2}$O, for example, the feature at 1.175 $\mu$m.
Al I, Fe I and Mn I; indicators of the M-L transition
-----------------------------------------------------
Figure 7 provides a more detailed view of the Al I doublet in order 58. In this plot, the spectra for the M6, M9 and L0 objects shown in Figures 1-3 are expanded and overlaid. Evidently, there is significant spectral structure in this part of the spectrum making it difficult to identify a true continuum level. All three spectral types show consistent features, in particular the wide depression containing the shorter wavelength Al I line. The equivalent width of the Al I lines clearly decreases from M6 to M9, but the change over these three spectral types is only about 25%. Because this region of spectrum is contaminated by 2o absorption, it is difficult to obtain accurate equivalent widths. A pseudo-equivalent width over a 4.2Å interval centered on each line was obtained relative to the local continuum in the troughs where the Al I lines are found. For the stronger line of the pair at 1.3127 $\mu$m, the measured values of equivalent width for the M6 and M9 dwarfs respectively are 420$\pm$20 mÅ and 300$\pm$40 mÅ. At L0, however, the pseudo-equivalent width of this line is $\le$40 mÅ. Clearly, at the transition from M9 to L0, both Al I lines vanish completely. Although only three objects bridging this transition were observed at high resolution, the conclusions given here are supported fully by the results of our low resolution BD spectroscopic survey (M03) where two objects of every spectral type from M6 to L5 was included.
As shown in Table 3, these lines arise from absorption from an energy level at 3.14 eV. Interestingly, the Na I line at 1.268 $\mu$m in order 60 is already absent in the M9, and careful inspection shows that a somewhat broadened Mn I line at 1.290 $\mu$m in order 59 persists through L0. The Na I line is excited from a high energy state at 3.6 eV whereas the Mn I line comes from a state at only 2.1 eV. Thus, the sequence in which the lines disappear is at least qualitatively consistent with thermal excitation. But the abrupt loss of Al I lines at the classical M-L boundary is too great to be explained by Boltzmann factors alone. For example, for a temperature change from 2850 K from the M6 to about 2400 K for the M9 say, the population of excited atoms in the upper level would drop by 51% and the equivalent width of the line might change from 420 mÅ to about 210 mÅ if all other factors remain the same. From M9 to L0 the change would be a further 13% assuming a change in effective temperature of 150 K. Thus, the line should still be measurable with an equivalent width of 140 $\pm$40 mÅ, or about one-third its value at M6. Yet, both lines disappear abruptly. It is likely, from the models of Lodders (2002), that aluminum has been sequestered in compounds such as hibonite (CaAl$_{12}$O$_{19}$) and that this abrupt change in absorption line strength is really caused by the sudden depletion of aluminum as an absorber due to a significant change in atmospheric chemistry, rather than simply a drop in effective temperature. Gas temperatures typical of this transition are near 2000 K (Lodders 2002).
It is also curious that the intensity ratio of the components of the Al I doublet is closer to 3:2 than the expected 2:1 ratio based on their statistical weights. However, as shown in Figure 7, this spectral region is highly complex with many overlapping transitions which makes it difficult to determine the true continuum level for each line. Alternatively, the peculiar line ratios may be a non-LTE effect, or the result of line blending.
Another element that is also important for understanding the temperature structure of these cool, dust-forming objects is iron. As previously mentioned, order 64 contains a remarkably strong pair of Fe I lines at 1.1886 and 1.1976 $\mu$m. The shorter wavelength Fe I line is a resolved double with a separation of 1.2Å in the M6 and M9 objects, but appears as a single broad feature at L0. By L5 the Fe lines are completely absent. Lower resolution studies (M03) also suggested that Fe disappeared around L2 or L3. These Fe I lines arise from low-lying energy levels near 2.2 eV, and gas phase iron requires temperatures above about 1700 K (Burrows et al. 2001). Combining the results that Al disappears at L0 and Fe is no longer present by L3, and using the chemistry temperature scale, suggests that there is about a 300 K temperature change from L0 to L3, which is shallower but still consistent with the interval of about 140 K per spectral type derived by Burgasser et al. (2002) as well as the effective temperature scale of Golimowski et al. (2004). As noted by Burgasser et al. (2002), temperatures derived from condensation chemistry tend to be systematically cooler by about 500 K than those derived from empirical determinations of T$_{eff}$ using objects with known parallax. These conclusions are not necessarily inconsistent if different spectral features probe a range of optical depths in the atmosphere.
Finally, we note the presence of several weak lines of Ti I that arise from energy states near 1.4 eV, even lower than those of the strong potassium lines. A strong Ti multiplet at 0.97 has also been seen in the spectra of M dwarfs up to at least M9 (Cushing et al. 2005). Unfortunately, these weak lines are impossible to trace after M9.
Fine-scale structure; the role of FeH and 2o
--------------------------------------------
The astronomical $J$-band is bounded by 2o absorption bands from terrestrial water vapor. It is therefore no surprise that high-temperature 2o (hot steam) transitions intrinsic to M, L and T dwarfs encroach far into the $J$-band from both the short and long wavelength ends. These so-called infrared water bands are difficult to model because millions of transitions are needed (Partridge & Schwenke 1997). Typically, models over-estimate the depth of the infrared water bands. In addition to 2o, some of the stronger non-atomic transition features are known to be attributable to FeH from lower resolution studies (Jones et al.1996; McLean et al. 2000). These features occur at 1.2091, 1.2113, 1.2135 and 1.2221 $\mu$m. Cushing et al. (2003) verified the features at 1.1939 and 1.2389 $\mu$m as the 0–1 and 1–2 band heads of the F$^{4}\Delta$ – X$^{4}\Delta$ system of FeH, and attributed a blended feature described by McLean et al. (2000) at 1.2221 $\mu$m as the F$^{4}\Delta_{7/2}$ – X$^{4}\Delta_{7/2}$ Q-branch. These authors also listed 24 other relatively strong features lying within the $J$-band. In the Cushing et al. list, no FeH features were identified in the wavelength interval covered by our order 65, which includes the strong shorter-wavelength doublet of K I, and only one feature (at 1.2464 $\mu$m) was tabulated for order 61, where the other K I doublet dominates.
To identify many more of the complex fine-scale features seen in the spectral sequences of Figures 1–6, we analyzed opacity (cross-section) data for both FeH and H$_{2}$O, (R. Freedman 2003, private communication) and utilized the FeH line list and transition catalog by Phillips et al. (1987). We are also grateful to Adam Burrows who provided CrH opacity data (Burrows et al. 2002) and Linda Brown who provided new opacity calculations for 4 in the $J$-band (L. Brown 2004, private communication).
Figure 8 is a detailed view of order 62 (1.221-1.239 $\mu$m) for the M9 object 2MASS 0140+27. Superimposed on the M9 dwarf spectrum is a normalized and scaled FeH opacity plot for a temperature of 2000 K and a pressure of 1 bar. We use this plot to assist in identifying spectral features. This interesting region of the $J$-band contains a feature which is seen in lower resolution spectra as a broad, flat-bottomed line at 1.2221 $\mu$m (McLean et al 2000a). Each transition in the opacity data can be correlated with either 0–1 or 1–2 transitions of the F$^{4}\Delta$ – X$^{4}\Delta$ system tabulated by Phillips et al. (1987). As shown in Figure 8, the feature observed at 1.2221 $\mu$m and attributed to the Q-branch by Cushing et al. (2003), is actually a composite of four Q-branch and three P-branch transitions plus one R-branch transition of FeH. Thus, at our higher resolution, the broad flat-bottomed feature seen in lower resolution spectra is completely resolved into eight separate transitions. These transitions are the following: Q(0-1) at 1.22137 and 1.221383 $\mu$m blended, 1.221934, 1.222504 and 1.22305 $\mu$m; P(0-1) at 1.22166, 1.22219 and 1.22244 $\mu$m; R(0-1) at 1.22218 $\mu$m. R-branch transitions tend to correspond closely to P-branch lines. For example, the P(0-1) 1.22219 $\mu$m line is only 0.1Å from the R(0-1) transition just given. Also, the Q-branch line at 1.222504 $\mu$m is blended with the nearby P-branch transition at 1.22244 $\mu$m.
The remainder of order 62 is dominated by 5 R-branch lines and about 30 P-branch transitions; all are members of the 0–1 band of F$^{4}\Delta$ – X$^{4}\Delta$ system. In this one order alone, there are now more identified FeH transitions than previously known for the entire $J$-band. Summing up across all NIRSPEC orders, we identify over 200 matches to the FeH opacity data base and the tables by Phillips et al. (1987). There is little doubt that FeH is a dominant molecular absorber in late M and early L spectral types.
Interestingly, there are several features that cannot be identified with FeH transitions in the given opacity tables. For example, a sharp line at 1.2227 $\mu$m contaminates the Q-branch feature, and there are other isolated groups near 1.225, 1.228, 1.231 and 1.234 $\mu$m which also are not attributable to FeH. Because these features have a dependence on spectral type that is similar to FeH, they may be unknown FeH transitions or transitions of CrH. To explore the latter possibility we also compared our spectra to CrH opacity calculations by Burrows et al. (2002). Although the CrH opacity data had a lower resolution than our spectra, there was good coincidence for the features at 1.225, 1.228 and 1.231 $\mu$m, but the features near 1.234 $\mu$m could not be identified with CrH. Although transitions of CrH contaminate those of FeH, the strongest fine-scale features in this part of the $J$-band are attributable to FeH rather than CrH. Stronger CrH bands occur at shorter wavelengths (Kirkpatrick et al. 1999).
We also investigated the fine-scale structure in orders 61 and 65 which contain the important K I doublets. Figure 9 provides a plot of the FeH opacity (at 2000 K and 1 bar pressure) across order 61, normalized for convenience, smoothed to match the resolution of our spectra, and over-plotted with the spectra of our M9 and L5 dwarfs. Essentially every feature can be traced to FeH, with a few notable exceptions. For example, absorption features at 1.2448 and 1.2458 $\mu$m are clearly associated with FeH, but two significant absorptions bands in between these limits do not correlate with FeH. Note also that the L5 spectral features are broader than the equivalent features in the M9 object.
As shown in Figures 1-6, there is a general increase in the strength of FeH from mid-M until about L4 and then a decrease in FeH features towards later spectral types. Figure 10 shows a plot for order 61 in which the L5 and the T4.5 dwarfs in our sample are compared. In this figure H$_{2}$O opacity curves (at 1000 K and 1 bar) are over-plotted. The T4.5 spectrum lacks FeH, and its shallow depressions and small features are consistent with the H$_{2}$O opacity. The FeH feature near 1.245 $\mu$m appears strongest in the early to mid Ls, while the 2o transition at 1.1752 $\mu$m seems to gain in strength systematically toward later spectral types.
In order 65, no transitions for FeH are listed in the opacity files at the shorter wavelengths. Therefore, in Figure 11, we plot the H$_{2}$O opacity (at 1000 K and 1 bar) for order 65, and over-plot with the spectrum of the L3.5 and T4.5 from our sample. This pair of spectral types provides distinct morphological samples. Most of the features in the T4.5 dwarf evidently correspond to H$_{2}$O. There is a particularly strong 2o feature at 1.1752 $\mu$m in the T4.5 dwarf, which is also present in the L3.5 object at a weaker level. In fact, weak H$_{2}$O absorption is already present even in late M objects.
We also looked for transitions associated with the $\phi$ bands of TiO (Galehouse 1980; Phillips 1973). These bands have been detected in young, low-gravity sub-stellar objects by McGovern et al. (2004), but are not usually apparent in low resolution spectra of older field dwarfs. A model spectrum kindly provided by D. Saumon shows the location of numerous $\phi$-band transitions in order 61, with noticeable band heads for the $\Delta\nu$=-1, (0-1) and (1-2) transitions at 1.25644 and 1.24239 $\mu$m respectively. Absorption features do occur at these wavelengths from M6-L5, but these also coincide with FeH features and are less likely to be TiO because none of the other TiO transitions of the $\phi$ band are seen.
Finally, using new opacity data for 4 (L. Brown 2004, private communication), we searched for evidence of methane transitions in the T dwarfs. The strongest features should occur in order 65 which also contains the very broad shorter-wavelength K I doublet. Unfortunately, this spectral region is already heavily blanketed by 2o absorption. Even in the high signal-to-noise spectra of 2MASS 0559-1404 (T4.5), there are no transitions that can be uniquely attributed to 4.
The potassium doublets
----------------------
Because the K I doublets dominate orders 61 and 65, and because these strong lines clearly persist throughout almost the entire M, L and T range, we plot 12 of the objects from Table 1 together in Figures 12 and 13 to provide a more continuous sequence for these features. For this figure, the L4 GD165B has been left out because the L3.5 2MASS 0036+18, with better signal-to-noise ratio, has been included; GD165B will be discussed separately. The other two objects omitted from these plots are the peculiar L2, G196-3B, and the peculiar T6 dwarf, 2MASS 0937+29; again, these objects are discussed separately later. Each spectrum in Figures 12 and 13 has been normalized to a continuum value of one at a common wavelength and shifted for clarity by an additive constant along the y-axis. The spectra have been ordered according to their published classification (Optical types for L-dwarfs: Kirkpatrick et al. 1999, 2000, 2001; NIR for T dwarfs: Burgasser et al. 2006) and aligned at the K I rest wavelengths in vacuum. As mentioned already, radial velocity determinations will be presented in a separate paper (Prato et al. 2006, in prep.).
As given in Table 3, the shorter wavelength K I lines (in order 65) correspond to the multiplet designation 4p $^{2}$P$^{o}$$-$3d $^{2}$D, and the order 61 pair to the 4p $^{2}$P$^{o}$$-$5s $^{2}$S multiplet. All are transitions between states at 1.61$-$2.67 eV. The pair of K I lines in order 65 have almost equal intensity at the line center, whereas in order 61, the 1.2436 $\mu$m line is always weaker than the 1.2526 $\mu$m line. For both K I doublets, the line ratios are similar throughout the spectral sequence.
One of the most prominent results, evident in both Figures 12 and 13, is the significant K I line broadening in the later type objects. Primarily because of its higher temperature, the M2.5 dwarf manifests very narrow lines with almost no wings. There is a slight contamination of the shortest wavelength K I line in order 65 from an Fe I line for spectral types before L4. For the M2.5, M6 and M9 dwarfs, separate peaks are clearly discernable for the longer wavelength line of the K I doublet in order 65 at 1.17761 $\mu$m and the secondary K I line at 1.17728 $\mu$m, but for later spectral types the 1.17761 $\mu$m lines are heavily blended and not discernable as separate features. In the earlier type objects, the pair of lines near 1.1786 $\mu$m in order 65 are attributable to Ti and Fe.
Analysis and Discussion
=======================
Correlating spectral features and spectral type
-----------------------------------------------
To quantify the changes in the K I lines illustrated in Figures 12 and 13 as a function of spectral type we calculate three quantities, a line depth, a line strength (equivalent width) and a line width. Line depth is defined in terms of the measured flux ($F_{\lambda}$) at the line center compared to the average value for the continuum. For convenience we construct the line depth as $1-F_{line}/F_{cont}$ at the line center; a weak line has almost the same flux in the line as in the continuum and its line depth measure is therefore almost zero. Deep lines have a depth index approaching unity. Determining the continuum or pseudo-continuum level introduces the largest uncertainty into this ratio. The continuum is estimated by fitting a sloping line across the feature between the highest points on either side lying within $\pm$ 50Å of the line center. Multiple trials obtained by varying the positions by a few Angstroms provide a mean value and an estimate of the uncertainty from fluctuations in the continuum level.
Equivalent width (in Å) is also determined by fitting a continuum line between two points on either side of the line, summing the residual intensities and multiplying by the width of a pixel. The same range and method is used as for the line depth. Line width (in km s$^{-1}$) is defined as the full width of the line at the intensity level halfway between the apparent local continuum, defined above for the line depth and equivalent width, and the minimum line intensity at the line center (FWHM).
Several sources contribute to the uncertainty in the derived quantities. The continuum level is difficult to identify, photon noise reduces the signal-to-noise ratio for the fainter sources, and contamination by numerous weak features is dependent on spectral type. Molecular line contamination results in additional fluctuations in the measured K I quantities and gives larger uncertainties, despite the good signal-to-noise ratio of most of the spectra. Our final plotted uncertainties are based on repetitive trial fits and photon noise estimates.
Table 4 and Table 5 provide the measured quantities for the 1.2525 and 1.2436 $\mu$m K I lines respectively. This pair of K I lines (order 61) is preferred because the shorter wavelength doublet is too heavily contaminated by H$_{2}$O absorption, and one of the K I lines is itself a blend (§3). Figures 14, 15, and 16 show, respectively, the line depth, equivalent width and velocity line width as a function of spectral type for the 1.2525 $\mu$m line.
Several of the trends mentioned in the previous section are confirmed. The K I line depth increases from a weak line in early M dwarfs to a strong deep line in the early/mid L dwarfs before decaying towards later spectral types. Interestingly, there is considerable scatter in line depth among similar spectral types. The equivalent width (designated by W) is better behaved. Variations in defining a consistent continuum may have a large effect on the line depth quantity, whereas the growth of the line wings may overcome such variations when calculating W and FWHM. Equivalent width increases to a broad maximum around L2-L4 and then decreases again, remaining essentially constant from L5 - T4.5. These results are consistent with our more extensive lower resolution studies (M03). In contrast, the velocity line width of the 1.2525 $\mu$m K I line (Figure 16) increases almost monotonically and steeply with spectral type. The change in velocity width is dramatic, ranging from $\sim$ 60 km s$^{-1}$ at M6 to almost 500 km s$^{-1}$ at T4.5. This change does not result from rotational broadening because the line develops extensive wings characteristic of pressure broadening as discussed in the next section. One L-dwarf appears to stand out from the others in this sequence and is indicated by a star symbol. This object is Kelu-1AB, and will be discussed separately in the next section. The behavior of the 1.2436 $\mu$m K I line (Table 5) is similar.
The changing behavior of the K I lines can be explained as follows. As the temperature decreases towards later spectral types, the transition levels become less populated and the line should become weaker. However, with decreasing temperature, dust grains settle below the photosphere and the transparency of the atmosphere at wavelengths where gas opacity is weak improves. Thus, line formation can be observed at much greater depths and pressures (Saumon et al. 2003). Higher pressures cause the development of the broad line wings through collision broadening, primarily with H$_{2}$ molecules (e.g. Burrows & Voloyubev 2003), and hence an increase in the FWHM of the lines. The greater column depth of K I in these transparent regions also serves to increase their equivalent widths. The behavior has been observed with the K I and Na I fundamental doublets (Reid et al. 2000), and is one of the unique properties of brown dwarf atmospheres. The pressure broadening of the K I lines at J-band follow quantitative expectations as well. The observed ratio between the K I line widths in the T4.5 and L5 spectra is about 2:1. As the T$_{eff}$ decreases, the pressure at $\tau$=2/3 increases as already described. According to models of cloudy atmospheres, kindly provided by D. Saumon (2003; private communication), the change in the photospheric pressure at these wavelengths is a factor of 5 from 1600 K to 1000 K. These effective temperatures were chosen to illustrate a prediction of the models over the typical range of L/T spectral types. This implies that the average kinetic energy of the molecules of H$_{2}$ is five times greater, or that the average speeds are $\sim$2.24 times larger. Pressure broadened line widths should be proportional to the average velocity of the disturbing atoms. Thus, the observed pressure broadening from L5 to T4.5 is consistent with the expected change.
In contrast to the K I lines, Figures 17 and 18 show the changing trends with spectral type for two representative FeH and H$_{2}$O features that fall within the orders containing the K I doublets; data are provided in Table 6. For FeH we plot the line depth index of the 1.245 $\mu$m feature in order 61, and for H$_{2}$O we use the strong 1.175 $\mu$m feature in order 65. The FeH line depth reaches a peak around L3-L4, whereas the H$_{2}$O line depth increases more-or-less monotonically from M6 to T4.5. Once again Kelu-1AB stands out, this time because the FeH bands are not as deep as expected for an L2.
Line shapes: rotation and pressure broadening effects
-----------------------------------------------------
In the previous section we found that the K I line width (FWHM) and the 2o line depth provided the best correlations with spectral type. The exception to the FWHM correlation is Kelu-1AB, which [*is*]{} known to be a rapid rotator (Basri et al. 2000), and which was recently discovered to be a $\sim$03 visual binary (Liu & Leggett 2005; Gelino et al. 2006), although the orbital velocity of 3 km s$^{-1}$ is only a fraction of the $Vsin~i$. As will be shown below, the K I lines are not really suitable for $Vsin~i$ studies.
Interestingly, Kelu-1AB is not abnormal in its 2o ratio. The one point in the 2o plot of Figure 18 that does seem slightly high is due in fact to GD165B (L4); this result is consistent with M03. As the companion to a white dwarf, GD165B is most likely to be an old L dwarf with a higher gravity. Burgasser et al. 2003 has shown that late-type subdwarfs, which are also presumably older and have higher surface gravities, can show stronger 2o compared to equivalently classified disk dwarfs.
The presence of a binary companion can impact the observed line properties discussed above in several ways. For example, a spectroscopic companion only partially resolved in velocity space might produce a spectrum with anomalously broadened lines at certain epochs. A spectroscopic companion of similar mass to the primary object should not only manifest itself in broadened lines, observed at a favorable phase, but also should almost double the expected brightness of the system. Fainter companions, although interesting in their own right, should not significantly impact the primary dwarf spectrum if the mass ratio, $M_2/M_1$, is less than $\sim$0.5 (Prato et al. 2002). Visual companions at separations too great to noticeably effect the spectrum will also cause the observed spectrum to appear brighter than expected for the distance and age of the target system. Four targets in Table 1 are visual binaries: 2MASS 0746+20 (L0.5), Kelu-1 (L2), DENIS 0205$-$11 (L7) and SDSS 0423$-$04 (T0). Signatures of velocity variations in these spectra will be addressed in a forthcoming publication.
Spectral features in Kelu-1AB appear to be significantly broader than other field dwarfs of similar temperature by a factor of 2-3, consistent with its high rotational velocity of $60\pm5$ km/s (Basri et al. 2000). In Figure 19 we compare the L2 dwarf, Kelu-1AB, with the L0.5 and L3.5 dwarfs in our sample. Because no feature is completely free of blending at this spectral resolution and wavelength throughout the late M to T sequence, measuring precise rotational velocities is challenging. Indeed, because of the increasing complexity of molecular features with cooler effective temperatures, most if not all of the lines in the mid-L to T objects may be blends. However, the wealth of fine-scale spectral structure attributable to FeH and 2o throughout the $J$-band does provide some possibility of detecting trends in $Vsin~i$, especially because pressure broadening should not be a large effect for molecular transitions.
Examining the rich FeH structure in order 61 (Figure 12), as well as orders 62 and 63 (Figures 1-4) where there are no strong alkali lines, it is clear that there is a sudden broadening of the FeH lines at the transition from M9 to L0, and that all of the FeH features remain broad after that transition. In fact, if the M9 spectra in orders 62 and 63 are smoothed by a factor of 3 using a simple moving average, the resulting spectrum is a very good match for the observed L0 spectrum. This sudden broadening could be caused by either increased pressure broadening or an increase in rotational velocities. Examination again of Figure 13 shows that many of the 2o spectral features that develop in the T dwarfs are relatively sharp, indicating that they are not formed at the same depth and pressure as the potassium line wings. Therefore, in the coolest objects, we are seeing evidence of vertical structure in the atmosphere. Like the FeH lines in the late M and early L dwarfs, the 2o features provide a better estimate of rotational velocities than the alkali lines. We can de-convolve the instrumental profile by using the observed widths of unresolved OH night sky lines and the very narrow features seen in the M2.5 object. If we interpret the change in line width at the M9-L0 boundary as a rotational velocity, then after removal of the instrumental profile, all of the early L dwarfs appear to be rapid rotators with $Vsin~i$ $\ga$ 30 km s$^{-1}$.
Rotational velocities have been reported for some of the objects in our target list. Basri et al. (2000) give $Vsin~i$ = 60$\pm$5 km s$^{-1}$ for Kelu-1AB and 22$\pm$5 km s$^{-1}$ for DENIS 0205$-$11AB and $\la$3 km s$^{-1}$ for Wolf 359 (Gl 406). The L3.5 dwarf 2MASS 0036+18 is reported to have a $Vsin~i$ of $\sim$15 km s$^{-1}$ by Schweitzer et al. (2001). Several of our targets were also observed by Zapatero Osorio et al. (2006) at high resolution, enabling them to measure the $Vsin~i$ of 2MASS 0036+18, SDSS 1254-01, and 2MASS 0559-14. We note with interest that the rotational velocities they measure for these sources, as well as for another dozen dwarfs, are all very close to 30 km s$^{-1}$ within the uncertainties. We will analyze in detail the rotational velocities of our sample in a future paper; however, we comment here that our Figures 12 and 13 suggest that the T0 and T4.5 dwarfs in our sample, SDSS 0423-04 and 2MASS 0559-14, appear to reflect larger values of $Vsin~i$. If there is a tendency towards more rapid rotation among the T dwarfs, it is not apparent in the measurements of Zapatero Osorio et al. (2006).
Clarke et al. (2003) report time resolved spectroscopy of Kelu-1AB used to search for variability in photospheric molecular species. They confirm the short rotation period of the system and find variable H$\alpha$ profiles. No evidence for a spectroscopic companion was detected and it appears to be a normal L dwarf apart from the high rotation rate (Clarke et al. 2003). The recently discovered 0.29binary contributes only about 3 km s$^{-1}$ in velocity shift and therefore has no impact on the measured $Vsin~i$. However, it is interesting to ask whether or not both components have the same $Vsin~i$. In the binary system Gl569B (Zapatero Osorio et al. 2004), line broadening is 2-3 times greater in one component than the other, although this has been attributed to a nested spectroscopic binary (Simon et al. 2006). It is possible that the higher rotation rate of Kelu-1AB is the result of age. A younger L dwarf just past the accretion phase might have a higher rotational velocity. On the other hand, as discussed in the next section, the young L2 dwarf G196-3B does not have such a rapid rotation, or is being viewed at a special angle.
Surface gravity effects
-----------------------
As shown by models (e.g. Burrows et al. 2001), the youngest brown dwarfs are expected to be hotter and more luminous. As a sub-stellar mass object cools over the first 100 Myr, its radius will contract by a factor of 2-10, depending on initial mass, and then remain almost constant at a value close to that of Jupiter’s radius. Consequently, a very young low-mass brown dwarf could be observed with a much earlier spectral type than it will have when older than 1 Gyr, and its surface gravity ($g = GM/R^{2}$) will be less than that of an old brown dwarf with the same observed spectral type. A lower surface gravity implies less pressure broadening (P$\sim$g/k$_{r}$, where k$_{r}$ is the Rosseland mean opacity) and therefore one expects the K I lines to be narrower in such objects. The effect of surface gravity on the K I lines is significant and has already been observed at lower resolution (McGovern et al. 2004; Kirkpatrick et al. 2006). High resolution infrared spectroscopy of brown dwarfs provides a means of measuring surface gravities and hence estimating mass. Because of their greater column depth the K I lines are much more sensitive to pressure and therefore to surface gravity, whereas the weaker FeH and 2o lines formed high in the atmosphere provide a better probe of rotational broadening.
G196-3B, the companion to the M2.5 G196-3A, is classified as an L2 dwarf (Kirkpatrick et al. 1999). The G196-3AB system is believed to be $\sim$20-300 Myr old, rather than the 1 Gyr thought to be typical of field L dwarfs (Rebolo et al. 1998). Figure 20 shows the order 61 spectrum of G196-3B plotted together with the L0.5 dwarf (2MASS 0746+20AB) and the L3.5 dwarf (2MASS 0036+18). Clearly, although the signal-to-noise ratio is poorer, all of the spectral features of G196-3B are much narrower than expected for an L2 dwarf. Using high resolution far-red optical spectra, Basri et al. (2000) measured 10 km s$^{-1}$ for the rotational velocity of G196-3B. The infrared K I lines have cores only half as wide as those of the L0.5 and L3.5 dwarfs, and much less pronounced pressure-broadened line wings. This behavior suggests that the line profiles result primarily from lower surface gravity, consistent with the age of this object.
Metallicity effects
-------------------
Burgasser et al. (2002) noted that the T6 dwarf 2MASS 0937+29 is peculiar because, despite having characteristics common to its spectral class, it appears to have no K I lines in a low-resolution $J$-band spectrum. If the K I lines were shallow and very broad, or sharp and weak, then they might escape detection at low spectral resolution. As shown in Figure 21 however, where we compare our high resolution spectrum of this T6 to that of the T4.5 dwarf 2MASS 0559$-$14, the K I lines are neither weak nor broad, they are completely absent. Given the good signal-to-noise ratio of this high resolution spectrum, there is no possibility that the K I lines were simply too weak to detect.
Disappearance of the K I lines is unexpected because other T6 dwarfs still show these features (Burgasser et al. 2002, M03). In fact, we have observed the T5.5 2MASS 2356$-$15 which exhibits broad K I lines similar to those of a T5. Strong 2o absorption features are clearly present in order 65 in the T6. These features show good correlation with the same features in the T4.5 but there is clear variation in the individual features. One possible reason for the absence of the K I lines in 2MASS 0937+29 is that this object has a different metallicity from other field dwarfs. Alternatively, the lines could be veiled by dust, but the trend at these spectral types is for dust to settle below the photosphere. In any case, there are other features in 2MASS 0937+29 that do not appear to be veiled.
2MASS 0937+29 was also classified as peculiar because of its extremely blue NIR colors and because it also has a very red optical spectrum for its type, as determined from CH$_{4}$ and 2o strengths. This combination of attributes led Burgasser et al. (2002) to propose that 2MASS 0937+29 is an old, metal-poor brown dwarf in which enhanced collision-induced H$_{2}$ absorption in the $K$-band gives the unusual blue NIR color. If this is an old T dwarf then it is also probably a fairly high-mass brown dwarf that has cooled to the temperature of a T6, and consequently it has a higher gravity, higher pressure atmosphere. The higher gravity could also result in increased pressure broadening of the 2o lines in 2MASS 0937+29 which would explain why these features seem muted and broader. In a recent study, Burgasser, Burrows and Kirkpatrick (2006) have found that the high surface gravity can result in a cooler effective temperature than equivalent T6 dwarfs. It is known (e.g., M03) that K I lines disappear around T7 or T8, hence it is possible that 2MASS 0937+29 is depleted in K I because of its low temperature rather than a low metallicity. Because the high-resolution spectra presented here show no residual trace of the K I lines, and because the 2o lines are unusually broad, we contend that 2MASS 0937+29 must be cooler and/or more metal poor than a normal T6 dwarf.\
Conclusions
===========
Using a sequence of high-resolution infrared $J$-band spectra (R$\sim$20,000 or 15 km s$^{-1}$) obtained with NIRSPEC on the Keck II telescope, we have studied the spectral morphology of objects from M6 to T6. The principal results of the survey are as follows:
\(1) Hundreds of small-scale spectral features are identified to be either FeH or H$_{2}$O absorption features. Over ten times as many FeH features as previously identified in brown dwarf spectra are now confirmed. A few features of CrH are also identified, but no convincing transitions of TiO or 4 at J-band are found in this sample. FeH features attain maximum strength in the mid-L dwarfs, while H$_{2}$O absorption becomes steadily stronger towards later spectral types.
\(2) FeH and 2o line widths are typically $\sim$20 km s$^{-1}$ for the late M dwarfs, but broaden abruptly by over a factor of two at the M to L transition. We interpret this effect as evidence for increased rotational velocities in L dwarfs.
\(3) The doublet of Al I at 1.31270 and 1.31543 $\mu$m is shown to be very sensitive to spectral type. This doublet weakens through M9 and then vanishes abruptly between M9 and L0. We suggest that this sudden disappearance is more a consequence of a transition in atmospheric chemistry than a simple decrease in atomic population levels resulting from a change in effective temperature. That is, the aluminum atoms are suddenly bound up into molecules and grains.
\(4) The wings and line widths of the K I doublets at 1.16934, 1.17761 $\mu$m and at 1.24357, 1.25256 $\mu$m increase systematically, while line depth weakens with later spectral type. The equivalent width (W) of the K I features reaches a maximum in the mid-L dwarfs, decreases and then remains almost constant through T4.5. The K I line profiles begin to exhibit pressure broadened wings as early as late M. Line widths (FWHM) range from $\sim$50 km s$^{-1}$ at M5 to almost 500 km s$^{-1}$ at T4.5. This effect is consistent with the much greater depth that is probed in cool T dwarf atmospheres at $J$-band.
\(5) As shown in Figure 12, a characteristic of the transition from L to T dwarfs is the decay of FeH spectral structure, resulting in a smooth spectrum at high resolution for late L dwarfs and early T dwarfs, before 2o dominant spectral structure develops.
\(6) The young L2, G196-3B, exhibits very narrow K I lines without extensive pressure-broadened wings, indicative of a lower gravity atmosphere.
\(7) Kelu-1AB, another L2, has exceptionally broad infrared lines, including FeH and H$_{2}$O features, confirming its status as a rapid rotator ($Vsin~i \sim$ 60 km s$^{-1}$).
\(8) Finally, the peculiar T6 dwarf 2MASS 0937+29 displays a complete absence of potassium, in contrast to other late T objects. We interpret this as either a metallicity effect (depletion of K atoms) or a cooler T$_{eff}$ for this high surface gravity object.
Although the sample of objects of different spectral types is relatively small, these high-resolution, high signal-to-noise spectra of M, L and T dwarfs should provide an important benchmark for the detailed development and improvement of model atmospheres.\
The authors wish to thank the staff of the Keck Observatory for their outstanding support. I.S.M. acknowledges the staff of the UCLA Infrared Laboratory and colleagues James Graham (UCB), James Larkin (UCLA) and Eric Becklin (UCLA) for their support throughout the development of the NIRSPEC instrument. We thank Adam Burrows, Katharina Lodders, Linda Brown, Didier Saumon, Richard Freedman, Travis Barman and Mark Marley for helpful discussions and opacity data. Finally, we thank the anonymous referee for a careful and complete critique of this paper. Work by S.S.K. was supported by the Astrophysical research Center for the Structure and Evolution of the Cosmos (ARCSEC) of Korea Science and Engineering Foundation through the Science Research Center (SRC) program. A.J.B. acknowledges support by NASA through Hubble Fellowship grant HST-HF-01137.01 awarded by the Space Telescope Science Institute, which is operated by the Association of universities for research in Astronomy, Inc., for NASA, under contract NAS 5-26555. This research has made use of the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This publication makes use of data from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center, funded by the National Aeronautics and Space Administration and the National Science Foundation. Finally, the authors wish to extend special thanks to those of Hawaiian ancestry on whose sacred mountain we are privileged to be guests.
Auman, J., Jr., 1967, ApJS, 14, 171 Basri, G. 2000, , 38, 485 Basri, G., Mohanty, S., Allard, F., Hauschildt, P. H., Delfosse, X., Mart[í]{}n, E. L., Forveille, T., & Goldman, B 2000, ApJ, 538, 363 Becklin, E.E., & Zuckerman, B. 1988, Nature, 336, 656 Bouy, H., Mart[í]{}n, E. L., Brandner, W., & Bouvier, J. 2005, AJ, 129, 511 Burgasser, A. J., et al. 2000, AJ, 120, 1100 Burgasser, A. J., et al. 2002, ApJ, 564, 421 Burgasser, A. J., McElwain, M. W., Kirkpatrick, J. D., Cruz, K. L., Tinney, C. G. & Reid, I. N. 2004, AJ, 127, 2856 Burgasser, A. J., Reid, I. N., Leggett, S. K., Kirkpatrick, J. D., Liebert, J. & Burrows, A. 2005, ApJ, 634, L177 Burgasser, A. J., Geballe, T. R., Leggett, S. K., Kirkpatrick, J. D., & Golimowski, D. A. 2006, ApJ, 637, 1067 Burgasser, A. J., Burrows, A., & Kirkpatrick, J. D. 2006, ApJ, 639, 1095 Burrows, A., Hubbard, W. B., Lunine, J. I., & Liebert, J. 2001, Rev. Mod. Phys., 73, 719 Burrows, A., Ram, R.S., Bernath, P., Sharp, C.M. & Milsom, J.A. 2002, ApJ, 577, 986 Burrows, A. & Voloyubev, M. 2003, ApJ, 583, 985. Clarke et al. 2003, MNRAS, 341, 239 Cushing, M. C., Rayner, J. T., Davis, S. P., & Vacca, W. D. 2003, ApJ, 582, 1066 Cushing, M. C., Rayner, J. T., & Vacca, W. D. 2005, ApJ, 623, 1115 Delfosse, X., et al. 1997, A & A, 327, L25 Delfosse, X., Tinney, C. G., Forveille, T., Epchtein, N., Borsenberger, J., Foque, P., Kimeswenger, S., Tiphene, D., 1999, A & AS, 135, 41 Galehouse, D. C., Davis, S. P., & Brault, J. W. 1980, ApJS, 42, 241 Geballe, T. R., Kulkarni, S. R., Woodward, C. E. & Sloan, G. C. 1996, ApJ, 467, L101 Geballe, T. R., et al. 2002, ApJ, 564, 466 Gelino, C. R., Kulkarni, S. R., & Stephens, D. C. 2006, PASP, 118, 611 Golimowski, D. A., et al. 2004, AJ, 127, 3516 Jones, H. R. A., Longmore, A. J., Jameson, R. F., & Mountain, C. M. 1994, MNRAS, 267, 413 Jones, H. R. A., Longmore, A. J., Allard, F., & Hauschildt, P. H. 1996, MNRAS, 280, 77 Kirkpatrick, J. D., Beichman, C. A., & Skrutskie, M. F. 1997, ApJ, 476, 311 Kirkpatrick, J. D., et al. 1999, ApJ, 519, 802 Kirkpatrick, J. D., et al. 2000, AJ, 120, 447 Kirkpatrick, J. D., Dahn, C. C., Monet, D. G., Reid, I. N., Gizis, J. E., Liebert, J., & Burgasser, A. J. 2001, AJ, 121, 3235 Kirkpatrick, J. D., Barman, T. S., Burgasser, A. J., McGovern, M. R., McLean, I. S., Tinney, C. G., & Lowrance, P. J. 2006, ApJ, 639, 1120 Leggett, S. K., et al. 2000, ApJ, 536, L35 Leggett, S. K., Allard, F., Geballe, T., Hauschildt, P. H., & Schweitzer, A. 2001, ApJ, 548, 908 Liu, M.C. & Leggett, S.K., 2005, ApJ, 634, 616 Lodders, K. 2002, ApJ, 577, 974. McGovern, M. R., Kirkpatrick, J. D., McLean, I. S., Burgasser, A. J., Prato, L., & Lowrance, P. J. 2004, ApJ, 600, 1020 McLean, I. S., et al. 1998, Proc. SPIE, 3354, 566 McLean, I. S., et al. 2000a, ApJ, 533, L45 McLean, I. S., Graham, J. R., Becklin, E. E., Figer, D. F., Larkin, J. E., Levenson, N. A., & Teplitz, H. I.. 2000b, Proc. SPIE, 4008, 1048 McLean, I. S., Prato, L., Kim, S. S., Wilcox, M. K., Kirkpatrick, J. D., & Burgasser, A. 2001, ApJ, 561, L115 McLean, I. S., McGovern, M. R., Burgasser, A. J., Kirkpatrick, J. D., Prato, L., Kim, S. S. 2003a, ApJ, 596, 561 (M03) McLean, I. S., McGovern, M. R., Prato, L., Burgasser, A. J., & Kirkpatrick, J. D. 2003b, in Brown Dwarfs, IAU Symp. 211, ed. E. Mart[í]{}n, ASP Conference Series. Mohanty, S., Basri, G., Jayawardhana, R., Allard, F., Hauschildt, P., & Ardila, D. 2004, ApJ, 609, 854 Partridge, H., & Schwenke, D. W. 1997, J. Chem. Phys., 106, 4618 Phillips, J.G. 1973, ApJS, 26, 313 Phillips, J.G., Davis, S.P., Lindgren, B., & Balfour, W.J. 1987, ApJS, 65,721 Prato, L., Simon, M., Mazeh, T., McLean, I.S., Norman, D. & Zucker, S. 2002, ApJ, 569, 863 Rebolo, R., Zapatero Osorio, M. R., Madruga, S., Bejar, V. J. S., Arribas, S., Licandro, J., 1998, Science, 282,1309 Reid, I. N., Kirkpatrick, J. D., Gizis, J. E., Dahn, C. C., Monet, D. G., Williams, R. J., Liebert, J., & Burgasser, A. J. 2000, AJ, 119, 369 Reid, I. N., Kirkpatrick, J. D., Liebert, J., Gizis, J. E., Dahn, C. C., & Monet, D. G., 2002, AJ, 124, 519 Ruiz, M. T., Leggett, S. K., & Allard, F. 1997, ApJ, 491, L107 Saumon, D., Marley, M. S., Lodders, K., & Freedman, R. S.2003, in Brown Dwarfs, IAU Symp. 211, ed. E. Mart[’i]{}n, ASP Conference Series. Schweitzer, A., Gizis, J. E., Hauschildt, P. H., Allard, F., & Reid, I. N. 2001, ApJ, 555, 368 Simon, M., Bender, C., & Prato, L. 2006, ApJ, 644, 1183 Testi, L., et al. 2001, ApJ, 522, L147 Tokunaga, A. T. 2000, in Allen’s Astrophysical Quantities, ed. A. N. Cox (4th ed.; New York: Springer), 151 Wilson, J. C., Kirkpatrick, J. D., Gizis, J. E., Skrutskie, M. F., Monet, D. G., & Houck, J. R. 2001, AJ, 122, 1989 Zapatero Osorio, M. R., Lane, B. F., Pavlenko, Ya., Mart[í]{}in, E. L., Britton, M., & Kulkarni, S. R. 2004, ApJ, 615, 958 Zapatero Osorio, M. R., Mart[í]{}n, E. L., Bouy, H., Tata, R., Deshpande, R., & Wainscoat, R. J. 2006, ApJ, 647, 1405
[lccccc]{}
G196-3A & 10 04 21.0 & 50 23 06 & M2.5 & 8.08$\pm$0.026 & 2002 Apr 23\
Wolf 359 (GJ 406) & 10 56 28.9 & 07 00 53 & M6 & 7.09$\pm$0.024 & 2002 Apr 23\
2MASSW J0140026+270150 & 01 40 02.6 & 27 01 50 & M9 & 12.49$\pm$0.021 & 2000 Dec 4\
2MASP J0345432+254023 & 03 45 43.2 & 25 40 23 & L0 & 14.00$\pm$0.027 & 2000 Dec 4\
2MASSI J0746425+200032 & 07 46 42.6 & 20 00 32 & L0.5 & 11.76$\pm$0.020 & 2002 Jan 1\
Kelu-1 & 13 05 40.2 &$-$25 41 06 & L2 & 13.41$\pm$0.026 & 2003 May 12\
G196-3B & 10 04 21.0 & 50 23 06 & L2 & 14.83$\pm$0.047 & 2002 Apr 23\
2MASSW J0036159+182110 & 00 36 16.2 & 18 21 10 & L3.5 & 12.47$\pm$0.027 & 2000 Dec 4\
GD165B & 14 24 39.1 & 09 17 10 & L4 & 15.69$\pm$0.078 & 2003 May 13\
2MASSW J1507476$-$162738 & 15 07 47.7 &$-$16 27 39 & L5 & 12.83$\pm$0.027 & 2000 Apr 25\
DENIS-P J0205.4$-$1159 & 02 05 29.4 &$-$11 59 30 & L7 & 14.59$\pm$0.030 & 2001 Oct 9\
SDSSp J042348.57$-$041403.5 & 04 23 48.6 &$-$04 14 04 & T0 & 14.47$\pm$0.027 & 2001 Oct 9\
SDSSp J125453.90$-$012247.4 & 12 54 53.9 &$-$01 22 47 & T2 & 14.89$\pm$0.035 & 2003 May 14\
2MASS J05591914$-$1404488 & 05 59 19.1 &$-$14 04 48 & T4.5 & 13.80$\pm$0.024 & 2001 Oct 9\
2MASSI J2356547$-$155310 & 23 56 54.8 &$-$15 53 11 & T5.5 & 15.82$\pm$0.057 & 2005 July 19\
2MASSI J0937347+293142 & 09 37 34.7 & 29 31 42 & T6p & 14.65$\pm$0.036 & 2003 May 12\
\
[lcccc]{}
58 & 1.30447–1.32370 & 192.3 & 0.188 & 84.5\
59 & 1.28262–1.30151 & 188.9 & 0.184 & 85.9\
60 & 1.26137–1.27999 & 186.2 & 0.182 & 87.5\
61 & 1.24081–1.25913 & 183.2 & 0.179 & 89.0\
62 & 1.22093–1.23899 & 180.6 & 0.176 & 90.7\
63 & 1.20168–1.21938 & 177.0 & 0.173 & 91.7\
64 & 1.18293–1.20011 & 171.8 & 0.168 & 91.9\
65 & 1.16496–1.18207 & 171.1 & 0.167 & 94.4\
\
[lccc]{} Al I & 1.3127007 & 4s $^{2}$S$_{1/2}$ - 4p $^{2}$P$_{3/2}$ & 3.143-4.087\
Al I & 1.3154345 & 4s $^{2}$S$_{1/2}$ - 4p $^{2}$P$_{1/2}$ & 3.143-4.085\
CrH & 1.18? & 6 bands of A$^{6}$ $\Sigma^{+}$ - X$^{6}$ $\Sigma^{+}$ &\
Fe I & 1.1693174 & a$^{5}$P$_{1}$ - z$^{5}$D$^{o}$$_{1}$ & 2.223-3.283\
Fe I & 1.1786490 & b$^{3}$P$_{2}$ - z$^{3}$D$^{o}$$_{3}$ & 2.831-3.884\
Fe I & 1.1886098 & a$^{5}$P$_{2}$ - z$^{5}$D$^{o}$$_{3}$ & 2.198-3.241\
Fe I & 1.1887337 & a$^{5}$P$_{1}$ - z$^{5}$D$^{o}$$_{2}$ & 2.223-3.266\
Fe I & 1.1976325 & a$^{5}$P$_{3}$ - z$^{5}$D$^{o}$$_{4}$ & 2.176-3.211\
FeH & 1.1939 band head & 0-1 band of F$^{4}$ $\Delta$-X$^{4}$ $\Delta$ &\
FeH & 1.2389 band head & 1-2 band of F$^{4}$ $\Delta$-X$^{4}$ $\Delta$ &\
H$_{2}$O & 1.135 & $\nu_{1}+\nu_{2}+\nu_{3}$ &\
H$_{2}$O & 1.331 & 2$\nu_{3}$ band &\
K I & 1.1693420 & 4p $^{2}$P$_{1/2}$ - 3d $^{2}$D$_{3/2}$ & 1.610-2.670\
K I & 1.1772861 & 4p $^{2}$P$_{3/2}$ - 3d $^{2}$D$_{3/2}$ & 1.617-2.670\
K I & 1.1776061 & 4p $^{2}$P$_{3/2}$ - 3d $^{2}$D$_{5/2}$ & 1.617-2.670\
K I & 1.2435675 & 4p $^{2}$P$_{1/2}$ - 5s $^{2}$S$_{1/2}$ & 1.610-2.607\
K I & 1.2525560 & 4p $^{2}$P$_{3/2}$ - 5s $^{2}$S$_{1/2}$ & 1.617-2.607\
Mn I & 1.290329 & a$^{6}$D$_{9/2}$ - z$^{6}$P$_{7/2}$ & 2.114-3.075\
Na I & 1.268261 & 3d $^{2}$D$_{5/2,7/2}$ - 5f $^{2}$F$_{5/2}$ & 3.617-4.595\
Na I & 1.268269 & 3d $^{2}$D$_{3/2}$ - 5f $^{2}$F$_{5/2}$ & 3.617-4.595\
Ti I & 1.1896124 & b$^{3}$F$_{2}$ - z$^{3}$D$^{o}$$_{1}$ & 1.430-2.472\
Ti I & 1.2674567 & b$^{3}$F$_{2}$ - z$^{3}$F$^{o}$$_{3}$ & 1.430-2.408\
Ti I & 1.2834947 & b$^{3}$F$_{2}$ - z$^{3}$F$^{o}$$_{2}$ & 1.430-2.396\
Ti I & 1.2850544 & b$^{3}$F$_{3}$ - z$^{3}$F$^{o}$$_{3}$ & 1.443-2.408\
[llccc]{}
M2.5 & G196-3A & 0.36$\pm$0.03 & 1.3$\pm$0.4 & 39$\pm$4\
M6 & Wolf 359 & 0.72$\pm$0.01 & 5.2$\pm$0.5 & 64$\pm$6\
M9 & 2MASS J0140+27 & 0.80$\pm$0.01 & 7.5$\pm$0.7 & 110$\pm$11\
L0 & 2MASS J0345+25 & 0.66$\pm$0.02 & 9.3$\pm$0.9 & 180$\pm$18\
L0.5 & 2MASS J0746+20AB & 0.71$\pm$0.01 & 11.5$\pm$1.1 & 210$\pm$21\
L2 & Kelu-1AB & 0.60$\pm$0.02 & 14.1$\pm$1.4 & 320$\pm$32\
L3.5 & 2MASS J0036+18 & 0.69$\pm$0.02 & 14.4$\pm$1.4 & 240$\pm$24\
L4 & GD165B & 0.80$\pm$0.01 & 12.6$\pm$1.3 & 230$\pm$23\
L5 & 2MASS J1507$-$16 & 0.68$\pm$0.02 & 10.0$\pm$1.0 & 240$\pm$24\
L7 & DENIS J0205$-$11AB & 0.47$\pm$0.05 & 8.6$\pm$1.3 & 290$\pm$29\
T0 & SDSS J0423$-$04AB & 0.47$\pm$0.06 & 9.3$\pm$1.4 & 240$\pm$24\
T2 & SDSS J1254$-$01 & 0.60$\pm$0.05 & 9.8$\pm$1.5 & 270$\pm$27\
T4.5 & 2MASS J0559$-$14 & 0.41$\pm$0.07 & 9.4$\pm$1.4 & 490$\pm$50\
[llccc]{}
M2.5 & G196-3A & 0.24$\pm$0.04 & 1.3$\pm$0.1 & 47$\pm$5\
M6 & Wolf 359 & 0.65$\pm$0.02 & 5.6$\pm$0.6 & 65$\pm$7\
M9 & 2MASS J0140+27 & 0.77$\pm$0.01 & 9.0$\pm$0.9 & 78$\pm$8\
L0 & 2MASS J0345+25 & 0.61$\pm$0.02 & 11.5$\pm$1.2 & 220$\pm$22\
L0.5 & 2MASS J0746+20AB & 0.67$\pm$0.02 & 14.1$\pm$1.4 & 230$\pm$23\
L2 & Kelu-1AB & 0.57$\pm$0.02 & 14.1$\pm$1.4 & 320$\pm$31\
L3.5 & 2MASS J0036+18 & 0.63$\pm$0.02 & 11.4$\pm$1.1 & 290$\pm$29\
L4 & GD165B & 0.75$\pm$0.01 & 14.0$\pm$1.4 & 150$\pm$15\
L5 & 2MASS J1507$-$16 & 0.62$\pm$0.02 & 14.7$\pm$1.5 & 270$\pm$27\
L7 & DENIS J0205$-$11AB & 0.36$\pm$0.06 & 8.2$\pm$1.2 & 390$\pm$39\
T0 & SDSS J0423$-$04AB & 0.36$\pm$0.08 & 7.5$\pm$1.1 & 300$\pm$30\
T2 & SDSS J1254$-$01 & 0.49$\pm$0.06 & 7.9$\pm$1.2 & 260$\pm$26\
T4.5 & 2MASS J0559$-$14 & 0.29$\pm$0.09 & 9.4$\pm$1.4 & 460$\pm$46\
[llcc]{}
M2.5 & G196-3A & 0.03$\pm$0.04 & 0.022$\pm$0.01\
M6 & Wolf 359 & 0.12$\pm$0.04 & 0.080$\pm$0.016\
M9 & 2MASS J0140+27 & 0.17$\pm$0.04 & 0.140$\pm$0.029\
L0 & 2MASS J0345+25 & 0.25$\pm$0.04 & 0.110$\pm$0.022\
L0.5 & 2MASS J0746+20AB & 0.29$\pm$0.04 & 0.100$\pm$0.02\
L2 & Kelu-1AB & 0.15$\pm$0.04 & 0.110$\pm$0.022\
L3.5 & 2MASS J0036+18 & 0.28$\pm$0.04 & 0.160$\pm$0.023\
L4 & GD165B & 0.35$\pm$0.05 & 0.35 $\pm$0.05\
L5 & 2MASS J1507$-$16 & 0.23$\pm$0.04 & 0.220$\pm$0.044\
L7 & DENIS J0205$-$11AB & 0.19$\pm$0.08 & 0.270$\pm$0.067\
T0 & SDSS J0423$-$04AB & 0.08$\pm$0.07 & 0.280$\pm$0.055\
T2 & SDSS J1254$-$01 & 0.16$\pm$0.07 & 0.380$\pm$0.075\
T4.5 & 2MASS J0559$-$14 & 0.06$\pm$0.09 & 0.620$\pm$0.093\
[^1]: Data presented herein were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation.
[^2]: http://www.astro.ucla.edu/$\sim$mclean/BDSSarchive/
[^3]: See http://www2.keck.hawaii.edu/inst/nirspec/redspec/index.html
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'New generalized equations of motion for the Weber-Clebsch potentials that describe both the Navier-Stokes and MHD dynamics are derived. These depend on a new parameter, which has dimensions of time for Navier-Stokes and inverse velocity for MHD. Direct numerical simulations (DNS) are performed. For Navier-Stokes, the generalized formalism captures the intense reconnection of vortices of the Boratav, Pelz and Zabusky flow, in agreement with the previous study by Ohkitani and Constantin. For MHD, the new formalism is used to detect magnetic reconnection in several flows: the $3$D Arnold, Beltrami and Childress (ABC) flow and the ($2$D and $3$D) Orszag-Tang vortex. It is concluded that periods of intense activity in the magnetic enstrophy are correlated with periods of increasingly frequent resettings. Finally, the positive correlation between the sharpness of the increase in resetting frequency and the spatial localization of the reconnection region is discussed.'
author:
- Carlos Cartes
- 'Miguel D. Bustamante'
- Annick Pouquet
- 'Marc E. Brachet'
title: 'Generalized Eulerian-Lagrangian description of Navier-Stokes and resistive MHD dynamics'
---
Introduction
============
The Eulerian-Lagrangian formulation of the (inviscid) Euler dynamics in terms of advected Weber-Clebsch potentials [@ConstantinLocal] was extended by Constantin [@Constantin] to cover the (viscous) Navier-Stokes dynamics. Ohkitani and Constantin (OC) [@Ohk] then performed numerical studies of this formulation of the Navier-Stokes equations. They concluded that the diffusive Lagrangian map becomes noninvertible under time evolution and requires resetting for its calculation. They proposed the observed sharp increase of the frequency of resettings as a new diagnostic of vortex reconnection.
We were able to recently complement these results, using an approach that is based on a generalized set of equations of motion for the Weber-Clebsch potentials that turned out to depend on a parameter $\tau$ which has the unit of time for the Navier-Stokes case [@CBB07] (the MHD case is different, see below Section \[sec: MPsec\]). The OC formulation is the (singular) $\tau \to 0$ limit case of our generalized formulation. Using direct numerical simulations (DNS) of the viscous Taylor-Green vortex [@TG1937] we found that for $\tau\ne0$ the Navier-Stokes dynamics was well reproduced at small enough Reynolds numbers [*without*]{} resetting. However, performing resettings allowed computation at much higher Reynolds number.
The aim of the present article is to extend these results to different flows, both in the Navier-Stokes case and in magnetohydrodynamics, and thereby obtain a new diagnostic for [*magnetic*]{} reconnection. Our main conclusion is that intense reconnection of magnetic field lines is indeed captured in our new generalized formulation as a sharp increase of the frequency of resettings. Here follows a summary of our principal results.
We first derive new generalized equations of motion for the Weber-Clebsch potentials that describe both the Navier-Stokes and MHD dynamics. Performing DNS of the Boratav, Pelz and Zabusky flow [@Peltz1992], that was previously used by Ohkitani and Constantin [@Ohk], we first check that our generalized formalism captures the intense Navier-Stokes vortex reconnection of this flow. We demonstrate the reconnection of vortices is actually occurring at the instant of intense activity in the enstrophy, near the lows of the determinant that trigger the resettings. We then study the correlation of magnetic reconnection with increase of resetting frequency by performing DNS of several prototypical MHD flows: the $3$D Arnold, Beltrami and Childress (ABC) flow [@Archontis] and the Orszag-Tang vortex in $2$D [@OT2D] and $3$D [@OT3D].
Theoretical Framework {#Theory}
=====================
General Setting
---------------
### Weber-Clebsch representation for a class of evolution equations
Let us consider a $3$D vector field $\mathbf{Z}$ depending on time and ($3$-dimensional) space, with coordinates $(x^1,x^2,x^3,t)$. Assume $\mathbf{Z}$ satisfies an evolution equation of the kind: $$\begin{aligned}
\label{eq:Z} \frac{D \mathbf{Z}}{D t}&=&- \nabla P + \sum_{\alpha=1}^3 u_\alpha \nabla {Z}_\alpha + \kappa \triangle \mathbf{Z}\,\\
\label{eq: divergence}\nabla \cdot \mathbf{Z} &=& 0\,,\end{aligned}$$ where greek indices $\alpha, \beta$ denote vector field components running from $1$ to $3$, $\mathbf{u}$ is a given $3$D velocity field and we have used the convective derivative defined by $$\nonumber
\frac{D}{D t} \equiv \frac{\partial}{\partial t}+ (\mathbf{u}\cdot
\nabla)\,.$$
In the following sections, two different cases will be considered. In section \[sec: NS case\] (Navier-Stokes case) the vector field $\mathbf{Z}$ will correspond to the velocity field $\mathbf{u}$, whereas in section \[sec: MHD case\] (MHD case) it will correspond to the magnetic vector potential $\mathbf{A}$.
Let us first recall that performing a change from Lagrangian to Eulerian coordinates on the Weber transformation [@Lamb] leads to a description of the Euler equations as a system of three coupled active vector equations in a form that generalizes the Clebsch variable representation [@ConstantinLocal].
Our starting point will be to apply this classical Weber-Clebsch representation to the field $\mathbf{Z}$:
$$\begin{aligned}
\label{eq:WC_Z} \mathbf{Z} &=& \sum_{i=1}^3 \lambda^i
\nabla \mu^i - \nabla \phi\,,\end{aligned}$$
where each element of the $3$ pairs of Weber-Clebsch potentials $(\lambda^i, \mu^i), \,\, i=1,2,3$ is a scalar function.
Performing a variation on the Weber-Clebsch representation (\[eq:WC\_Z\]) yields the relation $$\label{eq: delta(Clebsch transformation)}
\mathbf{\delta Z} = \sum_{i=1}^3 \left(\delta \lambda^i \nabla \mu^i - \delta \mu^i \nabla \lambda^i\right) - \nabla \left(\delta \phi-
\sum_{i=1}^3 \delta \mu^i \lambda^i \right)\,,$$ where the symbol $\delta$ stands for any (spatial or temporal) partial derivative. Taking into account the identity , it is straightforward to derive from (\[eq: delta(Clebsch transformation)\]) the following explicit expression for the convective derivative of the vector field $\mathbf{Z}$:
$$\label{eq: Dt(Clebsch transformation)}
\frac{D \mathbf{Z}}{D t} = \sum_{i=1}^3 \left(\frac{D \lambda^i}{D t} \nabla \mu^i - \frac{D \mu^i}{D t} \nabla \lambda^i\right) -
\sum_{\alpha=1}^3 Z_\alpha \nabla u_\alpha - \nabla \left(\frac{D \phi}{D t} - \sum_{i=1}^3 \frac{D \mu^i}{D t} \lambda^i \right)\,,$$
### Equations of motion for the potentials
Following steps that are similar to those presented in our previous paper [@CBB07], we now derive a system of equations of motion for the Weber-Clebsch potentials (\[eq:WC\_Z\]) that is equivalent to the original equation (\[eq:Z\]). If we use the RHS of equation (\[eq:Z\]) to replace the LHS of our general identity (\[eq: Dt(Clebsch transformation)\]), the resulting relation can be solved for the time derivative of the potentials: $$\begin{aligned}
\label{eq: evolution_tilde Clebsch_lambda}
\frac{D \lambda^i}{D t} &=& \kappa \triangle \lambda^i + \widetilde{L}^i[\lambda,\mu]\\
\label{eq: evolution_tilde Clebsch_mu} \frac{D\mu^i}{D t} &=&\kappa \triangle \mu^i + \widetilde{M}^i[\lambda,\mu]\,.\end{aligned}$$ Here $\widetilde{L}^i, \widetilde{M}^i$ obey the linear equation $$\label{eq: linear system tilde L M}
\sum_{i=1}^3 \left(\widetilde{L}^i \nabla \mu^i - \widetilde{M}^i \nabla \lambda^i\right) = \widetilde{\mathbf{f}}- \nabla \widetilde{G}\,,$$ where $$\label{eq: tilde f}
\widetilde{\mathbf{f}} = 2 \kappa \sum_{i=1}^3 \sum_{\alpha=1}^3
\partial_\alpha \lambda^i \partial_\alpha \nabla
\mu^i$$ and $\widetilde{G}[\lambda,\mu]$ is an arbitrary scalar related to the non-unique separation of a gradient part in eq.(\[eq: Dt(Clebsch transformation)\]): $$\label{eq: evolution_phi} \frac{D \phi}{D t} - P = \sum_{i=1}^3\lambda^i \widetilde{M}^i - \widetilde{G} - \mathbf{u} \cdot \mathbf{Z} .$$
The “divergence-less gauge” (\[eq: divergence\]) allows one to express $\phi$ in terms of $\lambda^i$ and $\mu^i$, as the solution of the linear equation $$\label{eq:lap_phi}
\triangle \phi = \sum_{i=1}^3 \nabla \cdot (\lambda^i \nabla\mu^i).$$ Thus there is no need to solve equation (\[eq: evolution\_phi\]) for the field $\phi$, since this equation is identically satisfied when $\phi$ is determined by eq. (\[eq:lap\_phi\]).
Equation (\[eq: linear system tilde L M\]) above is a system of $3$ linear equations for the $6$ unknowns $\widetilde{L}^i, \widetilde{M}^i$. When $\kappa=0$ there is a simple solution to (\[eq: linear system tilde L M\]): $\widetilde{L}^i = \widetilde{M}^i = \widetilde{G} =0$. In this case the evolution equations (\[eq: evolution\_tilde Clebsch\_lambda\]) and (\[eq: evolution\_tilde Clebsch\_mu\]) represent simple advection.
### Moore-Penrose solution and minimum norm {#sec: MPsec}
The linear system (\[eq: linear system tilde L M\]) is underdetermined ($3$ equations for $6$ unknowns). In order to find a solution to the system we need to impose extra conditions. Since $\widetilde{L}^i, \widetilde{M}^i$ appear in the equations on an equal footing, it is natural to supplement the system by a requirement of minimum norm, namely that $$\label{eq: vnorm}
\sum_{i=1}^3 (\widetilde{L}^i\widetilde{L}^i+{\tau^{-2}}\widetilde{M}^i\widetilde{M}^i)$$ be the smallest possible (this is the so-called general Moore-Penrose approach [@Moore; @Penrose; @Ben-Israel], see also our previous paper [@CBB07]). The parameter $\tau$ has physical units equal to $[\widetilde{M}/\widetilde{L}]$. Using eqs.(\[eq: evolution\_tilde Clebsch\_lambda\]),(\[eq: evolution\_tilde Clebsch\_mu\]) these are the units of $[\mu/\lambda]$. It will turn out (see equation (\[eq: mu def\]) below) that $[\mu]= L$ (length) and this implies from eq.(\[eq:WC\_Z\]) that $[\lambda]=[\mathbf{Z}]$. Therefore the units of $\tau$ are $$\nonumber [\tau] = \frac{L}{[\mathbf{Z}]}.$$ In the Navier-Stokes case (section \[sec: NS case\]) $[\mathbf{Z}]=[\mathbf{u}] = L T^{-1}$ and thus $[\tau]=T$, whereas in the MHD case (section \[sec: MHD case\]) $[\mathbf{Z}] = [\mathbf{A}] = L^2T^{-1}$ and thus $[\tau]=T L^{-1}$.
The Moore-Penrose solution to (\[eq: linear system tilde L M\]), that minimizes the norm (\[eq: vnorm\]), is explicitly given in equations (A6,A7) of reference [@CBB07]. Inserting this solution in (\[eq: evolution\_tilde Clebsch\_lambda\]),(\[eq: evolution\_tilde Clebsch\_mu\]) we finally obtain the explicit evolution equations $$\begin{aligned}
\label{eq: New_Alg L}
\frac{D \lambda^i}{D t} &=& \kappa \triangle \lambda^i + \,\,\,\,\nabla \mu^i \cdot \mathbb{H}^{-1}\cdot \left(\widetilde{\mathbf{f}} - \nabla \widetilde{G}\right)\\
\label{eq: New_Alg M} \frac{D\mu^i}{D t} &=&\kappa \triangle \mu^i - \tau^{2}\,\nabla \lambda^i \cdot \mathbb{H}^{-1}\cdot
\left(\widetilde{\mathbf{f}} - \nabla \widetilde{G}\right)\,,\end{aligned}$$ where $\widetilde{\mathbf{f}}$ is given in eq.(\[eq: tilde f\]), the dot product denotes matrix or vector multiplication of $3$-dimensional tensors, and $\mathbb{H}^{-1}$ is the inverse of the square symmetric $3\times 3$ matrix $\mathbb{H}$, defined by its components: $$\label{eq: mat}
\mathbb{H}_{\alpha \beta} \equiv \sum_{i=1}^3 \left(\tau^{2}\,\partial_\alpha \lambda^i
\partial_\beta \lambda^i+\partial_\alpha
\mu^i
\partial_\beta \mu^i\right)\,.$$ These evolution equations together with the particular choice for the arbitrary function $\widetilde{G}$ (see equation (A11) of reference [@CBB07]) $$\label{eq: explicit G}
\widetilde{G}= \triangle^{-1}\nabla\cdot \widetilde{\mathbf{f}},$$ is our new algorithm.
In the Navier-Stokes case, we showed in a previous paper [@CBB07] that the limit $\tau \to 0$ corresponds to the approach used by Ohkitani and Constantin [@Ohk]. In the general case (Navier-Stokes as well as MHD), we remark that the matrix $\mathbb{H}$ (see equation (\[eq: mat\])) can be written (using obvious notation) as $\mathbb{H} = (\nabla {{\mbox{\boldmath{$\mu$}}}})\cdot (\nabla {{\mbox{\boldmath{$\mu$}}}})^{\rm T}+\tau^2 (\nabla {{\mbox{\boldmath{$\lambda$}}}})\cdot
(\nabla {{\mbox{\boldmath{$\lambda$}}}})^{\rm T}$, which has a very simple structure in the limit $\tau\to 0$ . Because the condition $\det(\nabla {{\mbox{\boldmath{$\mu$}}}})=0$ is generically obtained at lower codimension than the condition $\det \mathbb{H} = 0$, the limit $\tau\to 0$ is [*singular*]{}.
Navier-Stokes equations {#sec: NS case}
-----------------------
The standard incompressible NS equations can be written in the form: $$\begin{aligned}
\nonumber
\frac{D \mathbf{u}}{D t}&=& -\nabla \left(p + \frac{1}{2}|\mathbf{u}|^2\right) + \sum_{\alpha=1}^3 u_\alpha \nabla u_\alpha + \nu \triangle \mathbf{u} \\
\nonumber \nabla \cdot \mathbf{u} &=& 0\,,\end{aligned}$$ which is indeed of the general form (\[eq:Z\]), (\[eq: divergence\]) with $ \mathbf{Z} = \mathbf{u}$, $\kappa = \nu$ and $P=p +
\frac{1}{2}|\mathbf{u}|^2$.
MHD equations {#sec: MHD case}
-------------
The standard incompressible MHD equations for the fluid velocity $\mathbf{u}$ and the induction field $\mathbf{b}$, expressed in Alfvenic velocity units, can be written in the form: $$\begin{aligned}
\label{eq:momentum}
\frac{D \mathbf{u}}{D t}&=& -\nabla p + \nu \triangle \mathbf{u} + (\nabla \times \mathbf{b}) \times \mathbf{b}\\
\label{eq:B}
\frac{D \mathbf{b}}{D t} &=& (\mathbf{b}\cdot
\nabla)\mathbf{u} + \eta \triangle \mathbf{b} \\
\label{eq: vinc} \nabla \cdot \mathbf{u} &=& 0\\
\nonumber \nabla \cdot \mathbf{b} &=& 0\,,\end{aligned}$$ where $\nu$ and $\eta$ are the viscosity and magnetic resistivity, respectively.
We introduce the vector potential in the Coulomb gauge: $$\begin{aligned}
\nonumber \mathbf{b} &=& \nabla \times \mathbf{A}\\
\nonumber \nabla \cdot \mathbf{A} &=& 0\,.\end{aligned}$$ Using the identity $\nabla \times (\sum_{\alpha=1}^3 u_\alpha \nabla {A}_\alpha - (\mathbf{u} \cdot \nabla) \mathbf{A}) = (\mathbf{b} \cdot
\nabla) \mathbf{u} - (\mathbf{u} \cdot \nabla) \mathbf{b} - (\nabla \cdot \mathbf{u}) \mathbf{b}\,$ and the incompressibility condition (\[eq: vinc\]), eq. (\[eq:B\]) can be written as $$\begin{aligned}
\nonumber
\frac{D \mathbf{A}}{D t}&=& - \nabla c + \sum_{\alpha=1}^3 u_\alpha \nabla {A}_\alpha + \eta \triangle \mathbf{A},\end{aligned}$$ which is indeed of the general form (\[eq:Z\]) with $ \mathbf{Z} = \mathbf{A}$, $\kappa = \eta$ and $P = c$.
Numerical Results {#Direct Numerical Simulations}
=================
Implementation
--------------
### Initial conditions in pseudo-spectral method
Spatially periodic fields can be generated from the Weber-Clebsch representation (\[eq:WC\_Z\]) by setting $$\label{eq: mu def}
\mu^i = x^i + \mu^i_{\rm p},$$ and assuming that $\mu^i_{\rm p}$ and the other fields $\lambda^i$ and $\phi$ appearing in (\[eq:WC\_Z\]) are periodic. Indeed, any given periodic field $\mathbf{Z}$ can be represented in this way by setting $$\begin{aligned}
\label{eq: resetting mu}
\mu^i_{\rm p} &=&0\\
\label{eq: resetting lambda}
\lambda^i&=&Z^i\\
\label{eq: resetting phi}
\phi &=&0.\end{aligned}$$ Note that the time independent non-periodic part of $\mu^i$ of the form given in (\[eq: mu def\]) is such that the [*gradients*]{} of $\mu^i$ are periodic. It is easy to check that this representation is consistent with the generalized equations of motions (\[eq: New\_Alg L\],\[eq: New\_Alg M\]). We chose to use standard Fourier pseudo-spectral methods, both for their precision and for their ease of implementation [@Got-Ors].
### Resettings and reconnection
Following Ohkitani and Constantin [@Ohk], we now define resettings. Equations (\[eq: resetting mu\]), (\[eq: resetting lambda\]) and (\[eq: resetting phi\]) are used not only to initialize the Weber-Clebsch potentials at the start of the calculation but also to [*reset*]{} them to the current value of the field $ \mathbf{Z}$, obtained from (\[eq:WC\_Z\]) and (\[eq:lap\_phi\]), whenever the minimum of the determinant of the matrix (\[eq: mat\]) falls below a given threshold $$\nonumber
\det \mathbb{H} \le\epsilon^2 .$$
It is possible to capture reconnection events using resettings. The rationale for this approach is that reconnection events are associated to localized, intense and increasingly fast activity which will drive the potentials to a (unphysical) singularity in a finite time. One way to detect this singularity is via the alignment of the gradients of the potentials, which leads to the vanishing of $\det \mathbb{H}$ at the point(s) where this intense activity or ‘anomalous diffusion’ is taking place. Now, the time scale of this singularity is much smaller than the time scale of the reconnection process itself [@Ohk], so when $\det \mathbb{H}$ goes below the given threshold and a resetting of the potentials is performed, the anomalous diffusion starts taking place again, more intensely as we approach the fastest reconnection period, driving the new (reset) potentials to a new finite-time singularity, in a time scale that decreases as we approach this period. Therefore, successive resettings will be more and more frequent near the period of fastest reconnection, and that is what we observe in the numerical simulations. This procedure will be used to capture reconnection events in particular flows in both the Navier-Stokes case ($\mathbf{Z} =
\mathbf{u}$, Section \[sec:NS\]) and the MHD case ($\mathbf{Z} = \mathbf{A}$, Sections \[sec:2DOT\] and \[sec:3DOT\]).
Navier-Stokes case: BPZ Flow, resettings and reconnection {#sec:NS}
---------------------------------------------------------
Ohkitani and Constantin (OC) [@Ohk] used a flow that initially consists of two orthogonally placed vortex tubes that was previously introduced in Boratav, Pelz and Zabusky (BPZ) [@Peltz1992] to study in detail vortex reconnection. Our previous numerical study of the generalized Weber-Clebsch description of Navier-Stokes dynamics [@CBB07] was performed using the Taylor-Green vortex, a flow in which vorticity layers are formed in the early stage, followed by their rolling-up by Kelvin-Helmholtz instability [@Brachet1]. It can be argued [@Ohk] that cut-and-connect type reconnections are much more pronounced in the BPZ flow than in the Taylor-Green flow. In this section we present comparisons, performed on the BPZ flow, of our $\tau \ne 0$ generalized algorithm with direct Navier-Stokes simulations and with OC original approach. The potentials are integrated with resettings in resolution $128^3$ for a Reynolds number of $R=1044$, which is the one used by BPZ and OC.
The BPZ initial data is explicitly given in [@Peltz1992].
### Comparison of Weber-Clebsch algorithm with DNS of Navier-Stokes
In order to characterize the precision of the $\tau \ne 0$ Weber-Clebsch algorithm, we now compare the velocity field $\mathbf{Z} = {\bf u}$ obtained from (\[eq:WC\_Z\]) and (\[eq:lap\_phi\]), by evolving the Weber-Clebsch potentials using (\[eq: New\_Alg L\])–(\[eq: explicit G\]), with the velocity field obtained independently by direct Navier-Stokes evolution from the BPZ initial data.
More precisely, we compare the associated kinetic enstrophy $\Omega(t)=\sum_{k}k^2 E(k,t)$ where the kinetic energy spectrum $E(k,t)$ is defined by averaging the Fourier transform ${\bf \hat u}({\bf k',t})$ of the velocity field (\[eq:WC\_Z\]) on spherical shells of width $\Delta k = 1$, $$\nonumber
E(k,t) = {\frac1 2} \sum_{k-\Delta k/2< |{\bf k'}| < k + \Delta k/2} |{\bf \hat u}({\bf k',t})|^2 \, .$$
![ \[EnstrophyBPZ\]](Enstrophy.eps){width="8.0cm"}
![ \[DeltaResBPZ\]](deltaRes.eps){width="8.0cm"}
Figure \[EnstrophyBPZ\] shows that the kinetic enstrophy is well resolved, independently of the choice of the parameter $\tau$.
### Time between resettings as a method for reconnection capture
In this section we study the influence of the parameter $\tau$ on the temporal distribution of the intervals $\Delta t_j = t_j - t_{j-1}$ between resetting times $t_j$, at fixed value of the resetting threshold $\epsilon^2 = 0.1$. Using the same Reynolds number and resolution that was used to create Fig. \[EnstrophyBPZ\], Figure \[DeltaResBPZ\] is a plot of $\Delta t$ as a function of time, for simulations with different values of $\tau$. In the same figure we also show the corresponding $\Delta t$ for a replica of the simulation performed by OC, that is in excellent agreement with our general case.
We see that, independently of $\tau$, there are sharp minima in $\Delta t$ during the periods of maximum enstrophy (see Fig. \[EnstrophyBPZ\]). Inspection of figure \[FiguraRecon\] demonstrates that the deepest minimum corresponds in fact to the time when reconnection is taking place. The main tubes in the left and right figures are isosurfaces of vorticity corresponding to $60\%$ of the maximum vorticity, which is attained inside each of the main tubes.
Figure \[FiguraDet\] (left) shows that the spatial region where the determinant $\det \mathbb{H}$ goes below the threshold before each resetting corresponds to a small, localized neighborhood between the main interacting vortices. This region is seen in the right figure as a bridge connecting the two vortices: this bridge is an isosurface of vorticity corresponding to $73\%$ of the maximum vorticity, which is attained inside the bridge. The main tubes correspond to isosurfaces of $30\%$ of the maximum vorticity. Note that this behavior of the determinant $\det
\mathbb{H}$ is also true for any value of $\tau$ (data not shown), confirming in this way the original rationale for the study of reconnection with the aid of resettings.
Figures \[FiguraRecon\] and \[FiguraDet\] were made using the VAPOR [@ncarVapor1; @ncarVapor2]visualization software.
![ \[FiguraRecon\]](ReconAntesDesp.eps){width="8.0cm"}
![ \[FiguraDet\]](ReconDurante.eps){width="8.0cm"}
MHD Flows
---------
In this section we study MHD flows with simple initial conditions. The magnetic potential $\mathbf{Z} = \mathbf{A}$ is obtained in terms of the Weber-Clebsch potentials from (\[eq:WC\_Z\]) and (\[eq:lap\_phi\]), and the Weber-Clebsch potentials are evolved using equations (\[eq: New\_Alg L\])–(\[eq: explicit G\]).
We treat the evolution of the velocity field in two different ways: (i) As a kinematic dynamo (ABC flow, Section \[sec:ABC\]), where the velocity is kept constant in time; (ii) Using the full MHD equations (Orszag-Tang $2$D and $3$D, Sections \[sec:2DOT\] and \[sec:3DOT\]), where the velocity field is evolved using the momentum equation (\[eq:momentum\]).
To compare with DNS of the induction equation (\[eq:B\]) for the magnetic field we proceed analogously as in the Navier-Stokes case. We compare the magnetic enstrophy [@Dahl89] $\Omega_m(t) = \sum_{k}k^2 E_m(k,t)$, where the magnetic energy spectrum $E_m(k,t)$ is defined by averaging the Fourier transform ${\bf \hat b}({\bf k',t})$ of the magnetic field $\mathbf{b} = \nabla \times \mathbf{A}$ (with $\mathbf{A}$ given by (\[eq:WC\_Z\])) on spherical shells of width $\Delta k = 1$, $$\nonumber
E_m(k,t) = {\frac1 2} \sum_{k-\Delta k/2< |{\bf k'}| < k + \Delta k/2} |{\bf \hat b}({\bf k',t})|^2 \, .$$ Note that magnetic dissipation is the square current.
Resettings will be performed with a resetting threshold $\epsilon^2 = 0.1$. We have checked that $\epsilon^2=0.4$ and $\epsilon^2=0.025$ give results that vary only slightly (figures not shown). This is an evidence of the robustness of the resetting method and a validation of the rationale for the use of resettings to diagnose reconnection.
### Kinematic dynamo: ABC Flow {#sec:ABC}
We have used the ABC [@Archontis] velocity: $$\begin{aligned}
u_x&=& B_0 \cos k_0 y + C_0 \sin k_0 z \nonumber\\
u_y&=& C_0 \cos k_0 z + A_0 \sin k_0 x \nonumber\\
u_z &=& A_0 \cos k_0 x + B_0 \sin k_0 y \,, \nonumber\end{aligned}$$ with $k_0=2$ and $A_0=B_0=C_0=1$. We used an initial magnetic seed that reads $$\begin{aligned}
A_x&=&0 \nonumber\\
A_y&=& 0 \nonumber\\
A_z &=& d_0 \sin x \sin y \nonumber\,.\end{aligned}$$ The magnetic resistivity has been chosen as $\eta = 1/12$ and we have set $d_0 = 1/100$ for simplicity (its value is unimportant in the kinematic dynamo).
Runs with resettings are compared for different values of the parameter $\tau$. It is seen in Fig. \[fig\_Magnetic\_Enstr\_ABC\] that the magnetic enstrophy $\Omega_m$ is well resolved for each case, at resolution $128^3$.
The resettings are quite regular in time and indeed they slow down as time goes by, at a regular rate which decreases with increasing resolution (figure not shown). There is no increase in the resetting frequency. This behavior is consistent with the monotonic behavior of the magnetic enstrophy and with the absence of localized or intense activity of the magnetic field.
![ \[fig\_Magnetic\_Enstr\_ABC\]](CurrentABC.eps){width="8.0cm"}
### Full MHD equations: 2D Orszag-Tang Vortex {#sec:2DOT}
In the rest of the paper, the full MHD equations of motion are integrated. The momentum equation for the velocity (\[eq:momentum\]) is integrated together with the Weber Clebsch evolution equations (\[eq: New\_Alg L\])–(\[eq: explicit G\]) where the magnetic potential $\mathbf{Z} = \mathbf{A}$ is obtained from (\[eq:WC\_Z\]) and (\[eq:lap\_phi\]).
We have chosen the following initial data for the $2$D Orszag-Tang (hereafter, OT) vortex [@OT2D]: $$\begin{aligned}
u_x&=& -2 \sin y \nonumber\\
u_y&=& 2 \sin x \nonumber\\
u_z &=& 0, \nonumber\\
A_x&=& 0 \nonumber\\
A_y&=& 0 \nonumber\\
A_z &=& 2\cos x \cos 2y \,. \nonumber\end{aligned}$$ The OT vortex has a magnetic hyperbolic X-point located at a stagnation point of the velocity, and is a standard test of magnetic reconnection, both in two dimensions [@add1] and in three dimensions [@add2], see below section \[sec:3DOT\].
We compare runs with resettings for different values of the parameter $\tau$. Figure \[Current2d\] shows that the magnetic enstrophy is well resolved in resolution $128^2$.
Figure \[DeltaResOT2d\] shows the time between resettings as a function of time, for runs performed with different values of $\tau$. It is apparent from the figure that there are periods of frequent resettings which coincide with the periods of high magnetic enstrophy from Fig. \[Current2d\]. This is a robust evidence of the utility of the resetting approach for $2$D magnetic reconnection.
We have also simulated the Orszag-Tang vortex in the so-called $2.5$D setting [@Montgomery82] (see also the DiPerna-Majda’s construction [@DiPerna87]), defined by the same initial data as the above $2$D Orszag-Tang vortex, but with $A_x= \sin y $ and $A_y=- \sin x$. We obtained (data not shown) a behavior of the resetting frequency which was very similar to that of the $2$D case.
![ \[Current2d\]](Current2dm.eps){width="8.0cm"}
![ \[DeltaResOT2d\]](dReset2d.eps){width="8.0cm"}
### Full MHD equations: 3D Orszag-Tang Vortex {#sec:3DOT}
For the $3$D Orszag-Tang vortex [@OT3D] the initial magnetic potential reads $$\begin{aligned}
A_x&=& c_0 \left(\cos y - \cos z \right) \nonumber\\
A_y&=& c_0 \left(-\cos x + \cos z \right) \nonumber\\
A_z &=& c_0 \left(\cos x + \cos 2y \right)\,, \nonumber\end{aligned}$$ with $c_0=0.8$. The initial velocity is given by $$\begin{aligned}
u_x&=& -\sin y \nonumber\\
u_y&=& \sin x \nonumber\\
u_z &=& 0 \,.\nonumber\end{aligned}$$ As in the $2$D case, we compare runs with resettings for different values of the parameter $\tau$. Figure \[OmegaM Orszag-Tang3d\] shows that the magnetic enstrophy is well resolved in resolution $128^3$, and Fig. \[DeltaResOT3d\] shows the time between resettings as a function of time. Again the periods of frequent resettings coincide with the periods of high magnetic enstrophy from Fig. \[OmegaM Orszag-Tang3d\], proving the utility of the resetting approach for $3$D magnetic reconnection.
![ \[OmegaM Orszag-Tang3d\]](OmegMOt3dm.eps){width="8.0cm"}
![ \[DeltaResOT3d\]](dreset3d.eps){width="8.0cm"}
Conclusions
===========
We have shown that the generalized Weber-Clebsch evolution equations allow to study reconnection events for both Navier-Stokes and MHD dynamics. We have checked for the Navier-Stokes BPZ flow that reconnection events can be viewed as periods of fast and localized changes in the geometry of the Weber-Clebsch potentials, leading to more and more frequent resetting of the potentials.
We have applied the new generalized Weber-Clebsch evolution equations to the study of magnetic reconnection in MHD. Taking as examples both the $2$D and $3$D Orszag-Tang vortices, we show a correlation of the reconnection events (associated to periods of high magnetic dissipation) with the periods of fast changes in the geometry of the Weber-Clebsch potentials, leading to frequent resettings of the potentials.
However, unlike the case of BPZ reconnection, in this case the frequency of resettings does not have a sharp peak but a smeared one. Notice that, in the Navier-Stokes case, the corresponding frequency of resettings for the Taylor-Green vortex has also a mild peak. [@CBB07] One can argue that the 2D and 3D Orszag-Tang flows are more similar to Taylor-Green than to BPZ. Indeed, both Orszag-Tang and Taylor-Green have initial conditions with just a few Fourier modes, therefore they are extended spatially, whereas the BPZ initial condition is spatially localized (two orthogonal vortex tubes).
This wide spatial extent of the vorticity in both Orszag-Tang and Taylor-Green vortices, as opposed to the localized extent of BPZ, might be the reason for the mildness in the shape of the minimum of the time between resettings. In both spatially extended cases one expects reconnection events to happen in relatively distant places at similar times, as opposed to the BPZ very localized cut-and-connect type of reconnection. In terms of the singularities of the Weber-Clebsch potentials and associated resetting, we should observe (to be studied in detail in future work) that the set of points where $\det \mathbb{H}$ goes below the threshold consists of an extended region, as opposed to BPZ where we have confirmed that these points belong to a very localized region in space. Consequently, the widely distributed events that lead to resetting in Orszag-Tang and Taylor-Green configurations would tend to be less correlated in time, leading to the smearing of the minimum of the curve for the time between resettings, which would otherwise be very sharp if the events were more localized and therefore more correlated in time.
[**Acknowledgments:**]{} We acknowledge very useful scientific discussions with Peter Constantin and Edriss S. Titi. One of the authors (MEB.) acknowledges support from an ECOS/CONICYT action. The computations were carried out at the Institut du Développement et des Ressources en Informatique Scientifique (IDRIS) of the Centre National pour la Recherche Scientifique (CNRS).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we develop an adaptive procedure for the numerical solution of semilinear parabolic problems, with possible singular perturbations. Our approach combines a linearization technique using Newton’s method with an adaptive discretization—which is based on a spatial finite element method and the backward Euler time stepping scheme—of the resulting sequence of linear problems. Upon deriving a robust *a posteriori* error analysis, we design a fully adaptive Newton-Galerkin time stepping algorithm. Numerical experiments underline the robustness and reliability of the proposed approach for various examples.'
address:
- 'Mathematics Institute, University of Bern, CH-3012 Switzerland'
- 'Mathematics Institute, University of Bern, Sidlerstrasse 5, CH-3012 Switzerland'
author:
- Mario Amrein
- 'Thomas P. Wihler'
bibliography:
- 'references.bib'
title: 'Fully Adaptive Newton-Galerkin Time Stepping Methods for Singularly Perturbed Parabolic Evolution Equations'
---
Introduction
============
Semilinear evolution problems appear in a wide range of applications including, for instance, ecology, (bio-)chemistry, quantum- and astro-physics, material science, or optics; see [@BarlesBurdeau:95; @BarlesPerthame:07; @BorisyukErmentroutFriedmanTerman:05; @CantrellCosner:03; @Chandrasekhar:39; @Edelstein-Keshet:05; @JagerLuckhaus:92; @Kelley:65; @OkuboLevin:01; @Perthame:07]. In this contribution we consider the numerical approximation of semilinear parabolic equations with a possibly degenerate diffusion coefficient $0<\varepsilon\ll 1$. Specifically, for a continuously differentiable nonlinearity $ f:\mathbb{R}\times \Omega \times (0,T] \rightarrow \mathbb{R}$, and an initial function $g \in L^{\infty}(\Omega)$, we study initial/boundary value problems of the form $$\label{heat}
\left\{ \begin{aligned}
\partial_{t}u(\bm x,t) - \varepsilon \Delta u(\bm x,t) &= f(u(\bm x,t),\bm x,t), & &(\bm x,t)\in \Omega\times(0,T],\\
u(\bm x,t)&=0, &&(\bm x,t)\in \partial\Omega\times(0,T],\\
u(\bm x,0)&=g(\bm x), & &\bm x\in\Omega.
\end{aligned} \right.$$ Here $ \Omega \subset \mathbb{R}^{d} $, with $d\in \{1,2\} $, is an open and bounded one dimensional ($1$d) interval or a two-dimensional ($2$d) Lipschitz polygon, respectively. Furthermore, $T\in (0,\infty)$ denotes the final time of the evolutionary process. In the singularly perturbed case, solutions of are known to exhibit boundary layers, interior shocks, or (multiple) spikes. The appearance of singular effects of this type constitutes a challenging issue when solving such problems numerically; see, e.g. [@RoStTo08; @Verhulst].
With the aim of designing an adaptive numerical procedure for , we follow our recent work on stationary (elliptic) PDE in [@AmreinWihler:14; @AmreinWihler:15] (see also [@Deuflhard:04; @El-AlaouiErnVohralik:11]). In particular, this includes the application of a Newton linearization framework to the nonlinear problem at hand, and, subsequently, the discretization of the resulting *sequence of linear evolution problems* by appropriate numerical schemes. It is worth mentioning that this methodology enables the use of numerical analysis techniques that originate from the treatment of *linear* problems; this is opposed to studying nonlinear discretizations schemes (see, e.g., [@georg; @KyzaMakridakis:11]). The challenge in deriving practically effective *a posteriori* error bounds within this setting is to provide a suitable splitting of the total residual into several computable quantities, each of which accounts for one of the different errors that have been committed during the discretization process: (1) a linearization residual, (2) a time discretization residual, and (3) a space discretization residual. Then, based on the resulting *a posteriori* error estimates, a fully adaptive Newton-Galerkin time stepping algorithm for the numerical solution of can be derived. Specifically, in order to obtain an efficient overall complexity of the scheme, we propose an interplay between the Newton linearization, time adaptivity, and spatial adaptivity. To do so, the algorithm will take into account the different residuals (1)–(3), and will perform a Newton step, a refinement of the current time step, or a refinement of the spatial mesh according to whatever residual is currently dominant.
In the context of this paper, a $\mathbb{P}_1$-finite element approach in space, and a backward Euler discretization in time will be applied. Our *a posteriori* error analysis proceeds along the lines of the theory presented in [@Verfuerthbook] on linear parabolic equations. Furthermore, in order to obtain $\varepsilon$-robust bounds, we will follow the papers [@Verfuerth] and [@AmreinWihler:15] on finite element discretizations for singularly perturbed linear and semilinear elliptic problems, respectively. By means of a series of numerical experiments we will demonstrate that the interactive application of temporal and spatial mesh refinements, together with a continued monitoring of the linearization effect, leads to an $\varepsilon$-robust decay of the residual even in the singularly perturbed regime.
Outline {#outline .unnumbered}
-------
The paper is organized as follows: In Section \[sc:linearizatio/discretization\] we begin by deriving the Newton linearization of , and formulate the discretization of the resulting sequence of linearized problems in the spatial and temporal variables by means of a finite element method and the backward Euler scheme, respectively. Furthermore, the goal of Section \[sec:aposteriori\] is to derive an $\varepsilon$-robust [*a posteriori*]{} error analysis. Moreover, in Section \[sec:numerics\], we develop a fully adaptive Newton-Galerkin time stepping algorithm. Furthermore, we present a series of numerical experiments illustrating the performance of the proposed adaptive procedure. Finally, we summarize our findings in Section \[sc:conclusions\].
Notation and Problem Formulation {#notation-and-problem-formulation .unnumbered}
--------------------------------
For the purpose of this paper, we define the space $Z:=H_{0}^{1}(\Omega)$, where $H^1_0(\Omega)$ is the standard Sobolev space of functions in $H^1(\Omega)~=W^{1,2}(\Omega)$, with zero trace on $\partial\Omega$. The space $Z$ is equipped with the singular perturbation norm $\|\cdot\|_Z:={\left|\!\left|\!\left|\cdot\right|\!\right|\!\right|}_{\varepsilon,\Omega}$, where, for any subset $D\subseteq\Omega$, we define $${\left|\!\left|\!\left|u\right|\!\right|\!\right|}_{\varepsilon,D}:=\Bigl(\varepsilon{\left\|\nabla u\right\|}_{0,D}^2 +{\left\|u\right\|}_{0,D}^2 \Bigr)^{\nicefrac{1}{2}},\qquad u\in H^1(D).$$ Here, $\|\cdot\|_{0,D}$ denotes the $L^2$-norm on $D$. Frequently, for $D=\Omega$, the subindex ‘$D$’ will be omitted. In the sequel, we will abbreviate $f(u,\bm x,t)$ by $f(u)$; note that, in the case of $f(u)=-u$, when is linear, the norm ${\left|\!\left|\!\left|\cdot\right|\!\right|\!\right|}_{\varepsilon,\Omega}$ is the natural energy norm on $Z$.
Moreover, we signify by $Z'=H^{-1}(\Omega)$ the dual space of $Z$; it is equipped with the norm $$\|\psi\|_{Z'}=\sup_{\genfrac{}{}{0pt}{}{z\in Z}{\|z\|_Z=1}}{{\left<\psi,z\right>}},$$ where $ {\left<\cdot,\cdot\right>}$ is the dual product in $Z'\times Z$. Furthermore, consider the Bochner spaces $Y:=L^{2}(0,T;Z')$, and $X:=\{u \in L^{2}(0,T;Z):\partial_t u \in Y\}$, with $\partial_{t}$ being the time derivative operator in the distributional sense. On $Y$ and $X$ we introduce the norms $${\left\|u\right\|}_{Y}:=\left(\int_{0}^{T}{{\left\|u(\cdot,t)\right\|}_{Z'}^{2}{\,{\mathsf{d}}t}}\right)^{\nicefrac{1}{2}},\qquad u\in Y,$$ and $${\left\|u\right\|}_{X}:=\left(\int_{0}^{T}\left\{\|u\|_Z^2+\|\partial_tu\|_{Z'}^2\right\}{\,{\mathsf{d}}t}\right)^{\nicefrac{1}{2}},\qquad u\in X,$$ respectively.
Defining the map ${\mathsf{F}}_{\varepsilon}: X\rightarrow Y$ through $$\label{eq:goal}
{\left<{\mathsf{F}}_{\varepsilon}(u),v\right>} := {\left<\partial_{t}u,v\right>}+\int_{\Omega}\{\varepsilon \nabla u \cdot \nabla v-f(u)v\}{\,{\mathsf{d}}\bm x}\qquad \forall v \in Z,$$ the above problem can be written as a nonlinear operator equation in $Y$: $$u \in X: \qquad {\mathsf{F}}_{\varepsilon}(u)=0 \qquad \forall t \in (0,T),$$ with $ u(\cdot,0)=g $.
Throughout this work we shall use the abbreviation $x \preccurlyeq y$ to mean $x \leq c y$, for a constant $c>0$ independent of the mesh size $h$ and of $\varepsilon>0 $.
Linearization and Discretization {#sc:linearizatio/discretization}
================================
Linearization
-------------
We note that the Fréchet derivative of $ {\mathsf{F}}_{\varepsilon} $ from at $u\in X$ is given by $$\left \langle {\mathsf{F}}_{\varepsilon}'(u)w,v\right \rangle = {\left<\partial_{t}w,v\right>}+\int_{\Omega}{\{\varepsilon \nabla w \cdot \nabla v-\partial_{u}f(u)wv\}{\,{\mathsf{d}}\bm x}},\qquad w\in X,v\in Z.$$ Then, starting from an initial guess $ u_0 \in X $, Newton’s method for is an iterative procedure by which we find $u_{N+1} \in X $ from $u_N \in X$, for $N=0,1,2\ldots$, such that there holds $$\label{eq:linearization2}
{\mathsf{F}}_{\varepsilon}'(u_{N})(u_{N+1}-u_{N})=-{\mathsf{F}}_{\varepsilon}(u_{N})$$ in $ Y$. Upon defining the increment $\delta_N:=u_{N+1}-u_{N}\in X$, and recalling , we note that $$\label{eq:equiv}
\begin{split}
{\left<\partial_{t}(u_{N}+\delta_{N}),v\right>}&+\int_{\Omega}\left\{\varepsilon \nabla(u_{N}+\delta_{N})\cdot \nabla{v}\right\}{\,{\mathsf{d}}\bm x}\\
&=\int_{\Omega}{\{f(u_{N})+\partial_{u}f(u_{N})\delta_{N}\}v{\,{\mathsf{d}}\bm x}},
\end{split}$$ for all $v\in Z$.
Finite Element Meshes and Spaces
--------------------------------
Let $ \mathcal{T}_{h}=\{K\}_{K\in\mathcal{T}_h}$ be a regular and shape-regular mesh partition of $\Omega $ into disjoint open simplices, i.e., any $K\in\mathcal{T}_h$ is an affine and non-degenerate image of the (open) reference simplex $\widehat K=\{\widehat x\in\mathbb{R}_+^d:\,\sum_{i=1}^d\widehat x_i<1\}$. By $h_K=\mathrm{diam}(K)$ we signify the element diameter of $K\in\mathcal{T}_h$, and by $h=\max_{K\in\mathcal{T}_h}h_K$ the mesh size of $\mathcal{T}_h$. Furthermore, by $\mathcal{E}_{h}$ we denote the set of all interior mesh nodes for $d=1$ and interior (open) edges for $d=2$ in $\mathcal{T}_{h}$. In addition, for $K\in\mathcal{T}_h$, we let $\mathcal{E}_h(K)=\{E\in\mathcal{E}_h:\,E\subset\partial K\}$. For $E\in\mathcal{E}_h$, we let $h_E$ be the mean of the lengths of the adjacent elements in 1d, and the length of $E$ in 2d.
We consider the finite element space of continuous, piecewise linear functions on $\mathcal{T}_{h}$ with zero trace on $\partial\Omega$, given by $$V_{0}^{h}:=\{\varphi\in H^1_0(\Omega):\,\varphi|_{K} \in \mathbb{P}_{1}(K) \, \forall K \in \mathcal{T}_{h}\},$$ where $\mathbb{P}_1(K)$ is the standard space of all linear polynomial functions on $K$. Moreover, for any function $ \varphi \in V^h_0 $ and a given edge $ E \in \mathcal{E}_{h}$ with $E=\mathcal{E}_h(K^\sharp)\cap\mathcal{E}_h(K^\flat) $ shared by two neighboring simplices $K^\sharp, K^\flat\in\mathcal{T}_h$, we denote by $\llbracket \varphi \rrbracket_{E} $ the (vector-valued) jump of $ \varphi $ across $E$: $$\llbracket \varphi \rrbracket_{E}(x)=\lim_{t\to 0^+}\varphi(\bm x+t\bm n^\sharp)\bm n^\sharp+\lim_{t\to 0^+}\varphi(\bm x+t\bm n^\flat)\bm n^\flat \qquad \forall \bm x \in E.$$ Here, $\bm n^\sharp$ and $\bm n^\flat$ denote the unit outward vectors on $\partial K^\sharp$ and $\partial K^\flat$, respectively.
Furthermore, for any $K\in\mathcal{T}_h$, we consider the element patch $$ \widetilde{w}_{K}:=\bigcup_{\genfrac{}{}{0pt}{}{K'\in\mathcal{T}_h:}{\overline K \cap \overline K' \neq \emptyset}}{K'}.
$$ Let us also define the following two quantities: $$\label{boundary}
\begin{split}
\alpha_K&:=\min(1,\varepsilon^{-\nicefrac12}h_K),\qquad K\in\mathcal{T}_h,\\
\alpha_E&:=\min(1,\varepsilon^{-\nicefrac12}h_E),\qquad E\in\mathcal{E}_h.
\end{split}$$ Then, we recall the following approximation result from [@AmreinWihler:15 Corollary 4.2]:
Let ${\mathsf{I}_h^{n}}:\,H_{0}^{1}(\Omega)\rightarrow V_{0}^{h} $ be the quasi-interpolation Clément operator (see, e.g., [@Verfuerthbook]). Then, for any element $K\in\mathcal{T}_h$, and any edge $E\in\mathcal{E}_h$, with $E=\mathcal{E}_h(K^\sharp)\cap\mathcal{E}_h(K^\flat)$ for some neighboring elements $K^\sharp, K^\flat\in\mathcal{T}_h$, and any $v\in H^1_0(\Omega)$, there hold the approximation bounds $$\begin{aligned}
{\left\|v-{\mathsf{I}_h^{n}}v\right\|}_{0,K}&\preccurlyeq \alpha_{K}{\left|\!\left|\!\left|v\right|\!\right|\!\right|}_{\varepsilon,\widetilde{w}_K},\\
{\left\|v-{\mathsf{I}_h^{n}}v\right\|}_{0,E}&\preccurlyeq \frac12\varepsilon^{-\nicefrac14}\alpha_E^{\nicefrac12}\left({\left|\!\left|\!\left|v\right|\!\right|\!\right|}_{\varepsilon,\widetilde{w}_{K^\sharp}}+{\left|\!\left|\!\left|v\right|\!\right|\!\right|}_{\varepsilon,\widetilde{w}_{K^\flat}}\right),\end{aligned}$$ where $\alpha_K$ and $\alpha_E$ are defined in .
Newton-Galerkin Backward Euler Discretization
---------------------------------------------
In order to provide a numerical approximation of , we will discretize the spatial and temporal variables in the formulation by means of a finite element method in space and the backward Euler scheme in time, respectively. In combination with the Newton iteration this results in a Newton-Galerkin time stepping approximation scheme.
We consider a time partition of the interval $(0,T)$ into $M\ge 1$ subintervals $ I_{n}=(t_{n-1},t_{n})$, $n=1,\ldots,M$, satisfying $$0=t_{0}<t_{1}<\ldots<t_{M-1}<t_{M}=T,$$ and define the time step lengths $k_{n}:=t_{n}-t_{n-1}$. We mark any quantities related to the finite element discretization on a given time interval $I_n$ by an index ‘$n$’; in particular, we denote by $ \mathcal{T}_{h}^{n} $ the corresponding spatial partition of $\Omega$, and by $$V_{0}^{h,n}:=\{\varphi\in H^1_0(\Omega):\,\varphi|_{K} \in \mathbb{P}_{1}(K) \, \forall K \in \mathcal{T}_{h}^{n}\}$$ the associated finite element space on a time subinterval $I_n$. Furthermore, by $ \Pi^{n} $ we signify the $L^{2}$-projection onto $ V_{0}^{h,n} $.
Applying the backward Euler time stepping scheme, the finite element discretization of the Newton iteration on each time interval $I_{n}$, $n=1,2,\ldots,M$, is to find $ \delta_{N}^{n} \in V_{0}^{h,n} $ from $ u_{N}^{n} \in V_{0}^{h,n} $ such that $$\label{disc}
\begin{aligned}
\int_{\Omega}\bigg\{v\frac{u_{N}^{n}+\delta_{N}^{n}-u^{n-1}}{k_{n}}&+\varepsilon \nabla(u_{N}^{n}+\delta_{N}^{n})\cdot \nabla v\bigg\}{\,{\mathsf{d}}\bm x}\\
=& \int_{\Omega}\left\{f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}\right\}v{\,{\mathsf{d}}\bm x}\qquad \forall v \in V_{0}^{h,n},
\end{aligned}$$ with the update $ u_{N+1}^{n}=u_{N}^{n}+\delta_{N}^{n}$, and with $$f^{n}(\cdot):=f(u,\bm x,t)|_{(\cdot,\bm x,t_{n})}, \qquad
\partial_{u}f^{n}(\cdot):=\partial_{u}f(u,\bm x,t)|_{(\cdot,\bm x,t_n)}.$$ Moreover, for $n=2,\ldots,M$, we denote by $u^{n-1}\in V_{0}^{h,n-1} $ the (space-dependent) discrete solution at the previous time node $t_{n-1}$ (resulting from a sufficient number of Newton iterations), and for $n=1$, we let $u^{0}:=\Pi^0g$ be a suitable approximation in $H^1_0(\Omega)$ of the initial condition $g\in L^\infty(\Omega)$. Furthermore, for the $n$-th time step, the initial guess $u_{0}^{n}\in V_{0}^{h,n}$ is defined by $$\label{eq:u0}
u_{0}^{n}:=\Pi^{n}u^{n-1},\quad 1\leq n \leq M.$$ We will denote the procedure of performing one Newton update, i.e., solving to obtain $ u_{N+1}^{n} $, by $$u_{N+1}^{n}={\mathtt{solve}}(k_{n},\mathcal{T}_{h}^{n},u_{N}^{n}).$$ Here, we make the assumption that we reinitiate the Newton iteration on each time step, i.e., for each $n=1,2,\ldots,M$, we start with $N=0$.
\[im\] The adaptive procedure based on the *a posteriori* analysis to be presented in the next sections enables the use of possible coarsening of some (spatial) elements with small error contributions. A coarsening strategy within the procedure of solving is not a trivial task. In fact, suppose we have solved up to time $t_{n-1} $ so that we are given $u^{n-1}\in V_{0}^{h,n-1}$. Then, for a time step $k_{n}>0 $ small enough, it is reasonable to assume that $u^{n-1} $ is located in an attracting $\epsilon$-ball $ B_{\epsilon}(u_{\infty}^{n})\subset \mathcal{A}(u_{\infty}^{n}) $ of $ u_{\infty}^{n}\in V_{0}^{h,n} $, where $\mathcal{A}(u_{\infty}^{n}) $ denotes the attractor of the Newton iteration corresponding to $u_{\infty}^{n}$. Thence, we have $$\label{attra}
{\left\|u_{0}^{n}-u_{\infty}^{n}\right\|}
\leq {\left\|\Pi^{n}u^{n-1}-u^{n-1}\right\|} + {\left\|u^{n-1}-u_{\infty}^{n}\right\|} \leq {\left\|\Pi^{n}u^{n-1}-u^{n-1}\right\|} + \epsilon,$$ for some suitable norm ${\left\|\cdot\right\|} $. Hence, if $\mathcal{T}_{h}^{n}$ is obtained from $\mathcal{T}_{h}^{n-1}$ by refinement only, then we see from that $ u_{0}^{n}\in \mathcal{A}(u_{\infty}^{n}) $ since $\Pi^{n}u^{n-1}=u^{n-1}$. If, however, there is a partial coarsening involved, we usually have $\Pi^{n}u^{n-1}\neq u^{n-1} $, and in consequence, the quantity ${\left\|\Pi^{n}u^{n-1}-u^{n-1}\right\|}$ in may be too large in order to guarantee for $u_{0}^{n} $ to stay within $ \mathcal{A}(u_{\infty}^{n}) $. Thus any coarsening strategy should only remove those degrees of freedom for which $ {\left\|\Pi^{n}u^{n-1}-u^{n-1}\right\|} $ remains of moderate size (cf. also [@georg]).
*A Posteriori* Error Analysis {#sec:aposteriori}
=============================
The goal of this section is to derive a residual based *a posteriori* error bound for the discretization scheme , which can be employed for the purpose of formulating an adaptive refinement procedure for the meshes and time steps in each Newton step. This leads to a fully adaptive Newton-Galerkin backward Euler discretization method for . In the subsequent *a posteriori* error analysis we follow closely the approach presented in [@Verfuerthbook].
Residuals
---------
The discrete problem generates a sequence $\{u_{N}^{n}\}_{N\ge 0}$ for each time step $n=1,\ldots,M$. We can thus define a function $u_{\mathcal{I}}\in C^0([0,T];H_{0}^{1}(\Omega))$ by $$u_{\mathcal{I}}|_{I_{n}}:=\frac{t_n-t}{k_n}u^{n-1}+\frac{t-t_{n-1}}{k_{n}}u_{N+1}^{n}, \qquad t \in [t_{n-1},t_{n}].$$ We remark that $u_{\mathcal{I}}$ is understood as a function in time that depends on the (varying) Newton iteration index $N$ on each subinterval $I_n$. For later purposes notice that $$\label{brevity}
q_{n}(t)\left(u^{n-1}-u_{N+1}^{n}\right)=u_{\mathcal{I}}-u_{N+1}^{n},$$ where $q_{n}(t):=k_n^{-1}(t_n-t)$, and, moreover, we observe that $$\partial_{t} u_{\mathcal{I}}|_{I_{n}}=\frac{u_{N+1}^{n}-u^{n-1}}{k_{n}},$$ for $n=1,\ldots,M$. Therefore, motivated by the linear case discussed in [@Verfuerthbook], we decompose the residual $ {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}|_{I_n}) $ from on each time interval $I_{n}$, $n=1,\ldots,M$, as $$\label{decomposition}
\left \langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}|_{I_n}),v \right \rangle= \left \langle {\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n}),v \right \rangle+
\left \langle {\mathsf{F}}_{\varepsilon}^{2}(u_\mathcal{I}|_{I_n}),v \right \rangle+\left \langle {\mathsf{F}}_{\varepsilon}^{3}(u_\mathcal{I}|_{I_n}),v \right \rangle,\qquad v\in Z.$$ Here, the parts ${\mathsf{F}}_{\varepsilon}^{i}(u_{\mathcal{I}}|_{I_n})$, $i =1,2,3$, are defined by $$\begin{aligned}
\left \langle {\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n}),v \right \rangle&:=\int_{\Omega}{\{v\partial_{t} u_{\mathcal{I}}|_{I_n}+\varepsilon \nabla u_{N+1}^{n}\cdot\nabla v-(f^n(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n})v\}{\,{\mathsf{d}}\bm x}},\\
\left \langle {\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n}),v \right \rangle&:=\int_{\Omega}\varepsilon \nabla(u_{\mathcal{I}}|_{I_n}-u_{N+1}^{n})\cdot\nabla v{\,{\mathsf{d}}\bm x}+ \int_{\Omega}{\{f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\}v{\,{\mathsf{d}}\bm x}},\\
\left \langle {\mathsf{F}}_{\varepsilon}^{3}(u_{\mathcal{I}}|_{I_n}),v \right \rangle&:=\int_{\Omega}{\{f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}-f^{n}(u_{N+1}^{n})\}v{\,{\mathsf{d}}\bm x}},\end{aligned}$$ for any $v \in Z$. We emphasize that this splitting is crucial when aiming at an adaptive algorithm that is able to identify the individual error contributions resulting from the time and space discretizations as well as from the Newton linearization. In accordance with the notation introduced in [@Verfuerthbook], we call ${\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})$ and ${\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n}) $ the *spatial* and the *temporal residuals*, respectively. Furthermore, ${\mathsf{F}}_{\varepsilon}^{3}(u_{\mathcal{I}}|_{I_n})$ is termed the *nonlinear residual*.
*A Posteriori* Error bound
--------------------------
Upon applying the triangle inequality to the decomposition we obtain that $$\label{eq:triangle}
\begin{split}
&{\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')}\\
&\quad\leq {\left\|{\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')}
+{\left\|{\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')}+{\left\|{\mathsf{F}}_{\varepsilon}^{3}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')},
\end{split}$$ on each time interval $I_{n}$, $n=1,\ldots,M$. We will now derive individual error bounds for each of the three residual terms ${\mathsf{F}}_{\varepsilon}^{i}(u_{\mathcal{I}}|_{I_n})$, $i=1,2,3$.
### Spatial Residual {#spatial-residual .unnumbered}
We note the fact that the spatial residual ${\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})$ is constant with respect to time. It can thus be estimated as in the stationary case [@AmreinWihler:15 Theorem 4.4]. In fact, observing that $${\left\|{\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_n;Z')}=\sqrt{k_n}\|{\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})\|_{Z'},$$ we infer the estimate $$\label{esf}
{\left\|{\mathsf{F}}_{\varepsilon}^{1}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_n;Z')}^2 \preccurlyeq k_{n}\sum_{K\in \mathcal{T}_{h}^{n}}{\eta_{n,K,N}^{2}},$$ for any time interval $I_n$, $n=1,\ldots,M$. Here, for any $K\in \mathcal{T}_{h}^{n} $, the quantities $$\label{eq:eta}
\begin{aligned}
\eta_{n,K,N}^{2}:&=\alpha_{K}^{2}{\left\|f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}+\varepsilon \Delta u_{N+1}^{n}-\partial_{t}u_{\mathcal{I}}|_{I_n}\right\|}_{0,K}^2\\
&\quad+\frac{1}{2}\sum_{E\in \mathcal{E}_{h}^{n}(K)}{\varepsilon^{-\nicefrac12}\alpha_{E}{\left\|\varepsilon \llbracket \nabla u_{N+1}^{n}\rrbracket\right\|}_{0,E}^{2}}
\end{aligned}$$ are computable error indicators, with $\alpha_K$ and $\alpha_E$ being defined in . We emphasize that the bound is robust with respect to the singular perturbation parameter $\varepsilon$.
### Temporal Residual {#temporal-residual .unnumbered}
Using the identity we have that $$\begin{aligned}
\left \langle {\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n}),v \right \rangle &= q_n(t)\int_{\Omega}{\{\varepsilon \nabla (u^{n-1}-u_{N+1}^{n})\cdot \nabla v\}{\,{\mathsf{d}}\bm x}}\\
&\quad+\int_{\Omega}{\{f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\}v{\,{\mathsf{d}}\bm x}},
\end{aligned}$$ from which, by application of the Cauchy-Schwarz inequality, we obtain $$\|{\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n})\|_{Z'}^{2} \leq q_{n}(t)^{2}\varepsilon\|\nabla(u^{n-1}-u_{N+1}^{n})\|_0^{2}+{\left\|f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\right\|}_{0}^{2}.$$ Moreover, since $\int_{t_{n-1}}^{t_n}{q_{n}(t)^2dt} = \nicefrac{k_n}{3}$, we arrive at the bound $$\label{eq:time}
\begin{split}
{\left\|{\mathsf{F}}_{\varepsilon}^{2}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')}^{2} &\leq \frac{\varepsilon k_{n}}{3}\|\nabla(u^{n-1}-u_{N+1}^{n})\|_{0}^{2}\\
&\quad+{\left\|f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};L^{2}(\Omega))}^{2},
\end{split}$$ on each time interval $I_{n}$, $n=1,\ldots,M$.
Applying the trapezoidal rule we see that we can approximate $${\left\|f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};L^{2}(\Omega))}^{2}
\approx k_{n}{\left\|f^{n}(u_{N+1}^{n})-f^{n-1}(u^{n-1})\right\|}_{0}^{2},$$ up to an error of order $\mathcal{O}(k_{n}^2)$, for each time interval $I_{n}$, $n=1,\ldots,M$.
### Nonlinear Residual {#nonlinear-residual .unnumbered}
We immediately infer $$\|{\mathsf{F}}_{\varepsilon}^{3}(u_{\mathcal{I}}|_{I_n})\|_{Z'} \leq {\left\|f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}-f^{n}(u_{N+1}^{n})\right\|}_0,$$ and hence, $$\label{nonlin}
{\left\|{\mathsf{F}}_{\varepsilon}^{3}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};Z')}^2 \leq k_n{\left\|f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}-f^{n}(u_{N+1}^{n})\right\|}_{0}^2,$$ on each $I_{n}$, $n=1,\ldots,M$.
Combining the bounds , , , and leads to the following result.
\[lam:1\] On each time interval $I_{n}$, $n=1,2,\ldots,M$, there holds the *a posteriori* error bound $${\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_n;Z')}^2
\preccurlyeq
k_{n}\sum_{K\in \mathcal{T}_{h}^{{n}}}\left\{
\eta_{n,K,N}^{2}+
\vartheta^{2}_{n,K,N} + \Upsilon^{2}_{n,K,N}\right\},$$ where, for $K\in\mathcal{T}^n_h$, we recall the spatial error indicators $\eta_{n,K,N}$ in , and let $$\label{eq:definition}
\begin{aligned}
\vartheta_{n,K,N}^{2}&:= k_n^{-1}{\left\|f^{n}(u_{N+1}^{n})-f(u_{\mathcal{I}}|_{I_n})\right\|}_{L^{2}(I_{n};L^{2}(K))}^{2}+\frac{1}{3}\varepsilon\|\nabla(u^{n-1}-u_{N+1}^{n})\|_{0,K}^{2},\\
\Upsilon_{n,K,N}^{2}&:={\left\|f^{n}(u_{N}^{n})+\partial_{u}f^{n}(u_{N}^{n})\delta_{N}^{n}-f^{n}(u_{N+1}^{n})\right\|}_{0,K}^2.
\end{aligned}$$
For later reference, in addition to the previously introduced local *a posteriori quantities*, we define the corresponding global error indicators as follows: $$\label{eq:global}
\begin{split}
\eta_{n,\Omega,N}^2&:=\sum_{K\in\mathcal{T}^n_h}\eta^2_{n,K,N},\qquad
\vartheta_{n,\Omega,N}^2:=\sum_{K\in\mathcal{T}^n_h}\vartheta^2_{n,K,N},\\
\Upsilon_{n,\Omega,N}^2&:=\sum_{K\in\mathcal{T}^n_h}\Upsilon^2_{n,K,N}.
\end{split}$$
Residual and Error Norm {#sc:RN}
-----------------------
Under certain conditions on the nonlinearity $f$ in it can be shown that the residual ${\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}) $ defined in and the error $u- u_{\mathcal{I}}$ are equivalent. For example, suppose that the nonlinearity $f$ is Lipschitz continuous with Lipschitz constant $L>0$, i.e., $$|f(u,\bm x,t)-f(v,\bm x,t)|\le L|u-v|,$$ and that it satisfies the monotonicity condition $$(f(u,\bm x,t)-f(v,\bm x,t))(u-v)\leq 0,$$ for all $u,v\in\mathbb{R}, \bm x\in\Omega, t\in[0,T]$. Then, if $u$ is the exact solution of , we have that $$\begin{aligned}
-\langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}),u-u_{\mathcal{I}} \rangle
&= \frac{1}{2}\frac{{\mathsf{d}}}{{\,{\mathsf{d}}t}}{\left\|u-u_{\mathcal{I}}\right\|}_{0}^{2}+\varepsilon {\left\|\nabla(u-u_{\mathcal{I}})\right\|}_{0}^{2}\\
&\quad-\int_{\Omega}{(f(u)-f(u_{\mathcal{I}}))(u-u_{\mathcal{I}}){\,{\mathsf{d}}\bm x}} \\
&\geq \frac{1}{2}\frac{{\mathsf{d}}}{{\,{\mathsf{d}}t}}{\left\|u-u_{\mathcal{I}}\right\|}_{0}^{2}+\varepsilon {\left\|\nabla(u-u_{\mathcal{I}})\right\|}_{0}^{2},
\end{aligned}$$ for any $t\in(0,T)$. Proceeding as in [@AmreinWihler:15 Eq. (4.13)] it holds that $$-\left \langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}),u-u_{\mathcal{I}} \right \rangle
\geq \frac{1}{2}\frac{{\mathsf{d}}}{{\,{\mathsf{d}}t}}{\left\|u-u_{\mathcal{I}}\right\|}_{0}^{2}+C_\varepsilon{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}^{2},$$ for a constant $C_\varepsilon>0$. Furthermore, choosing $\delta >0 $ such that $C_{\delta,\varepsilon}:=C_\varepsilon-\nicefrac{\delta}{2}>0$, and using that $$\left|\left \langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}),u-u_{\mathcal{I}} \right \rangle\right| \leq \frac{1}{2}\left(\delta^{-1}\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\|_{Z'}^{2}+\delta{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}^{2} \right),$$ we conclude $$\label{imp}
\frac{1}{2}\frac{{\mathsf{d}}}{{\,{\mathsf{d}}t}}{\left\|u-u_{\mathcal{I}}\right\|}_{0}^{2}+C_{\delta,\varepsilon}{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}^{2}
\leq \frac{1}{2\delta}\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\|_{Z'}^{2}.$$ Integrating with respect to $t$ yields $$\label{final1}
\begin{aligned}
\int_0^t{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}^{2}{\,{\mathsf{d}}t}& \leq \frac{\max{\{1,\delta^{-1}\}}}{2C_{\delta,\varepsilon}}\left({\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}^{2}+{\left\|g-\Pi^{0}g\right\|}_{0}^{2}\right).
\end{aligned}$$ In addition, invoking again , we obtain $$\label{eq:final2}
{\left\|u-u_{\mathcal{I}}\right\|}_{L^{\infty}(0,t;L^{2}(\Omega))}^{2}
\leq \max\{1,\delta^{-1}\} \left({\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}^{2}+{\left\|g-\Pi^{0}g\right\|}_{0}^{2}\right).$$ Moreover, applying the Lipschitz continuity of $f$, we observe, for $ v\in Z $ and $t\in(0,T]$, that $${\lvert \langle \partial_{t}(u-u_{\mathcal{I}}),v \rangle \rvert}
\leq {\lvert \left \langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}),v \right \rangle \rvert}+\int_{\Omega}{\{\varepsilon {\lvert \nabla(u-u_{\mathcal{I}}) \rvert}{\lvert \nabla{v} \rvert}
+L{\lvert u-u_{\mathcal{I}} \rvert}{\lvert v \rvert}\}{\,{\mathsf{d}}\bm x},}
$$ i.e., there holds $${\lvert \left \langle \partial_{t}(u-u_{\mathcal{I}}),v \right \rangle \rvert} \leq {\lvert \left \langle {\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}}),v \right \rangle \rvert}+\max{(1,L)}{\left|\!\left|\!\left|v\right|\!\right|\!\right|}_{\varepsilon}{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}.$$ Thus, $$\label{eq:aux1}
\|\partial_{t}(u-u_{\mathcal{I}})\|_{Z'}\leq \|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\|_{Z'}+\max{(1,L)}{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}.$$ Taking the square in the previous inequality, integrating over $(0,t)$, and recalling leads to $$\label{her}
{\left\|\partial_{t}(u-u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}^{2}\le C_{\delta,\varepsilon,L}\left({\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}^{2}+{\left\|g-\Pi^{0}g\right\|}_{0}^{2}\right),$$ for a constant $C_{\delta,\varepsilon,L}>0$. Combining , , and , we finally get that $$\begin{aligned}
E(t;u_{\mathcal{I}},g):&={\left\|g-\Pi^{0}g\right\|}_{0}^2+{\left\|u-u_{\mathcal{I}}\right\|}_{L^{\infty}(0,t;L^{2}(\Omega))}^{2}\\
&\quad+\int_0^t\left\{{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}^{2}+\|\partial_{t}(u-u_{\mathcal{I}})\|^2_{Z'}\right\}{\,{\mathsf{d}}t}\\
&\le \widetilde C_{\delta,\varepsilon,L} \left({\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}^{2}+{\left\|g-\Pi^{0}g\right\|}_{0}^{2}\right),\end{aligned}$$ for a constant $\widetilde C_{\delta,\varepsilon,L}>0$, and any $t\in(0,T]$. Conversely, proceeding as in , it is possible to show that $$\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\|_{Z'}\leq \|\partial_{t}(u-u_{\mathcal{I}})\|_{Z'}+\max{(1,L)}{\left|\!\left|\!\left|u-u_{\mathcal{I}}\right|\!\right|\!\right|}_{\varepsilon}.$$ Integrating over $(0,t)$ shows the equivalence of the residual term ${\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}_{L^{2}(0,t;Z')}+{\left\|g-\Pi^{0}g\right\|}_{0}$ and of the error $E(t;u_{\mathcal{I}},g)$.
A Fully Adaptive Newton-Galerkin Algorithm
------------------------------------------
We will now propose a procedure that will combine a Newton method with automatic spatial finite element mesh and time step refinements based on the *a posteriori* error estimate from Theorem \[lam:1\]. Recalling our derivations in the previous Section \[sc:RN\], it is reasonable to control the quantity $$E^n(u_{\mathcal I},g):={\left\|{\mathsf{F}}_{\varepsilon}(u_{\mathcal{I}})\right\|}^2_{L^{2}(0,t_n;Z')}+{\left\|g-\Pi^{0}g\right\|}_{0}^2,$$ for $n=1,\ldots,M$. Then, by means of Theorem \[lam:1\], we have that $$\label{eq:En}
E^n(u_{\mathcal I},g)
\preccurlyeq \eta_0^2+\sum_{j=1}^nk_j\left(\eta_{j,\Omega,N}^2+\vartheta^2_{j,\Omega,N}+\Upsilon^2_{j,\Omega,N}\right),$$ where $\eta_0:=\|g-\Pi^0g\|_0$, and $\eta_{j,\Omega,N}$, $\vartheta_{j,\Omega,N}$, and $\Upsilon_{j,\Omega,N}$ are defined in . Given a final time $T>0 $, and some positive tolerances $ \varepsilon_{0}, \varepsilon_{\eta}, \varepsilon_{\vartheta}, \varepsilon_{\Upsilon}>0$, we define $$\label{eq:localtolerances}
\varepsilon_{\text{loc},\eta}:=\frac{\varepsilon_{\eta}}{\sqrt{T}},\quad \varepsilon_{\text{loc},\vartheta}:=\frac{\varepsilon_{\vartheta}}{\sqrt{T}},\quad \varepsilon_{\text{loc},\Upsilon}:=\frac{\varepsilon_{\Upsilon}}{\sqrt{T}}.$$ Suppose that $$\label{eq:eta0}
\eta_0\le \varepsilon_0,$$ and that, for any $n=1,\ldots,M$, there holds $$\eta_{n,\Omega,N}\le\varepsilon_{\text{loc},\eta},\quad
\vartheta_{n,\Omega,N}\le\varepsilon_{\text{loc},\vartheta},\quad
\Upsilon_{n,\Omega,N}\le\varepsilon_{\text{loc},\Upsilon}.$$ Then, we conclude $$\label{Finale}
E^M(u_{\mathcal I},g)\preccurlyeq \varepsilon_{T}^2,$$ where $$\begin{aligned}
\varepsilon_{T}^{2}:=\varepsilon_{0}^{2}+\varepsilon_{\eta}^{2}+\varepsilon_{\vartheta}^{2}+\varepsilon_{\Upsilon}^{2}.
\end{aligned}$$ We now formulate a possible realization of a time-space-Newton-Galerkin adaptive algorithm which aims to generate a numerical solution $ u_{\mathcal{I}} $ that satisfies the error bound . The basic idea is to exploit the structure of the error bound to provide an interplay between adaptive finite element space refinements (and derefinements), automatic selection of the time steps, and an appropriate resolution of the Newton linearization error. More precisely, our adaptive procedure identifies whichever of the computable *a posteriori* quantities occurring in is currently dominant, and performs a corresponding refinement. In this way, the scheme follows along the lines of our previous approach in [@AmreinWihler:14; @AmreinWihler:15] on stationary problems, with the additional feature that the temporal errors are now taken into account, too. Our fully adaptive Newton-Galerkin method is outlined in Algorithm \[spacetime\].
Initialization: Input a final time $T>0$. Prescribe the overall error $\varepsilon_{T}>0$, a lower bound for the time steps $k_{\min}>0$, an initial time step $k_{0}\ge k_{\min}$, and two time mesh parameters $\kappa>1$ (derefinement) and $\sigma \in (0,1)$ (refinement). Set $n\gets 1$. Find an approximation $\Pi^0g\in H^1_0(\Omega)$ of the initial condition $g$ such that $\varepsilon_0:=\|g-\Pi^0\|_0<\varepsilon_T$. Split the tolerance $\varepsilon_{T}$ into contributions $\varepsilon_{\eta}, \varepsilon_{\vartheta}, \varepsilon_{\Upsilon}>0$ such that there holds $\varepsilon_{T}^{2}:=\varepsilon_{0}^{2}+\varepsilon_{\eta}^{2}+\varepsilon_{\vartheta}^{2}+\varepsilon_{\Upsilon}^{2}$.
As already emphasized in Remark \[im\], step $(6)$ in the above Algorithm \[spacetime\] may be delicate to realize if derefinements of the spatial mesh are taken into account. Indeed, in step (5), any coarsening procedure should be moderate in order to prevent the Newton iteration from leaving the current basin of attraction.
We also remark that the parameter $ k_{\min}>0$ in step (4) guarantees that the step size $k_{n}$ does not become overly small. This restriction needs to be relaxed when resolving *finite time blow-up problems*, where the adaptivity with respect to the time evolution requires arbitrarily small step sizes $k_n$ close to the blow-up time; see, e.g., [@JanssenWihler:15; @BandleBrunner:98; @Nakagawa:75] for details.
Numerical Experiments {#sec:numerics}
=====================
We will now illustrate and test the above Algorithm \[spacetime\] by means of a number of numerical experiments. We choose the initial spatial meshes to be sufficiently fine (as to fulfill ). Elements $K\in \mathcal{T}_{h}^{n} $ are derefined whenever $ \eta_{h,K,N}<0.1 \overline{\eta}_{h,\Omega,N} $, where $ \overline{\eta}_{h,\Omega,N}$ signifies the mean of all $\eta_{h,K,N}$, $K\in\mathcal{T}^n_h$; see .
\[ex:1\] On $\Omega=(0,1)$ let us consider the [*linear*]{} singularly perturbed initial/boundary value problem $$\begin{aligned}
\partial_{t}u -\varepsilon u''&=\exp(t) \ &&\text{on} \ \Omega\times(0,T],\\
u&=0 \ &&\text{on} \ \{0,1\}\times(0,T], \\
u(0,\cdot)&=g_{\varepsilon}&&\text{in}\ \Omega,
\end{aligned}$$ where $g_{\varepsilon}$ is the solution of the elliptic boundary value problem $$-\varepsilon g_{\varepsilon}'' + g_{\varepsilon} = 1, \qquad
g_\varepsilon(0)=g_\varepsilon(1)=0.$$ Note that $g_{\varepsilon}$ exhibits boundary layers at $ x\in \{0,1\}$; cf. [@AmreinWihler:15]. Since the problem is linear, the Newton iteration is redundant in this example. We prescribe the time derefinement/refinement parameters $ \kappa =2, \sigma=\nicefrac{1}{2}$. Moreover, we compute a numerical solution up to the final time $T=1$, and set the local error tolerances from (as well as $ \varepsilon_{0} $ given in ) to $ 10^{-3}$. Furthermore, the initial time step $k_0$ is chosen to be $ \nicefrac{1}{10}$.
Our goal here is to test the robustness of the *a posteriori* error analysis with respect to $\varepsilon$ as $\varepsilon\to 0$. To this end, we quantify the performance of our algorithm by comparing the true error $ {\left\|u-u_{\mathcal{I}}\right\|}_{L^{2}(0,t_n;Z)}^{2}+{\left\|u-u_{\mathcal{I}}\right\|}_{L^{\infty}(0,t_n;L^{2}(\Omega))}^{2} $ with the estimated error (i.e., the right-hand side of ), and compute the time-dependent efficiency indices (defined by the ratio of the estimated error and the true error $ {\left\|u-u_{\mathcal{I}}\right\|}_{L^{2}(0,t_n;Z)}^{2}+{\left\|u-u_{\mathcal{I}}\right\|}_{L^{\infty}(0,t_n;L^{2}(\Omega))}^{2} $ for $ n \in \{1,2,\ldots,M\}$); the results are displayed in Figure \[bild2\] for $ \varepsilon = 10^{-p}$, with $p \in \{1,2,3,4,5\} $. They show that the boundary layers close to $\{0,1\}$ are properly resolved, and clearly highlight the robustness of the efficiency indices with respect to $\varepsilon$ as $\varepsilon \to 0$.
![Example \[ex:1\]: Numerical solution (with ’$\circ$’ indicating the mesh points) vs. exact solution at $T=1$ (left), and efficiency indices for various choices of $\varepsilon$ (right).[]{data-label="bild2"}](figures/LinearSolution "fig:"){width="47.00000%"} ![Example \[ex:1\]: Numerical solution (with ’$\circ$’ indicating the mesh points) vs. exact solution at $T=1$ (left), and efficiency indices for various choices of $\varepsilon$ (right).[]{data-label="bild2"}](figures/effizienzLunendlich "fig:"){width="46.00000%"}
\[ex:2\] Furthermore, on $\Omega = (0,1)$ consider the following nonlinear singularly perturbed initial/boundary value problem: $$\label{boundarylayers}
\left\{ \begin{aligned}
\partial_{t}u - \varepsilon u'' &= -u^{4}+\sin(t), && \text{on} \ \Omega\times(0,T],\\
u&=0, &&\text{on} \ \{0,1\}\times(0,T],\\
u(0,\cdot)&=g && \text{in} \ \Omega.
\end{aligned} \right.$$ When evolving in time, problem exhibits boundary layers for $ 0<\varepsilon \ll 1 $; see Figure \[bild22\] (left), and [@Verhulst] for a detailed discussion of this problem. We consider $\varepsilon=10^{-5}$, and choose the local error tolerances from (as well as $\varepsilon_{0}$ from ) to be $10^{-3}$, and the initial time step as $k_{0}=\nicefrac{1}{4}$. In Figure \[bild22\] (right) we depict a log/log plot of the estimated error from up to the final time $T=2$. Notice that the slope $\nicefrac{1}{2}$ in the log/log plot is due to the fact that $$\begin{aligned}
\sqrt{E^{n}(u_{\mathcal{I}},g)} & \preccurlyeq \left(\eta_0^2+\sum_{l=1}^nk_l\left(\eta_{l,\Omega,N}^2+\vartheta^2_{l,\Omega,N}+\Upsilon^2_{l,\Omega,N}\right)\right)^{\nicefrac{1}{2}}\\
& \leq \left(\varepsilon_{0}^{2}+t_n(\varepsilon_{\text{loc},\eta}^{2}+\varepsilon_{\text{loc},\vartheta}^{2}+\varepsilon_{\text{loc},\Upsilon}^{2})\right)^{\nicefrac{1}{2}},\end{aligned}$$ i.e., for sufficiently small $\varepsilon_0>0$, we expect the error to grow of order $ \mathcal{O}(t_n^{\nicefrac12})$ as time evolves.
![Example \[ex:2\]: Snapshots of the numerical solution (as time is evolving) corresponding to problem with $\varepsilon = 10^{-5}$ (left), and the estimated error for $\sqrt{E^{n}(u_{\mathcal{I}},g)}$ (right).[]{data-label="bild22"}](figures/VerhulstGraph "fig:"){width="47.90000%"} ![Example \[ex:2\]: Snapshots of the numerical solution (as time is evolving) corresponding to problem with $\varepsilon = 10^{-5}$ (left), and the estimated error for $\sqrt{E^{n}(u_{\mathcal{I}},g)}$ (right).[]{data-label="bild22"}](figures/PerformanceVerhulst "fig:"){width="47.00000%"}
\[ex:3\] Finally, we consider the nonlinear problem $$\label{blow up}
\left\{ \begin{aligned}
\partial_{t}u - \varepsilon \Delta u &= u^{\beta}, && \text{on} \ \Omega\times(0,T],\\
u&=0, &&\text{on} \ \{0,1\}\times(0,T],\\
u(0,\cdot)&=g && \text{in} \ \Omega.
\end{aligned} \right.$$ A detailed discussion of problems with power-type source terms can be found, for instance, in the monograph [@SamarskiiGalaktionovKurdyumovMikhailov:95]. In particular, for $\beta>1$, the solution of will become unbounded in finite time provided the initial data $u(0,\cdot)=g\geq 0 $ is suitably chosen.
On the left of Figure \[bild1\] we show the numerical solution of for $ \Omega = (0,4)$, $ \beta=4$, $\varepsilon=10^{-3}$, and $T\approx 0.1$. The local error tolerances from are set to $10^{-2}$, and $k_{0}=10^{-3}$. On the right in Figure \[bild1\], we present a log/log plot of the estimated error (i.e. the right-hand side of ) corresponding to the numerical solution shown on the left in Figure \[bild1\]. We clearly observe that the estimated error from increases with slope $=\nicefrac{1}{2}$, as in Example \[ex:2\]. Moreover, as time evolves, we see that the adaptive procedure is able to resolve properly the spike located around $x=2$.
![Example \[ex:3\]: Snapshots of the numerical solution (as time is evolving) corresponding to problem with $\varepsilon=10^{-3}
$, $\beta = 4 $, $ \Omega = [0,4] $ (left), and the estimated error for $\sqrt{E^{n}(u_{\mathcal{I}},g)}$ (right).[]{data-label="bild1"}](figures/BlowUpSolution "fig:"){width="47.40000%"} ![Example \[ex:3\]: Snapshots of the numerical solution (as time is evolving) corresponding to problem with $\varepsilon=10^{-3}
$, $\beta = 4 $, $ \Omega = [0,4] $ (left), and the estimated error for $\sqrt{E^{n}(u_{\mathcal{I}},g)}$ (right).[]{data-label="bild1"}](figures/BlowUpPerformance "fig:"){width="47.10000%"}
Conclusions {#sc:conclusions}
===========
The aim of this paper is the development of a reliable and computationally efficient procedure for the numerical solution of semilinear parabolic boundary value problems with possible singular perturbations. The key idea is to employ Newton’s method to locally linearize the problem, and to apply an automatic (spatial) finite element mesh refinement approach as well as an adaptive time stepping control procedure. The numerical scheme is studied within the context of a robust (with respect to the singular perturbations) [*a posteriori*]{} residual-oriented error analysis, and a corresponding adaptive mesh refinement scheme is developed. Our numerical experiments clearly illustrate the ability of the proposed methodology to reliably find solutions, and to robustly resolve the singular perturbations at the expected rate.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Both complete protocol and optical setup for experimental realization of quantum teleportation of unknown single-photon wave packet are proposed.'
---
17.0cm 25.cm -3.5cm -2.0cm
**Experimental scheme for quantum teleportation of a single-photon packet**
S.N.Molotkov
*Institute of Solid State Physics of Russian Academy of Sciences,*
*Chernogolovka, Moscow distr., 142432 Russia*
PACS numbers: 03.67./a, 03.65.Bz, 42.50.Dv
E-mail: [email protected]
Quantum mechanics prohibits cloning (copying) an unknown quantum state (no cloning theorem \[1\]). Is it possible to transmit to a distant user a previously unknown quantum state without sending that state itself? Strictly speaking, any measurement aimed at obtaining classical information to be sent to another observer changes the state itself without providing complete information about it. Production of a large number of identical copies which can be measured many times to obtain complete information on their common quantum state is prohibited by the no-cloning theorem. Thus, it is impossible to transmit information on a quantum state employing only a classical communication channel.
Quantum teleportation lifts this restriction if a quantum communication channel is used in addition to the classical one. The idea of quantum teleportation for the case of discrete quantum states (e.g. a spin-1/2 particle in an unknown state) was first proposed in Ref.\[3\]. A quantum channel is realized through the non-local EPR-correlations \[2,3\][^1]. An EPR-pair is a pair of particles described by an entangled state. Entanglement constitutes a special kind of quantum correlations which do not have any classical analogies.
The quantum teleportation protocol described in Ref.\[4\] has the following form. To teleport an unknown quantum state from user [*A*]{} to a distant user [*B*]{}, user [*A*]{} generates an EPR-pair. One of the particles of that EPR-pair remains with the user [*A*]{} while the second one is sent to the distant user [*B*]{}. User [*A*]{} performs a joint measurement over the particle in the unknown state to be teleported and his particle from the EPR pair thus obtaining classical information. Because of the non-local correlations inherent in the EPR-pair, the measurement outcome uniquely determines the resulting state of the second particle in the EPR-pair sent to user [*B*]{}. The state of the second particle coincides with the unknown state to within a unitary rotation. Classical information obtained in the measurement is sent by user [*A*]{} to [*B*]{} and is used by the latter to determine which unitary transformation should be performed to obtain a new state identical to the original unknown state. In the course of teleportation user [*A*]{} obtains no information on the teleported unknown state.
Quantum teleportation has recently been demonstrated experimentally for a photon in an unknown polarization state \[5,6\].
The problem of teleportation of the wave function of a particle in the one-dimensional case where position and momentum are the two continuous dynamic variables was investigated in Ref.\[7\] using the wave function proposed in Ref.\[2\] to describe an EPR pair. Studied in a recent work \[8\] was the quantum teleportation of a quantum state described by the dynamical variables (the unknown state in Ref.\[8\] corresponds to a single-mode photon state) for the case of non-ideal EPR-correlations. A quadrature-squeezed state was used as an EPR-state while the measurement procedure actually corresponded to the homodyne detection.
Proposed in the present paper is a new scheme (a complete protocol and its experimental realization) for the teleportation of a multi-mode state, i.e. a single-photon wave packet, employing an EPR-pair in an entangled state with respect to the energy–time variables.
To simplify further formulas, we shall assume that the packet polarization state is known. The discussion below is also applicable to the case of unknown polarization which can be accounted for introducing an additional subscript. The state of a single-photon wave packet can be written as (see, e.g. Ref.\[9\]) $$%%%%%%%%%%%%%1
|1\rangle_3 =
\int_{0}^{\infty}d\omega f(\omega)\hat{a}^{+}(\omega)|0\rangle =
\int_{0}^{\infty}d\omega f(\omega)|\omega\rangle_3,$$ $$[\hat{a}(\omega),\hat{a}^+(\omega')]=I\delta(\omega-\omega'),\quad
\int_{0}^{\infty} |f(\omega)|^2 d\omega =1,$$ where $\hat{a}^+(\omega)$, $\hat{a}(\omega)$ are the creation and annihilation operators of a single-mode Fock state $|\omega\rangle_3$, $|0\rangle$ is the vacuum state, $f(\omega)$ is the packet amplitude Subscript 3 labels the channel number (see Fig.1). The density matrix at an arbitrary time is $$%%%%%%%%%%%%%%%2
\rho(3)=
\left(\int_{0}^{\infty}d\omega e^{-i\omega t} f(\omega)|\omega\rangle_3\right)
\left(\int_{0}^{\infty}d\omega'{}_3\langle \omega| e^{i\omega' t}f^*(\omega')
\right)$$ In our case the state of an EPR-pair can be written as (these photon EPR-pairs are produced in the parametric energy down-conversion processes \[10\]) $$%%%%%%%%%%%%%%3
|\psi_{EPR}\rangle_{1,2} =
\int_{0}^{\infty}d\omega
|\omega\rangle_1\otimes|\Omega-\omega\rangle_2,\quad
\rho_{_{EPR}}(1,2)=|\psi_{EPR}\rangle_{1,2}\mbox{ } {}_{1,2}\langle \psi_{EPR}|,$$ where $\Omega$ is the pumping frequency and 1, 2 are the channel numbers (Fig.1). The normalization of the state (3) is insignificant for further analysis.
According to the general scheme \[11–13\], quantum mechanical measurements are described by positive operators realizing the identity resolution. Measurements of the variables corresponding to self-adjoint operators are associated with the orthogonal resolutions. Parameters (such as time and rotation angle) are not associated with any self-adjoint operators so that the corresponding measurements are described by the non-orthogonal identity resolution \[11-13\].
In the present paper the basic idea of proposed teleportation scheme consists in the usage of a joint (entangled) time–energy measurement performed on a pair of photons one of which belongs to the EPR-pair and the second photon in the unknown state to be teleported. The measurement is given by the non-orthogonal identity resolution \[14\] $$%%%%%%%%%%%% 4
\int\int M(dtd\Omega_{+})=
\int\int R^{+}R\mbox{ }(dtd\Omega_{+})=I,$$ where $R$ is the “reduction” operator, $M(dtd\Omega_{+})$ describes a quantum operation and is a positive operator valued measure, POVM, (the details can be found, e.g. in Refs.\[15,16\]); we have \[14\] $$%%%%%%%%%%%%%%%%%% 5
M(dtd\Omega_{+})=$$ $$\left(\int d\omega_{-} e^{i\omega_{-}t}
|\omega_{+}+\omega_{-}\rangle_1
\otimes|\omega_{+}-\omega_{-}\rangle_3\right)
\left(\int d\omega_{-}'e^{-i\omega_{-}'t}
{}_3\langle\omega_{+}-\omega_{-}'|
\otimes{}_1\langle\omega_{+}+\omega_{-}'|\right)
\frac{\textstyle dtd\omega_{+}}{\textstyle 2\pi},$$ $\omega_{\pm}=\frac{\Omega_{\pm}}{2}$. The integration covers the frequency ranges corresponding to positive arguments of the Fock states. It should be emphasized that the frequency $\omega_{+}$ is common to bra- and ket-states.
According to the general concepts of the quantum measurement theory \[10–13,15,16\], application of a quantum operation (measurement) to a system described by the density matrix $\rho$ transforms it to a new state $$\rho\rightarrow
\frac{\textstyle R_i\rho R^{+}_{i} }
{ \textstyle \mbox{Tr} \{ R_i\rho R^{+}_{i} \} }.$$ The probability of the $i$-th outcome is given by the formula $\mbox{Pr}=\mbox{Tr} \{ R_i\rho R^{+}_{i} \}$, where $E_i=R^{+}_{i}R_i$ is the POVM element.
In our case, after the measurement performed by user [*A*]{}, the state of the second photon from the EPR-pair observed by user [*B*]{} is given by the density matrix $$%%%%%%%%%%%%%%%%%% 6
\tilde{\rho}(2)=
\frac{\textstyle \mbox{Tr}_{1,3}
\{ \rho_{_{EPR}}(1,2)\otimes\rho(3)M(dt d\Omega_{+}) \} }
{ \textstyle \mbox{Pr}\{dt d\Omega_{+}\} },$$ $$%%%%%%%%%%%%%%%%%% 7
\mbox{Pr}\{dt d\Omega_{+}\}=
\mbox{Tr}_{1,2,3} \{ \rho_{_{EPR}}(1,2)\otimes\rho(3)M(dt d\Omega_{+}) \},$$ $$%%%%%%%%%%%%%%%%%% 8
\tilde{\rho}(2)=
\left(\int d\omega e^{-i(\Omega_{+}/2+\omega) t}
f(\Omega_{+}-\omega)|\Omega-\omega\rangle_2\right)
\left(\int d\omega'{}e^{i(\Omega_{+}/2+\omega') t}
{}_{2}\langle\Omega- \omega'| f^*(\Omega_{+} - \omega')
\right)$$ Formally, the measurement (4) corresponds to the situation where the measurement moment $t$ and frequency $\Omega_{+}$ are chosen by the experimentator, and the positive result probability is given by Eq. (7). The teleportation will be ideal if the chosen $\Omega_{+}=\Omega$ (detector registration frequency $\omega_{+}$ coincides with the pumping frequency $\Omega$). In that case it follows from Eq. (6) that the state $\tilde{\rho}(2)$ coincides with $\rho(3)$ to within a phase factor which can be eliminated by user [*B*]{} if user [*A*]{} sends to him the registration time $t$ via a classical channel. Note that the registration time $t$ does not depend on unknown state $\rho(3)$ if $\Omega_{+}=\Omega$.
Physically, the measurement (4) can be understood in the following way. User [*A*]{} has a continuum of detectors “tuned” to the frequencies in the ($0,\infty$) range, and each of them can fire at an arbitrary time $t$, formally in the infinite ($-\infty,\infty$) interval. The probability for the detector tuned to frequency $\Omega_+$ to fire at time $t$ is given by Eq. (7). The firing probability does not depend on time only for the detector tuned to frequency $\Omega$. The teleportation will only be ideal if the detector tuned to $\Omega$ fires. In that case it follows from Eq. (7) that the firing probability does not depend on the unknown input state. User [*A*]{} does not obtain any information on the teleported state.
Since the measurements can be performed at spatially separated points, all the times in the above formulas should be understood as the reduced time-of-flight corrected times ($t\rightarrow t-x/c$). It will be seen below that this point is insignificant in the case of ideal teleportation.
The major problem in realizing the single-photon packet teleportation is the implementation of the measurement given by Eq. (4). The measurement (4) on a pair of photons is an intermediate case between the time and energy measurements; its experimental realization is described below. The idea is to convert a photon pair into a single photon which then is measured by narrow-band photodetector. The latter can be easily realized experimentally.
The experimental setup is presented in Fig.1. The first non-linear crystal with the second-order susceptibility $\chi$ and a narrow-band filter tuned to the frequency $\Omega$ serve to generate the EPR-pair in channels 1 and 2. The channel 3 is used to feed the unknown single-photon packet. The latter can be prepared by exciting a two-level system with a $\pi$-pulse a long time ago. The second non-linear crystal, the narrow-band filter behind it tuned to frequency $\Omega$, and then a standard photodetector realize the measurement (4). The teleported state arises in channel 2.
Consider step by step the input state evolution in the optical scheme. After the first narrow-band filter in front of the first non-linear crystal the state is described by a monochromatic state with the density matrix $$%%%%%%%%%%%%%%%%%% 9
\rho_{in}(in)=|\Omega\rangle_{in}\mbox{ }{}_{in}\langle\Omega|,$$ which can be obtained by cutting a narrow band by the first filter from an auxiliary single-photon packet fed into the input channel $in$ (Fig.1). The photon-photon interaction in the nonlinear crystal is described in the interaction representation by the following Hamiltonian (details can be found in Refs. \[17,18\]) $$%%%%%%%%%%%%%%% 10
H_1(t)=\chi\int d{\bf x}E_{in}^{(+)}({\bf x},t)E_{1}^{(-)}({\bf x},t)
E_{2}^{(-)}({\bf x},t) +h.c.,$$ all the insignificant constants are assumed to be included in the definition of $\chi$, which as usually \[17,18\] will be assumed to be frequency-independent. It is convenient to present the electric field operators in the form \[9\] $$%%%%%%%%%%%%%%% 11
E_{i}^{(-)}({\bf x},t)=
\frac{1}{\sqrt{2\pi}}
\int_{0}^{\infty}d\omega e^{i(\omega t-{\bf kx})}
\hat{a}^{+}(\omega)|0\rangle_{i}=
\frac{1}{\sqrt{2\pi}}
\int_{0}^{\infty}d\omega e^{i(\omega t-{\bf kx})}|\omega\rangle_{i},$$ where $i$ is the channel number. Analogously for $E_{i}^{(+)}({\bf x},t)$. Taking into account Eq. (11) we have $$%%%%%%%%%%%%%%% 12
H_1(t)=\frac{\chi}{(2\pi)^{3/2}}
\int\int\int
d\omega_1d\omega_2 d\omega_{in}
e^{ it( \omega_1+\omega_2-\omega_{in} ) }
|\omega_1\rangle_1\otimes |\omega_2\rangle_2\mbox{ }{}_{in}\langle\omega_{in}|
\int_{vol}d{\bf x} e^{-i{\bf x}({\bf k_1+k_2-k_{in}})}+h.c.$$ In the second integral the integration is performed throughout the entire crystal yielding the $\delta$-symbol with respect to momentum, and results in the phase synchronism conditions \[18\] (${\bf k_1+k_2}={\bf k_{in}}$), which will be assumed satisfied (below it means that ${\bf k_2}\|{\bf k_3}$, Fig.1). In the subsequent formulas the quantity $\chi$ is understood as the renormalized constant corrected for the additional factors arising from the second integral. The first-order susceptibility which is always present can be neglected in our analysis since the corresponding terms in the Hamiltonian do not contribute to the [*out*]{} channel.
The state after the first crystal in channels 1 and 2 is described by the density matrix $$%%%%%%%%%%%%%%% 13
\rho_{EPR}(1,2)=S(t)\rho_{in}(in)S^{-1}(t),$$ where $S(t)$ is the $S$-matrix $$%%%%%%%%%%%%%%% 14
S(t)=e^{i\int_{-\infty}^{t}H_1(t')dt'}=1+S^{(1)}+S^{(2)}+\ldots.$$ In the first order with respect to $\chi$ one has $$%%%%%%%%%%%%%%% 15
S^{(1)}=i\chi\int\int
d\omega_1 d\omega_{in} |\omega_1\rangle_1\otimes
|\omega_{in}-\omega_1\rangle_2
\mbox{ }{}_{in}\langle\omega_{in}| +h.c.$$ The upper integration limit in the exponent in $S$ can be replaced by $\infty$ which is physically actually related to the fact that the input state is a monochromatic one (roughly speaking infinitely extended in time) and the teleportation process is formally stationary. To within an unimportant normalization constant, the state in channels 1 and 2 is described by the density matrix $$%%%%%%%%%%%%%%% 16
\rho_{EPR}(1,2)=\chi^2
\left(\int_{0}^{\infty}d\omega|\omega\rangle_1\otimes|\Omega-\omega\rangle_2
\right)
\left(\int_{0}^{\infty}d\omega'{}_{1}\langle\omega'|\otimes
{}_{2}\langle\Omega-\omega'|\right).$$ Measurement by a photodetector which has a narrow-band filter tuned to the frequency $\Omega$ installed in front of it is formally described by the projector $P(\Omega)=|\Omega\rangle_{out}\mbox{ }{}_{out}\langle\Omega|$.
The teleported state in channel 2 after the registration by the photodetector is described by the density matrix (again to within the normalization constant) $$%%%%%%%%%%%%%%% 17
\tilde{\rho}(2)=\mbox{Tr}_{out}\left\{
S(t)\rho_{in}(in)\otimes\rho(3)S^{-1}(t)P(\Omega)\right\},$$ where $S(t)$ is now the full $S$-matrix of the entire optical scheme, $$%%%%%%%%%%%%%%% 18
S(t)=e^{i\int_{-\infty}^{t}[H_1(t') + H_2(t')]dt'}=1+S^{(1)}+S^{(2)}+\ldots,$$ where $H(t)_2$ is th second non-linear crystal Hamiltonian which coincides to within the subscripts interchange with $H(t)_1$ in Eq. (12). Contributing to the teleportation processes are the $S$-matrix terms of the form $$%%%%%%%%%%%%%%% 19
S^{(2)}\propto$$ $$\chi^2
\left(\int\int d\omega_1
d\omega_{in} |\omega_1\rangle_1 \otimes|\omega_{in}-\omega_1\rangle_2
\mbox{ }{}_{in}\langle\omega_{in}|\right)
\left(\int\int d\omega'_1 d\omega_{out}
|\omega_{out}\rangle_{out}\mbox{ } {}_{1}\langle\omega'_{3}|\otimes
{}_{in}\langle\omega_{out}-\omega'_{1}|\right).$$ Taking into account Eq. (19), the density matrix in channel 2 to within the normalization constant coincides with the initial density matrix of he unknown wave packet $$%%%%%%%%%%%%%%% 20
\tilde{\rho}(2)=\chi^4
\left(\int_{0}^{\infty}d\omega f(\omega)|\omega\rangle_3\right)
\left(\int_{0}^{\infty}d\omega'{}_3\langle \omega| f^*(\omega')
\right)$$ It also follows from Eq. (17) that the detection probability in the output channel ($out$) does not depend on the unknown state and is proportional to $$%%%%%%%%%%%%%%% 21
\mbox{Pr}=\mbox{Tr}_{2,out}\left\{
S(t)\rho_{in}(in)\otimes\rho(3)S^{-1}(t)P(\Omega)\right\}
\propto \chi^4.$$ In the outlined scheme the classical channel is used to inform the distant user of the fact that a photodetector fired, and in that case the teleportation is assumed to be successful. The probability (efficiency) of the teleportation process is small to the measure of $\chi^4$. The fraction of false photodetection firings when a wrong state will be teleported due to the terms of higher orders in $\chi$ in the $S$-matrix has an additional smallness in $\chi^2$. Note that formally the teleportation process is stationary (requires infinite time), since a monochromatic state should be prepared. In that case the probability of the photodetector firing in the $out$ channel does not depend on time and does not depend on the input state. In that case user [*A*]{} obviously acquires zero information on the teleported state.
Of course, quantum teleportation does not allow information transmission faster than the speed of light. In the present scheme an intuitive and qualitative explanation is as follows. Since the input state is monochromatic and always intuitively assumed to be non-localized (infinitely extended), the latter means that the field is “prepared in advance” throughout the entire space, including the locations of both distant users [*A*]{} and [*B*]{}. The measurement performed by user [*A*]{} transforms the entire system to a new state—reduces the state vector “immediately” and “everywhere” for the entire system. This assumption is usually considered as counterintuitive. However, these “immediately” and “everywhere” do not result in faster than light communication. To transmit classical information from [*A*]{} to [*B*]{} with the help of a teleported state, one requires a classical channel from [*A*]{} to [*B*]{} to tell that the detector fired and the teleportation was successfully completed. Classical communication channel assumes sending a classical object from [*A*]{} to [*B*]{} whose velocity cannot exceed that of light. The problem of the field “prepared in advance” everywhere is closely related to the photon localizability (to be more precise, non-localizability) (e.g. see Refs.\[19–23\]). As far as I know this problem has not yet been discussed in detail in the context of quantum teleportation.
It should be noted that the teleportation process can be reformulated in the Feynman diagram language, the averaging being performed over the stationary state $|\Omega\rangle\otimes \int_{0}^{\infty}d\omega f(\omega)|\omega\rangle$ corresponding to the input monochromatic state and the single-photon state. In the case of ideal teleportation this state is also the output state. In that sense the process is stationary and the averaging is performed over the stationary state which should not necessarily be the ground state. In that case the diagrammatic technique is developed in the same way as in the Keldysh method \[24\].
The author is grateful to B.A.Volkov, M.V.Lebedev, S.S.Nazin, and S.T.Pavlov for discussions. The work was supported by the Russian Fund for Basic Research (Grant \# 96-02-19396), and the Program “Advanced Technologies in Micro- and Nanoelectronics” (Grant \# 02.04.329.895.3).
[99]{} W.K.Wootters, W.H.Zurek, Nature, [**299**]{}, 802 (1982). A.Einstein, B.Podolsky, N.Rosen, Phys. Rev., [**47**]{}, 777 (1935). J.S.Bell, [*Speakable and Unspeakable in Quantum Mechanics*]{}, (Cambridge University Press, Cambridge, England, 1988). C.H.Bennett, G.Brassard, C.Crepeau, R.Jozsa, A.Peres, W.K.Wootters, Phys. Rev. Lett., [**70**]{}, 1895 (1993). B.Boumeester, Jian-Wei Pan, K.Mattle, M.Eibl, H.Weinfurter, A.Zeilinger, Nature, [**390**]{}, 575 (1997). D.Boschi, S.Branca, F.De Martini, L.Hardy, S.Popescu, Phys. Rev. Lett., [**80**]{}, 1121 (1998). L.Vaidman, Phys. Rev., [**A49**]{}, 1473 (1994). S.Braunstein, H.J.Kimble, Phys. Rev. Lett., [**80**]{}, 869 (1998). R.A.Campos, B.E.A.Saleh, M.C.Teich, Phys. Rev., [**A42**]{}, 4127 (1990). P.G.Kwait, K.Mattle, H.Weinfurter, A.Zeilinger, A.V.Sergienko, Y.H.Shin, Phys. Rev. Lett., [**75**]{}, 4337 (1995). A.S.Holevo, [*Probabilistic and Statistical Aspects of Quantum Theory*]{}, “Nauka”, M. 1980 (in Russian). P.Busch, M.Grabowski, P.J.Lahti, [*Operational Quantum Physics*]{}, Springer Lecture Notes in Physics, [**v. 31**]{}, 1995. K.Krauss, [*States, Effects and Operations*]{}, Springer-Verlag, Berlin, 1983. S.N.Molotkov, in [*LANL E-Print Archive, http:// xxx.lanl.gov/ quant-ph/9805045*]{}. B.W.Schumacher, Phys. Rev., [**A54**]{}, 2614 (1996). M.A.Nielsen, C.M.Caves, Phys. Rev., [**A55**]{}, 2547 (1997). C.K.Hong, L.Mandel, Phys. Rev., [**A31**]{}, 2409 (1985). P.W.Milonni, H.Fearn, A.Zeilinger, Phys. Rev., [**A53**]{}, 4556 (1996). T.D.Newton, E.P.Wigner, Rev. Mod. Phys., [**21**]{}, 400 (1949). A.S.Wightman, Rev. Mod. Phys., [**34**]{}, 845 (1962). K.Kraus, [*Position observable of photon, in: *The Uncertainty Principle and Foundations of Quantum Mechanics**]{}, Eds. W.C.Price and S.S.Chissik, John Wiley& Sons, New York, p.293, (1976). H.Bacry, [*Localizability and Space in Quantum Physics, in: Lecture Notes in Physics*]{}, Eds. H.Araki et al, Springer-Verlag, Vol.308, (1988). J.A.Brooke, F.E.Schroek, J. of Math. Phys., [**37**]{}, 5958 (1996). L.V.Keldysh, ZhETF, 47 (1964) 1515 (Sov. JETP, 20 (1965) 1018).
[^1]: This term stems from the well-known Einstein-Podolsky-Rosen effect \[2\]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
T. Boku$^{a}$, $^{b,c}$, Y. Kuramashi$^{a,c,f}$, K. Minami$^{c}$, Y. Nakamura$^{c}$, F. Shoji$^{c}$, D. Takahashi$^{a}$, M. Terai$^{c}$, A. Ukawa$^{a}$, T. Yoshié$^{a,f}$\
$^{a}$Center for Computational Sciences, University of Tsukuba, Tsukuba, Ibaraki 305-8577, Japan\
$^{b}$Graduate School of Science, Hiroshima University, Higashi-Hiroshima, Hiroshima 739-8526, Japan\
$^{c}$RIKEN Advanced Institute for Computational Science, Kobe, Hyogo 650-0047, Japan\
$^{f}$Faculty of Pure and Applied Sciences, University of Tsukuba, Tsukuba, Ibaraki 305-8571, Japan\
E-mail:
title: 'Multi-block/multi-core SSOR preconditioner for the QCD quark solver for K computer'
---
Lattice QCD on the K computer
=============================
The K computer has been developed by RIKEN and Fujitsu since 2006 as the national leadership computer of Japan to promote science and technology [@KcomputerAICS]. The construction has been finished on July 2012 and the system is now provided to the strategic programs of the High Performance Infrastructure (HPCI) and to the research users those who apply the computer resource through the HPCI [@HPCI]. The strategic programs consist of five fields: ([1]{}) Predictable life science, healthcare and drug discovery foundation, ([2]{}) New Materials and Energy Creation, ([3]{}) Projection of Planet Earth Variations for Mitigating Natural Disasters, ([4]{}) Next-generation manufacturing technology, and ([5]{}) The origin of matter and the universe. Lattice QCD is contained in the fifth strategic program [@HPCISPFFive].
The K computer consists of over 80,000 computational nodes connected by the so called “Tofu” network. The Tofu network topology is six dimensional topology with 3D-mesh times 3D-torus shape, with which robustness against the node failure and high node flexibility and availability are ensured. Each node has a single CPU chip called “[SPARC64$^{\mathrm{TM}}$ VIIIfx ]{}”. Equipping 8 cores with SIMD enabled 256 registers and 6MB shared L2 cache, the CPU can achieve a high efficiency for scientific applications (for the detailed system structure see [@FSTJ]). Lattice QCD is one of the suitable applications for this kind of massively parallel system architecture.
In this paper we present the algorithmic and performance tuning of the $O(a)$ improved Wilson quark solver for the K computer. We focus on the solver algorithm preconditioned by the domain-decomposed Schwartz alternating procedure (SAP) with mixed precision [@LuscherSAPHMC]. In the next section we briefly introduce the SAP preconditioner and the nested BiCGStab algorithm [@NestedBiCGStab]. The optimization of the algorithm and tuning methods for the computational kernels of the SAP are presented in section \[seq:Tunning\]. We give the performance benchmark results in section \[seq:Results\] and summarize the paper in the last section.
Nested BiCGStab with Lüscher’s SAP preconditioner
=================================================
Our target problem is solving the following linear equation: $$D x = b,
\label{eq:LinEQ}$$ where $D$ is the clover term preconditioned $O(a)$-improved Wilson-Dirac operator: $$\begin{aligned}
D^{a,b}_{\alpha,\beta}(n,m) &=& \delta^{a,b}\delta_{\alpha,\beta}\delta(n,m)
- \kappa F^{a,c}_{\alpha,\gamma}(n) \sum_{\mu=1}^{4}
\left[
( 1-\gamma_{\mu})_{\gamma,\beta} (U_{\mu}(n))^{c,b} \delta(n+\hat{\mu},m)\right.\nonumber\\
&&\left.\hspace{12em}
+( 1+\gamma_{\mu})_{\gamma,\beta} ((U_{\mu}(m))^{b,c})^* \delta(n-\hat{\mu},m)
\right].\end{aligned}$$ Where $F(n)$ is the inverse clover term $(1-(c_{\mathrm{SW}}\kappa/2)\sigma_{\mu\nu}F_{\mu\nu}(n))^{-1}$, $(n,m)$ are lattice site, $a,b$ color, $\alpha,\beta$ are spin indexes.
Lüscher has introduced the Schwartz alternating procedure (SAP) to further preconditioning the Eq. (\[eq:LinEQ\]) [@LuscherSAPHMC]. By dividing the whole lattice into two colored blocks in checkerboard manner, Eq. (\[eq:LinEQ\]) can be rewritten in the following $2 \times 2$ block form: $$\left(
\begin{array}{cc}
D_{EE} & D_{EO}\\
D_{OE} & D_{OO}
\end{array}\right)
\left(
\begin{array}{c}
x_{E}\\ x_{O}
\end{array}\right)=
\left(
\begin{array}{c}
b_{E}\\ b_{O}
\end{array}\right),$$ where $D_{EE}$ ($D_{OO}$) is a block restricted operator in even-domain (odd-domain), while $D_{EO}$ ($D_{OE}$) contains hopping operations from odd-domain to even-domain (and vice versa). The SAP practitioner $M_{\mathrm{SAP}}$ is introduced as $$M_{\mathrm{SAP}} = K\sum_{j=0}^{N_{\mathrm{SAP}}-1} (1-DK)^{j},\quad\quad \mbox{with} \quad\quad
K= \left(
\begin{array}{cc}
A_{EE} & 0 \\
-A_{OO}D_{OE}A_{EO} & A_{OO} \\
\end{array}
\right),$$ where $A_{EE}$ ($A_{OO}$) can be any approximation for $(D_{EE})^{-1}$ ($(D_{OO})^{-1}$). When $A_{EE}$ and $A_{OO}$ are exact, $DK$ becomes block triangular and is expected to be well preconditioned. In the case $|DK|<1$, $M_{\mathrm{SAP}}$ converges to $D^{-1}$ when $N_{\mathrm{SAP}}\rightarrow \infty$.
The nested BiCGStab solver is designed to be flexible against the preconditioner changing iteration by iteration [@NestedBiCGStab] using an inner-outer strategy. The outer BiCGStab contains the inner BiCGStab solver as the flexible preconditioner. The flexibility ensures the double precision accuracy of the solution vector even if we use the single precision for the inner BiCGStab solver [@mixedprec]. We apply the SAP preconditioner $M_{\mathrm{SAP}}$ to the inner single precision BiCGStab solver. The use of single precision has a merit as it requires less system resources than double precision. We target the single precision part of the solver as the tuning part for the K computer. In the following section we focus on the tuning of the single precision $DK$ of $M_{\mathrm{SAP}}$.
Performance tuning for the K computer {#seq:Tunning}
=====================================
#### Inverse of block operator and OpenMP threading
The approximate inverse of the block operator $A_{EE}$ ($A_{OO}$) are the important part of the SAP. The exactness is not required for the SAP, however, the better approximation with less computational cost is preferred for $A_{EE}$ ($A_{OO}$). The even-odd site preconditioning has been used in [@LuscherSAPHMC] and the SSOR preconditioning has been used in [@PACSCSLDDHMC]. The latter has a better performance than the former at the same computational cost. The SSOR preconditioning is derived by decomposing the original operator into the sum of an upper and a lower triangular matrices. The preconditioning is achieved by solving the upper and lower triangular parts through forward and backward substitutions. The decomposition and the efficiency of the SSOR depend on the site ordering. It is observed that the SSOR with natural ordering have a better performance [@PACSCSLDDHMC]. However the SSOR with natural ordering is not suitable for the multi-core CPU architecture because the natural ordering has a global data recurrence pattern and less parallelism in the forward and backward substitutions. To extract 8 core parallelism for the K computer, we further divide the block into 16 sub-blocks via the [*locally-lexicographical*]{} ordering ([*ll*]{}-ordering) described in [@LLSSOR].
Figure \[fig:BSSOR\] shows an example of $6^4$ lattice in a node. The block in a node is divided into $2^4$ sub-blocks. Each single core contains two sub-blocks (two sub-blocks with size $3^4$ adjacent in temporal direction). The numbering on the sites is the ordering according to the [*ll*]{}-ordering. The arrows on the links represent the data recurrence direction. The most of the recurrence are limited in each sub-block and there are little data reference on the surface of sub-blocks. Based on this ordering we wrote the forward and backward solver to construct $A_{EE}$ ($A_{OO}$) with the SSOR. The parallelization in a node is achieved by explicit OpenMP threading. 8 threads are invoked and two sub-blocks are assigned to each OpenMP thread. The spatial division is mapped to OpenMP 8 threads, and the temporal division is dedicated to the loop unrolling in a single thread. To resolve the recurrence dependency among threads, we carefully insert explicit [**omp barrier**]{}’s at the sites that require data on other threads.
[c]{}[0.59]{} 
\[fig:BSSOR\]
The sub-block size affects the performance of the SAP. It has been reported that the size of $2^4$ for block still has a gain against the even-odd site ordering on a larger size lattice [@LLSSOR]. Although we apply the [*ll*]{}-ordering to a small lattice (corresponds to a block in a node), we observed that the use of the SSOR with sub-blocking via [*ll*]{}-ordering for $A_{EE}$ still has a gain against the even-odd site ordering.
#### SIMDzation
The [SPARC64$^{\mathrm{TM}}$ VIIIfx ]{}CPU has 256 double precision (64bit) FP registers per core. Each core can issue two independent fused multiply add (FMAD) on paired registers (SIMD) in a cycle resulting in 8 floating operations per cycle. To fully make use of these core infrastructure we have to extract enough independent data-computation stream in the kernel code $D_{EE}$ and $A_{EE}$.
The SIMD FMAD operation can be explicitly invoked using intrinsic functions defined in the C/C++ language provided by Fujitsu. Although our entire program codes were developed in the Fortran90 at first, we totally write the single precision inner BiCGStab solver using the SIMD intrinsic functions in the C language. The details of the intrinsics are not available publicly, while these have one-to-one correspondence to the mnemonic of the [SPARC64$^{\mathrm{TM}}$ VIIIfx ]{} [@SPARCACE]. The linkage between the Fortran90 and the C codes is realised through the `ISO_C_BINDING` feature of the Fortran 2003 which is partly included in the Fujitsu Fortran. We use double precision intrinsic functions for the SIMD computation because the single precision FLOPS performance is identical to that in double precision in this CPU architecture. The data are converted from single precision to double precision (and vice verse) at the data loading (storing) from (to) memory. We employ the $SU(3)$ reconstruction method in the kernel code where the link variables are stored in compressed form (keeps only two-columns of the $SU(3)$ matrix) and the third column is reconstructed using the unitarity condition on the fly.
[r]{}[0.45]{}
\[23pt\]\[0pt\] \[0pt\]\[0pt\][Real/Flop]{}
----------------- ------ ----- -----------------------------
Yes 2232 372 0.1667
No 1896 420 0.2215
\[tab:flopc\]
To further improve the kernel performance we unroll the temporal direction loop, which is the innermost site loop, by three times for $D_{EE}$. For the forward/backward solver in $A_{EE}$ the two independent sub-blocking in temporal direction is used as unrolling. This is based on the technique of the register blocking which hide the data load latency. The unrolling also contains branch elimination by condition merging. This loop unrolling with huge loop body is possible thanks to the many registers of the [SPARC64$^{\mathrm{TM}}$ VIIIfx ]{}.
In table \[tab:flopc\] we summarize the total flop and load/store count per site for the hopping matrix of $D$. These numbers are estimated for bulk sites in a block (not for the surface sites). We use the Dirac representation for $\gamma_{\mu}$ ($\gamma_4$ is diagonal) in the spin projection, and use the chiral representation for the inverse clover term $F(n)$. When we multiply $F(n)$, it is converted to the Dirac representation on the fly. The required byte/flop is 0.667 with the $SU(3)$ reconstruction method in single precision. The theoretical system byte/flop is 64 \[GByte/s\]/128 \[GFlops\] = 0.5. Thus our computation is still limited by the memory bandwidth. The maximum theoretical performance is estimated to be $\sim$ 96 GFlops from the system memory bandwidth of 64 GByte/s. Excluding the redundant flop count caused by the $SU(3)$ reconstruction, the effective performance is $\sim$ 82 GFlops, which is still better than that without the $SU(3)$ reconstruction ($\sim 72$ GFlops). This performance number could be reduced or enhanced by the cache system, site loop control, conditional branch for site location, thread controlling etc. We test and benchmark the $D_{EE}$ and $A_{EE}$ kernels.
[r]{}[0.49]{}
$x_{E} = A_{EE} v_{E}$ and [**MPI\_Irecv**]{} for $w_O = D_{OE} x_{E}$. $w_E = D_{EE} x_E$ for $w_O = D_{OE} x_{E}$. $f_O = v_O - w_O$ $x_{O} = A_{OO} f_O$ and [**MPI\_Irecv**]{} for $f_E = D_{EO} x_{O}$. $f_O = D_{OO} x_O$ $w_O = w_O + f_O$ for $f_E = D_{EO} x_{O}$. $w_E = w_E + f_E$
#### Communication hiding
The internode communication in the SAP kernel $DK$ is arranged in the $D_{EO}$ and $D_{OE}$ kernels. The structure of $D_{EO}$ ($D_{OE}$) is basically organized as follows: ([1]{}) spin projection, link $U^{\dag}_{\mu}(m)$ multiplication and data packing, ([2]{}) sending data, ([3]{}) receiving data, ([4]{}) link $U_{\mu}(n)$ multiplication, spin reconstruction and accumulation. Other computation is possible during the step ([2]{}) and ([3]{}).
We organize the SAP kernel $DK$ to hide the communication time of $D_{EO}$ ($D_{OE}$) behind the computation of $D_{EE}$ ($D_{OO}$) as shown in Alg. \[alg:MatMulDK\]. To hide the communication we employ the non-blocking MPIs: [**MPI\_Isend**]{}, [**MPI\_Irecv**]{}, and [**MPI\_Wait**]{}. The steps ([1]{})-([2]{}) are done at the lines 2 and 7, and the steps ([3]{})-([4]{}) are at the lines 4 and 10. We benchmark the communication performance in the weak-scaling test by comparing the performance of $D_{EE}$ and $DM_{\mathrm{SAP}}$, where the latter contains the internode communication while the former does not.
Results {#seq:Results}
=======
We benchmark the single precision BiCGStab with the SAP preconditioner on the K computer. The lattice sizes benchmarked are $12^3\times 24$, $24^3\times 48$, and $48^3\times 96$. The block size of the SAP is kept fixed at $6^4$ and the local lattice size in a node is $6^3\times 12$. The sub-block size for the SSOR is thus $3^4$. The number of nodes used for the benchmark is 16, 256 and 4096 nodes. The performance is measured using the profiler provided by the Fujitsu’s compiler system.
Figure \[fig:SIMD\] shows the SIMD rate in the instruction executed in benchmarking runs. We achieve 90% SIMD rate for the kernel $D_{EE}$ and $A_{EE}$ by explicitly using the SIMD intrinsic functions. At the solver level it reduces to 85% as it contains internode communication. Figs \[fig:EFF\] and \[fig:FLOPS\] represents the weak-scaling property of the flops performance. The efficiency (Fig. \[fig:EFF\]) is almost at constant for the kernels resulting an ideal weak-scaling (Fig. \[fig:FLOPS\]) from 16 nodes to 4096 nodes.
[r]{}[0.48]{}
The efficiency reaches over 50% for the $D_{EE}$ kernel, while the $A_{EE}$ kernel has lower 35%.
There are two reasons for the lower performance for $A_{EE}$. One is the load imbalance of the computation among the threads in $A_{EE}$. For example, the number of red arrows is more than that of purple in the left panel of figure \[fig:BSSOR\]. This means that the core \#0 has much computation than that of the core \#3. The second is the less computational density in the loop body compared to that of $D_{EE}$. As mentioned in previous section the loop of $D_{EE}$ is unrolled by three times, while $A_{EE}$ is by two with sub-blocking. The forward/backward substitution of $A_{EE}$ only has one sided hopping operation for bulk sites. This also reduces the computational density in the loop body. The theoretical peak efficiency for $D_{EE}$ is estimated be 75%($=$96 \[GFlops\]/128 \[GFlops\]), while the observed 50% efficiency does not reach the theoretical peak efficiency. A reason for this could be the conditional branch to distinguish the sites in the bulk or on the surface of the block. We still need a further investigation on the gap between the theoretical peak efficiency and the observed efficiency.
The performance efficiency of the single precision inner solver is at $\sim$26% in all. We observe an ideal weak-scaling property as shown in Fig. \[fig:FLOPS\]. Note that the measured performance presented in the figures contains redundant floating operations coming from the $SU(3)$ reconstruction and the use of FMAD for the spin projection. We have to roughly multiply 0.8 on the Flops numbers in Fig. \[fig:FLOPS\] to exclude the redundant operations and to obtain the effective performance.
Summary
=======
We have benchmarked the single-precision BiCGStab solver for the $O(a)$-improved Wilson fermion on the K computer. The solver code has been successfully optimized and tuned for the system. We have applied the SSOR with the [*ll*]{}-ordered sub-blocking for the approximate block inverse of the SAP preconditioner to enhance the parallelism for the multi-core architecture. Thanks to the science specific architecture of th K computer we could achieve the ideal weak-scaling performance and $\sim 26\%$ efficiency for the single-precision BiCGStab solver. It partly remains unclear the origin of the gap between the theoretical system performance and the measured performance. We expect a further improvement on the internode communication by using the “Tofu” specific optimization which is not covered in this study.
#### Acknowledgement {#acknowledgement .unnumbered}
This work has been done in the “collaborative research for the performance analysis of large scale simulations on next-generation supercomputers” between RIKEN and University of Tsukuba. We thank the members of the next-generation Technical Computing Unit of Fujitsu for giving us the technical advice and support, and for tuning the computational kernels. Part of the results is obtained by using the K computer at the RIKEN Advanced Institute for Computational Science (Proposal numbers hp120108, hp120153, hp120170, hp120281), T2K-Tsukuba System in Center for Computational Sciences, University of Tsukuba, and a PC-cluster of HPCI strategic program Field 5. This work is supported in part by Grants-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science and Technology (Nos. 22244018, 24540276).
[99]{}
RIKEN Advanced Institute for Computational Science (AICS), \[<http://www.aics.riken.jp/en/>\].
High Performance Computing Infrastructure (HPCI), \[<https://www.hpci-office.jp/index.html>\].
HPCI Strategic Program Field 5, *The origin of matter and the universe*, \[<http://www.jicfus.jp/field5/en/>\].
FUJITSU SCIENTIFIC & TECHNICAL JOURNAL (FSTJ), *The K computer*, 2012-7 (Vol.48, No.3) \[<http://www.fujitsu.com/global/news/publications/periodicals/fstj/>\].
M. Lüscher, Comput. Phys. Commun. [**165**]{} (2005) 199 \[arXiv:hep-lat/0409106\]; JHEP [**0305**]{} (2003) 052 \[hep-lat/0304007\].
J.A. Vogel, Appl. Math. Comput, [**167**]{} (2005) 1004-1025; H. Tadano and T. Sakurai, LSSC’07, Lec. Notes Comput. Sci. [**4818**]{} (2008) 721.
A. Buttari, J. Dongarra, J. Kurzak, P. Luszczek, and S. Tomov, *Using mixed precision for sparse matrix computations to enhance the performance while achieving 64-bit accuracy*, ACM Trans. Math. Soft., [**34**]{} (2008) 1.
S. Aoki [*et al.*]{} \[PACS-CS Collaboration\], Phys. Rev. D [**81**]{} (2010) 074503 \[arXiv:0911.2561 \[hep-lat\]\]; Phys. Rev. D [**79**]{} (2009) 034503 \[arXiv:0807.1661 \[hep-lat\]\].
N. Eicker, W. Bietenholz, A. Frommer, T. Lippert, B. Medeke and K. Schilling, Nucl. Phys. Proc. Suppl. [**73**]{} (1999) 850 \[arXiv:hep-lat/9809038\]; S. Fischer, A. Frommer, U. Glassner, T. Lippert, G. Ritzenhofer and K. Schilling, Comput. Phys. Commun. [**98**]{} (1996) 20 \[arXiv:hep-lat/9602019\].
Fujitsu Limited, “SPARC64$^{\mathrm{TM}}$ VIIIfx Extensions” Version 15, 2010.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Emanuele Olivetti, Thien Bao Nguyen and Paolo Avesani'
date: March 2012
title: The Approximation of the Dissimilarity Projection
---
[10]{}
Maria-Florina Balcan, Avrim Blum, and Nathan Srebro. . , 72(1):89–112, August 2008.
P. J. Basser, J. Mattiello, and D. LeBihan. , 66(1):259–267, January 1994.
Yihua Chen, Eric K. Garcia, Maya R. Gupta, Ali Rahimi, and Luca Cazzanti. . , 10:747–776, March 2009.
E. W. Forgy. . , 21:768–769, 1965.
E. Garyfallidis. . PhD thesis, University of Cambridge, 2012.
Dorit S. Hochbaum and David B. Shmoys. . , 10(2):180–184, May 1985.
Nathan Linial, Eran London, and Yuri Rabinovich. . , 15(2):215–245, June 1995.
Susumu Mori and Peter C. M. van Zijl. . , 15(7-8):468–480, 2002.
Emanuele Olivetti and Paolo Avesani. . In [*Proceedings of the First international conference on Similarity-based pattern recognition*]{}, SIMBAD’11, pages 261–274, Berlin, Heidelberg, 2011. Springer-Verlag.
E. Pekalska, R. Duin, and P. Paclik. . , 39(2):189–208, February 2006.
Elzbieta Pekalska, Pavel Paclik, and Robert P. W. Duin. . , 2:175–211, 2002.
D. Turnbull and C. Elkan. . , 17(4):580–584, April 2005.
Xiaogang Wang, Grimson, and Carl-Fredrik Westin. . , 54(1):290–302, January 2011.
Song Zhang, S. Correia, and D. H. Laidlaw. . , 14(5):1044–1053, 2008.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The correlation properties of the cold system consisting of the electron-hole plasma interacting with the exciton gas are analyzed. It is shown that the homogeneous state of the system is unstable and in the stationary state the densities of the electron-hole plasma and exciton gas are modulated.'
author:
- 'V. S. Babichenko'
title: 'Correlation Properties of the Electron-Hole Plasma Interecting with the Exciton Gas and the Formation of Inhomogeneous State. '
---
The problem of the existence of the Bose condensed state of the exciton gas in semiconductors is the subject of the experimental and theoretical investigations for a long time. However, there is no undoubted proof of the existence of this state until now.
Recently, the properties of the cold exciton gas are investigated intensively, both experimentally and theoretically, in double 2D quantum wells, besides, electrons are localized on one plane of these wells and the holes are localized on the other plane. The state of the cold exciton system being observed in the experiments of different experimental groups [@B1], [@Sn1], [@Sn2], [@Tim], [@B2], [@Sn3], [@Sn4], [@Tim1], [@Tim2], [@Tim3] is nonhomogeneous, and the density of the system in these experiments is modulated in space.
The theory of this phenomenon based on the Turing kinetic mechanism of the instability [@Tur] have been proposed in works [@B2], [@B3]. In this theory the system of nonlinear diffusion equations with sources of particles for electrons, holes and excitons are proposed and solved. The solution of these equations demonstrates the state with the periodically modulated density of the exciton system only if the constant describing the decay of excitons depends on their density. However, in the low density approximation this dependence should be neglected. Moreover, this mechanism has the classical character and does not describe the systems with the quantum coherency, but for the sufficiently small temperatures one can suppose the existence of the coherency in the exciton system [@K1], [@K2], [@K4].
Another theory, based on the supposition of the existence of the attraction between excitons at large distances and the formation of the liquid phase in the low-density exciton system due to this attraction, has been proposed in the work [@Sug]. The supposition of the existence of the liquid phase in the low-density exciton system is baseless. The existence of the attractive part of the interparticle interaction potential in 2D Bose systems results in the collapse of the system to the densities of the order of the squared inverse radius of the repulsive core [@Bab]. However, for these densities excitons can not exist as individual particles, they are destroyed and form the electron-hole liquid. Note that in the usual 3D case, as it was shown in the work [@K1], the scattering amplitude for the isotropic excitons is positive and is equal to the value $a=\frac{13\pi}{3}a_{B}$, where $a_{B}$ is the effective Bohr radius, and just the scattering amplitude, but not the bare potential, determines the properties of the system in the case of small densities. The effective Bohr radius has the form $a_{B}=\frac{\hbar^{2}}{m^{\ast}\left( e^{\ast}\right) ^{2}}$; where $\hbar$ is the Planck constant; $m^{\ast}=\frac{m_{e}m_{h}}{m_{e}+m_{h}}$ is the reduced mass, $m_{e,h}$ are electron (e) and hole (h) masses, which are supposed to be isotropic and of the same order $m_{e}\sim m_{h}$; $e^{\ast}$ is the effective charge in the semiconductor $e^{\ast}=\left( \frac{e^{2}}{\varkappa_{0}}\right) ^{1/2}$; where $e$ is the bare electron charge and $\varkappa_{0}$ is the static dielectric constant of the semiconductor. As the result, the 3D exciton system of small density can exist only in the gas phase. The repulsion for the interaction between excitons in 2D system, where electrons and holes are localized on different planes, is sharply defined than in the usual 3D case due to the geometrical constraint. In this case the formation of the exciton liquid phase is possible only for the densities $n_{ex}\geq
1/a_{ex}^{2}$, where $a_{ex}$ is the size of the exciton. The size of the exciton is equal to the effective Bohr radius $a_{ex}=a_{B}$ in the case of the small distance $l\ $between the planes of the quantum wells $l<<a_{B}$, and $a_{ex}>>a_{B}$ for the large distance between planes when $l>>a_{B}$. If the exciton density is sufficiently large $n_{ex}\geq1/a_{ex}^{2}$ the excitons can not be considered as the individual particles. In this case the electron-hole system represents the strongly interacting electron-hole liquid.
In the present work the Coulomb correlations in the cold electron-hole plasma as well as the quantum effects and the coherency of the exciton subsystem in quantum wells are taken into account. We suppose that the external source is stationary, has the frequency larger than the semiconductor gap $\Delta_{g}$ and creates electrons and holes being localized on different planes. Moreover, the region of the action of the electromagnetic field is supposed to be sufficiently small. In this case the electron-hole plasma being created by the external source propagates along 2D planes, besides, the electrons propagate along one plane and the holes along the other plane. During this propagation the electron-hole plasma loses the energy giving away the energy to the phonon system and comes to the equilibrium state with the small temperature T. Note that the parameters of the stationary state, such that the densities of the electron-hole plasma $n$ and the exciton gas $n_{ex}$, are defined by the equilibrium between incoming rate of the electrons, holes and excitons and outgoing rate of these particles.
Thus, the model being analyzed represents the quasi-equilibrium electron-hole plasma from which the exciton gas is created as the result of the formation of the electron-hole bound states. In the present work we consider the properties of the stationary state of this system, namely, the properties of the electron-hole plasma of the small density $na_{B}^{3}<<1$ interacting with the exciton gas of small density $n_{ex}a_{B}^{3}<<1$.
The small density of the electron-hole plasma and the exciton gas are necessary if we suppose the existence both electron-hole plasma and the exciton gas. In the opposite case of the large density $na_{B}^{3}>>1$ the excitons are destroyed and can not exist as the individual particles. This case is not under consideration in this work. Note that the experiments [@B1], [@Tim], [@B2] demonstrate the existence of the excitons, and this means the existence of small density of the electron-hole plasma in these experiments.
We analyze the correlation properties of the electron-hole plasma of the small density and the interaction of this plasma with the exciton gas. The energy of the ground state of the electron-hole plasma has a minimum value as the function of the density at some density $n_{0}$ such that $n_{0}a_{B}^{2}\sim1$. As a result, the electron-hole plasma tends to form the liquid electron-hole drops and the homogeneous state of the system becomes unstable. The inhomogeneous state of the electron-hole plasma creates the periodic mean field which acts on the exciton gas and makes the density of the exciton system modulated in space.
We suppose that the system consisting of the electron-hole plasma interacting with the exciton gas is in the stationary state with some small temperature T obeying the inequality $T<<\varepsilon_{F}^{\left( e,h\right) }<<E_{b}^{\left( ex\right) }$; where $\varepsilon_{F}^{\left( e,h\right) }$ are the Fermi energies of the electrons and the holes correspondingly $\varepsilon_{F}^{\left( e,h\right) }=\frac{p_{F}^{2}}{2m_{e,h}}$; the Fermi momentum $p_{F}$ in 2D case is expressed by the equality $p_{F}=\left( 2\pi
n\right) ^{1/2}$; the average density $n$ of the electron-hole plasma is equal to the density of each component $n=n_{e}=n_{h}$; $E_{b}^{\left(
ex\right) }$ is the exciton bound energy $E_{b}^{\left( ex\right) }=\frac{\hbar^{2}}{2m^{\ast}a_{B}^{2}}$. Below we use the system of units in which $\hbar=m^{\ast}=e^{\ast}=1$.
Effective action
================
To analyze the properties of the exciton system interacting with the electron-hole plasma the effective action for the slow fields is derived. In this paper the properties of the stationary state of this system is under consideration, but in future the kinetics of the formation of this state is supposed to be analyzed. For this reason the Keldysh-Schwinger technique for the nonequilibrium processes [@K3], [@Schw] in the functional integral formulation (see [@BK], for example) is used for the derivation of the effective action of the slow fields. The generation functional for the considering system can be written as $$Z=\int{\displaystyle\prod\limits_{\alpha}}
D\psi_{\alpha}D\overline{\psi}_{\alpha}\exp\left( iS_{e-h}+i{\displaystyle\oint}
dtd^{2}r\left( \overline{J}_{\alpha}\psi_{\alpha}+\overline{\psi}_{\alpha
}J_{\alpha}\right) \right) \label{Z}$$
The action of the electron-hole plasma in the electron-hole representation has the form $$S_{eh}={\displaystyle\oint}
d^{3}xd^{3}x^{\prime}\left\{
\begin{array}
[c]{c}{\displaystyle\sum\limits_{\alpha=e,h}}
\overline{\psi}_{\alpha}\left( x\right) \left[ i\partial_{t}+\mu_{\alpha
}-\varepsilon_{\alpha}\left( \widehat{\overrightarrow{p}}\right) \right]
\psi_{\alpha}\left( x\right) \delta\left( x-x^{\prime}\right) -\\
-\frac{1}{2}{\displaystyle\sum\limits_{\alpha,\beta}}
\left( -1\right) ^{\alpha+\beta}\overline{\psi}_{\alpha}\left( x\right)
\psi_{\alpha}\left( x\right) \widehat{U}_{C}\overline{\psi}_{\beta}\left(
x^{\prime}\right) \psi_{\beta}\left( x^{\prime}\right) \delta\left(
t-t^{\prime}\right)
\end{array}
\right\} \label{Seh}$$
here $x=\left( t,\overrightarrow{r}\right) $ denotes the collection of time t and space coordinates $\overrightarrow{r}$ in 2D; $\psi_{e}$ and $\psi_{h}$ are fermion (Grassmann) fields of electrons and holes; $\varepsilon_{e}\left(
\widehat{\overrightarrow{p}}\right) $ and $\varepsilon_{h}\left(
\widehat{\overrightarrow{p}}\right) $ are the dispersive laws of the electron and hole bands $\varepsilon_{\alpha}\left( \widehat{\overrightarrow{p}}\right) =\frac{1}{2}\Delta_{g}+\frac{1}{2m_{\alpha}}\widehat{\overrightarrow
{p}}^{2}$, where $\Delta_{g}$ is the semiconductor gap; $\widehat
{\overrightarrow{p}}=-i\overrightarrow{\nabla}$ is the momentum operator; $\widehat{U}_{C}=U_{C}\left( \overrightarrow{r}-\overrightarrow{r}^{\prime
}\right) =\frac{e^{2}}{\varepsilon_{0}\mid\overrightarrow{r}-\overrightarrow
{r}^{\prime}\mid}=\frac{1}{\mid\overrightarrow{r}-\overrightarrow{r}^{\prime
}\mid}$ is the Coulomb interaction. In the Keldysh-Schwinger technique the time variable t changes along the double time contour with return.
The effective action for the electron-hole plasma with the formation of the exciton system have been obtained in the functional-integral technique in [@BKex], [@BKex1]. In this paper we give more simple derivation of the effective action and introduce some simplification transforming to the density functional for the Fermi subsystem. Integrating over the rapid $\psi_{\alpha}$-fields, changing on the scale much smaller than the average distance between the charge carriers $n^{-1/3}$, in the ladder approximation we can obtain the action for the smooth fields in the form $$S_{eh}^{\left( s\right) }={\displaystyle\oint}
d^{3}xd^{3}x^{\prime}\left\{
\begin{array}
[c]{c}{\displaystyle\sum\limits_{\alpha=e,h}}
\overline{\psi}_{\alpha}\left( i\partial_{t}-\xi_{\alpha}\right)
\psi_{\alpha}\delta\left( x-x^{\prime}\right) -\\
-\frac{1}{2}\left( \overline{\psi}_{e}\psi_{e}-\overline{\psi}_{h}\psi
_{h}\right) \widehat{U}_{C}\left( \overline{\psi}_{e}\psi_{e}-\overline
{\psi}_{h}\psi_{h}\right) \delta\left( t-t^{\prime}\right) -\frac{1}{2}{\displaystyle\sum\limits_{\alpha,\beta=e,h}}
\overline{\psi}_{\alpha}\psi_{\alpha}\Gamma_{\alpha\beta}\overline{\psi
}_{\beta}\psi_{\beta}\end{array}
\right\} \label{Sgam}$$
here the fields $\psi_{\alpha}$ are the smooth Fermi fields. The vertex $\Gamma_{\alpha\beta}$ is the sum of the ladder diagrams. The internal Green functions of these diagrams have the large momentums, much larger than $p_{F}$, and the external lines of these diagrams have the small momentums, much smaller than the internal lines. The value of the boundary momentum $\Lambda$ separating the rapid and the smooth fields obeys the inequality $\frac{\hbar
}{a_{B}}>>\Lambda>>p_{F}$. The vertex $\Gamma$ can be considered as independent of the external momentums or frequencies due to the large values of the momentums corresponding to the internal lines of this vertex, on the assumption that this vertex does not contain the pole. The vertexes $\Gamma_{ee}$ and $\Gamma_{hh}$ do not contain the pole parts, but $\Gamma_{eh}$ has the pole part corresponding to the formation of the bound state of the electron and the hole, i.e. the exciton, $$\Gamma_{eh}\left( P,k,k^{\prime}\right) =\left( E-\frac{\overrightarrow
{k}^{2}}{2m^{\ast}}\right) \int\frac{d^{2}\overrightarrow{q}}{\left(
2\pi\right) ^{2}}\left\{ \left[
{\displaystyle\sum\limits_{n}}
\frac{\psi_{n}\left( \overrightarrow{k}\right) \psi_{n}^{\ast}\left(
\overrightarrow{q}\right) }{E-E_{ex}^{\left( n\right) }+i\delta}+\int
\frac{d^{2}\overrightarrow{p}}{\left( 2\pi\right) ^{2}}\frac{\psi
_{\overrightarrow{p}}\left( \overrightarrow{k}\right) \psi_{\overrightarrow
{p}}^{\ast}\left( \overrightarrow{q}\right) }{E-E_{\overrightarrow{p}}+i\delta}\right] U_{C}\left( \overrightarrow{q}-\overrightarrow{k}^{\prime
}\right) \right\} \label{Ga}$$
The wave functions $\psi_{n}\left( \overrightarrow{k}\right) $ are the relative motion wave functions of the discreet part of the spectrum of the electron-hole pair, $\psi_{\overrightarrow{p}}\left( \overrightarrow
{k}\right) $ are the wave functions of the continuous part of the spectrum of the electron-hole pair; the energies $E_{ex}^{\left( n\right) }$ are the spectrum of the bound energies of the exciton; here we denote $E=\Omega
+\mu_{e-h}-\overrightarrow{P}^{2}/2M$, where $\mu_{e-h}=p_{F}^{2}/2m^{\ast}$. The incoming $\overrightarrow{p}_{e}$, $\overrightarrow{p}_{h}$ and outgoing $\overrightarrow{p}_{e}^{\prime}$, $\overrightarrow{p}_{h}^{\prime}$ momentums of the ladder diagrams can be represented in the form $$\begin{aligned}
\overrightarrow{p}_{e} & =\frac{m_{e}}{M}\overrightarrow{P}+\overrightarrow
{k};\text{ \ \ \ \ \ \ }\overrightarrow{p}_{h}=\frac{m_{h}}{M}\overrightarrow
{P}-\overrightarrow{k}\\
\overrightarrow{p}_{e}^{\prime} & =\frac{m_{e}}{M}\overrightarrow
{P}+\overrightarrow{k}^{\prime};\text{ \ \ \ \ \ \ }\overrightarrow{p}_{h}^{\prime}=\frac{m_{h}}{M}\overrightarrow{P}-\overrightarrow{k}^{\prime}$$
where the momentum $\overrightarrow{P}=\overrightarrow{p}_{e}+\overrightarrow
{p}_{h}$ is the exciton momentum, i.e. the total momentum of the electron-hole pair forming the bound state. The wave functions $\psi_{n}\left(
\overrightarrow{r}\right) $ and $\psi_{\overrightarrow{p}}\left(
\overrightarrow{r}\right) $ of the discreet and continuous parts of the internal exciton spectrum obey the Schredinger equation $$\begin{aligned}
\left( -\frac{\overrightarrow{\nabla}^{2}}{2m^{\ast}}-\frac{1}{\mid
\overrightarrow{r}\mid}\right) \psi_{n}\left( \overrightarrow{r}\right) &
=E_{ex}^{\left( n\right) }\psi_{n}\left( \overrightarrow{r}\right) \\
\left( -\frac{\overrightarrow{\nabla}^{2}}{2m^{\ast}}-\frac{1}{\mid
\overrightarrow{r}\mid}\right) \psi_{\overrightarrow{p}}\left(
\overrightarrow{r}\right) & =E_{ex}\left( \overrightarrow{p}\right)
\psi_{\overrightarrow{p}}\left( \overrightarrow{r}\right)\end{aligned}$$
The normalized solution of this equation in 2D for the lower energy state can be written in the momentum representation as $$\psi_{0}\left( \overrightarrow{p}\right) =\frac{4\sqrt{\pi}}{\left(
1+\overrightarrow{p}^{2}\right) ^{3/2}}$$
The normalization of this wave function has the form $\int\frac{d^{2}p}{\left( 2\pi\right) ^{2}}\psi_{0}^{2}\left( \overrightarrow{p}\right)
=1$.
We suppose that the energy $E$ is near the lower bound energy of the exciton $E_{ex}^{\left( 0\right) }$. Separating the pole and the non-pole terms we can write the vertex $\Gamma_{eh}$ as $$\Gamma_{eh}\left( P,\overrightarrow{k},\overrightarrow{k}^{\prime}\right)
=\frac{\left( E_{ex}^{\left( 0\right) }-\overrightarrow{k}^{2}/2m^{\ast
}\right) \left( E_{ex}^{\left( 0\right) }-\left( \overrightarrow
{k}^{\prime}\right) ^{2}/2m^{\ast}\right) \psi_{0}\left( \overrightarrow
{k}\right) \overline{\psi}_{0}\left( \overrightarrow{k}^{\prime}\right)
}{\Omega+\mu_{e-h}-\frac{\overrightarrow{P}^{2}}{2M}-E_{ex}^{\left( 0\right)
}+i\delta}+\Gamma_{eh}^{\left( c\right) } \label{Ga1}$$
where the term $\Gamma_{eh}^{\left( c\right) }$ is the non-pole part of the vertex $\Gamma_{eh}$, which can be considered as a constant of the order of unity $\Gamma_{eh}^{\left( c\right) }\sim1$; the collection of the summarized frequency and the summarized momentum is $P=\left( \Omega
,\overrightarrow{P}\right) $. Due to the smallness of the incoming and outgoing momentums $\mid\overrightarrow{p}_{e,h}\mid,\mid\overrightarrow
{p}_{e,h}^{\prime}\mid<<\hbar/a_{B}=1$ the vertex $\Gamma_{eh}$ can be represented in the form $$\Gamma_{eh}\left( P\right) =\frac{F}{\Omega+\mu_{e-h}-\frac{\overrightarrow
{P}^{2}}{2M}-E_{ex}^{\left( 0\right) }+i\delta}+\Gamma_{eh}^{\left(
c\right) } \label{Ga2}$$
where the constant $F$ in the case of the small incoming and outgoing momentums can be represented as $$F=\left( E_{ex}^{\left( 0\right) }\psi_{0}\left( 0\right) \right) ^{2}$$
Near the pole of the vertex$\ \Gamma_{eh}$ the term $\Gamma_{eh}^{\left(
c\right) }\sim1$ can be neglected. The vertex $\Gamma_{eh}$ can be decoupled by the introduction of the virtual exciton field $b$, and the action of the system can be written in the form $$S_{e,h,ex}^{\left( s\right) }={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}{\displaystyle\sum\limits_{\alpha=e,h}}
\overline{\psi}_{\alpha}\left( i\partial_{t}-\xi_{\alpha}\right)
\psi_{\alpha}-\frac{1}{2}\int d^{2}r^{\prime}\left( \overline{\psi}_{e}\psi_{e}-\overline{\psi}_{h}\psi_{h}\right) \widehat{U}_{C}\left(
\overline{\psi}_{e}\psi_{e}-\overline{\psi}_{h}\psi_{h}\right) +\\
+\overline{b}\left( i\partial_{t}-\xi_{ex}\right) b-\sqrt{F}\left(
\overline{b}\psi_{e}\psi_{h}+\overline{\psi}_{h}\overline{\psi}_{e}b\right)
\end{array}
\right\} \label{Sehex}$$
where $$\xi_{ex}=\frac{\widehat{p}^{2}}{2M}-\mu_{e-h}+E_{ex}^{\left( 0\right) }
\label{ExcSpec}$$ $$E_{ex}^{\left( 0\right) }=-E_{B}$$
$E_{B}$ is the bound energy of the internal exciton ground state. As it will be seen lower, the chemical potential for excitons is connected with $\mu_{e-h}$ as $$\mu_{ex}=\mu_{e-h}-E_{ex}^{\left( 0\right) }-<T_{e,h}^{\left( 0\right) }>
\label{EexcR}$$
at that, the value $<T_{e,h}^{\left( 0\right) }>$ can be written as $$<T_{e,h}^{\left( 0\right) }>=G_{e}^{\left( 0\right) }\left( x-x^{\prime
}\right) G_{h}^{\left( 0\right) }\left( x-x^{\prime}\right) ={\displaystyle\int\limits_{\mid\overrightarrow{p}\mid>\Lambda}}
\frac{d^{2}p}{\left( 2\pi\right) ^{2}}\frac{1}{p^{2}/2m^{\ast}} \label{T0}$$
It is convenient to transfer from the Fermi fields of electron-hole plasma to the density variables, i.e. the density functional with respect to the Fermi fields. This transition can be accomplished by the introduction of the functional $\delta$-function to the statistical sum $$\begin{aligned}
Z & =\int D\psi_{\alpha}D\overline{\psi}_{\alpha}DbD\overline{b}\exp\left(
iS_{e,h,ex}\right) =\\
& =\int D\psi_{\alpha}D\overline{\psi}_{\alpha}DbD\overline{b}Dn_{\alpha}\exp\left( iS_{e,h,ex}\right)
{\displaystyle\prod\limits_{\alpha,x}}
\delta\left[ \overline{\psi}_{\alpha}\psi_{\alpha}-n_{\alpha}\right]\end{aligned}$$
Using the Fourier representation for the functional $\delta$-function $${\displaystyle\prod\limits_{x}}
\delta\left[ \overline{\psi}_{\alpha}\psi_{\alpha}-n_{\alpha}\right] =\int
DV_{\alpha}\exp\left( -i{\displaystyle\oint}
d^{3}xV_{\alpha}\left( x\right) \left[ \overline{\psi}_{\alpha}\psi
_{\alpha}-n_{\alpha}\right] \right)$$
we can rewrite the action in the form $$S_{e,h,ex}\left[ \psi_{\alpha},b,V_{\alpha},n_{\alpha}\right] ={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}\left(
\begin{array}
[c]{cc}\overline{\psi}_{e}, & \psi_{h}\end{array}
\right) \left(
\begin{array}
[c]{cc}i\partial_{t}-\xi_{e}-V_{e} & b\\
\overline{b} & i\partial_{t}-\xi_{h}-V_{h}\end{array}
\right) \left(
\begin{array}
[c]{c}\psi_{e}\\
\overline{\psi}_{h}\end{array}
\right) +\\
+{\displaystyle\sum\limits_{\alpha}}
V_{\alpha}n_{\alpha}-\frac{1}{2}\left( n_{e}-n_{h}\right) \widehat{U}_{C}\left( n_{e}-n_{h}\right) +\overline{b}\left( i\partial_{t}-\xi
_{ex}\right) b
\end{array}
\right\} \label{Sex0}$$
where $$Z=\int D\psi_{\alpha}D\overline{\psi}_{\alpha}DbD\overline{b}Dn_{\alpha
}DV_{\alpha}\exp\left( iS_{e,h,ex}\left[ \psi_{\alpha},b,V_{\alpha
},n_{\alpha}\right] \right) \label{Zex}$$
It is convenient to introduce new variables instead of $n_{\alpha}$ and $V_{\alpha}$, namely, $$\begin{aligned}
n_{e} & =\frac{n+\delta n}{2};\text{ \ \ \ \ \ \ \ }n_{h}=\frac{n-\delta
n}{2}\label{Vn}\\
V_{e} & =V+\delta V;\text{ \ \ \ \ \ \ \ }V_{h}=V-\delta V\nonumber\end{aligned}$$
Note that the fields $\delta n$ and $\delta V$ correspond to the fluctuating violation of the electro-neutrality in the electron-hole plasma. The integrals over the fields $\psi_{\alpha}$ and $\delta n$ in Eq.(\[Zex\]) are Gauss integrals and can be calculated. As a result we obtain the action in the form $$\begin{aligned}
S_{ex}\left[ b,V_{\alpha},n_{\alpha}\right] & =-iSp\ln\left[ \left(
\begin{array}
[c]{cc}i\partial_{t}-\xi_{e}-\left( V+\delta V\right) & b\\
\overline{b} & i\partial_{t}-\xi_{h}-\left( V-\delta V\right)
\end{array}
\right) \right] +\label{Sex}\\
& +{\displaystyle\oint}
d^{3}x\left\{ Vn+\frac{1}{2}\delta V\widehat{U}_{C}^{-1}\delta V+\overline
{b}\left( i\partial_{t}-\xi_{ex}\right) b\right\} \nonumber\end{aligned}$$
The supposition of the low exciton density gives the possibility to expand in powers of the field $b$. Integrating over the fields $\delta V$ and neglecting the powers higher than $b^{4}$ we obtain $$\begin{aligned}
S_{ex}\left[ b,V_{\alpha},n_{\alpha}\right] & =-iSp\ln\left[ \left(
\begin{array}
[c]{cc}i\partial_{t}-\xi_{e}-V & 0\\
0 & i\partial_{t}-\xi_{h}-V
\end{array}
\right) \right] +\label{Sex2}\\
& +iSp\left\{ \frac{1}{2}\left[ \left(
\begin{array}
[c]{cc}0 & b\\
\overline{b} & 0
\end{array}
\right) \widehat{G}^{\left( 0\right) }\left[ V\right] \right]
^{2}\right\} +iSp\left\{ \frac{1}{4}\left[ \left(
\begin{array}
[c]{cc}0 & b\\
\overline{b} & 0
\end{array}
\right) \widehat{G}^{\left( 0\right) }\left[ V\right] \right]
^{4}\right\} -\nonumber\\
& -E_{corr}\left[ V\right] +{\displaystyle\oint}
d^{3}x\left\{ Vn+\overline{b}\left[ i\partial_{t}-\xi_{ex}\right] b\right\}
\nonumber\end{aligned}$$
The functional $E_{corr}\left[ V\right] $ is the correlation energy. This functional is represented as the sum of all simply connected closed diagrams dressed by the internal Coulomb interaction lines and the lines of the external field V. In the case of constant V the correlation energy $E_{corr}\left[ V\right] $ can be represented in the well-known form $$E_{corr}\left[ V\right] ==-i\frac{1}{2}Sp\ln\left[ 1-\Pi\left[ V\right]
\widehat{U}_{C}\right] =i\frac{1}{2}{\displaystyle\int\limits_{0}^{1}}
dg\int\frac{d\omega d^{2}k}{\left( 2\pi\right) ^{3}}\frac{U_{C}\left(
k\right) \Pi\left[ V\right] }{1-gU_{C}\left( k\right) \Pi\left[
V\right] }\label{Ecorr}$$
where $\Pi\left[ V\right] $ is the total polarization operator. The simple transformations of the second and third terms in (\[Sex2\]) give the action $S_{ex}$ in the form $$\begin{aligned}
S_{ex}\left[ V,n\right] & =-iSp\ln\left[ \left(
\begin{array}
[c]{cc}i\partial_{t}-\xi_{e}-V & 0\\
0 & i\partial_{t}-\xi_{h}-V
\end{array}
\right) \right] -E_{corr}\left[ V\right] +\label{Sex3}\\
& +{\displaystyle\oint}
d^{3}x\left\{ Vn+\overline{b}\left[ i\partial_{t}-\xi_{ex}\right]
b-\frac{1}{2}\gamma_{ex}\left( \overline{b}b\right) ^{2}-\gamma
_{ex-e,h}\left( \overline{b}b\right) n\right\} \nonumber\end{aligned}$$
The value $<T_{e,h}^{\left( 0\right) }>$ has the form (\[T0\]). The coupling constant $\gamma_{ex}$ is the exciton-exciton scattering amplitude; and $\gamma_{ex-e,h}=\gamma_{ex-e}+\gamma_{ex-h}$ where $\gamma_{ex-e}$, $\gamma_{ex-h}$ are the exciton-electron and exciton-hole scattering amplitudes, respectively.
The first order expansion of the first term in Eq.(\[Sex3\]) over the field V gives the change $Vn\rightarrow V\left( n-<n>\right) $, where $<n>$ is the average density of the electron-hole plasma in stationary state $<n>=\frac
{\int d^{2}rn\left( t,\overrightarrow{r}\right) }{Vol}$, where $Vol$ is the volume of the system. The second order expansion of the first term in Eq.(\[Sex3\]) in series of the field $V$ and the integration over this field gives $$S_{ex}\left[ b,n\right] ={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}\left[ \frac{1}{2}\left( n-<n>\right) \left( \Pi^{\left( 0\right)
}\right) ^{-1}\left( n-<n>\right) -E_{corr}\left[ n\right] \right] +\\
+\overline{b}\left[ i\partial_{t}+\mu_{ex}-\mu_{e-h}-\xi_{ex}\right]
b-\frac{1}{2}\gamma_{ex}\left( \overline{b}b\right) ^{2}-\gamma
_{ex-e,h}\left( \overline{b}b\right) n
\end{array}
\right\} \label{Seff}$$
where $\Pi^{\left( 0\right) }$ is the polarization operator of the form $\Pi^{\left( 0\right) }\left( x-y\right) =-i{\displaystyle\sum\limits_{\alpha}}
G_{\alpha}^{\left( 0\right) }\left( x-y\right) G_{\alpha}^{\left(
0\right) }\left( y-x\right) $. Expending in series of the field V we suppose the smallness of this field. This supposition is correct when the fluctuations of the density n are small compared with $<n>$. This supposition is correct for the beginning of the density fluctuations growth. The derivation of the effective action in the supposition of the smooth variation of the density n at the scales lager than the average distance between particles without supposition of the smallness of the density fluctuations will be considered in the next paper.
Instability of the system and the density modulation of the electron-hole plasma
================================================================================
Using the effective action for the interacting electron-hole plasma and the exciton gas $S_{ex}\left[ b,n\right] $ Eq.(\[Seff\]) we show that the homogeneous state of this system is unstable, and due to the Coulomb correlations the low density electron-hole plasma should form the inhomogeneous state. The formation of the inhomogeneous state of the low density electron-hole plasma is the result of the tendency of the low density plasma to collapse to the liquid electron-hole drops. The interaction of this plasma with the exciton gas results in the existence of the self-consistent nonhomogeneous periodic field acting on the exciton system as the periodic external field. This periodic field results in the density modulation of the exciton system.
To analyze the effective action for simplicity we put $m=m_{e}=m_{h}=1$. The polarization operator $\Pi^{\left( 0\right) }\left( \omega,\overrightarrow
{k}\right) $ for $\mid\overrightarrow{k}\mid<<p_{F}$ is calculated as $$\Pi^{\left( 0\right) }\left( \omega,\overrightarrow{k}\right) =\left\{
\begin{array}
[c]{c}\frac{m}{2\pi}\frac{\left( kV_{F}\right) ^{2}}{\omega^{2}}\text{
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ for }\omega>>kV_{F}\\
-\frac{1}{\pi}m\left( 1+i\pi\left( \frac{\omega}{kV_{F}}\right) \right)
\text{ \ \ for }\omega<<kV_{F}\end{array}
\right\} \label{Pai}$$
The correlation energy $E_{corr}\left[ n\right] $ for small densities $na_{B}^{2}<<1$ can be calculated as $$E_{corr}\left[ n\right] =-An^{4/3}\label{Ecorr1}$$
where the constant A is estimated as $A\sim1$. The main contribution to the integral for $E_{corr}\left[ n\right] $ (\[Ecorr\]) is given by the momentums $k_{0}\sim n^{1/3}>>p_{F}\sim n^{1/2}$ and the frequencies $\omega_{0}\sim k_{0}^{2}$. Note that the Hartree-Fock energy for this system is estimated as $$E_{H-F}\left[ n\right] \sim-\frac{n^{2}}{p_{F}}\sim-n^{3/2}\label{Ehf}$$
Thus, the contribution of the Hartree-Fock energy can be neglected compared with the correlation energy $E_{H-F}\left( n\right) <<E_{corr}\left(
n\right) $ for $na_{B}^{2}<<1$.
The substitution of the polarization operator $\Pi^{\left( 0\right) }\left(
\omega,\overrightarrow{k}\right) $ Eq.(\[Pai\]) for $\omega>>kV_{F}$ to Eq.(\[Seff\]) gives $$S_{ex}\left[ n\right] ={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}\left[ \left( \Delta n\right) \left[ \frac{1}{2\pi}\frac{\omega^{2}}{\left( kV_{F}\right) ^{2}}\right] \left( \Delta n\right) -E_{corr}\left( n\right) \right] +\\
+\overline{b}\left[ i\partial_{t}-\xi_{ex}\right] b-\frac{1}{2}\gamma
_{ex}\left( \overline{b}b\right) ^{2}-\gamma_{ex-e,h}\left( \overline
{b}b\right) n
\end{array}
\right\}$$
where $\Delta n=n-<n>\ $ and expanding in $\Delta n$ the functional $E_{corr}\left[ n\right] $ we obtain $$S_{ex}\left[ \delta n\right] ={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}\left( \Delta n\right) \left[ \frac{1}{2\pi}\frac{\omega^{2}}{\left(
kV_{F}\right) ^{2}}-\frac{1}{2}E_{corr}^{\prime\prime}\left( <n>\right)
\right] \left( \Delta n\right) +\overline{b}\left[ i\partial_{t}-\xi
_{ex}\right] b-\\
-\frac{1}{2}\gamma_{ex}\left( \overline{b}b\right) ^{2}-\gamma
_{ex-e,h}\left( \overline{b}b\right) <n>-\gamma_{ex-e,h}\left( \overline
{b}b\right) \Delta n
\end{array}
\right\} \label{Sn1}$$
In these expressions the frequency $\omega$ should be considered as $\omega=i\partial_{t}$. The term linear over $\Delta n$, namely, $E_{corr}^{\prime}\left( <n>\right) \Delta n$, can be omitted if the total number of particles is fixed, i.e., $\int\Delta n=0$. The substitution of $E_{corr}^{\prime\prime}\left( <n>\right) $ (\[Ecorr1\]) to (\[Sn1\]) gives $$S_{ex}\left[ \delta n\right] ={\displaystyle\oint}
d^{3}x\left\{
\begin{array}
[c]{c}\Delta n\left[ \frac{1}{2\pi}\frac{\omega^{2}}{\left( kV_{F}\right) ^{2}}+\frac{2}{9}A<n>^{-2/3}\right] \Delta n+\overline{b}\left[ i\partial
_{t}-\xi_{ex}\right] b+\\
-\frac{1}{2}\gamma_{ex}\left( \overline{b}b\right) ^{2}-\gamma
_{ex-e,h}\left( \overline{b}b\right) <n>-\gamma_{ex-e,h}\left( \overline
{b}b\right) \Delta n
\end{array}
\right\} \label{Sn2}$$
From the first term in Eq. (\[Sn2\]) it can be seen that the system is unstable relative to the phonon oscillations of the electron-hole plasma density. The sound velocity of the density oscillations has the imaginary value $\mid c\mid^{2}=\left( \frac{4\pi}{9}A<n>^{-2/3}V_{F}^{2}\right) $ . Note that the sound character of the considering density fluctuations is not surprising, since these fluctuations do not violate the electro-neutrality of the electron-hole plasma. This instability results in the growth of the density fluctuations of the plasma with the increment of the growth $$\zeta=\left( \frac{4\pi}{9}A\right) ^{1/2}<n>^{-1/3}kV_{F}\sim
k<n>^{1/6}>>kV_{F}$$
This increment of the growth corresponds to the frequency which is much larger than $kV_{F}$, $\zeta>>kV_{F}$. Due to this inequality we can use the asymptotic of the polarization operator $\Pi^{\left( 0\right) }\left(
\omega,k\right) \sim\frac{\left( kV_{F}\right) ^{2}}{\omega^{2}}$ being correct for the large frequencies $\omega>>kV_{F}$.
As the result, the stationary state of the electron-hole plasma should be nonhomogeneous, besides, the modulation of the density has the periodic character.
The modulated density of the electron-hole plasma results in the existence of the self-consistent field acting on the exciton system. This field has the form (\[Sn2\]) $$V\left( \overrightarrow{r}\right) =\gamma_{ex-e,h}\Delta n\left(
\overrightarrow{r}\right)$$
Note that this external for the exciton system potential plays the essential role, if the amplitude of this potential is larger than the chemical potential of the exciton gas. This condition can be obeyed for the sufficiently large density of the electron-hole plasma, i.e., for the sufficiently large pump.
At the conclusion we estimate the radius of the character non-homogeneity of the electron-hole plasma from the kinetic consideration [@KKin]. We suppose that the kinetics of the non-homogeneity formation is analogous to the kinetics of the electron-hole liquid drop formation or the phase immiscibility and can be described by the equation $$\partial_{t}\left( \pi R^{2}n_{0}\right) =2\pi RV_{F}n+2\pi RV_{ex}n_{ex}-\pi R^{2}\frac{n_{0}}{\tau} \label{KinDrop}$$
The first term in the right hand side of the equation (\[KinDrop\]) is the incoming term and the second one is the outgoing term; $V_{F}n$ is the current of Fermi particles incoming into the drop of the radius R; $V_{ex}n_{ex}$ is the current of excitons incoming into the drop, $V_{ex}$ is the average velocity of excitons which can be written as $V_{ex}\sim\frac{\sqrt{\gamma_{ex}n_{ex}}}{m}$; n is the density of the electron-hole plasma in the gas phase, $n_{ex}$ is the density of excitons, and $n_{0}$ is the density of the electron-hole liquid within the drop; $\tau$ is the lifetime of the carriers. There are no vapor terms in Eq.(\[KinDrop\]) since T=0, and particles can not overcome work function from the liquid drop. The equilibrium value of the radius R can be found from Eq.(\[KinDrop\]) as $$R_{0}=2\frac{V_{F}n+V_{ex}n_{ex}}{n_{0}}\tau\label{R0}$$
Note that the radiuses larger than $R_{0}$ are not profitable due to the different dependence on R of the incoming and outgoing terms - the first one is the surface term and the second one has the volume character. From the point of view of the phase immiscibility the liquid phase is the electron-hole liquid drops and the gas phase is the mixture of the exciton gas and the electron-hole plasma of the small density.
Finally, the set of drops forms the periodic structure. We suppose that this is the consequence of the homogeneity of the system in average. This periodic structure forms the external periodic field for the exciton gas. If this Bose gas could be considered as the ideal, it should be localized at the minima of the periodic external potential.
[99]{}
L.V. Butov, A.C. Gossard, D.S. Chemla, Nature 418, 751(2002), cond-mat/0204482.
D. Snoke, S. Denev, Y. Liu, L. Pfeiffer, K. West, Nature 418, 754 (2002).
D. Snoke, Science 298, 1368 (2002).
A.V. Larionov, V.B. Timofeev, JETP Lett, 73, 342 (2001)
A.V. Larionov, V.B. Timofeev, P.A. Ni, S.V. Dubonos, I. Hvam, and K. Soerensen, JETP Lett. 75, 570 (2002).
A. V. Gorbunov, V.B. Timofeev, JETP Lett, 84, 390 (2006).
A. V. Gorbunov, V.B. Timofeev, Usp. Fiz. Nauk, 176, 652 (2006)
L.V. Butov, L.S. Levitov, A.V. Mintsev, B.D. Simons, A.C. Gossard, and D.S. Chemla, Phys. Rev. Lett. 92, 117404 (2004), cond-mat/0308117.
L.S. Levitov, B.D. Simons, and L.V. Butov, Phys. Rev. Lett. 94, 176404 (2005), cond-mat/0503628.
D. Snoke, S. Denev, Y. Liu, L. Pfeiffer, and K. West, Solid State Comm. 127, 187 (2003).
R. Rapaport, G. Chen, D. Snoke, S.H. Simon, L. Pfeiffer, K. West, Y. Liu, and S. Denev, cond-mat/0308150, Phys. Rev. Lett. 92, 117408 (2004).
A.M.Turing, Phil. Trans. R. Soc. London, Ser.B 327, 37 (1952).
A.A. Chernyuk, V.I. Sugakov, Phys. Rev. B 74, 085303 (2006), A.A. Chernyuk, V.S. Kopp, V.I. Sugakov, cond-mat/ 0702392v1.
V.S. Babichenko, JETP 64, 612 (1973)
L.V. Keldysh, and A.N. Kozlov, JETP 27, 521 (1968).
L.V. Keldysh, Problems of Theoretical Physics (Nauka, Moscow, 1972), p.433.
L.V. Keldysh, JETP 47, 1515 (1964).
L.V. Keldysh, Bose-Einstein Condensation (Cambridge University Press, 1995), p.246.
J. Schwinger, J. Math. Phys., 2, 407 (1961).
V.S. Babichenko, and A.N. Kozlov, Solid State Comm. 59, 39 (1986).
E.M. Lifshits, L.P. Pitaevskii, Physical Kinetics, (Nauka, Moscow, 1979).
L. V. Keldysh, “Excitones in Semiconductors”, (Nauka, Moscow, 1971).
V. S. Babichenko, M. N. Kiselev, J. Mosc. Phys. Soc., 2, 311 (1993).
V. S. Babichenko, M. N. Kiselev, JETP Lett., 57, 174 (1993).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we present a geometric control law for position and line-of-sight stabilization of the nonholonomic spherical robot actuated by three independent actuators. A simple configuration error function with an appropriately defined transport map is proposed to extract feedforward and proportional-derivative control law. Simulations are provided to validate the controller performance.'
author:
- |
Krishna Chaitanya Kosaraju, Arun D. Mahindrakar, Vijay Muralidharan\
Anup K. Ekbote and Ramkrishna Pasumarthy [^1] [^2]
bibliography:
- 'ref.bib'
title: '**Position and line-of-sight stabilization of spherical robot using feedforward proportional-derivative geometric controller**'
---
Introduction
============
The application of Lie groups in Mechanics has been the subject of interest to the control community as it provides a rich platform for the application of geometric control techniques. The textbook [@bullo], provides comprehensive treatment of geometric methods for mechanical systems defined on manifolds. In [@Jason], the authors present a geometric PD controller for a double-gimbal mechanism that evolves on the torus. An output tracking for aggressive maneuvers involving various flight modes is presented in [@uav] for an unmanned quadrotor. Mechanical systems when subjected to motion constraints, particularly nonholonomic was presented in [@bloch]. In this paper, we consider a nonholonomic mechanical system involving the spherical robot rolling on a horizontal plane. The control design for spherical robot initiated with motion planning and open-loop steering input designs with Euler-angle parameterizations. A few notable examples are [@sph23; @sph13; @sph1]. The study of the geometric properties of spherical robot is a recent interest. A steering control for full state reconfiguration based on the geometry of the sphere was proposed in [@sph40]. Euler-Poincaré equations using a coordinate-free approach were obtained in [@sph3; @sph4; @sph9] for various actuator configurations. Geometric open-loop control algorithms were developed in [@sph3] for steering the spherical robot to the origin. Stabilizing control inputs were designed in [@sph4] using the geometric model of the spherical robot for two independent objectives, a finite-time position stabilization and a finite-time attitude stabilization.
The control laws reported in literature are obtained by observations on the mathematical model of the spherical robot, we intend to identify a control objective which can be accomplished by the currently established tools in geometric control design [@bullo]. The negative result of Brockett [@brockett] for nonholonomic systems rules out asymptotic stabilization to an equilibrium point using smooth geometric control laws. We identify that position and line-of-sight stabilization problem is achievable within the framework of smooth geometric control. The notion of configuration error function and the associated transport map are the necessary prerequisites in applying the geometric tools developed in [@bullo]. In this direction, we propose a novel potential function for the spherical robot model to meet the control objective of position and line-of-sight stabilization. In doing so, we design a transport map that paves the way for the synthesis of a feedforward proportional-derivative geometric control law.
Preliminaries
=============
Let the orientation of a rigid body be denoted by $R(t) \in {SO\left( 3 \right)}$ relative to the reference inertial frame, where ${SO\left( 3 \right)} =\{R|R^\top R=I, det(R)=1 \}$. $\dot{R}(t) \in T_R{SO\left( 3 \right)}$, the tangent space to ${SO\left( 3 \right)}$ at $R$. ${SO\left( 3 \right)}$ is a Lie group and $T_{I}{SO\left( 3 \right)}\simeq {\mathfrak{so}\left( 3 \right)}$ is the Lie algebra of the group, where $I$ is the identity element of the group ${SO\left( 3 \right)}$, ${\mathfrak{so}\left( 3 \right)}$ is a vector space formed by skew-symmetric matrices. Since $so(3)$ is isomorphic to $\mathbb{R}^3$, we denote wedge operation by = \[wedge\] for $x \in {\mathbbm{R}}^3$. Further, $\vee $ be the inverse of the wedge operation and the Lie algebra isomorphism between $({\mathbbm{R}}^3, \times)$ and $({\mathfrak{so}\left( 3 \right)},[\cdot,\;\cdot])$ is \^=v, , v, \^3.\[so3\_R3\_iso\] The dual of ${\mathfrak{so}\left( 3 \right)}$ can be identified with $\mathbb{R}^3$ using the map $\wedge^{\ast}:{\mathfrak{so}\left( 3 \right)}^{\ast}\rightarrow \mathbb{R}^3$. For $\eta\in {\mathfrak{so}\left( 3 \right)}^{\ast}$ and $\hat{\rho}\in {\mathfrak{so}\left( 3 \right)}$, the action of $\eta$ on $\hat{\rho}$ can be identified with the usual inner product $\textquoteleft\cdot\textquoteleft$ in $\mathbb{R}^3$ as $\eta(\hat{\rho})=\wedge^{\ast}(\eta)\cdot \rho$. Let $R,R_1 \in
{SO\left( 3 \right)}$, the left translation map $L_R: {SO\left( 3 \right)}\rightarrow {SO\left( 3 \right)}$ is defined as $L_R(R_1)=RR_1.$ In a similar way, the right translation map $R_R: {SO\left( 3 \right)}\rightarrow
{SO\left( 3 \right)}$ as $R_R(R_1)=R_1R.$ From here unless stated as constant, all the variable are assumed to be time varying. A vector field $X(R) \in T_R{SO\left( 3 \right)}$ is left invariant if $
X(RR_1)=RX(R_1), $ and similarly right invariant if $X(R_1R)=X(R_1)R. $
Body angular velocities of a rigid body are left invariant vector fields, while the spatial angular velocities are right invariant. They can be identified using their velocity at the group identity $I$ of ${SO\left( 3 \right)}$. Let $R\in {SO\left( 3 \right)}$, $X(R)\in T_RSO(3)$, $X(I)=\hat{v}\in T_I{SO\left( 3 \right)}\simeq
{\mathfrak{so}\left( 3 \right)}$. If $v$ is body angular velocity then $X(R)=RX(I)=R\hat{v}$, while if $v$ is spatial angular velocity then $X(R)=X(I)R=\hat{v}R$. The velocity $\dot{R}=R\hat{v}$ at point $R$, which is equivalent to $T_R\dot{R}$, can be defined using the map $T_IL_R:{\mathfrak{so}\left( 3 \right)}\rightarrow T_R{SO\left( 3 \right)}$ as $T_IL_R\hat{v}$. Accordingly, the dual of $T_IL_R$ is the map $(T_I
L_R)^{\ast}:T_R{SO\left( 3 \right)}^{\ast}\rightarrow {\mathfrak{so}\left( 3 \right)}^{\ast}$. Let $\beta_R\in (T_R{SO\left( 3 \right)})^{\ast}$. Then the action of $\beta_R$ on $T_IL_R\hat{\omega}$ can identified with the inner product $<\cdot,\cdot>_{{\text{Tr}}}$ by $<(T_IL_R)^{\ast}\beta_R,\hat{\omega}>_{{\text{Tr}}}$, where $<\cdot ,
\cdot>_{{\text{Tr}}}$ on $\mathbb{R}^{n \times n}$ is defined as $<A,B>_{{\text{Tr}}} =\frac{1}{2}{\text{Tr}}(A^{\top}B)$ for $A,B \in
\mathbb{R}^{n \times n}$.
The Riemannian metric $\mathbb{G}(R): T_R{SO\left( 3 \right)}\times T_R{SO\left( 3 \right)}
\rightarrow \mathbb{R}$, a $(0,2)-$tensor on ${SO\left( 3 \right)}$ defined as $\mathbb{G}(R)(X(R),Y(R))=X(R)^{\top}\mathbb{G}(R)Y(R)$ is left invariant if (R)(X(R),Y(R))=(R(I)R\^[-1]{})(X(R),Y(R))where $X(R),\;Y(R)\in T_RSO(3)$. Therefore it can be seen that for left invariant vector fields $X(R), Y(R)$, &&(R)(X(R),Y(R))\
&&=(R(I)R\^)(X(R),Y(R))\
&&=(R(I)R\^)(RX(I),RY(I))\
&&=(R\^(R(I)R\^)R)(X(I),Y(I))\
&&=(I)(X(I),Y(I)) \[left invariant Rmetric\] which is a constant. Since $X(I), Y(I) \in T_I{SO\left( 3 \right)}\simeq {\mathfrak{so}\left( 3 \right)}$, $J
\deff \mathbb{G}(I)$, a $(0,2)-$tensor on ${\mathfrak{so}\left( 3 \right)}$.
For $\hat{w}\in {\mathfrak{so}\left( 3 \right)}$ the adjoint map ${\operatorname{Ad}}: {SO\left( 3 \right)} \times {\mathfrak{so}\left( 3 \right)}\rightarrow {\mathfrak{so}\left( 3 \right)}$ is defined as \_[R]{}()=RR\^ =. The following general facts involving matrix operations will be useful. For $A, B, C \in \mathbb{R}^{n \times n}$, we denote the trace of $A$ as ${\text{Tr}}(A)$, the symmetric component of $A$ by $sym(A)=\frac{A+A^{\top}}{2}$ and the skew-symmetric component as $skew(A)=\frac{A-A^{\top}}{2}$ and if $A=A^{\top}$, $B=-B^{\top}$ then ${\text{Tr}}(AB)=0$. For $a,b \in \mathbb{R}^3$, ${\text{Tr}}(\hat{a}\hat{b})=-2(a^{\top}b)$. It then follows that
[lcl]{} (C )&=& ((sym(C)+skew(C)))\
&=& ((sym(C)))+ (skew(C)))\
&=& 0 + (skew(C)))\
&=& -2 ((skew(C))\^ a)
Therefore $<\hat{a},\hat{b}>_{{\text{Tr}}} = a\cdot b$.
Modeling of spherical robot
===========================
![Schematic of the spherical robot[]{data-label="schematic"}](spherical1){width=".6\linewidth"}
The spherical robot schematic shown in Figure \[schematic\] consists of a spherical shell of radius $r$ and mass $m$ moving in a horizontal plane. The center-of-mass of the robot is assumed to coincide with the geometric center. The position coordinates of the spherical robot are denoted by $(x,y)$, which are the coordinates of the point $O_1$ with respect to $O$. Let $J=\text{diag}(J_1,J_2,J_3) \in\mathbb{R}^{3\times 3}$ be the moment-of-inertia matrix of the robot with respect to the body frame centered at $O_2$. We make the following assumption.
The principal moments of inertia satisfy $0<J_1<J_2<J_3$. \[assum\]
The sphere has three independent torques acting on the body-coordinate frame. The orientation of body frame $(X_b,Y_b,Z_b)$ of the robot with respect to an inertial frame $(X_i,Y_i,Z_i)$ is given by a matrix $R\in \mathrm{SO}(3)$. The no-slip constraints are given by v== =r Re\_3, \[cons1\] where, $\omega\in \mathbb{R}^3$ denotes the body angular velocity and $\Omega\in \mathbb{R}^3$ is the spatial angular velocity of the robot. Denoting the rows of $R$ by $r_1, r_2, r_3$, the kinematics of the spherical robot is given by
[lcl]{} &=&r(r\_2)\
&=&-r(r\_1)\
&=&R.
\[kin\] Let $X,Y \in T_R{SO\left( 3 \right)}$, an Levi-Civita affine connection on ${SO\left( 3 \right)}$ is left invariant if it satisfies \_[T\_IL\_RX(I)]{}T\_IL\_RY(I)&=&T\_IL\_R\_[X(I)]{}Y(I) \[leftic\]for all $R \in {SO\left( 3 \right)}$ and let $\{e_1, e_2,
e_3\}$ span ${\mathbbm{R}}^3$. Since ${\mathbbm{R}}^3$ is naturally isomorphic to ${\mathfrak{so}\left( 3 \right)}$, it implies that $\mathrm{span}\{\hat{e}_1, \hat{e}_2,
\hat{e}_3\}= {\mathfrak{so}\left( 3 \right)}$. It then follows for $X^i, Y^i \in {\mathbbm{R}}$ we define $X(R)=X^i\hat{e}_i(R), Y(R)=Y^i\hat{e}_i(R)$ and can be simplified as follows
[lcl]{} &&\_[X(R)]{}Y(R)\
&&=T\_IL\_R\_[X(I)]{}Y(I)\
&&=R.(\_[X\^i\_i(I)]{}Y\^i\_j(I))\
&&=R.( DY(I).X(I) + X\^iY\^j\_[\_i(I)]{}\_j(I))
\[lic1\] where $DY$ is Jacobian of $Y$. From we see that $\mathbb{G}(I)\simeq
J$ represents an inner product on ${\mathfrak{so}\left( 3 \right)}$, and $G(I)(\hat{e}_i,\hat{e}_j)$ has a constant value which renders ${\stackrel{g}{\nabla}}:so(3) \times so(3)\mapsto so(3)$ a bilinear map. It now follows as &&\_[X(R)]{}Y(R)\
&&=T\_IL\_R(DY(I).X(I)+X\^iY\^j\_[\_i]{}\_j)\
&&=T\_IL\_R(DY(I).X(I)+\_[X\^i\_i]{}Y\^j\_j)\
&&=T\_IL\_R(DY(I).X(I)+\_[X(I)]{}Y(I)) \[lic\] In , we observe that $X(I)$, $Y(I)$ and ${\stackrel{g}{\nabla}}_{X(I)}Y(I) \in {\mathfrak{so}\left( 3 \right)}$. By letting $\hat{\omega}=\omega^i\hat{e}_i$, $\dot{R}=T_IL_{R}\hat{\omega}(t)=\omega^i\hat{e}_i(R)$, where $\omega^i\in {\mathbbm{R}}$ also known as pseudo velocities. Let $\hat{\tau}\in ({\mathfrak{so}\left( 3 \right)})^{\ast}$ be the covector representing the external torque acting on the robot. Next, the covariant derivative of $\dot{R}$ is
[lcl]{} \_&=&R(+\_)\
&=&T\_IL\_R J\^[-1]{}.
\[cov\_der\_R\] From , we obtain the well-known attitude dynamics governed by Euler-Poincaré equations of motion &=& -J\^[-1]{} (J ) + J\^[-1]{}\[dyn\] where $\tau\in {\mathbbm{R}}^3$ is the external torque about the body-axis of the robot.
Position and line-of-sight stabilizing controller
=================================================
Without loss of generality we assume that the desired position of the robot is the origin and the line-of-sight is $Z_b$. The control objective is to stabilize the position of the robot to the origin and the line-of-sight (fixed to the body) $Z_b$ to coincide with the $Z_i$-axis of the inertial frame. In other words, the objective is to stabilize the closed loop system to submanifold $E=\{(x,y,R,\omega)\in {\mathbbm{R}}^2\times {SO\left( 3 \right)}\times {\mathbbm{R}}^3 :x=0$, $y=0$ and $\omega= R^\top e_3\}$. We note that $\omega= R^\top e_3 \Rightarrow \dot{\omega}=0$.
Before we proceed to derive the control to meet the aforementioned objective, consider the configuration error function $\psi:\mathbb{R}^2\rightarrow \mathbb{R}$ (x,y)=k\_p(x\^2+y\^2), k\_p>0 . Using $\psi$, the position of the robot can be stabilized to the origin of the $(X,Y)$ plane. The controller synthesis can proceed as follows.
The derivative of $\psi(x,y)$ with respect to time along the trajectories of is given by,
[lcl]{} (x,y)&=&k\_p(x+y)\
&=&k\_p(xr\_2-yr\_1)
\[error\_function\_posi\] Equation can be rewritten as
[lcl]{} (x,y)&=&k\_p(x+y)\
&=&k\_pr(x(r\_2-r\_2r\_3)-y(r\_1-r\_1r\_3))\
&=&k\_pr(xr\_2(-r\_3)-yr\_1(-r\_3))\
&=&k\_pr(x r\_2-y r\_1)(-r\_3)
which implies that $\frac{d}{dt}\psi=d\psi e_{\omega}$, where $e_{\omega}\deff (\omega-r_3)$ is the velocity error. Hence the error function $\psi$ is compatible with $e_{\omega}$. If $\omega_d=r_3$, then $\Omega_d=e_3$, where the subscript $d$ refers to the desired values.
The right transport map ${\cal T}: {SO\left( 3 \right)} \times T_{R_d}{SO\left( 3 \right)}
\rightarrow T_{R}{SO\left( 3 \right)}\times {SO\left( 3 \right)}$ is defined as (R,R\_d)(\_d)= \_dR\_d\^R. \[rtm\] Here, $\widehat{\Omega}_d=\dot{R}_dR_d^\top$ and $R_d$ satisfies $R_de_3=e_3$. Next, we define the velocity error using the transport map ${\cal T}$.
[lcl]{} [T]{}(\_d)&=&\_d R\
&=&RAd\_[R\^]{}\_d.
The following derivatives are useful in deriving the covariant derivative of right transport map. For $\hat{v}\in {\mathfrak{so}\left( 3 \right)}$, \_[R]{} &=& R R\^\
&=& R(R\^-R\^)R\^\
&=& R( - )R\^\
&=& R\[ ,\]R\^\
&=& \_[R]{}\[, \] and $\frac{d}{dt}{\operatorname{Ad}}_{R^\top R_d}\hat{\omega}_d$ can be expressed as
[lcl]{}&=& ((R\^R\_d))\_d (R\^\_dR)+(R\^R\_d)\_d((R\^\_dR))\
&&+\_[R\^R\_d]{}\_d\
&=& ((\_[R\^R\_d]{}\_d )(R\^)-(R\^)(\_[R\^R\_d]{}\_d )\
&&+\_[R\^R\_d]{}\_d\
&=& +\_[R\^R\_d]{}\_d.
Thus, the covariant derivative of the right transport map\
${\stackrel{\mathbb{G}}{\nabla}}_{\dot{R}} {\cal T}(\dot{R}_d)$ is &=&\_RAd\_[R\^]{}\_d\
&=&R(Ad\_[R\^]{}\_d+\_[ ]{}Ad\_[R\^]{}\_d)\
&=&R(+Ad\_[R\^]{}\_d+\_Ad\_[R\^]{}\_d)\
&=&R(+\_Ad\_[R\^]{}\_d)\
&=&R \_[ff]{} \[ff\] The last step follows by noting that ${\Omega}_d=e_3$.
We next present the feedforward and proportional-derivative controller in $\mathbb{R}^3$. For $v,\omega \in \mathbb{R}^3$, the following holds (\_)\^ &=& (v)+J\^[-1]{}(v J-Jv ) and from it follows (-)=\_[[( 3 )]{}]{} &=& \_[\^3]{}=. Thus $f_{ff}$ in and $f_{pd}$ can be written as
[lcl]{} f\_[ff]{}&=&R\^e\_3+ (R\^e\_3.\
&&.+J\^[-1]{}(J R\^e\_3-JR\^e\_3))\
\[ff\_control\]
[lcl]{} f\_[pd]{}&=&-J\^[-1]{}(k\_pd+k\_v e\_[ ]{})\
&=&-J\^[-1]{}(k\_p r R\^(x e\_2-y e\_1)\
&&+k\_v (-R\^e\_3)).
\[pd\_control\] With $\tau=J(f_{ff}+f_{pd})$, the closed-loop dynamics , and is &=& -J\^[-1]{} (J ) + f\_[ff]{}+f\_[pd]{}.\[closed\_loop\_dyn\]
Consider a spherical robot satisfying assumption \[assum\]. Then, the closed-loop system is asymptotically stable with respect to $(x,y, R^\top e_3)$ uniformly in $\omega$.
Let $e_R=R\hat{e}_w$, Consider the candidate Lyapunov function V&=&(R)(e\_R,e\_R)+(x,y)\
&=&(R)(R\_,R\_)+(x,y)\
&=&(I)(\_(I),\_(I))+(x,y). The derivative of $V$ with respect to time along the trajectories of the closed-loop system is &=&(I)(\_(I),\_\_(I))+(x,y)\
&=&(I)(\_(I),\_(-Ad\_[R\^]{}\_3)(I))+\
&=&G(I)(\_(I),(\_-\_Ad\_[R\^]{}\_3)(I))+\
&=&J(e\_,(+\_-\_[ff]{})\^)+\
&=&J(e\_,f\_[pd]{})+\
&=&J(e\_,-J\^[-1]{}(d+k\_v e\_))+\
&=&I(e\_,-d-k\_v e\_)+\
&=&-k\_ve\_\^e\_-e\_\^d+\
&=&-k\_ve\_\^e\_0. Let $L\deff \{(x,y,R,\omega)\in {\mathbbm{R}}^2\times {SO\left( 3 \right)}\times
{\mathbbm{R}}^3:V(x,y,R,\omega)\le c, c>0\}$ is compact, connected and contains $E$. Consider the residual set $S\deff\{(x,y,R,\omega)\in
L:\dot{V}=0\}$. Let $(x,y,R,\omega)\in S \implies
\omega=R^\top e_3,\dot{\omega}=0$. Since $r_1$ and $r_2$ are independent, from it follows that $k_p(xr_2-yr_1)=0$ if and only if $x=0$ and $y=0$. Thus the largest invariant set in $S$ is $E$. Thus, by LaSalle’s invariance principle, all trajectories originating in $L$ approach $E$ asymptotically.
Thus the controller stabilizes the robot to the origin of the $(X,Y)$ plane at which the robot spins about its local vertical axis ($Z_b$-axis) at a constant angular velocity.
SIMULATIONS
===========
The system parameters used for simulation is $r=0.4\
\mathrm{m},J=\mathrm{diag}(0.3,0.4,0.5)\ \mathrm{kg} \mathrm{m}^2$. The control gains in are chosen as $k_p=5,k_v=1$. The time-response of the closed-loop with the initial condition $x(0)=4\ \mathrm{m},y(0)=3\ \mathrm{m},
R(0)=\left[\begin{array}{ccc}
1 & 0 & 0 \\
0 & \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \\
0 & \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}}
\end{array} \right],
\omega (0) = (0,0,0)\ \mathrm{rad/s} $ is shown in Figure \[fig:dynstab\] and the $(x,y)$ trajectory is shown in Figure \[fig:dynstab2\].
![Time-response of attitude dynamics[]{data-label="fig:dynstab"}](ic1rwt)
![$(x,y)$ trajectory[]{data-label="fig:dynstab2"}](ic1xy)
The simulation is repeated with $R(0)=\left[\begin{array}{ccc}
1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & -1
\end{array} \right] $ while all other initial condition remaining the same. The time-response is shown in Figure \[fig:dynstab3\] and the $(x,y)$ trajectory is shown in Figure \[fig:dynstab4\].
![Time-response of attitude dynamics[]{data-label="fig:dynstab3"}](ic2rwt)
![$(x,y)$ trajectory[]{data-label="fig:dynstab4"}](ic2xy)
A consequence of the control law is the regulation of $\Omega$ to $e_3$, which implies a) $\omega=e_3,r_3=e_3$ as seen in Figure \[fig:dynstab\] or b) $\omega=-e_3,r_3=-e_3$ as seen in Figure \[fig:dynstab3\].
Conclusions
===========
In this paper we have presented a smooth geometric controller to asymptotically stabilize the system to a smooth submanifold. This results in the robot reaching the origin of the plane while the robot spins with constant angular velocity about its local spin-axis, which by design is the body $Z_b$-axis coincident with the inertial $Z_i$-axis. This control strategy can be used in line-of-sight application for payload pointing, such as a camera mounted inside the sphere.
[^1]: Chaitanya, Vijay and Anup are graduate students in the Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai-600036, India. [[email protected], m\_vijay\[email protected], [email protected]]{}
[^2]: Arun and Ramkrishna are with the Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai-600036, India [arun\_dm,[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'J. Milli'
- 'P. Hibon'
- 'V. Christiaens'
- 'É. Choquet'
- 'M. Bonnefoy'
- 'G. M. Kennedy'
- 'M. C. Wyatt'
- 'O. Absil'
- 'C. A. Gómez González'
- 'C. del Burgo'
- 'L. Matrà'
- 'J.-C. Augereau'
- 'A. Boccaletti'
- 'C. Delacroix'
- 'S. Ertel'
- 'W. R. F. Dent'
- 'P. Forsberg'
- 'T. Fusco'
- 'J. H. Girard'
- 'S. Habraken'
- 'E. Huby'
- 'M. Karlsson'
- 'A.-M. Lagrange'
- 'D. Mawet'
- 'D. Mouillet'
- 'M. Perrin'
- 'C. Pinte'
- 'L. Pueyo'
- 'C. Reyes'
- 'R. Soummer'
- 'J. Surdej'
- 'Y. Tarricq'
- 'Z. Wahhaj'
bibliography:
- 'biblio.bib'
date: 'Received: 15 October 2016; accepted 28 November 2016 '
title: 'Discovery of a low-mass companion inside the debris ring surrounding the F5V star HD206893'
---
[Uncovering the ingredients and the architecture of planetary systems is a very active field of research that has fuelled many new theories on giant planet formation, migration, composition, and interaction with the circumstellar environment. We aim at discovering and studying new such systems, to further expand our knowledge of how low-mass companions form and evolve.]{} [We obtained high-contrast H-band images of the circumstellar environment of the F5V star HD206893, known to host a debris disc never detected in scattered light. These observations are part of the SPHERE High Angular Resolution Debris Disc Survey (SHARDDS) using the InfraRed Dual-band Imager and Spectrograph (IRDIS) installed on VLT/SPHERE.]{} [We report the detection of a source with a contrast of $3.6\times 10^{-5}$ in the H-band, orbiting at a projected separation of 270 milliarcsecond or 10au, corresponding to a mass in the range 24 to 73[$M_{\text{Jup}}$]{} for an age of the system in the range 0.2 to 2 Gyr. The detection was confirmed ten months later with VLT/NaCo, ruling out a background object with no proper motion. A faint extended emission compatible with the disc scattered light signal is also observed.]{} [The detection of a low-mass companion inside a massive debris disc makes this system an analog of other young planetary systems such as $\beta$ Pictoris, HR8799 or HD95086 and requires now further characterisation of both components to understand their interactions.]{}
Introduction
============
Through direct imaging, instruments fed with adaptive optics (AO) have enabled the detection and characterisation of a few tens of low-mass companions, either giant planets (hereafter GP) or brown dwarfs (BD), probing a parameter space in the mass vs orbital radius still inaccessible with other indirect techniques such as radial velocities or transits. The direct detection of the thermal emission of such substellar objects brings precious information for understanding their formation mechanisms and physical properties [see @Bowler2016 for a recent review]. In addition, many of the GP/BD systems discovered in high-contrast imaging are associated to a debris disc, generally detected through its infrared or submillimetre emission (e.g. HR8799, @Marois2006; HD95086, @Rameau2013_discovery; HR3549 @Mawet2015; HR 2562, @Konopacky2016). In only three cases, this disc was also resolved in scattered light ($\beta$ Pictoris, HD106906 and Fomalhaut), enabling to study the interactions with the companion.
This letter presents the discovery of a low-mass BD in orbit around the nearby F5V star HD206893 located at $38.3\pm0.8$ pc (see details in Table \[tab\_prop\]). The star is known to host a debris disc detected through its large infrared excess [$L_{dust}/L_\star=2.3 \times 10^{-4}$, @Moor2006], characterised through its Spectral Energy Distribution (SED) with Spitzer/IRS-MIPS [@Chen2014], and marginally resolved with Herschel/PACS (as detailed in this letter). This study also presents the putative scattered light signal of the disc, at low signal-to-noise (S/N) due to the faintness of its emission. The age of the system is not well constrained as it is not known to belong to a moving group. @Zuckerman2004 estimated an age of 200 Myr based on X ray, radial velocity and proper motion measurements. This age is also inferred by @Holmberg2009 with an upper limit of 1.2 Gyr from Padova stellar evolution models. More recently, @David2015 suggest a median age of 2.1 Gyr using a Strömgren photometry fit to stellar atmosphere models in a Bayesian framework while @Pace2013 derives $860 \pm 710$ Myr based on the chromospheric activity calibrated against the Geneva-Copenhagen survey.
Observation and data reduction {#sec_data_red}
==============================
The companion was first detected with VLT/SPHERE in 2015. HST/NICMOS archival data from 2007 showed that the companion cannot be a background object without proper motion and data from VLT/NaCo redetected the object in 2016.
SPHERE
------
The SPHERE High Angular Resolution Debris Disc Survey (SHARDDS) is a high resolution imaging survey aimed at resolving and characterising new debris discs never detected in scattered light [PI: Milli, 096.C-0388, 097.C-0394, see also @Wahhaj2016]. This programme is a search for discs around stars within 100pc having an infrared excess greater than $10^{-4}$, with the IRDIS subsystem [@Dohlen2008] in broad band H and the apodised Lyot coronagraph of diameter 185mas. Each target is observed in pupil-stabilised mode to allow angular differential imaging [ADI, @Marois2006]. On 5 October 2015, we observed the star HD206893. Over the 40min effective on-source integration time, we obtained $50^\circ$ field rotation. The atmospheric conditions were average with a mean seeing of 0.9" and a coherence time of 2.8ms, resulting in a Strehl ratio of 85% in the H band. The raw frames were sky-subtracted, flat-fielded and bad-pixel corrected using the SPHERE Data Reduction and Handling pipeline [@Pavlov2008], resulting in a temporal cube of 576 frames with individual integration time 4s. The frames were thereafter re-centred using the four satellite spots imprinted in the image during the centring sequence obtained before and after the 576 frames. With broad-band filters, these satellite spots are elongated. We fitted a 2D Gaussian to each spot and evaluated the star location as the intersection of the two lines joining the centres of opposite satellite spots, as explained in @Wertz2016, which yields an absolute centring accuracy of 0.2px or 2.5mas. The individual frames of the cube were not re-centred relative to one another because an active centring using the SPHERE differential tip-tilt sensor is dealing with this to an accuracy smaller than what can be obtained from an individual frame-to-frame recentring [@Wertz2016]. We reduced the images using the principal component analysis algorithm [PCA, @Soummer2012], as implemented in the Vortex Image Processing pipeline (VIP[^1], @Gomez2017, Fig. \[fig\_IRDIS\]). In this algorithm, the only free parameter is the number of modes removed. We detect a point source with an S/N of 14 (Fig. \[fig\_IRDIS\]) at a projected separation of 270.4$\pm 2.6$ mas or 10.4 au.
NICMOS {#sec_data_red_nicmos}
------
HD206893 was observed on 12 June 2007 with the NICMOS instrument on the Hubble Space Telescope (HST). We re-analysed the data (see Appendix \[app\_nicmos\]) and do not detect any point source.
NaCo
----
HD206893 was observed with VLT/NaCo on 8 August 2016, taking advantage of its AGPM coronagraph [e.g. @Mawet2005; @Mawet2013]. This observation was part of programme 095.C-0937(B) (PI: O. Absil). HD206893 was observed in pupil-stabilised mode using the 27.1 mas/pixel plate scale and the [$\text{L}^\prime$]{}-band filter (3.8). The seeing (0.7-0.8) and the coherence time ($\sim$5 ms) were stable throughout the observation. A total of 90 science data cubes of 0.3s (DIT) exposure and 200 (NDIT) frames were obtained, corresponding to 1h30 on-source and a $107^\circ$ total parallactic angle variation. Sky data cubes were obtained every 10-12min. The star was carefully re-centred behind the coronagraph after each sky observation to a $\sim$ 0.2-0.3 pixel accuracy. Four data cubes were also obtained with a shorter exposure time and the star offset from the coronagraph centre (but still behind the AGPM substrate), to obtain unsaturated PSF images. These data cubes were used for photometric calibration and to generate fake companions.\
After standard calibrations (sky subtraction, flat-fielding and bad pixel correction), the frames were re-centred by fitting a negative gaussian to the AGPM central hole as done in @Absil2013. The frame selection process, essential to reach the best contrast, kept the 12879 most correlated frames and with the lowest level of residual speckle noise out of the 18000 original frames. These 12879 frames were binned four by four, to yield a final ADI cube of 3219 frames. The final cube was reduced using algorithms based on the PCA implemented in VIP, as shown in Fig. \[fig\_NaCo\] with seven principal components removed, which was found to optimise the companion S/N (approximately six).
![SPHERE H-band coronagraphic image reduced with ADI + PCA, showing the detection of the companion HD206893B at 270 mas with a S/N of 14.[]{data-label="fig_IRDIS"}](SPHERE_HD206893_only.pdf){width="0.9\hsize"}
Analysis {#sec_analysis}
========
Detection of the companion HD 206893 B
--------------------------------------
The astrometry and photometry of the point source detected in the SPHERE and NaCo data set are described in Table \[tab\_prop\]. The ADI algorithms affect the astrometry and photometry of a detected companion. A robust way to estimate them is to use the negative fake companion algorithm [hereafter NEGFC; e.g. @Lagrange2010]. For both SPHERE and NaCo, we used NEGFC as implemented in VIP [@Wertz2016] with the exploration of the three parameters (radial separation, $PA$ and contrast) performed with a simplex algorithm minimising the residual standard deviation in an aperture at the location of the companion. The uncertainties given in Table \[tab\_prop\] combine that on the instrument (plate-scale, north alignement, filter transmission), the centring uncertainty and the measurement uncertainty due to the presence of speckle noise, as detailed in @Wertz2016 for SPHERE and @Absil2013 for NaCo. We used a true north offset of $-1.75\pm0.08^\circ$ for SPHERE [@Maire2016] and re-calibrated the NaCo true north against that of SPHERE using the common astrometric field 47 Tuc observed in September 2016 which yielded a value of $0.58\pm0.10^\circ$. The measurement uncertainty was computed by injecting fake companions at the separation of the point source and various azimuths, and by retrieving their astrometry and photometry with the NEGFC algorithm. It dominates the NaCo error budget. For SPHERE, the budget of error is dominated by the 2.5mas conservatively attributed to the centring accuracy. The contrast curves obtained from SPHERE, NaCo and NICMOS are shown in Fig. \[contrast\]. They take into account the penalty term coming from small-sample statistics at small separations [@Mawet2014].
This target is at high galactic latitude ($-44^\circ$) and therefore a background contamination within 0.3of the star is unlikely. To confirm the object is truly bound, we computed its expected position if it was a background object with no significant proper motion for the date corresponding to the NICMOS and NaCo data. We find a separation and PA of ($0.995\arcsec \pm 0.006$; $85.3^\circ \pm 0.58^\circ$) for the NICMOS epoch in 2007 and ($0.174\arcsec \pm 0.003\arcsec$ ; $61.0^\circ \pm 0.6^\circ$) for the NaCo epoch in 2016 also shown in Fig. \[fig\_NaCo\] right, using a star proper motion of 94.2 mas/yr at PA $89.9^\circ$. In the NICMOS data, no point source is detected at high confidence level at the position where the candidate would have been in 2007 assuming it is a background star. Fig. \[contrast\] shows the 5$\sigma$ radial detection limit measured on the combined image. We also repeated the same processing steps described in Section \[sec\_data\_red\_nicmos\] after injecting a synthetic NICMOS PSF in the raw data at the background star position and at the $3.6\times10^{-5}$ contrast measured on our candidate in the SPHERE data. The injected point source is detected at 7$\sigma$ as shown in Fig. \[fig\_NICMOS\], demonstrating that the candidate found with SPHERE would have been detected at a high confidence level, if it were a background object. In addition, the star was also observed in two other exoplanet surveys: the International Deep Planet Survey [IDPS, @Galicher2016] with the Gemini North/NIRI instrument and the Gemini Planet-finding campaign with the Subaru/NICI instrument [@Wahhaj2013]. No detection is reported as their discovery spaces start from 0.3and 0.5respectively (Fig. \[contrast\]) but they would have been sensitive to a background object with the same magnitude at 1.
In the NaCo data, the position of the companion is clearly not compatible within error bars with a background object (Fig. \[fig\_NaCo\] right). We interpret the changes in projected separation and position angle between the IRDIS and NaCo data as due to the orbital motion of the companion (see section \[sec\_orbital\_motion\]). We can thus confidently assert that this object is bound to HD206893.
Companion physical properties
-----------------------------
The companion HD206893B has a very red colour, with $3.19^{+0.18}_{-0.16}$ mag difference between the H and [$\text{L}^\prime$]{} band. Fig. \[cmd\] compares its position in a colour-magnitude diagram to that of other young companions and field dwarfs [@Legett2010; @Legett2013]. As the age is debated, we overplotted three isochrones using LYON evolutionary tracks [@Chabrier2000; @Baraffe2003] for 200 Myr, 800Myr and 2 Gyr. HD206893B lies among the L5-L9 field dwarf objects, with a similar [$\text{L}^\prime$]{}magnitude as 2MASS0122-2439B [@Bowler2013] but a redder colour. This makes it the reddest object among young and dusty L dwarfs in the field, which is likely due to a dusty atmosphere although interstellar or disc reddening cannot be ruled out. Using an age of 200Myr (respectively 800Myr, 2Gyr) and the AMES-Cond models [@Baraffe2003], the H-band contrast of HD206893B implies an object of 24[$M_{\text{Jup}}$]{} and effective temperature 1230K (50[$M_{\text{Jup}}$]{}with 1330K, and 75[$M_{\text{Jup}}$]{} with 1420K respectively).
Relation to the debris disc and orbital motion {#sec_orbital_motion}
----------------------------------------------
The debris disc is marginally resolved with Herschel/PACS with an inclination of $40^\circ\pm10^\circ$ along the PA $60^\circ \pm10^\circ$. This is presented in Appendix \[app\_disc\] along with a modelling concluding on a disc inner radius of 50au. With a projected separation of $10.4\pm0.1$au, the companion appears therefore to be interior to the disc, with a PA consistent with an orbit in the same plane as the disc. As shown in Fig. \[fig\_NaCo\] right, the companion moved during the 306 days separating the SPHERE and NaCo detections. We applied the methods laid in @Pearce2015 to constrain the orbit of a companion imaged over short orbital arcs. With the two epochs, the linear sky motion is $0.05 \pm 0.01$/year at a position angle of $-27{^\circ}^{+13}_{-17}$. We derived thereafter the parameter $B$ and $\phi$ of @Pearce2015, that proves useful to constrain the possible orbits. Including astrometric, distance, and 10% stellar mass uncertainties we find $B=0.33^{+0.19}_{-0.14}$ and $\phi=90^\circ \pm 15^\circ$. The constraints are weak as the NaCo astrometry is relatively poor; in particular we cannot constrain the ascending node to test whether the orbit is indeed coplanar with the disc. If we assume a low eccentricity orbit, the semi-major axis is about 12au (37yr period). Conversely, for a high eccentricity, if the apocentre distance is restricted to be below 50au (i.e. the approximate disc inner edge), this implies that the companion is currently at apocentre, that is with a semi-major axis below 10au. If we restrict the companion to lie within a projected separation of 500mas in June 2007 to avoid detection by NICMOS (see Fig. \[contrast\]), the marginalised constraints on specific parameters do not improve much. If the companion truncates the disc, then the semi-major axis is expected to be $\sim$1/3 of the inner edge radius [@Holman1999], so 15-20 au (0.4-0.5). Thus, while there is no evidence to suggest that the companion’s orbit is unusual, more precise astrometry is required to constrain the orbital elements and draw conclusions about the relation between the companion’s orbit and the disc.
This architecture is strikingly similar to the HR2562 system [@Konopacky2016], where a $30\pm15$ [$M_{\text{Jup}}$]{} brown dward was detected at 20au of its host star, within a debris disc of inner radius $\sim75$au.
[p[2.5cm]{} >p[2.5cm]{} >p[3.3cm]{} >p[1.cm]{} ]{} Property & HD206893 & HD206893B & Ref. [\
]{} Distance (pc) & & 1 [\
]{} & & 1 [\
]{} & & 1 [\
]{}Age (Gyr) & & 2 [\
]{}Spectral type & F5V & L5-L9 dwarf & 1 / 3,4[\
]{}H mag & 5.69 & $16.79 \pm 0.06$ & 1 / 3 [\
]{}[$\text{L}^\prime$]{}mag & 5.52 & $13.43^{+0.17}_{-0.15}$ & 1 / 4 [\
]{}Mass & $1.24 M_{\odot}$ & 24 / 50 / 73 [$M_{\text{Jup}}$]{}& 5 / 3 [\
]{}$T_\text{eff}$ (K) & 6486 & 1200 / 1310 / 1380 & 2 / 3 [\
]{} & & $270.4 \pm 2.6$ & 3 [\
]{} & & $268.8 \pm 10.4$ & 4 [\
]{} & & $69.95 \pm0.55$ & 3 [\
]{} & & $61.6 \pm 1.9$ & 4 [\
]{}
![Left: Image of NaCo reduced with PCA-ADI after removing seven principal components, which was found to optimise the companion S/N ($\sim 6$). Right: common proper motion analysis. The black line displays the motion of the companion if it was a background source. The uncertainty is given at $1\sigma$.[]{data-label="fig_NaCo"}](NaCo_ppm_v2.pdf){width="\hsize"}
![ Constrast curves at $5\sigma$. The red and blue dots show the positions of the companion in the NaCo and SPHERE data respectively. []{data-label="contrast"}](contrast_curves_combined_w_idps.pdf){width="0.9\hsize"}
![ \[cmd\] Colour-magnitude diagram, obtained from @Galicher2014 with the new photometry of HD206893B (yellow star), that of 51 Eri [@Macintosh2015] and 2MASS0122-2439B [@Bowler2013]. The lines show the isochrones for different ages and evolution models.](Lp_HLp.pdf){width="0.9\hsize"}
Conclusions {#sec_conclusion}
===========
This letter presents the detection of a low-mass companion orbiting at a projected separation of 10 au around the F5V star HD206893 as part of the SHARDDS survey, thanks to VLT/SPHERE high contrast capabilities. The object was confirmed by VLT/NACO and proven not to be a background source by HST/NICMOS. With an H-band contrast of 11 mag, evolutionary models suggest the object could be a 25[$M_{\text{Jup}}$]{} brown dwarf if the system is 200 Myr, or twice as massive for a 800Myr system. Along with its [$\text{L}^\prime$]{}contrast of 7.9, the object appears very red, and the closest to L5-L9 field dwarfs in a colour-magnitude diagram. The orbital motion is detected and suggests an orbital period of $\sim37$ yr in case of low eccentricty.
In addition, we report the detection of the disc, through its thermal emission with Herschel/PACS at a position angle of $\sim60^\circ$ almost aligned with the projected position of the companion. This system is therefore reminiscent of the cases of HR8799, HD95086 or $\beta$ Pictoris where one of several GPs have been detected in orbit inside a Kuiper belt analog. It is the second brown dwarf detected in the inner hole of a debris disc after HR2562.
Several aspects make this system very attractive for future characterisation. The contrast is well within range of current extreme AO instruments, enabling spectral identification. The orbital motion is fast enough to allow orbit monitoring which can bring constraints on the dynamical mass of the object. Deeper observations may detect the scattered light of the disc and confirm the faint emission seen in our image, to understand if the companion is responsible for the inner truncation of the disc at about 50 au, or possibly reveal asymetries and clumps resulting from interactions between the disc and the brown dwarf or possible yet undiscovered planets.
JM is supported by the ESO fellowship programme. He thanks R. Galicher for his re-reduction of the NIRI data, and O. Wertz for his help with VIP and NEGFC. EC is supported by NASA through Hubble Fellowship grant HST-HF2-51355 and HST-AR-12652 awarded by STScI, operated by the AURA, Inc., for NASA under contract NAS5-26555. OA is a F.R.S.-FNRS Research Associate. The research leading to these results was partly funded by the European Research Council under the European Union’s Seventh Framework Programme (ERC Grant Agreement n. 337569), and by the French Community of Belgium through an ARC grant for Concerted Research Action. GMK is supported by the Royal Society as a Royal Society University Research Fellow. MCW and LM are supported by the European Union through ERC grant 279973. VC acknowledges J. Smoker and M. Espinoza for their help with NaCo. VC is supported by the Millennium Science Initiative (Chilean Ministry of Economy) through grant RC130007. CdB acknowledges support from the Mexican CONACyT research grant CB-2012-183007.
NICMOS non detection {#app_nicmos}
====================
HD206893 was observed on 12 June 2007 with the NICMOS instrument on the *Hubble Space Telescope* (HST), as part of a survey looking for debris discs around nearby stars (PI: J.Rhee, GO-11157). The data were obtained with the mid-resolution NIC2 channel (plate scale 0.07565/px) with a coronagraph of radius 0.3, in two filters F160W and F110W. The target was observed at two orientations of the telescope separated by $\sim30^\circ$, to enable PSF subtraction with roll differential imaging. We reprocessed the F160W archival dataset (centre wavelength 1.6006, FWHM 0.4012) with the same PSF subtraction method as used in the Archival Legacy Investigations for Circumstellar Environment (ALICE) programme [@Soummer2014; @Choquet2016]. We used the KLIP algorithm [@Soummer2012] on PSF libraries composed of images from multiple reference stars. After bad pixel correction, we selected the 454 images (from 78 reference stars) the most correlated with each of HD206893’s exposures, out of a reference star library assembled with the ALICE pipeline. This selection favoured images from 78 different stars chosen mostly from the two dominant HST programmes in the initial library (programmes 11157 and 10176). Fig. \[fig\_NICMOS\] (left) shows the combined image after subtracting synthetic PSFs computed from the 55 strongest eigenmodes of the library. No point source is detected at the position where the candidate would have been in 2007 if it were a background object without proper motion. Our detection limits (Fig \[contrast\] and \[fig\_NICMOS\] right show that the point source would have been detected in that case. This re-analysis shows that the point source is co-moving with the star.
![F160W NICMOS 2007 coronagraphic images, processed with RDI + PCA without injecting a fake companion (left) and after injection of a fake companion of contrast 11.1mag at the location where the candidate would have been in 2007 assuming it is a background object with no proper motion (white circle).[]{data-label="fig_NICMOS"}](NICMOS_image.jpg){width="\hsize"}
The disc around HD 206893 {#app_disc}
=========================
Archival Herschel PACS data were obtained and modelled in the same way as @Kennedy2012_99Her [@Kennedy2012_coplanar_disk]. The 70 image is shown in Fig. \[fig\_disk\_herschel\] (top image). The subtraction of a scaled calibration observation (PSF) clearly shows that the source HD206893 is not point-like (bottom image) and reveals the approximate disc extent and position angle. The debris disc is only a few beams across so constraints on the disc properties are relatively poor. Our disc best fitting model assumes a temperature $T_{\rm disk}=288 r^{-0.4}$, with $r$ in au, and has a decreasing power-law surface density distribution of $\Sigma \propto r^{-0.5}$ extending from 50 to 200au. This is compatible with previous modelling by @Moor2011 who proposed a modified blackbody model for the disc with a temperature of 49K and a radius of 49 au. This is also compatible with the colder (48K) dust population in the double component disc model of @Chen2014. The disc is inclined (from face-on) by about 40$^\circ$ at a position angle of $\sim60^\circ$ (East of North). The uncertainties on these angles are on the order of 10$^\circ$
To try to reveal the disc in scattered light, we reduced the SPHERE images using classical ADI (Fig. \[fig\_disk\_IRDIS\] showing the whole $12\arcsec\times12\arcsec$ field of view of SPHERE/IRDIS) to maximise the sensitivity in the backround and limit flux losses induced with more aggresive reductions [@Milli2012] and binned the pixels by a factor two. We see a faint extended emission along the PA $\sim60^\circ$ with a surface brighntess of $\sim0.05 $mJy/arcsec$^2$. This faint emission is detected from $\sim1.5\arcsec$ (60 au) up to $\sim$4-5(150-190au) where the background noise starts to dominate.We could confidently rule out spurious emission along that PA coming from the diffraction pattern from the spiders or the elongation of the PSF due to the wind at the ground level or at higher altitudes. Furthermore the PA is compatible with the Herschel/PACS residual image. We therefore tentatively attribute this signal to the scattered light of the debris disc with a S/N of approximately one.
![Herschel / PACS image of the star HD206893 at 70 (top). The bottom image shows the residuals after subtraction of the Herschel PSF, showing the source is not point-like.[]{data-label="fig_disk_herschel"}](obser0-70.pdf "fig:"){width="0.9\hsize"} ![Herschel / PACS image of the star HD206893 at 70 (top). The bottom image shows the residuals after subtraction of the Herschel PSF, showing the source is not point-like.[]{data-label="fig_disk_herschel"}](peaksub0-70.pdf "fig:"){width="0.9\hsize"}
![SPHERE H-band image in its complete $12\arcsec\times12\arcsec$ field of view after a classical ADI reduction, in a linear colour scale in ${\mu}$Jy/arcsec$^2$. It shows a faint and extended emission along the same PA as the Herschel residual image likely coming from the disc.[]{data-label="fig_disk_IRDIS"}](SPHERE_HD206893_disk.pdf){width="\hsize"}
[^1]: available at <https://github.com/vortex-exoplanet/VIP>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We explore the degeneracy between mass and spin in gravitational waveforms emitted by black-hole binary coalescences. We focus on spin-aligned waveforms and obtain our results using phenomenological models that were tuned to numerical-relativity simulations. A degeneracy is known for low-mass binaries (particularly neutron-star binaries), where gravitational-wave detectors are sensitive to only the inspiral phase, and the waveform can be modelled by post-Newtonian theory. Here, we consider black-hole binaries, where detectors will also be sensitive to the merger and ringdown, and demonstrate that the degeneracy persists across a broad mass range. At low masses, the degeneracy is between mass ratio and total spin, with chirp mass accurately determined. At higher masses, the degeneracy persists but is not so clearly characterised by constant chirp mass as the merger and ringdown become more significant. We consider the importance of this degeneracy both for performing searches (including searches where only non-spinning templates are used) and in parameter extraction from observed systems. We compare observational capabilities between the early ($\sim$2015) and final (2018 onwards) versions of the Advanced LIGO detector.'
author:
- Emily Baird
- Stephen Fairhurst
- Mark Hannam
- Patricia Murphy
bibliography:
- 'refs.bib'
title: 'Degeneracy between mass and spin in black-hole-binary waveforms'
---
Introduction
============
The advanced LIGO and Virgo detectors are likely to allow us to observe numerous black-hole-binary coalescences in the coming years [@Abadie:2010cf; @Dominik:2012vs]. While the detectors are being installed, the challenge is to devise search and source extraction methods that will identify all black-hole-binary signals in the data whilst also allowing for the most accurate estimation of the source parameters possible.
Our understanding of gravitational waveforms emitted by black-hole mergers has improved dramatcially over the past five years, with the aid of a large number of numerical simulations, from which several models of the waveforms have been constructed (see e.g., Ref [@Ohme:2011rm] for an overview). A large fraction of the simulations have focussed on non-precessing systems (where the black holes are either non-spinning or have spins aligned with the orbital angular momentum); see, for example, Refs. [@Hannam:2009rd] and [@Ajith:2012tt]. Other than for high black-hole spins (greater than $a\sim0.8$), this space has been rather well covered for comparable-mass binaries, and several waveform models are available [@Ajith:2009bn; @Santamaria:2010yb; @Pan:2009wj]. The models characterize waveforms based on the masses of the two components and, in the case of the models we will consider in this paper, the total spin of the system. We will make use of these waveform models to investigate degeneracies in the waveforms. Since we are only considering a three-dimensional subspace of the full eight-dimensional space of binary masses and spins, many effects, most notably precession, cannot be probed. However, it is interesting to consider this smaller space because it is very likely that any degeneracies will persist in the full parameter space. Furthermore, recent results have shown that the waveform for a precessing binary can be factorized into precessional effects and a non-precessional part that is well modelled by the three parameters we consider here [@Schmidt:2012rh].
In post-Newtonian (PN) theory, there is a well known degeneracy between the black holes’ mass ratio and the total spin, which arises at 1.5PN order, while a combination of the total mass and mass ratio (the “chirp mass”) remains relatively well constrained [@Cutler:1994ys; @Poisson:1995ef]. Here, our focus is on higher-mass waveforms that include a merger and ringdown. We demonstrate that the degeneracy persists in the full waveforms and also up to higher masses where only the later part of the inspiral and the merger/ringdown are in the detector’s sensitivity band. However, for the high-mass binaries, the degeneracy in mass and spin is not so clearly characterized by the chirp mass.
A degeneracy in the emitted gravitational waveform across the parameter space has numerous consequences for gravitational wave (GW) searches, both good and bad. The positive effect is that a degeneracy in the parameter space will reduce the volume that needs to be searched, thereby reducing the computational cost and trials factor associated with the search. The majority of searches of LIGO and Virgo data have made use of non-spinning components in the template waveforms (see e.g., Refs. [@Collaboration:S6CBClowmass; @Collaboration:S5HighMass]). A degeneracy between mass and spin means that the template waveforms have covered a larger fraction of the parameter space than might be naively expected. Furthermore, it has recently been argued that a two-dimensional bank of templates is sufficient to cover the space of spinning neutron-star binaries [@Brown:2012qf]. In this paper, we investigate the effect of using non-spinning waveforms in a search for black-hole binaries with spins and show that the mass-spin degeneracy renders the search more sensitive to spinning systems than might be expected. Nevertheless, a search using waveforms incorporating spin effects would be a significant improvement and the degeneracy should make it computationally feasible.
On the other hand, a degeneracy has a negative effect on parameter estimation. To make the most of gravitational-wave observations, accurate extraction of the physical parameters is of paramount importance. There are detailed multi-dimensional methods under development to accurately recover the parameters [@Sluys:2008a; @Sluys:2008b; @Veitch:2010; @Feroz:2009]. However, there’s nothing that can be done about a real degeneracy in the emitted waveforms — there is no way of telling them apart. We evaluate the effects of mass-spin waveform degeneracy on our ability to accurately recover masses and spins and discuss the astrophysical implications. In the process we introduce a simple method to estimate the parameter-estimation confidence intervals.
Throughout the paper, we will provide sample results using a number of waveforms at different masses and mass ratios. We also make use of a number of noise curves for the advanced detectors, to illustrate how this effect is likely to change as the detectors approach their final, design sensitivity. No attempt has been made to perform an exhaustive study across the full parameter space, and this is left as a future project.
The layout of the paper is as follows: in Sec. \[sec:model\] we outline the waveform models that we use, and discuss some of the assumptions that went into producing them and their range of applicability. We also discuss the interpretation of results in terms of mismatches between waveforms and introduce the detector noise curves used in our studies. In Sec. \[sec:spin\_eta\] we focus on the degeneracy between mass ratio and spin for low-mass binaries, giving the post-Newtonian argument for this degeneracy and illustrating the degeneracy for a number of cases and different detector sensitivities, and discuss implications for searches. Then, in Sec. \[sec:high\_mass\] we extend the results to higher mass binaries and show that the degeneracy persists, although in a different form. In Sec. \[sec:parameters\] we discuss implications for the accurate estimation of parameters and astrophysical inference.
Model and methods {#sec:model}
=================
The phenomenological waveform models {#sec:phenom}
------------------------------------
We describe the GW signal from black-hole binaries with non-precessing spins (i.e., the spins are aligned or anti-aligned with the binary’s orbital angular momentum) using the phenomenological models presented in Refs. [@Ajith:2009bn] and [@Santamaria:2010yb]. For consistency with the labelling used within the LIGO-Virgo Collaboration [@LAL] we refer to these models respectively as “PhenomB” and “PhenomC”. (“PhenomA” refers to an earlier model of non-spinning binaries [@Ajith:2007qp; @Ajith:2007kx; @Ajith:2007xh].) In both models the waveforms are parametrized by their total mass $M = m_1 + m_2$, mass ratio $\eta = m_1 m_2 / M^2$, and an effective total spin parameter, $$\begin{aligned}
\chi & = & \frac{1}{M} (m_1 \chi_1 + m_2 \chi_2) \nonumber \\
& = & \chi_{s} + \delta \chi_{a},\end{aligned}$$ where $\chi_i = S_i / m_i^2$ for each black hole with angular momentum $S_i$, $\delta = (m_1 - m_2)/M$, and the symmetric and anti-symmetric combinations of the spins are $\chi_s = (\chi_1 + \chi_2)/2$ and $\chi_a = (\chi_1 - \chi_2)/2$. The phenomenological models incorporate a PN description of the inspiral, while the merger and ringdown regimes are tuned using the results of numerical simulations.
Both models have the same basic structure. The waveform is represented in the Fourier domain as $h(f) = A(f) e^{i \Psi(f)}$. The amplitude $A(f)$ and phase $\Psi(f)$ are modelled separately, using input from PN theory (inspiral), the observed properties of NR waveforms (plunge-merger), or results from perturbation theory (ringdown). The models are all power series in the frequency $f$, and the coefficients in the model are written as polynomials in the two physical parameters $\eta$ and $\chi$ (the total mass is an overall scale factor), and it is the coefficients of these polynomials that are then calibrated to hybrids of PN and NR waveforms. For both models, the amplitude is constructed in a similar manner, with expressions for each of the inspiral, plunge-merger, and ringdown portions of the waveform modelled independently. In PhenomB [@Ajith:2009bn] the three parts are connected as piecewise functions, and in PhenomC [@Santamaria:2010yb] smooth $\tanh$-function interpolation is used. To obtain an expression for the phasing, PhenomB uses a single series expansion, matching the coefficients beyond leading order to hybrid waveforms, as well as results from the test mass ($\eta \rightarrow 0$) limit. PhenomC, on the other hand, uses the complete TaylorF2 PN inspiral phasing, and only the late inspiral/merger phase is fitted in a narrow frequency range to numerical simulations, while the ringdown waveform is obtained from analytically derived quasi-normal mode expressions for the frequency and attached continuously to the merger phase. There are 54 free parameters in PhenomB, and 45 in PhenomC, although the final models are both functions of only $\{M,\eta,\chi\}$. There are three notable differences between the two models: (1) the PhenomB PN-NR hybrids are produced in the time domain, using TaylorT1 for the PN part, and the PhenomC hybrids are produced in the frequency domain, using TaylorF2 for the PN part [@Ajith:2007jx]; (2) PhenomB incorporates information about the test-mass limit; (3) in PhenomC the phase evolution during inspiral incorporates PN calculations up to 3.5PN order (although the spin terms are complete only up to 2.5PN), while in PhenomB only the leading-order PN inspiral term is fixed, and the remaining terms up to 3.5PN order are tuned to the hybrid waveforms.
There is good qualitative agreement between the two models [@Santamaria:2010yb], although no detailed quantitative comparison has yet been performed. We choose to work with PhenomB for generating the results presented in this paper. We have also cross-checked some of the calculations against PhenomC, and we comment further on this in Sec. \[sec:high\_mass\].
Detectors and noise curves {#sec:detectors}
--------------------------
![ \[fig:NoiseCurves\] Noise curves for the advanced LIGO detector configurations that we consider in this paper: “early”, “no-SRM” and “final”, which corresponds to the zero-detuned, high-power configuration. See text for more details. ](NoiseCurves.pdf){width="50.00000%"}
In this paper we compare spinning- and non-spinning-binary signals with reference to the expected sensitivity of the Advanced LIGO detector (aLIGO) [@Abbott:2007kv; @Shoemaker:aLIGO; @2010CQGra..27h4006H]. The sensitivity curves we use are shown in Fig. \[fig:NoiseCurves\]. During the early science runs, expected around 2015, the advanced LIGO detector is unlikely to be at its full design sensitivity. Consequently, we use the “early aLIGO” noise curve [@G1000176] to give results indicative of what may be achieved in the early runs. At its optimum sensitivity several years later, the anticipated sensitivity is given by the “zero-detuned high-power” noise curve [@T0900288]. In this paper we will take that as the “final” design sensitivity of the detector. Over the parameter space of binaries that we study in this paper, the early aLIGO curve represents a sensitivity of roughly five times greater in the signal-to-noise ratio (SNR) than in the initial LIGO detectors during their final S6 science run, which corresponds to an increase in potential sources of two orders of magnitude. The final curve represents a further factor of three improvement in SNR, or about 30 times as many potential sources over early aLIGO. Finally, we also consider the “no signal recycling” (no-SRM) configuration of the detector, that could be achieved by the aLIGO detector operating without a signal recycling cavity. This has comparable low-frequency sensitivity to the final configuration, but significantly worse sensitivity at high frequencies. Although it is unlikely to be an observational mode, the non-signal-recycled curve offers a means to compare the effects of low and high frequency sensitivity upon our results.
Waveform Mismatches {#sec:matches}
-------------------
We use the standard inner product between two waveforms, $h_1(f)$ and $h_2(f)$ with respect to the power spectral density $S_n(f)$ of a detector [@Cutler94], $$(h_1 | h_2 ) = 4 \, {\rm Re} \int_{0}^{\infty} \frac{h_1(f) h_2^*(f)}{S_n(f)} df.$$ The match between two waveforms is defined as their normalized inner product, maximized over time and phase shifts of the waveform, $$M(h_1,h_2) = \max_{\Delta t, \Delta \phi} \frac{(h_2 | h_2)}{ | h_1 | | h_2| }.$$
Alternatively, we can consider the situation of a true waveform produced by a physical source, $h_T$, and a model waveform $h_M$ that we will use to search for the real signal in detector data. If we normalize both waveforms, then we can split the model waveform into a component parallel to the true waveform, plus a component that is orthogonal to the true waveform, in the sense of our inner product. In other words, we have $$\hat{h}_M = \sqrt{1 - x^2} \hat{h}_T + x \hat{h}_E,$$ where $\hat{h}_E$ is the normalized “error” waveform that satisfies $(h_T | h_E) = 0$. In this form, the match is $M(h_T,h_M) = \sqrt{1 - x^2} \approx 1 - x^2/2$. Thus, the match is directly related to the (relative) amplitude of the “error” waveform.
In a GW search, a template bank of model waveforms is constructed [@Babak:2006dw] such that the match between every point in the waveform parameter space and the nearest template in the bank is at least 0.97. Assuming that the model waveforms are physically correct, such a template bank ensures that we will lose no more than 10% of signals in our search. (A match of 0.97 means that the sensitivity range of the detector is only 97% of its optimum, and the detector is therefore sensitive to only $0.97^3 \approx 0.9$ of its optimum volume, and so we lose about 10% of signals.)
If the physical waveforms do not agree exactly with the model waveforms, this will lead to an additional loss in match between a signal and the best matched template, and consequently a reduction in the number of signals observed above threshold. In addition, to counter non-stationary detector noise, a number of signal consistency tests are included into analyses to distinguish signals from non-stationarities or “glitches” in the data [@Babak:2012zx]. These tests are used to either remove completely any transients that do not match the templates or else to down-weight their significance. Since searches performed to date have made use of non-spinning waveform families these thresholds have been set relatively loosely (and tested with spinning signals) to ensure they are not removing signals. For the high matches between signal and template we consider in this paper, the effect of signal consistency tests will be minimal, and we will not consider it further.
In the next sections we will identify the regions of the non-spinning waveform parameter space that provide a match of greater than 0.97 with the chosen waveform (which will usually incorporate spin). In doing so, we densely sample the non-spinning waveforms to identify all points with a match above 0.97. When performing a search, there will then be two contributions to the mismatch between the signal and the best matched template: one due to the difference between the waveform and search space and a second arising from the discrete sampling of the template space. The match between the signal and the closest template is guaranteed to be above 0.94 as the mismatches add linearly in this case (see e.g. [@Lindblom:2008ha] for details). We are therefore requiring that the potential loss of SNR due to a mismatch between the model waveform and true waveform is no greater than the maximum possible loss due to the discreteness of the template bank.
We will also investigate how the parameters of the best-match *non-spinning* waveform vary when the signal corresponds to a non-precessing binary. In Sec. \[sec:parameters\] we will see how the match can be used to give an estimate of the parameter estimation accuracy.
Degeneracy between $\eta$ and $\chi$ at low masses {#sec:spin_eta}
==================================================
We first consider the degeneracy between the symmetric mass ratio $\eta$ and the effective total spin $\chi$ of the binary. This effect is already well known in PN theory [@Cutler:1994ys; @Poisson:1995ef], which we will discuss first, and then we will look at inspiral-merger-ringdown (IMR) signals at three different values of the total mass, $M = \{20, 50, 100\} M_\odot$.
Degeneracy in PN theory {#sec:PNtheory}
-----------------------
The phase evolution of a compact binary in PN theory has been calculated up to 3.5PN order in the non-spinning terms, and up to 2.5PN in the spin effects; see Ref. [@Ajith:2007jx] for a recent summary of PN treatments of the phase. Up to the leading order that includes spin, the phase for non-precessing binaries is given in the frequency domain by $$\begin{aligned}
\Psi(f) & = & \frac{3}{128 \eta v^5} \left\{ 1 + v^2 \left[ \frac{3715}{756} + \frac{55 \eta}{9} \right] \right. \\
&& \ \ \ \ \ \ \ \ \ \ \ \ \left. - v^3 \left[ 16 \pi - \left( \frac{113}{3} - \frac{76 \eta}{3} \right) \chi_s - \frac{113 \delta}{3} \chi_a \right] \right\}
\label{eqn:PN}\end{aligned}$$ where $v = (\pi M f)^{1/3}$. If we define the chirp mass of the binary as ${\cal M} = M \eta^{3/5}$, then we see that the leading factor is proportional to $1/({\cal M} \pi f)^{5/3}$, and the phase evolution is dominated by the chirp mass. This motivates the observation that in GW searches we expect to measure the chirp mass with high accuracy, and we will see examples of this in Sec. \[sec:IMRlow\]. At the next-to-leading order, the phase evolution depends on the mass ratio $\eta$, and dependence on the spins enters at the following order. We can absorb all of the spin effects at this order into an effective spin term, $\chi_{\rm PN} =
\chi_s + \delta \chi_a - (76\eta/113) \chi_s$, and this is what is proposed in the inspiral template family discussed in Ref. [@Ajith:2011ec] (and this is proportional to the leading-order spin-orbit parameter $\beta$ used in Refs. [@Cutler:1994ys; @Poisson:1995ef]). The phenomenological models [@Ajith:2009bn; @Santamaria:2010yb] use the simpler effective spin $\chi = \chi_s + \delta \chi_a$ to describe full inspiral, merger and ringdown waveforms.
If we adopt for the moment the PN effective spin term, we can write the deviation from the leading-order term as $$\Delta \Psi(f) = \frac{3}{128 \eta v^3} \left[ \frac{3715}{756} + \frac{55 \eta}{9}
+ v \left( \frac{113 \chi_{\rm PN}}{3} - 16 \pi \right) \right].
\label{eqn:PNdeg}$$ Thus, it is possible to mimic the effect of spin by modifying the mass ratio (while keeping the chirp mass constant). It is clear, however, that the required modification to $\eta$ will vary as $v$ changes over the course of the chirp, as expected as this is only an approximate degeneracy. The majority of the power from a binary merger is accumulated between around 30 to 300 Hz, and over this range $\nu$ changes by only a factor of two. This explains why the approximation is reasonable across the whole signal, and this leads to the common claim that non-spinning templates can be used to detect GW signals from spinning binaries, but there will be an offset in the measurement of the mass ratio and the total mass. It may also be possible to exploit this degeneracy to search for spinning binaries using a non-spinning model, but with $\eta$ extended to unphysical values, $\eta > 0.25$, in order to cover more of the spinning-binary parameter space; a similar idea has already been suggested to extend inspiral-only searches beyond their expected region of validity [@Boyle:2009dg]. It is the degeneracy between mass ratio and spin that we will investigate here.
Degeneracy in full IMR models {#sec:IMRlow}
-----------------------------
We would now like to investigate whether the degeneracy from post-Newtonian theory is present in full IMR waveform models.
We consider a set of spinning signals and those non-spinning waveforms (i.e. the phenomenological model with $\chi = 0$) that provide a good match to the signal. The efficacy of a waveform model in a GW search can be estimated by calculating the best match (fitting factor) between any member of the waveform model, and the target signal. Therefore, this will give an immediate indication of the merits (or otherwise) of using a non-spinning waveform model to search for mergers of spinning black holes. We consider fitting factors above 0.97 to be adequate as discussed previously.
![ \[fig:M20early\] Match between spinning $20 M_{\odot}$ waveforms with non-spinning templates in early aLIGO. Each of the strips shows the region of $M$-$\eta$ space for which the non-spinning waveform has a match of 0.97 or higher with a spinning signal. The total spin is indicated next to each region. For spins above 0.5 or below -0.7 the best overlap with a non-spinning waveform is less than 0.97. The curve of constant chirp mass is indicated by a dotted line.](M20mismatchEarly.pdf){width="50.00000%"}
Fig. \[fig:M20early\] shows the results for a binary with a total mass of 20$M_\odot$ and mass ratio 1:4 ($\eta = 0.16$), i.e. a $4M_\odot$–$16M_\odot$ binary, and matches calculated using the early aLIGO spectrum. We consider sources with a total spin in the range \[-1,1\] in steps of 0.1. For each configuration, we identify non-spinning waveforms which give a match of greater than 0.97 with the spinning source. When the binary spin is zero, the non-spinning model matches the target signal at the correct parameters (indicated by a star), and along a strip in the parameter space with a width of 5% in mass, and 10% in $\eta$. When the binary contains spinning black holes, the non-spinning model matches the signal, but with a bias in the mass and mass ratio. The binary spins are indicated with different colours, labeled by the total spin of the binary. For large-spin signals there are no non-spinning waveforms that have a match above 0.97. When using the early aLIGO noise curve, the range of spins for which non-spinning waveforms have a match above 0.97 is $[-0.7,0.5]$. However, for spins above about $\chi = 0.3$, the best-match waveform has an unphysical value of the mass ratio $\eta$. The curve of constant chirp mass $M \eta^{3/5}$ is indicated by a dotted line.
The first thing to note is that, in all cases, the chirp mass of the best matched non-spinning waveform is very similar to the true chirp mass of the system. We can see from Fig. \[fig:M20early\] that the high-match regions follow a line of constant chirp mass. In Fig. \[fig:M20matchesChirp\] we re-parameterize the results of Fig. \[fig:M20early\] in terms of $\eta$ and ${\cal M}$. We see that, even though there is a strong degeneracy between $\eta$ and $\chi$, the correct chirp mass ${\cal M} = 6.66 M_{\odot}$ falls within the high-match region for most values of spin. The deviation from the correct chirp mass is no more than $\sim3$% for large anti-aligned spins.
![ \[fig:M20matchesChirp\] Same results as in Fig. \[fig:M20early\], but now parametrized in terms of chirp mass and symmetric mass ratio. This figure reinforces the accuracy with which the chirp mass can be recovered. ](M20mismatchEarlyChirp.pdf){width="50.00000%"}
The second important point relates to the mass-spin degeneracy. For aligned-spin binaries ($\chi > 0$), the best match is obtained with a lower-mass, higher-mass-ratio model, while for anti-aligned spins, the best matched non-spinning waveform has a higher mass and lower mass-ratio. This is to be expected as aligned spins are expected to cause the system to “hang up” and chirp more slowly, mimicking a lower-mass signal.
It is informative to return to the post-Newtonian phasing given in Eq. (\[eqn:PNdeg\]) to see whether the observed degeneracy matches what is theoretically expected. In low-mass binaries, the signal is dominated by the inspiral, which can be represented in a PN expansion, as described in Sec. \[sec:PNtheory\]. The chirp mass is determined to high accuracy by the leading-order term in the PN expansion of the phase in the Fourier domain, and we can then solve Eq. (\[eqn:PN\]) (with fixed chirp mass) to find the symmetric mass ratio that mimics the effect of the spin. We do this by solving (\[eqn:PNdeg\]) as a function of frequency (in the detector’s sensitive band) and then averaging over the values of $\eta$ that we obtain. Fig. \[fig:degeneracy\] shows the symmetric mass ratio that corresponds to each value of the spin, as predicted from Eq. (\[eqn:PN\]), and as found in the mismatch analysis of the phenomenological models that is shown in Fig. \[fig:M20early\]. We see that the PN estimate is remarkably close to that found for the full IMR models, at least for the $20\,M_\odot$ binaries used in this example.
![ \[fig:degeneracy\] Comparison of the $\chi$-$\eta$ degeneracy as found for the phenomenological model (stars), and from the requirement that $\Delta \Psi = \Delta\Psi(\chi_{PN}=0)$ in Eq. (\[eqn:PNdeg\]) for fixed chirp mass, for signals with $M =
20\,M_\odot$ and $\eta = 0.16$. The naive PN prediction agrees well with the full IMR results for low spins, and for most anti-aligned spins. ](ChiEtaDegeneracy.pdf){width="50.00000%"}
![ \[fig:M20matches\] Matches of spinning $20 M_{\odot}$ waveforms with non-spinning templates in aLIGO at design sensitivity. The upper panel shows the results for the no-SRM configuration, while the bottom shows results for the final sensitivity ](M20mismatchNoSRM.pdf "fig:"){width="50.00000%"} ![ \[fig:M20matches\] Matches of spinning $20 M_{\odot}$ waveforms with non-spinning templates in aLIGO at design sensitivity. The upper panel shows the results for the no-SRM configuration, while the bottom shows results for the final sensitivity ](M20mismatch "fig:"){width="50.00000%"}
Evolution of noise curves
-------------------------
We present results for two design aLIGO configurations in Fig. \[fig:M20matches\]: the no-signal-recycling (no-SRM) configuration and the final high-power zero-detuned sensitivity curve (final). Qualitatively, the results are quite similar for the no-SRM and final noise curves. Again spinning signals are recovered with a match above 0.97 using non-spinning templates, with a decrease in recovered mass for positive spin systems. The chirp mass is still well recovered. However, the detector’s extra sensitivity makes it easier to distinguish the spinning signals with a non-spinning model. The range of spins for which we can find a non-spinning signal with matches greater than 0.97 is now only $\chi \in [-0.4,0.3]$ for no-SRM and $\chi \in [-0.3,0.3]$ for the final configuration. We also see that each match region shrinks, although its location is unchanged. In moving from the early noise curve to no-SRM, the most significant sensitivity improvement is at low frequency, while the final configuration offers much greater high frequency sensitivity. The results for these two curves are quite comparable (and significantly better than the early curve) indicating that it is low frequency sensitivity that provides the biggest improvement in breaking the degeneracy. With the later noise curves, the variation of chirp mass between the spinning signal and non-spinning template is less than 1%. [^1]
Mass-spin degeneracy at higher masses {#sec:high_mass}
=====================================
For black-hole binaries with masses greater than $20 M_{\odot}$ the merger and ringdown parts of the signal become increasingly important and contribute an ever increasing fraction of the signal-to-noise ratio. At higher masses, we do not expect the chirp mass to determine the waveform to such an extent as for the $20 M_{\odot}$ system and there is no *a priori* reason to expect that a degeneracy between mass and spin will persist.
We begin by considering a $50 M_{\odot}$ binary with mass ratio 1:4. At this mass, the merger and ringdown will provide a significant fraction of the signal to noise ratio. As a crude estimate of the effect, imagine that the inspiral part of the waveform is valid up to the innermost stable circular orbit (ISCO) of a point particle orbiting a Schwarzschild black hole, which is at 90Hz for a $50\,M_{\odot}$ binary. For the early aLIGO noise curve, the inclusion of the merger and ringdown will increase the SNR by a factor of 4. At the final aLIGO sensitivity, the improved low frequency response increases the contribution of the inspiral but the merger and ringdown still contribute a comparable SNR to the inspiral (so that the overall SNR is $\sim50$% higher, because they add in quadrature). Thus, there is no reason to expect that the chirp mass is still well recovered or that the degeneracy discussed previously still holds.
![ \[fig:M50matches\] Matches of spinning $50 M_{\odot}$ waveforms of mass ratio 4:1 with non-spinning templates in early and final aLIGO. Each of the strips shows the region of $M$-$\eta$ space for which the non-spinning waveform has a match of 0.97 or higher with a spinning signal.](M50mismatch.pdf){width="48.00000%"}
In Fig. \[fig:M50matches\] we again show the regions of the non-spinning parameter space which give a match greater than 0.97 with a spinning binary (with values of $ \chi$ from -1 to 1 in steps of 0.1). Here, we show the results for both the early and final aLIGO spectra. The results are qualitatively similar to the $20 M_{\odot}$ binary with spinning signals well recovered by the non-spinning waveforms, and the chirp mass accurately recovered. The sizes of the regions are roughly consistent with the lower mass system, with a mass accuracy of $\sim10$% and range in $\eta$ of 0.05-0.1 for the early noise curve. For both early and design curves, signals with spins between 0.2 and -0.4 have matches above 0.97 with non-spinning waveforms. Interestingly the degeneracy still roughly follows the line of constant chirp mass, even though the merger and ringdown contributed significantly to the SNR of the signal.
Next, we increase the mass to $100 M_{\odot}$ and repeat the analysis. At this mass, the point-particle ISCO is at 40 Hz, so there is essentially no power from the inspiral in the initial detectors, and only a small amount in the early aLIGO. Even at the final sensitivity, the merger and ringdown provide the vast majority of the SNR (a factor of 4 more than the inspiral). Thus one expects that it is really the merger that dictates the waveform as seen by the detector. Figure \[fig:M100matches\] shows results for 100$M_\odot$, again with $\eta = 0.16$. For aLIGO at final design sensitivity, the curves are again lying pretty much along the line of constant chirp mass.
We do not show results for the early aLIGO sensitivity curve in Fig. \[fig:M100matches\], because the results depend strongly on the choice of model, either PhenomB or PhenomC. The variation of the two models with respect to the physical parameters is sufficiently large through merger and ringdown that they lead to qualitatively different results in our mismatch studies, when we use the early aLIGO noise curve, where the merger and ringdown contribute essentially all of the SNR. The two models were developed as pioneering models of aligned-spin IMR waveforms for use in searches, and for this we expect that they are sufficient; but their fidelity with respect to parameter estimation at high masses has not yet been tested, and should be a focus of future work to refine them.
![ \[fig:M100matches\] Matches of spinning $100 M_{\odot}$ waveforms of mass ratio 4:1 with non-spinning templates in design aLIGO. ](M100mismatch.pdf){width="48.00000%"}
Parameter recovery {#sec:parameters}
==================
Once a black hole merger has been detected, we wish to extract the signal parameters as accurately as possible. Parameter estimation proceeds by identifying regions of parameter space which give a signal that is most consistent with the data. It stands to reason that these regions will contain waveforms that have a good match with the observed signal. Thus we might expect that confidence regions in parameter estimation are associated with regions of high match between signal and template. In this section, we show this expectation to hold in detail in the high-SNR limit and derive an expression for the value of the match that corresponds to a given confidence at a given SNR.
We begin by presenting the argument, using the Fisher matrix formulation, and then we use this connection to re-interpret our earlier results in terms of parameter recovery. We follow the formalism used Ref. [@2008PhRvD..77d2001V], but provide only a brief discussion of the Fisher matrix formalism, and refer the reader to the Ref. [@2008PhRvD..77d2001V] (and references therein) for more details.
Connection between mismatch and confidence regions
--------------------------------------------------
Let us assume that a signal $h_{0}$ is present in the data. The detector data is given as $$\label{eq:data}
s(t) = h_{0}(t) + n(t) \, .$$ To investigate signal recovery and parameter extraction at leading order, we Taylor expand the signal in a region of the true parameters ($\theta = 0$) as $$\label{eq:wf_expansion}
h(\theta) = h_{0} + \theta^{i} h_{i} + \ldots$$ Here $h_i = \partial_{i} h$ is used to denote the derivative of the waveform with respect to the paramters $\theta_{i}$, which will include $M$, $\eta$, $\chi$ as well as the amplitude, phase and coalescence time that are maximized over in the match calculations.
The likelihood, for a given set of $\theta_{i}$, is $$p(s| \theta) \propto \exp \left\{- \frac{(s - h(\theta) | s - h(\theta) )}{2} \right\} \, .$$ Substituting the expressions for $s$ (\[eq:data\]) and $h(\theta)$ (\[eq:wf\_expansion\]) into the above and keeping leading order terms gives $$\label{eq:like}
p(s| \theta) \propto \exp \left\{ - \frac{(n|n)}{2} + \theta_{i} (n|h_{i})
- \frac{\theta_{i} \theta_{j} (h_i | h_j)}{2} \right\}$$
In the context of Bayesian parameter estimation, this can be recast in terms of a posterior probability distribution for the parameters $\theta_{i}$ using Bayes theorem: $$p(\theta | s) \propto p(s | \theta) p(\theta)$$ where $p(\theta)$ is the prior probability distribution for the parameters. In what follows, we use a uniform prior on the parameters. In general, such a prior will not be physically motivated but, for a detectable signal, the likelihood will be peaked in a small enough region of parameter space to make this approximation reasonable.
Given the above, we are interested in calculating the expected offset between the true parameters and the mean value from the posterior distribution. We also want to evaluate the size of a confidence region containing a given fraction $p$ of the posterior probability. These quantities give us two different measures of the expected accuracy of parameter recovery.
We begin by calculating the mean of the parameter $\theta_{i}$ as $$\label{eq:mean}
\langle \theta_{i} \rangle = \int d\theta \theta_{i} \, p(\theta | s)
= (h_{i} | h_{j} )^{-1} (n | h_{j}) \, .$$ Thus, the mean of the posterior distribution will be offset from the true parameter values due to the presence of noise. One way to characterize this is the expected size of the error waveform, $h_{E} = h(\langle \theta_{i} \rangle) - h_{0}$ as $$\langle h_{E}^{2} \rangle_{n} = \langle (n | h_{i}) (h_{i} | h_{j} ) (h_{j} | n) \rangle_{n}
= k \, ,$$ where $k$ is the dimension of the parameter space and $\langle \rangle_{n}$ indicates the expectation value over many noise realizations. Thus, on average the difference between the true signal and “best fit” waveform will have an amplitude of $\sqrt{k}$.
Next, we turn our attention to confidence regions in parameter space — a region $\Theta$ of parameter space that contains a given probability $p$ of the posterior distribution, $$p = \int_{\Theta} d\theta p(\theta | s) \, .$$ There are many ways to construct such a region, and one typically also requires the smallest possible region. To calculate confidence regions in the Fisher approximation, we begin by observing that the covariance between parameters is given as $$\label{eq:covariance}
\langle \theta_{i} \theta_{j} \rangle = (h_{i} | h_{j} )^{-1} \, .$$ Using this expression and Eq. (\[eq:mean\]), we can re-express the posterior distribution as $$\begin{aligned}
\label{eq:posterior}
p(\theta | s) &\propto& \exp \left\{- \frac{1}{2} (\theta_{i} - \langle \theta_{i} \rangle) (h_{i} | h_{j})
(\theta_{j} - \langle \theta_{j} \rangle) \right\} \nonumber \\
&\simeq& \exp \left\{- \frac{1}{2} ( | h(\theta) - h(\langle \theta \rangle) |^{2} ) \right\} \, .\end{aligned}$$ Then, the minimum volume region which contains a fraction $p$ of the posterior probability is the one for which $$\label{eq:conf}
| h(\theta) - h(\langle \theta \rangle) |^{2} < \chi^{2}_{k}(1-p) \, ,$$ where $\chi^{2}_{k}(1-p)$ is the chi-square value for which there is $1-p$ probability of obtaining that value or larger, and $k$ denotes the degrees of freedom, determined by the number parameters included in $\theta_{i}$. At leading order, the confidence interval contains all points for which the amplitude of the difference between the model and best fit waveforms lies below the given threshold.
There are six parameters in the aligned-spin waveform model: $M$, $\eta$, $\chi$, $A$, $\phi_{c}$ and $t_{c}$. When calculating matches, we have maximized over the latter three parameters (amplitude, phase and time) and reported match over the three dimensional space of mass, mass ratio and spin. Thus, we have calculated the three dimensional matches $M(h_1 | h_2)$, maximized over A, $\phi_c$ and $t_c$. It is straightforward to re-cast our earlier results in terms of mismatches. To do this, we note that $$\label{eq:diff_as_match}
\min_{A_2} | h_{1} - h_{2} |^{2} =
|h_{1}|^{2} \left\{1 - \frac{(h_{1} | h_{2})^{2}}{|h_{1}|^2 |h_{2}|^{2}}\right\} \, .$$ When we restrict to the subspace of masses and spins, we can write the final term as the match between the waveforms.
By substituting Eq. (\[eq:diff\_as\_match\]) into Eq. (\[eq:conf\]), we see that the confidence region is defined by all points in the parameter space for which the match satisfies $$\label{eq:match_thresh}
M(h(\theta), h(\langle \theta \rangle) \ge 1 - \frac{\chi^{2}_{k}(1-p)}{2 \rho^2}$$ where the value of $k$ is given by the dimension of the remaining parameter space. This expression gives us a straightforward way to re-interpret the match calculations presented earlier. For example, with a three dimensional parameter space, the 90% confidence region at a given SNR is given by: $$M(h(\theta), h(\langle \theta \rangle) \ge 1 - \frac{3.12}{\rho^2}$$ which has the nice property that a match of 0.97 corresponds to a 90% confidence region at an SNR of 10. For a two dimensional parameter space, the right hand side is $1 - 2.3/\rho^2$, meaning that a match of 0.97 corresponds to 90% confidence region at an SNR of 9.
We should note that there are numerous assumptions used in the derivation of these results. Most significantly, all of the Fisher matrix results consider only leading order effects and become less reliable at lower SNR, see e.g., Ref [@2008PhRvD..77d2001V] for a detailed discussion of the issues. When actually calculating confidence regions for a signal, detailed parameter estimation analyses calculate the posterior distribution (\[eq:posterior\]) and integrate to find the region containing 90% of the probability. Here, we are using a hybrid approach: we use the result of the Fisher matrix calculation to decide the threshold on match required to define the given confidence region, but then calculate the match between waveforms exactly, without recourse to any approximations. Furthermore, we are maximizing the match over three dimensions, analytically for $A$ and $\phi$ and using a Fourier transform to search over time. Thus, we are only applying the Fisher result to the three dimensional subspace of mass, mass ratio and spin. Even there, we are merely using the calculation to determine the appropriate match threshold: the matches are calculated exactly. Consequently, the match regions should be in good agreement with the 90% confidence regions. We verified that for an SNR of 20, the region identified by our match criteria did contain 90% of the probability to within $\pm0.5\%$.
Implications for detectability and parameter estimation
-------------------------------------------------------
Let us now return to the results of Sec. \[sec:spin\_eta\] and \[sec:high\_mass\] and use the relationship derived above to re-interpret the results in terms of confidence intervals. We can interpret Fig. \[fig:M20early\] in two ways: in the context of either a non-spinning or spin-aligned templated search. If we were to perform a search with non-spinning templates and observe a $M = 20 M_\odot$, $\eta = 0.16$ binary with non-spinning components with SNR 9, then the 90% confidence region would lie along a strip in the parameter space with a width of 5% in mass, and 10% in $\eta$. For a binary containing the same mass black holes with spins of -0.5, the 90% confidence region would be of approximately the same size (5% in mass, and 10% in $\eta$) but centred at $M = 27 M_\odot$, $\eta = 0.10$.[^2] The same holds for other values of the spins: the reported statistical uncertainty in parameters is relatively small, while the systematic errors can be significant.
Alternatively, we can consider the three dimensional space of $M$-$\eta$-$\chi$. In that case, a match above 0.97 corresponds to a 90% confidence region at SNR of 10. Thus, for example, Fig. \[fig:M20early\] shows that the 90% confidence region for a signal with $m_{1} = 16, m_{2} = 4$ and $\chi \in [-0.7, 0.5]$ would contain waveforms with non-spinning components. Figures \[fig:M20matchesChirp\] to \[fig:M100matches\] can be interpreted in a similar manner. For higher SNRs, the confidence intervals shrink. At SNR 10 they correspond to matches of 0.97, SNR 20 to 0.992 and SNR 30 to 0.9965.
![ \[fig:3Dconfidence\] 90% confidence intervals for a 20$M_\odot$, 1:4 ($\eta = 0.16$) signal with spin $\chi = 0.2$, using the design aLIGO noise curve. The top panel shows the full three-dimensional confidence region for SNR=10; the gray surface indicates the $\chi = 0$ plane, and corresponds to the $\chi = 0.2$ region in Fig. \[fig:M20matches\]. The lower two panels show the same data projected onto the $M$-$\chi$ and $m_1$-$m_2$ planes, and also indicate the 90% confidence regions for SNRs 20 (blue) and 30 (green). The true physical parameters are indicated by a ball (top panel), or star (lower panels). Regions with unphysical $\eta$ ($>0.25$) are not shown. ](3DmatchesP90p02B.pdf "fig:"){width="48.00000%"} ![ \[fig:3Dconfidence\] 90% confidence intervals for a 20$M_\odot$, 1:4 ($\eta = 0.16$) signal with spin $\chi = 0.2$, using the design aLIGO noise curve. The top panel shows the full three-dimensional confidence region for SNR=10; the gray surface indicates the $\chi = 0$ plane, and corresponds to the $\chi = 0.2$ region in Fig. \[fig:M20matches\]. The lower two panels show the same data projected onto the $M$-$\chi$ and $m_1$-$m_2$ planes, and also indicate the 90% confidence regions for SNRs 20 (blue) and 30 (green). The true physical parameters are indicated by a ball (top panel), or star (lower panels). Regions with unphysical $\eta$ ($>0.25$) are not shown. ](MatchesP90p02Ab.pdf "fig:"){width="48.00000%"} ![ \[fig:3Dconfidence\] 90% confidence intervals for a 20$M_\odot$, 1:4 ($\eta = 0.16$) signal with spin $\chi = 0.2$, using the design aLIGO noise curve. The top panel shows the full three-dimensional confidence region for SNR=10; the gray surface indicates the $\chi = 0$ plane, and corresponds to the $\chi = 0.2$ region in Fig. \[fig:M20matches\]. The lower two panels show the same data projected onto the $M$-$\chi$ and $m_1$-$m_2$ planes, and also indicate the 90% confidence regions for SNRs 20 (blue) and 30 (green). The true physical parameters are indicated by a ball (top panel), or star (lower panels). Regions with unphysical $\eta$ ($>0.25$) are not shown. ](MatchesP90p02Bb.pdf "fig:"){width="48.00000%"}
![ \[fig:3DconfidenceB\] Same as Fig. \[fig:3DconfidenceB\], but now the signal has spin $\chi = -0.2$. ](3DMatchesP90m02B.pdf "fig:"){width="48.00000%"} ![ \[fig:3DconfidenceB\] Same as Fig. \[fig:3DconfidenceB\], but now the signal has spin $\chi = -0.2$. ](MatchesP90m02Ab.pdf "fig:"){width="48.00000%"} ![ \[fig:3DconfidenceB\] Same as Fig. \[fig:3DconfidenceB\], but now the signal has spin $\chi = -0.2$. ](MatchesP90m02Bb.pdf "fig:"){width="48.00000%"}
We can use the analysis of the previous section to estimate the 90% confidence intervals if we employ an aligned-spin model for parameter estimation. For signals with an SNR of 10, the three-dimensional confidence region in $(M,\eta,\chi)$ will correspond to matches greater than 0.97. In these cases we calculate the matches between a given aligned-spin signal, and all other aligned-spin model waveforms with varying mass, mass-ratio and total spin. In Fig. \[fig:3Dconfidence\] we show the 0.97 match volume for a 20$M_\odot$ 1:4 binary with $\chi = 0.2$, using the final aLIGO noise curve. The top panel shows the full three-dimensional confidence region, and in the lower panels the results are projections onto the $M$-$\chi$ and $m_1$-$m_2$ planes, to aid the interpretation. Figure \[fig:3DconfidenceB\] shows similar results, but for a binary with $\chi = -0.2$.
Since our waveform model now includes spin, the best match is unity at the correct parameters. But from the figures we see that the 90% confidence region extends well beyond the correct parameters, and is far from the naive image conjured by the term “error ellipse”. We note in particular that this volume includes a region of the non-spinning binary parameter space; the intersection with the $\chi = 0$ plane is consistent with the lower panel of Fig \[fig:M20matches\]. If we were to estimate the parameters of this signal with an aligned-spin waveform model, we would find that the best-fit parameters indicated that the black holes were spinning, but could not rule out that they might in fact be non-spinning — or indeed have spins with the opposite orientation.
The reason for this large uncertainty in the parameters is that the SNR is low; it is only 10. For an SNR of 20, the 90% confidence interval corresponds to a match of above 0.992, and for an SNR of 30, the match must be above 0.9965. We see that for these higher SNRs, the confidence region shrinks. However, only at SNR 30 is a non-spinning signal excluded from the 90% confidence region for the $\chi = 0.2$ case, and even at this SNR (which may be quite rare in aLIGO observations), the smaller mass is determined to within only 25%. When $\chi = -0.2$ (Fig. \[fig:3DconfidenceB\], the 90% confidence region includes spins as high as $\chi = -0.7$, even when the SNR is 20. Note also that in these figures we have included only those portions of the confidence regions with $\eta \leq 0.25$, i.e., physically acceptable values; the confidence region would extend further in the upper two panels if the model were not constrained.
Conclusions {#sec:conclusion}
===========
We have investigated the degeneracy between mass-ratio and spin in gravitational waveforms, going beyond inspiral models to include merger and ringdown signals from black-hole-binary mergers. We have used phenomenological inspiral-merger-ringdown (IMR) models to study the subset of the full binary parameter space that includes non-spinning black holes, and black holes with spins parallel or anti-parallel to the orbital angular momentum of the binary.
A degeneracy between mass-ratio and spin is already known for inspiral signals, which are relevant to ground-based gravitational-wave detectors for masses $M \leq 12$$M_\odot$ [@Buonanno:2009zt; @Brown:2012qf]. We find that this degeneracy persists at higher masses, where the detectors are also sensitive to the merger and ringdown. This means that in a GW search that uses only non-spinning binary templates (and this is computationally cheaper than including spin), the signal may still be detected, but the best-match template will have strongly biassed parameters. The mass-ratio–spin degeneracy follows lines of constant chirp mass, so the chirp mass will be recovered with reasonable accuracy, even up to high masses and for spinning signals. However, the total mass and mass-ratio of the best-matched template will be biased. As shown in Fig. \[fig:M20early\], binaries with higher aligned spins will be recovered with a higher mass ratio, while those with anti-aligned spins will be recovered with a lower mass ratio. If we restrict the search to physical values of the symmetric mass ratio $\eta$, then comparable-mass binaries whose spins are aligned with the orbital angular momentum will tend to be missed in the search. We show how these results will change as the detector evolves towards its final sensitivity in Fig. \[fig:M20matches\], and for higher masses in Figs. \[fig:M50matches\] and \[fig:M100matches\].
We also demonstrated that it is possible to use match calculations to estimate the confidence intervals in parameter estimation. For example, the 90% confidence region for the three-dimensional parameter space of ($M$, $\eta$, $\chi$) for signals with SNR 10 is given by the region with matches above $\sim0.97$. This allows us to estimate how the mass-rato–spin degeneracy will be reflected in parameter estimation. We show that for modest SNRs ($\sim10$) it may be difficult to determine whether a binary includes spin, and even for high SNRs ($\sim30$), the mass-ratio–spin degeneracy impairs the accurate recovery in the individual black-hole masses; see Figs. \[fig:3Dconfidence\] and \[fig:3DconfidenceB\]. These results will affect the astrophysical conclusions that can be drawn from GW observations, and this will be explored in more detail in future work. It would also be interesting to exploit our knowledge of the mass-ratio–spin degeneracy in a jump proposal for Bayesian parameter-estimation codes [@Sluys:2008a; @Sluys:2008b; @Veitch:2010].
Our results are restricted to aligned-spin binaries, and will be affected by the inclusion of precession effects. At present, no complete inspiral-merger-ringdown model for precessing systems exists. It seems unlikely, however, that the inclusion of precession effects will reduce the size of the confidence regions. In general it is far more likely that increasing the dimensionality of the model parameter space will increase the parameter uncertainty, because the confidence regions will now correspond to lower matches. The recent results in Ref. [@Schmidt:2012rh] show that precession effects have only a weak impact on the phasing, which suggests that while the inclusion of precession is unlikely to break the degeneracy discussed here, it is also unlikely to introduce an *additional* degeneracy in these three parameters. On the other hand, the inclusion of higher harmonics, which were ignored in this study, may improve the accuracy of the parameters. The net impact of these two effects remains to be studied in more detail in future work.
Acknowledgments {#acknowledgments .unnumbered}
===============
S.F. and M. H. were supported by Science and Technology Facilities Council grants ST/H008438/1 and ST/I001085/1. S.F. also acknowledges the support of the Royal Society. We are particularly grateful to P. Ajith, Sascha Husa and Frank Ohme for sharing with us their implementations of the PhenomB and PhenomC waveform families and match-calculation code. We also thank Chris Messenger and John Veitch for useful discussions.
[^1]: The figures may seem to indicate that a $20 M_{\odot}$ binary with total spin of 0.5 that would be detected by a non-spinning search in the early aLIGO configuration, but not later ones. This is of course not true: the final detector is roughly three times more sensitive at these masses, and so a signal with SNR 10 in early aLIGO would have an SNR of 30 in the final configuration. Even with a match of 0.9, this would still give an SNR of 27, and the match would be sufficient that it would pass any signal consistency tests.
[^2]: Strictly speaking, if the best match between the signal and a model waveform is less than unity, the relation between match and confidence region in equation (\[eq:match\_thresh\]) must be modified to reflect this. The 90% confidence region will contain all points with a match of 0.97 or greater with the *best fit* model waveform.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'When neutrino masses arise from the exchange of neutral heavy leptons, as in most seesaw schemes, the effective lepton mixing matrix $N$ describing neutrino propagation is non-unitary, hence neutrinos are not exactly orthonormal. New CP violation phases appear in $N$ that could be confused with the standard phase $\delta_{\text{CP}}$ characterizing the three neutrino paradigm. We study the potential of the long-baseline neutrino experiment DUNE in probing CP violation induced by the standard CP phase in the presence of non-unitarity. In order to accomplish this we develop our previous formalism, so as to take into account the neutrino interactions with the medium, important in long baseline experiments such as DUNE. We find that the expected CP sensitivity of DUNE is somewhat degraded with respect to that characterizing the standard unitary case. However the effect is weaker than might have been expected thanks mainly to the wide neutrino beam. We also investigate the sensitivity of DUNE to the parameters characterizing non-unitarity. In this case we find that there is no improvement expected with respect to the current situation, unless the near detector setup is revamped.'
author:
- 'F. J. Escrihuela $^1$'
- 'D. V. Forero $^2$'
- 'O. G. Miranda $^3$'
- 'M. Tórtola $^1$'
- 'J. W. F. Valle $^1$'
title: |
Probing CP violation with non-unitary mixing\
in long-baseline neutrino oscillation experiments:\
DUNE as a case study
---
Introduction
============
Following the celebrated discovery of neutrino oscillations [@Kajita:2016cak; @McDonald:2016ixn] subsequent accelerator and reactor studies have brought neutrino physics to the mature phase of precision studies. Sensitive laboratory oscillation studies not only play a key role in confirming the neutrino oscillation hypothesis, but also rule out exotic solutions, establishing the robustness of the simplest three neutrino paradigm.
Given its importance, more than ever it has become relevant to critically assess with improved sensitivity the robustness of the determination of the three-neutrino oscillation parameters within recent and current studies [@Adamson:2016tbq; @An:2016ses; @Kim:2016yvm; @An:2012eh; @Ahn:2012nd; @Abe:2012gx; @Abe:2012tg; @Abe:2015awa] as well as future experiments [@Acciarri:2015uup]. This includes the scrutiny of the uncertainties associated with neutrino fluxes, propagation and interactions. These may arise, for example, from helioseismology [@christensen-dalsgaard:2002ur; @Balantekin:2013tqa], solar chemical composition and solar fusion reactions [@Adelberger:2010qa], density fluctuations deep within the Sun [@balantekin:1996pp; @nunokawa:1996qu] as well as magnetic fields in the radiative [@burgess:2003fj; @burgess:2003su; @burgess:2002we] and convective zones [@miranda:2000bi; @barranco:2002te; @Miranda:2003yh; @Miranda:2004nz]. On the other hand the subleading role of neutrino non-standard interactions upon oscillations has been considered in various contexts and can also bring new sources of CP violation [@fornengo:2000sr; @GonzalezGarcia:2001mp; @Miranda:2004nb; @deGouvea:2015ndi; @Forero:2016cmb]. These issues have been widely explored, so here we focus on the impact of non-unitarity of the lepton mixing matrix upon neutrino propagation and the resulting expected sensitivities on the three–neutrino CP phase determination [@Goswami:2008mi; @Escrihuela:2015wra; @Miranda:2016wdr].
Non-unitarity of the lepton mixing matrix constitutes a most generic feature of schemes where neutrino masses arise from the exchange of fermionic messengers [@valle:1987gv; @Miranda:2016ptb] such as the type-I seesaw mechanism [@GellMann:1980vs; @Yanagida:1979as; @mohapatra:1980ia; @Schechter:1980gr]. Indeed there is a large class of low-scale variants of the seesaw mechanism, such as inverse and linear seesaw [@Mohapatra:1986bd; @gonzalezgarcia:1989rw; @Akhmedov:1995vm; @Malinsky:2005bi], where these right-handed neutrino messengers are not-so-heavy, as their masses could lie within reach of the LHC experiments. In this case one expects sizeable departures from unitarity in the lepton mixing matrix characterizing the light neutrino sector [@Goswami:2008mi; @Escrihuela:2015wra]. This brings in CP violation associated to the messenger sector into the physics describing the propagation of the light neutrinos [@Miranda:2016ptb]. The presence of unitarity violation makes it difficult to extract reliable information on leptonic and, indeed, first quantitative studies indicate the existence of a potentially serious ambiguity in probing CP violation in neutrino oscillations in such case [@Miranda:2016wdr]. As a result, dedicated leptonic CP violation studies taking into account the non-unitarity of the lepton mixing matrix will be necessary. Such studies can shed light on the seesaw scale, and thereby provide valuable insight on the scale of new physics responsible for neutrino mass generation.
In this paper we focus on the possible ambiguities in the CP phase determination for the upcoming DUNE experiment, including matter effects in a consistent way. The paper is organized as follows. In order to set up the framework in section \[sec:prel-updat-bounds\] we compile and update the bounds on the relevant parameters. These follow, for instance, from weak universality tests and short–distance neutrino oscillation searches. In section \[sec:non-unit-oscill\] we discuss the neutrino effective matter potential in the presence of non-unitarity and present the corresponding results for the oscillation probabilities in matter. In section \[sec:nuDUNE\] we study the sensitivity of the DUNE experiment to non-unitary neutrino mixing. First we discuss the determination of the standard three-neutrino CP phase $\delta_{\text{CP}}$ and the possible confusion with the seesaw phase. Finally we analyze the potential capability of DUNE in further constraining the non-unitarity of the light neutrino mixing matrix. We find that, although the potential to probe CP violation is somewhat degraded with respect to the unitary case, the effect is weaker than expected [@Miranda:2016wdr] thanks to the good statistics expected and the relatively wide band neutrino beam at DUNE. Further discussion and conclusions are summarized in section \[sec:conclusions\].
preliminaries: prior constraints {#sec:prel-updat-bounds}
=================================
Within a large variety of seesaw schemes the lepton mixing matrix describing the propagation of the light neutrinos is effectively non-unitary, hence these neutrino states are not exactly orthonormal [@valle:1987gv]. The description of this situation can be readily obtained by truncating the symmetrical parametrization of the full rectangular lepton mixing matrix characterizing general seesaw schemes, first given in [@Schechter:1980gr]. The resulting form can be written as [@Escrihuela:2015wra] $$N
=
N^{NP} U
=
\left\lgroup
\begin{array}{ccc}
\alpha_{11} & 0 & 0\\
\alpha_{21} & \alpha_{22} & 0\\
\alpha_{31} & \alpha_{32} & \alpha_{33}
\end{array}
\right\rgroup U \,,
\label{eq:NU}$$ where $U$ is the conventional unitary mixing matrix describing neutrino propagation in the standard case, and the pre-factor parametrizes the deviations from unitarity. This convenient description is general and holds for any number of extra neutrino states [@Schechter:1980gr; @Rodejohann:2011vc]. It involves three real parameters $\alpha_{11},\alpha_{22}$ and $\alpha_{33}$ (all close to one) plus three small complex parameters $\alpha_{21},\alpha_{31},\alpha_{32}$ containing extra CP violation. The resulting form provides the most general framework to describe neutrino oscillations relaxing the unitarity approximation.
In order to set the stage for our analysis we first give a brief review on the constraints on non-unitarity parameters. In what follows we update the discussion given in [@Escrihuela:2015wra; @Miranda:2016ptb], e.g. by including recent results for observables coming from pion decay studies [@Aguilar-Arevalo:2015cdf]. We also discuss the interplay, as well as the complementarity, of various “prior” restrictions with what can be learned by direct neutrino studies. The bottom-line of our discussion will be that few of these constraints are of general validity, most are model-dependent.
Weak interaction without universality: formalism {#sec:weak-inter-with}
------------------------------------------------
Here we show how, from general considerations, the constraints from weak no-universality translate into restrictions on non-diagonal $\alpha_{ij}$ parameters. In order to see this, we consider the parametrization of the non-unitary lepton mixing matrix in Eq. \[eq:NU\]. The diagonal entries of the pre-factor matrix are given as a product of cosines [@Escrihuela:2015wra]: $$\alpha_{11} \: = \: c_{1\, n}\: c_{1\,n-1}c_{1\, n-2}\ldots c_{14} ,$$ while the non-diagonal parameters are expressed as [@Escrihuela:2015wra]: $$\begin{aligned}
\alpha_{21} \: &=& \:
c_{2\, n}\: c_{\,2n-1}\ldots c_{2\, 5}\:{\eta}_{24}\bar{\eta}_{14}\: +\:
c_{2\, n}\: \ldots c_{2\, 6}\:{\eta}_{25}\bar{\eta}_{15}\:c_{14} \nonumber\\
&+&\: \ldots\:+ {\eta}_{2n}\bar{\eta}_{1n}\:c_{1n-1}\:c_{1n-2}\:\ldots\:c_{14} \end{aligned}$$ with the phase factors ${\eta}_{ij}=e^{-i\phi_{ij}}\,\sin\theta_{ij}$ and $\bar{\eta}_{ij}=-e^{i\phi_{ij}}\,\sin\theta_{ij}$ [@Schechter:1980gr]. Since the “heavy” iso-singlet admixture is assumed to be small, within the framework of seesaw schemes, as well as from experimental evidence [@Astier:2003gs], we now treat unitarity violation as a perturbation, making use of an small-angle expansion in $\theta_{i\beta}$, with $\beta > 3$, so that $$\alpha^2_{11} \: \simeq
\: 1-\sum^N_{i=4}\theta^2_{1\,i}\:.$$ On the other hand one can show that $$\begin{aligned}
\alpha_{21} \: &\simeq& \:
-\theta_{24}\theta_{14}e^{-i(\phi_{24}-\phi_{14})}
\: -\:
\theta_{25}\theta_{15}e^{-i(\phi_{25}-\phi_{15})}
\ldots -
\theta_{2n}\theta_{1n}e^{-i(\phi_{2n}-\phi_{1n})} \end{aligned}$$ so that $$\begin{aligned}
|\alpha_{21}|^2
\: &\simeq& \:
| \theta_{24}\theta_{14}e^{-i(\phi_{24}-\phi_{14})}
\: +\:
\theta_{25}\theta_{15}e^{-i(\phi_{25}-\phi_{15})}
\ldots: +
\theta_{2n}\theta_{1n}e^{-i(\phi_{2n}-\phi_{1n})} |^2 \nonumber \\
&\leq& \sum^N_{i=4} |\theta_{2i}\theta_{1i}e^{-i(\phi_{2i}-\phi_{1i})} |^2
~=~ \sum^N_{i=4} \theta^2_{2i}\theta^2_{1i} \nonumber\end{aligned}$$ Now from the triangle inequality relation one can write $$\sum^N_{i=4} \theta^2_{2i}\theta^2_{1i} \leq
\left(\sum^N_{i=4}\theta^2_{2\,i} \right)
\left(\sum^N_{i=4}\theta^2_{1\,i} \right)$$ which implies the relation $$|\alpha_{21}| \le \sqrt{(1-\alpha_{11}^2)(1-\alpha_{22}^2)} \label{eq:CScond21} \\$$ and similar relations will hold for $\alpha_{31}$ and $\alpha_{32}$, namely, $$\begin{aligned}
|\alpha_{31}| &\le& \sqrt{(1-\alpha_{11}^2)(1-\alpha_{33}^2)} \label{eq:CScond}\\
|\alpha_{32}| &\le& \sqrt{(1-\alpha_{22}^2)(1-\alpha_{33}^2)} \nonumber\end{aligned}$$ One sees that in the limit of small heavy singlet messenger admixture one has, $$\alpha_{ii} \sim 1~~~\mathrm{and}~~~ \alpha_{ij} \ll1.$$ These relations imply additional restrictions on non-diagonal entries coming from constraints on the diagonal ones. In the next section we will see that bounds on diagonal entries are relatively strong reinforcing the bounds on non-diagonal ones. This also implies that and rates in the charged sector are constrained mainly by universality restrictions, not by the smallness of neutrino masses themselves. This important observation has previously been made in a number of papers and reviews [@bernabeu:1987gr; @gonzalez-garcia:1992be; @rius:1989gk]. In the next subsection we compile bounds from universality as well as from the relevant neutrino oscillation experiments.
Universality constraints {#sec:univ-constr}
------------------------
The non-unitarity of the light neutrino mixing matrix can be constrained by several observables related to weak universality.
- As has been widely discussed in the literature [@Gronau:1984ct; @Langacker:1988ur; @GonzalezGarcia:1990fb; @Nardi:1994iv; @Abada:2012mc; @Atre:2009rg; @Abada:2013aba; @Abada:2014nwa; @Antusch:2014woa], the comparison of measurements of muon and beta decay rates can constrain the non-unitarity of the neutrino mixing matrix. For example, the Fermi constant value for muon and beta decay will be proportional to different non-unitary parameter combinations: $$\label{eq1}
G_{\mu}=G_F\, \sqrt{(NN^\dagger)_{11}(NN^\dagger)_{22}}\,
= G_F\, \sqrt{\alpha_{11}^2(\alpha_{22}^2+|\alpha_{21}|^2)},$$ and $$\label{eq1A}
G_{\beta}=G_F\, \sqrt{(NN^\dagger)_{11}}\,
= G_F\, \sqrt{\alpha_{11}^2}.$$ This will imply that the CKM elements $V_{ud}$ and $V_{us}$, proportional to the Fermi constant $G_{\mu}$, should be corrected by the corresponding factor and expressed as [@Langacker:1988ur; @GonzalezGarcia:1990fb; @Nardi:1994iv]: $$\label{eq3}
\sum_{i=1}^{3} |V_{ui}|^2=
\left(\frac{G_\beta}{G_\mu}\right)^2 =
\left(\frac{G_F\sqrt{(NN^\dagger)_{11}}}{G_F\sqrt{(NN^\dagger)_{11}(NN^\dagger)_{22}}}\right)^2
=\frac{1}{(NN^\dagger)_{22}},$$ The experimental value of this expression is given by [@Olive:2016xmw]: $$\label{eq11}
\sum_{i=1}^{3} |V_{ui}|^2=\frac{1}{\alpha_{22}^2 + |\alpha_{21}|^2}
=0.9999\pm 0.0006,$$
- The mass of the $W$ boson, $M_W$, is related with the values of the weak mixing angle, $s_W$, and the Fermi constant. Including radiative corrections, in the On-Shell renormalization scheme, this relation can be written as [@Olive:2016xmw]: $$M_{W} = \frac{A_{0}}{s_{W} (1-\Delta r)^{1/2}}
\, ,$$ with $$\begin{aligned}
A_{0} &=& \left(\frac{\pi \alpha}{\sqrt{2}G_{F}}\right)^{1/2}
\, , \\
s^2_{W} &=& 0.22336 \pm
0.00010 \, , \\
\Delta r &=& 0.03648 \pm 0.00031 \, ,\end{aligned}$$ where $\Delta r$ includes the radiative corrections relating $\alpha$, $\alpha(M_{Z})$, $G_F$, $M_W$ and $M_Z$. In the non-unitary case, the Fermi constant should take into account the corresponding corrections and the prescription for $A_0$ will be:
$$A_{0} =
\left(\frac{\pi \, \alpha \, \sqrt{\alpha^{2}_{11}(\alpha^{2}_{22} + |\alpha^{2}_{21}|)}}{\sqrt{2}G_{\mu}}\right)^{1/2}
\, .$$
- The couplings between leptons and gauge bosons are dictated by gauge symmetry. For the standard case of lepton unitarity these are flavor independent. This feature is no longer true in the presence of non-unitarity. As a result, the ratio between two different semileptonic decay rates would constrain non-unitarity parameters. For example, for the case of pion decay we have [@Abada:2012mc]: $$R_{\pi} = \frac{\Gamma (\pi^+ \to e^+\nu)}{\Gamma (\pi^+ \to \mu^+\nu)}=
\frac{(NN^\dagger)_{11}}{(NN^\dagger)_{22}} =
\frac{\alpha_{11}^2}{\alpha_{22}^2 + |\alpha_{21}|^2}.$$ Here we include the updated measurement from Ref. [@Aguilar-Arevalo:2015cdf] and theoretical prediction in Refs. [@Cirigliano:2007xi; @Bryman:2011zz]: $$\label{eq10}
r_\pi = \frac{R_\pi}{R_\pi^{SM}} =
\frac{(1.2344\pm 0.0029)\times 10^{-4}}{(1.2352\pm 0.0002)\times 10^{-4}}=
0.9994\pm0.0030 .$$ Notice that this constraint is more restrictive than the previously reported value, $r_\pi=0.9956\pm0.0040$ [@Czapek:1993kc]. One also has the corresponding bound from Kaon-decay [@Abada:2012mc] $$\label{eq10}
r_K= \frac{R_{K}}{R^{SM}_{K}} =
\frac{(NN^\dagger)_{11}}{(NN^\dagger)_{22}} =
\frac{\alpha_{11}^2}{\alpha_{22}^2 + |\alpha_{21}|^2} =
\frac{(2.488\pm 0.010)\times 10^{-5}}{(2.477\pm 0.001)\times 10^{-5}}=1.004\pm0.010.$$ However, this limit does not play a significant role, since the pion decay measurements are more restrictive.
- Likewise, for the case of $\mu-\tau$ universality there are restrictions that follow from the ratio of the decay of the meson ($\pi^{-}$ or $K^{-}$) to a muon plus a muon neutrino, or from the tau decay to a meson and a tau neutrino [@Nardi:1994iv; @Escrihuela:2015wra; @Pich:2013lsa]:
$$R_{\tau / P} = \frac{\Gamma (\tau^- \to P^-\nu_{\tau})}
{\Gamma (P^- \to \mu^+\nu_{\mu})} \propto
{\left|\frac{g_{\tau}}{g_{\mu}}\right|}^{2} =
\frac{\alpha_{33}^2 + |\alpha_{32}|^2 + |\alpha_{31}|^2
}{\alpha_{22}^2 + |\alpha_{21}|^2},$$
where $P^{-}$ stands for either $\pi^{-}$ or $K^{-}$ mesons. Several ratios can be considered and included in the analysis. In particular we have considered the results reported in Ref. [@Pich:2013lsa].\
- On the other hand, for the $e$ - $\tau$ sector, we have considered only pure leptonic decays as well as direct leptonic decays of $W$ boson, which lead to $${\left|\frac{g_e}{g_{\tau}}\right|}^2 = \frac{\alpha_{11}^2}{\alpha_{33}^2 + |\alpha_{32}|^2 + |\alpha_{31}|^2
} \, .
\label{eq:g_e_tau}$$ The value of $|g_{e}/g_{\tau}|$ ratio for each process was presented in Ref. [@Pich:2013lsa].
- Non-unitarity can affect the neutral current couplings. As noted in [@Schechter:1980gr], these are no longer “trivial” as in the [standard model ]{}since the couplings of light neutrinos to the Z-boson can be non-diagonal in the mass basis. Moreover the diagonal coupling strengths are smaller than in the [standard model ]{}thereby decreasing the invisible Z width, well measured at LEP and reported to be slightly smaller than three ($2.9840\pm0.0082$) [@ALEPH:2005ab]. However, neutral currents have a more complex structure that will depend both on the values of the $\alpha$ parameters as well as on the values of the three by three matrix $U^{3\times3}$. Given this complexity and the quadratic dependence on the $\alpha$, it is safe not to include this observable into the analysis.
Concerning searches for lepton flavor and CP violating processes we notice that these do not give us any independent robust constraint on unitarity violation. Indeed, such processes may proceed in the absence of neutrino mass and are only restricted by weak universality tests [@bernabeu:1987gr; @branco:1989bn; @rius:1989gk].
[ccc|cc]{} &\
&\
& $90$% C.L. & $3\sigma$ & $90$% C.L. & $3\sigma$\
\
$\alpha_{11}>$ &$0.9974$ &$0.9963$&$0.9961$ &$0.9952$\
$\alpha_{22}>$ &$0.9994$&$0.9991$&$0.9990$&$0.9987$\
$\alpha_{33} >$ & $0.9988$ & $0.9976$& $0.9973$ & $0.9961$\
$|\alpha_{21}|<$ & $\;1.7\times10^{-3}\;$ & $\;2.5\times10^{-3}\;$& $\;2.6\times10^{-3}\;$ & $\;4.0\times10^{-3}\;$\
$|\alpha_{31}|<$ & $2.0\times10^{-3}$ & $4.4\times10^{-3}$& $5.0\times10^{-3}$ & $7.0\times10^{-3}$\
$|\alpha_{32}|<$ & $1.1\times10^{-3}$ & $2.0\times10^{-3}$& $2.4\times10^{-3}$ & $3.4\times10^{-3}$\
\
$\alpha_{11}>$ & $0.98$& $0.95$ & $0.96$ & $0.93$\
$\alpha_{22}>$ & $0.99$ & $0.96$ & $0.97$ & $0.95$\
$\alpha_{33}>$ & $0.93$ & $0.76$ & $0.79$ & $0.61$\
$|\alpha_{21}|<$ & $1.0 \times10^{-2}$ & $2.6\times10^{-2}$ & $2.4\times10^{-2}$ & $3.6 \times10^{-2}$\
$|\alpha_{31}|<$ & $4.2 \times10^{-2}$ & $9.8\times10^{-2}$ & $9.0\times10^{-2}$ & $1.3\times10^{-1}$\
$|\alpha_{32}|<$ & $9.8 \times10^{-3}$ & $1.7\times10^{-2}$ & $1.6\times10^{-2}$ & $2. 1 \times10^{-2}$\
Neutrino oscillation constraints
--------------------------------
Direct constraints on the non-diagonal elements of the $N$ matrix come from the so-called zero distance effect in the conversion probability [@valle:1987gv]. For example, the conversion probability from muon to electron neutrinos can be written as [@Escrihuela:2015wra], $$P_{\mu e} \simeq
(\alpha_{11}\alpha_{22})^2 P^{3\times3}_{\mu e}
+ \alpha_{11}^2 \alpha_{22}|\alpha_{21}| P^{I}_{\mu e}
+ \alpha_{11}^2|\alpha_{21}|^2 ,
\label{eq:Pmue}$$ after neglecting cubic products of the small parameters $\alpha_{21}$, $\sin\theta_{13}$ and $\Delta m^2_{21}$. Here, $P^{3\times3}_{\mu e}$ stands for the standard conversion probability in the unitary case, while the interference probability term $P^{I}_{\mu e}$ depends on the non-unitarity parameters, including an additional $CP$ phase. Finally, the last term in this expression is a constant factor, independent of the distance travelled by the neutrino and its energy. Therefore, any neutrino appearance experiment in the $\nu_\mu \to \nu_e$ channel would be sensitive to this zero distance (0d) contribution: $$P^{0d}_{\mu e} = \alpha_{11}^2 |\alpha_{21}|^2 .$$ There is a similar expression for the conversion probability in the $\nu_e \to \nu_\tau$ and $\nu_\mu \to \nu_\tau$ channels. In the latter case, the oscillation probability formula is slightly more complicated, but at leading order in the non-unitary parameters, one can approximate both zero-distance appearance probabilities by $$\begin{aligned}
P^{0d}_{e \tau} & = \alpha_{11}^2 |\alpha_{31}|^2 \, , \\
P^{0d}_{\mu \tau} & \simeq \alpha_{22}^2 |\alpha_{32}|^2 \, .\end{aligned}$$ We have used these expressions to obtain direct constraints on the parameters $|\alpha_{21}|$, $|\alpha_{31}|$ and $|\alpha_{32}|$ using the negative searches from the NOMAD and CHORUS short-baseline experiments. NOMAD [@Astier:2003gs; @Astier:2001yj] reported limits on the search for $\nu_{\mu} \to \nu_{e}$ as well as $\nu_{\mu} \to \nu_{\tau}$ oscillations in a predominantly $\nu_{\mu}$ neutrino beam produced by the SPS at CERN, while CHORUS [@Eskut:1997ar; @Eskut:2007rn] used the same beam to search for $\nu_{\mu} \to \nu_{\tau}$ oscillations. Additionally, from the contamination of electron neutrinos in the beam, they were also able to constrain the oscillation channel $\nu_e \to \nu_{\tau}$. The stronger bounds at 90% C.L. from these experiments, obtained by NOMAD, can be summarized as: $$\begin{aligned}
P^{0d}_{\mu e} & < &7.0\times 10^{-4} \nonumber \\
P^{0d}_{\mu \tau} & < &1.6\times 10^{-4} \\
P^{0d}_{e \tau} & < & 0.74\times 10^{-2} \nonumber
\label{eq:0d-bounds}\end{aligned}$$ Similar constraints can also be obtained from the NuTeV data [@Avvakumov:2002jj]. Note that, in addition to the short-baseline experiments discussed above, there are also nontrivial constraints arising from medium and long-baseline experiments in combination with atmospheric and solar neutrino data [@Parke:2015goa; @Qian:2013ora]. For maximal values of the diagonal parameters $\alpha_{ii}$, one can summarize the bounds obtained in [@Parke:2015goa] in terms of 3$\sigma$ limits on the non-diagonal parameters: $$\begin{aligned}
|\alpha_{21}| < & 0.03 \nonumber \\
|\alpha_{31}| < & 0.11 \\
|\alpha_{32}| < & 0.12 \nonumber
\label{eq:osc-bounds}\end{aligned}$$
We stress that the above constraints coming from neutrino oscillations are independent of the mass scale of the heavy neutrinos. Therefore, they can be used to constrain the non-unitarity of the lepton-mixing matrix independent on the heavy mass scale. These are the only fully model-independent constraints. Therefore, such neutrino-data-only bounds play a special role and for this reason have been separated as the lower part in Table \[tab:bounds\]. This Table summarizes all the available non-unitarity bounds discussed in this section. Clearly, as seen from the upper part of Table \[tab:bounds\], one can see that universality tests provide strong constraints on the diagonal parameters, $\alpha_{ii}$, that are very close to unity, independently of the number of degrees of freedom considered. In addition, one can also combine with the relations in Eqs. (\[eq:CScond21\]) and (\[eq:CScond\]) in order to obtain stronger constraints on the non-diagonal $\alpha$ parameters. Indeed, by combining universality bounds with these relations one finds that the constraints on the non-diagonal parameters are of order $10^{-3}$.
However we note that these limits are all derived from charged current induced processes under the restrictive assumption that there is no new physics other than that of non-unitary mixing. As an example, we note that the presence of neutrino-scalar Yukawa interactions, absent in the [standard model ]{}but present in models with extra Higgs bosons, such as multi-Higgs schemes (e.g. incorporating flavor symmetries), would potentially avoid these bounds. Likewise, the presence of right-handed charged current contributions expected within a left-right symmetric seesaw scheme would have the same effect. This happens if the extra scalar or vector-mediated contributions compensate the unitarity violation effect [^1]. Of course one may go beyond the above well-motivated assumptions and consider, for the sake of generality, the most general Lorentz structure for the charged weak interactions [^2]. In such case these limits would be invalidated, leading us to regard them as fragile. In contrast, the constraints from neutrino experiments provide a direct restriction on the non-unitarity $\alpha$ parameters. These bounds are significantly less stringent, of the order of $10^{-2}$ for the non-diagonal $\alpha_{ij}$, and correspond to the lower entries in Table \[tab:bounds\].
Finally, there are also direct bounds from searches for neutral heavy leptons. These depend on the mass of the heavy neutrinos, and do not apply beyond the kinematical reach of the high energy experiments, such as LEP [@Dittmar:1989yg; @Akrawy:1990zq; @abreu:1997pa] and LHC [@Klinger]. All mass-dependent limits on light and heavy singlet neutrinos have been compiled in Refs. [@Escrihuela:2015wra; @Drewes:2015iva; @Alekhin:2015byh; @Miranda:2016ptb; @Drewes:2016jae].
In short, at this stage one may adopt two approaches:
- to use as reference the more restrictive bounds coming from charged current weak processes (upper part of Table \[tab:bounds\])
- to use as benchmarks bounds taken strictly from the neutrino sector (lower part of Table \[tab:bounds\]).
While the top limits on the $\alpha$’s are stronger, they are not robust enough for our purposes, so we would recommend to focus on the most direct constraints coming from the bottom part of Table \[tab:bounds\]). In any case in our simulations for the DUNE experiment in order to evaluate its potential in probing leptonic CP violation in the presence of unitarity violation we include as benchmark values not only the conservative, but also the model-dependent bounds, for comparison. The bottom-line is that the DUNE experiment will have the potential of providing independent and robust probes of neutrino properties beyond standard oscillations, properties which can not be probed otherwise in a model-independent way.
Non-unitary oscillation probabilities in matter {#sec:non-unit-oscill}
===============================================
In [@Escrihuela:2015wra; @Miranda:2016wdr] we have given the analytic expressions for the neutrino oscillation probabilities in vacuum [^3]. This approach is valid to study oscillation experiments where matter effects are not very relevant. However, in order to obtain direct sensitivities on the non-unitarity of the lepton mixing matrix from upcoming long-baseline neutrino oscillation experiments such as DUNE or NO$\nu$A, one must have a consistent way to describe matter effects appropriate to this situation. In order to quantify the impact of non-unitary mixing in such experiments, where matter effects are expected to play an important role, one should take into account how the effective matter potential for neutrinos gets modified in the presence of non-unitary three-neutrino mixing. We discuss this issue in the next subsection.
Neutrino effective matter potential in the presence of non-unitarity {#sec:neutr-effect-matt}
---------------------------------------------------------------------
The standard derivation of the effective potential that neutrinos feel when traversing a material medium assumes unitary mixing between the light neutrino species [^4]. In order to derive the neutrino potential in matter for a model with neutral heavy leptons, we note that the complete expression for the neutrino in a flavor state will be given by $$\nu_{\alpha} = \sum_{i}^n K_{\alpha i} \nu_{i} \, ,
\label{eq:nu_field_rel}$$ with the $\alpha$ subscript indicating flavor and $i$ mass eigenstates.
![Feynman diagram illustrating the matter potential associated to the charged current [@Wolfenstein:1977ue]. []{data-label="fig:Feyn_CC"}](Feynman_CC.pdf){width="79.00000%"}
Charged current matter effects in neutrino propagation are illustrated in the Feynman-like diagram in Fig. (\[fig:Feyn\_CC\]) and will be proportional to $$K_{\alpha i} K^{*}_{e i} K_{e j} K^{*}_{\beta j} =
(KK^{\dagger})_{\alpha e} \, (KK^{\dagger})_{e \beta}\, .
\label{eq:KK_rel}$$ Therefore, the charged current potential will be given by $$V^{\alpha \beta}_{CC} =
\sqrt{2} \,G_{F} N_{e} \left(KK^{\dagger}\right)_{\alpha e}
\left(KK^{\dagger}\right)_{e \beta} \, .
\label{eq:V_cc_mass}$$ where $N_e$ is the number density of electrons in the medium and $G_F$ is the Fermi constant. However, the heavy states will not take part in a long baseline neutrino oscillation set up. As a result the sum in Eq. (\[eq:nu\_field\_rel\]) must be performed only up to the third mass eigenstate. Therefore, effectively, one has: $$\nu_{\alpha} = \sum_{i}^3 K_{\alpha i} \nu_{i} = \sum_{i}^3 N_{\alpha i} \nu_{i} \, ,
\label{eq:nu_rel_3}$$ and the effective CC potential in the presence of non-unitarity will be given by: $$V^{\alpha \beta}_{CC} =
\sqrt{2} \,G_{F} N_{e} \left(NN^{\dagger}\right)_{\alpha e}
\left(NN^{\dagger}\right)_{e \beta} \, .
\label{eq:V_cc_mass2}$$ which is expressed in terms of the $\alpha$ parameters as: $$(NN^\dagger)_{\alpha e}(NN^\dagger)_{e \beta}= \alpha_{11}^2
\left(
\begin{array}{ccc}
\alpha _{11}^2 & \alpha _{11} \alpha _{21}^* & \alpha _{11} \alpha _{31}^* \\
\alpha _{11} \alpha _{21} & |\alpha _{21}|^2 & \alpha _{21} \alpha _{31}^* \\
\alpha _{11} \alpha _{31} & \alpha _{21}^* \alpha _{31} & |\alpha _{31}|^2 \\
\end{array}
\right)\, .
\label{eq:CC_alpha_corr}$$ Clearly in the unitary limit ($\alpha_{ii} = 1$ and $\alpha_{ij}=0$), one recovers the well-known Wolfenstein form for the effective CC potential: $$V^{\alpha\beta}_{CC} = \sqrt{2}G_F N_e \delta_{\alpha e}\delta_{\beta e}
\label{eq:V_cc_SM}$$
For the neutral current case we proceed in a similar way. Again we consider the Feynman-like diagram of the NC process as described in Fig. \[fig:Feyn\_NC\] and the neutral current potential is given by $$V^{\alpha \beta}_{NC} =
- \sum_{\rho} \frac{1}{\sqrt{2}}G_{F} N_{n} \left( KK^{\dagger}
\right)_{\alpha \rho} \left( KK^{\dagger} \right)_{\rho \beta}\, .
\label{eq:V_NC}$$
![Feynman diagram illustrating the matter potential associated to the neutral current [@valle:1987gv]. []{data-label="fig:Feyn_NC"}](Feynman_NC.pdf){width="79.00000%"}
After truncating the rectangular $K$ matrix into the square matrix $N$, we obtain that the NC contribution to the matter potential is given by: $$V^{\alpha \beta}_{NC}=-\sqrt{2}G_F \frac{N_n}{2}
\sum_{\rho}(NN^\dagger)_{\alpha \rho}(NN^\dagger)_{\rho
\beta}=-\sqrt{2}G_F \frac{N_n}{2} \left[(NN^\dagger)^2
\right]_{\alpha \beta} \, ,
\label{eq:V_NC_corr}$$ where the matrix product $(NN^\dagger)^2 $ at leading order in the non-diagonal $\alpha$’s is given by: $$\label{eq:matt-pot-app}
\left(
\begin{array}{ccc}
\alpha_{11}^4 & \alpha_{11} \alpha^*_{21} \left(\alpha_{11}^2+\alpha_{22}^2\right) & \alpha_{11} \alpha^*_{31} \left(\alpha_{11}^2+\alpha_{33}^2\right) \\
\alpha_{11} \alpha_{21} \left(\alpha_{11}^2+\alpha_{22}^2\right) & \alpha_{22}^4 & \alpha_{22} \alpha^*_{32} \left(\alpha_{22}^2+\alpha_{33}^2\right) \\
\alpha_{11} \alpha_{31} \left(\alpha_{11}^2+\alpha_{33}^2\right) & \alpha_{22} \alpha_{32} \left(\alpha_{22}^2+\alpha _{33}^2\right) & \alpha_{33}^4 \\
\end{array}
\right)$$ So, one sees how, starting with diagonal CC and NC potentials, due to the non-unitarity one ends up in general with non-diagonal forms for the effective matter potentials.
Notice that the non-unitarity parameters $\alpha_{31}$, $\alpha_{32}$ and $\alpha_{33}$, which do not enter in the expression of $P_{\mu e}$ in vacuum [@Escrihuela:2015wra; @Miranda:2016ptb], do appear in the calculation of $P_{\mu e}$ in matter due to the form of the effective matter potential. The effect of the non-diagonal parameters $\alpha_{31}$ and $\alpha_{32}$ is not as important as the role of $\alpha_{21}$. The $\alpha_{31}$ parameter enters linearly in the CC and NC potential in the 13 entry. Its effect will be analogous to that of the parameter $\epsilon_{e\tau}$ in the case of non-standard interactions, so that the resulting degeneracy with the reactor angle $\theta_{13}$ [@Huber:2001de; @Huber:2002bi] will imply a deterioration of the sensitivity to CP violation [@Forero:2016cmb]. In contrast, $\alpha_{32}$ will enter only in the neutral current potential in the 23 entry and, therefore, is expected to have a negligible impact.\
\[-0.2cm\]
Adding the two contributions to the effective potential in matter we will have [^5]: $$V^{\alpha \beta} = V^{\alpha \beta}_{CC} + V^{\alpha \beta}_{NC} = \sqrt{2} \,G_{F} N_{e} \left(NN^{\dagger}\right)_{\alpha e}
\left(NN^{\dagger}\right)_{e \beta}
-\sqrt{2}G_F \frac{N_n}{2}
\sum_{\rho}(NN^\dagger)_{\alpha \rho}(NN^\dagger)_{\rho
\beta}
\label{eq:V_tot}$$ where $\alpha$ and $\beta$ stands for the initial and final neutrino flavor, respectively, and $\rho$ implies a sum over the three active flavors. $N_e$ is the electron density in the medium while $N_n$ is the neutron density. In matrix form one has the following expression for the matter potential in the presence of non-unitarity: $$V_{NU} = \left(NN^{\dagger}\right)
\left[ \sqrt{2} \,G_{F} N_e
\left(\begin{array}{ccc}
1 & 0 & 0\\
0 & 0 &0 \\
0 & 0 & 0
\end{array}\right) - \frac{\sqrt{2}}{2} \,G_{F} N_n
\left(\begin{array}{ccc}
1 & 0 & 0\\
0 & 1 &0 \\
0 & 0 & 1
\end{array}\right)\right] \left(NN^\dagger\right)
\label{eq:V_tot2}$$ leading to a very simple compact form $$V_{NU} = \left(NN^{\dagger}\right)
V_\mathrm{unitary}(NN^\dagger)~.
\label{eq:V_tot3}$$ Notice that, in contrast to the standard procedure used in the three–neutrino unitary case, the contribution of the neutral current potential can no longer be neglected when treating the non-unitary case.
One can also see how to get this result from the truncation of the $N\times N$ mixing matrix, $U$. Therefore, the Hamiltonian in matter in the flavour basis will be given by: $$H_{NU} = N
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & \frac{\Delta m^2_{21}}{2E} &0 \\
0 & 0 & \frac{\Delta m^2_{31}}{2E}
\end{array}\right) N^{\dagger} +
\left(NN^{\dagger}\right)
\left(\begin{array}{ccc}
V_{cc} +V_{nc} & 0 & 0\\
0 & V_{nc} &0 \\
0 & 0 & V_{nc}
\end{array}\right) \left(NN^\dagger\right)
\label{eq:V_tot2}$$ with $V_{cc} = \sqrt{2} G_F N_e$ and $V_{nc} = - \frac{\sqrt{2}}{2} G_F N_n$.
Non-unitary neutrino mixing in DUNE {#sec:nuDUNE}
====================================
Here we explore the expected sensitivities to the non-unitarity of the neutrino mixing matrix within the upcoming DUNE experiment. Previous studies have already considered the impact of non-unitarity upon the CP-phase sensitivity at T2K [@Ge:2016xya]. Here we present a dedicated study for the DUNE experiment, whose longer baseline implies that matter effects are more relevant than for the cases of T2K and NOvA and therefore the formalism described above is crucial.
DUNE simulation with non-unitary neutrino mixing {#sec:dune-simul}
------------------------------------------------
DUNE is a long-baseline neutrino oscillation experiment that will measure neutrino oscillations over a broad energy range, from hundreds of MeV to few tenths of GeV. This experiment will detect neutrinos and anti-neutrinos produced in the NuMI beam line at Fermilab $1300$ km away from the source, with relevant matter effects in the neutrino propagation. The effect of non-unitary neutrino mixing in the DUNE simulation will modify the standard calculation of neutrino oscillation probabilities. Besides the non-unitary neutrino mixing matrix, to calculate the neutrino conversion probability in DUNE one must take into account the modified matter potential affecting neutrino propagation through the Earth, as discussed in the previous section. Using the neutrino Hamiltonian with matter effects as given by Eq. (\[eq:V\_tot2\]), we have solved numerically the evolution equation, obtaining the corresponding conversion probability from muon to electron neutrino in the case of DUNE. We illustrate the behavior of the modified neutrino appearance probability $P_{\mu e}$ for the DUNE experiment in Fig. \[fig:dune\]. The left panel corresponds to neutrino probability and the right panel to antineutrino probability. Each band corresponds to a different value of the standard CP phase, $\delta_{\text{CP}}$, while the width of the band is due to the variation over the non-unitary phase $\phi_{21} = Arg(\alpha_{21})$. The only deviation from unitarity in this calculation comes from the $\alpha_{21}$ parameter, set to 0.02. The overlap of the different bands indicates the presence of degeneracies in the neutrino oscillation probability in DUNE. This ambiguity, present at the probability level has already been noticed in Ref. [@Miranda:2016ptb; @Miranda:2016wdr]
![Neutrino electron appearance probability in DUNE for the neutrino (left) and antineutrino (right panel) channel, with $|\alpha_{21}| = 0.02$ and $\phi_{21}$ free for fixed value of $\delta_{CP}$.[]{data-label="fig:dune"}](fig_probability2.pdf){width="\textwidth"}
The DUNE experimental setup assumed for this analysis corresponds to a $40$–ton liquid argon far detector with optimized neutrino fluxes, cross sections, detector efficiency and energy resolution effects as provided in the form of GLoBES [@Huber:2004ka; @Huber:2007ji] files in Ref. [@Alion:2016uaj]. The calculation of the neutrino oscillation probabilities in the presence of non-unitarity has been implemented in the GLoBES package with an adequate modification of its probability engine. In our analysis, we have used the spectral event information from the four neutrino oscillation channels: electron (anti)neutrino appearance and muon (anti)neutrino disappearance. To statistically quantify the effect of the non-unitary lepton mixing parameters we have used the usual $\chi^2$ definition adding penalties on the ‘unitary’ oscillation parameters $\theta_{ij}$ and $\Delta
m^2_{k1}$ [@Forero:2014bxa]. The relative error on these parameters and the systematic uncertainties in the normalization of signal and background for each oscillation channel, ranging from $0.2\%$ to $20\%$ depending on the channel, were set to the values given in Ref. [@Alion:2016uaj].
For the rest of this section we denote the three mixing angles collectively as a vector $\vec{\lambda}=\{\theta_{ij},\Delta m^2_{k1}\}$. We have included penalties to the $\chi^2$ accounting for the allowed values of the $\vec{\lambda}$ parameters. Likewise, we denote the non-unitarity parameters in a compact form as $\vec{\alpha}=\{\alpha_{ii},\alpha_{ij}\}$, including both their diagonal and non-diagonal components. Note that we treat the CP phase $\delta_{CP}$ separately.
DUNE sensitivity to CP violation {#sec:DUNE-CP}
--------------------------------
![DUNE sensitivity to CP violation for non-unitary neutrino mixing. For comparison the black solid line shows the CP-sensitivity in DUNE, for the standard unitary case. The reduced sensitivities for non-zero $\alpha_{21}$ ($\alpha_{31}$) are shown in the left (right) panel. All undisplayed parameters have been marginalized over, including the non-unitarity phases as well as the diagonal parameters $\alpha_{ii}$.[]{data-label="fig:CPsens"}](fig4_a21_a31.pdf){width="80.00000%"}
In this section we analyze how DUNE sensitivity to the standard CP violation is affected by the presence of non-unitarity. To the oscillation parameters present in the standard unitary scenario this analysis implies the addition of nine real parameters describing the non-unitary mixing: the three real $\alpha_{ii}$ plus the three complex non-diagonal $\alpha_{ij}$.
In order to simplify the analysis, we consider however only five non-zero non-unitary parameters at a time: the three diagonal ones, plus one of the non-diagonal ones, with its complex phase at a time. The resulting CP sensitivity in the presence of non-unitarity is shown in Fig. \[fig:CPsens\]. As in the standard $\delta_{\text{CP}}$-sensitivity plot, the CP-violation hypothesis is tested with respect to a CP-conserving scenario [@Acciarri:2015uup]: $$\label{eq:chi2}
\Delta \chi^2 (\delta_\text{CP}^\text{true})=\text{Min}\left[\Delta \chi^2_{CP}(\delta_\text{CP}^\text{true},\delta_{\text{CP}}^\text{test}=0),\Delta \chi^2_{CP}(\delta_\text{CP}^\text{true},\delta_{\text{CP}}^\text{test}=\pi)\right],$$ The remaining standard $\vec{\lambda}$ as well as the non-unitarity parameters $\vec{\alpha} \equiv \{\alpha_{ii},\phi_{ij}\}$, which are included in both the simulated and reconstructed event rates in DUNE, $n(\vec{\lambda},\delta_{\text{CP}};\vec{\alpha})$ , have been marginalized over. The left panel has been obtained for different values of $|\alpha_{21}|$, while the right panel corresponds to the results of the corresponding analysis performed for the non-diagonal non-unitarity parameter $\alpha_{31}$. One sees from the left panel that the sensitivity to the Dirac CP phase decreases in the presence of non-unitarity with respect to the standard ‘unitary’ case, shown in the black-solid line. The remaining lines correspond to the non-unitary case with different values for the $\alpha_{21}$ parameter, as indicated. We have selected three different benchmark values, the smaller one, $0.003$, consistent with the upper part of Table \[tab:bounds\] and $0.010$ and $0.025$ consistent with sensitivities displayed in the lower part of the table, obtained from neutrino data only. Even taking at face value the “aggressive” sensitivity $|\alpha_{21}|=0.003$, the significance of a CP-violation measurement decreases by $0.85\sigma$, compromising the possibility of testing any range of values of the CP phase at $5\sigma$. For more conservative and reasonable choices $|\alpha_{21}|$ at the 1% level one sees that the presence of non-unitarity precludes our ability to probe CP violation at $3\sigma$ for nearly all of the $\delta_{CP}$ range. One sees that probing maximal CP violating values $\pm \pi/2$ with high significance in the presence of non-unitarity for ‘large’ $|\alpha_{21}|$ constitutes a big challenge for DUNE. In the right panel, we show the results of the same analysis for the non-diagonal parameter $\alpha_{31}$. As we discussed before, the impact of this parameter on the neutrino oscillation probabilities in DUNE is significantly less relevant in comparison with $\alpha_{21}$. As a result, one can see from the figure that even if the fraction of CP at $5\sigma$ is largely reduced respect to the unitary case, the reduction in the significance of CP tests is much smaller than for $|\alpha_{21}|$. This result holds for relatively large values of the parameter compatible with the bounds in the lower part of Table \[tab:bounds\], namely $|\alpha_{31}|=0.05$. The effect of the third non-diagonal parameter, $\alpha_{32}$, is not displayed in the figure. We have checked that it plays nearly no role in the analysis, which confirms our discussion in Section \[sec:non-unit-oscill\].
Here we note that Ref. [@Dutta:2016vcc] has also discussed the possibility of probing CP violation with T2K, NOvA and DUNE in the presence of non-unitarity. Although we have a qualitative agreement in the loss of CP-sensitivity due to the presence of non-unitarity, our results show some quantitative differences. We ascribe these discrepancies to the treatment of the DUNE simulation. Here we are using the official description released by the DUNE Collaboration, and we have validated our method against the official DUNE CP-sensitivity result for the standard (unitary) oscillation analysis.
DUNE sensitivity to non-unitary neutrino mixing {#sec:DUNE-a21}
-----------------------------------------------
In this section we analyze the potential of DUNE in constraining the non-unitarity of the neutrino mixing matrix. As we have discussed in Section \[sec:prel-updat-bounds\], the most robust and direct of these constraints come from neutrino oscillation experiments and are not very strong. Therefore, we wish to explore the capability of DUNE in further constraining non-unitarity. For this purpose we will focus in the analysis of the neutrino signal at the DUNE far detector. The capability of the near detector will be analyzed in the future. As we have discussed in the previous subsection, the parameter with the most impact on the DUNE sensitivity to CP violation is $\alpha_{21}$. As a result we will focus on $\alpha_{21}$ as the key parameter to be constrained in order to characterize the loss of sensitivity in CP searches at DUNE. Following the usual procedure in analyzing the sensitivity of a given experiment to an unknown parameter (the non-unitary parameter $\alpha_{21}$ in this case), we have simulated DUNE events under the hypothesis of unitary mixing $n^{\text{true}}(\vec{\lambda},\delta_{\text{CP}}^\text{true})$. Afterwards, we have tried to reconstruct DUNE data in terms of the non-unitary neutrino mixing ansatz, $n^{\text{test}}(\vec{\lambda},\delta_{\text{CP}}^\text{test};\vec{\alpha})$. It is worth noticing that the treatment of the non-unitarity here is different from the analysis performed in the previous subsection and therefore a direct comparison between the results presented in Figs. \[fig:CPsens\] and \[fig:dcpsensitiv\] is not straightforward. For this analysis, the true value of the Dirac CP phase has been fixed to its current preferred value, $\delta_{\text{CP}}^{\text{True}}=-\pi/2$. After marginalizing over the diagonal non-unitary parameters and all the oscillation parameters but $\delta_{\text{CP}}$, we obtained the allowed parameter regions (at $1-4 \sigma$ for 2 d.o.f) shown in Fig. \[fig:dcpsensitiv\]. In the left panel, the allowed regions in the $\delta_{\text{CP}}-|\alpha_{21}|$ plane show that DUNE is sensitive to values of $|\alpha_{21}|$ at the percent level at $1\sigma$. As expected, the best fit point for the Dirac CP phase is equal to the assumed ‘true’ value. However, for large enough values of $|\alpha_{21}|$, degenerate solutions around $\delta_{\text{CP}}= \pm \pi$ appear at higher C.L.
![Testing non-unitary neutrino mixing against the standard case, when only the off–diagonal parameter $\alpha_{21}$ is present. In the left (right) panel, the $\alpha_{21}$–$\delta_{CP}$ ($\phi_{21}$–$\delta_{CP}$) allowed parameter space is shown. The 3$\sigma$ upper bound on $\alpha_{21}$ from neutrino data is indicated by a red line. This limit has also been included as a prior in the results shown in the right panel. Here $\delta_{CP}^{\text{True}}$ is fixed to $-\pi/2$ and the additional undisplayed parameters have been marginalized over. The allowed regions, darkest to lightest, correspond to $1\sigma$ to $4 \sigma$ for 2 d.o.f., respectively.[]{data-label="fig:dcpsensitiv"}](fig_phases_v2.pdf){width="80.00000%"}
Finally, we present an estimate of the absolute sensitivity of DUNE to the non-unitary parameter $\alpha_{21}$. In order to do this, we extended our previous analysis, considering all the possible values of $\delta_{\text{CP}}^{\text{True}}$ and marginalizing over $\delta_{\text{CP}}$ and $\phi_{21}$. Fig. \[fig:a21sensitiv\] shows the $\chi^2$ profile obtained as a function of $|\alpha_{21}|$ after marginalizing over all the remaining parameters, including $\delta_{\text{CP}}^{\text{True}}$. The best fit point, denoted by a black point in the figure, is obtained for $|\alpha_{21}| = 3\times 10^{-4}$. Nevertheless, the preference over the unitary hypothesis is not significant at all, as can be seen from the figure. The shaded band in Fig. \[fig:a21sensitiv\] indicates the three benchmark values of $\alpha_{21}$ used in the analysis of CP sensitivity in DUNE (see the left panel of Fig. \[fig:CPsens\]), while the horizontal dotted black line defines the parameter region allowed by DUNE at $90\%$ of C.L., corresponding to the limit $|\alpha_{21}|<0.046$. This bound is somewhat weaker than the constraints derived from neutrino oscillation searches, indicating that the analysis of long-baseline neutrino oscillations in DUNE is not expected to improve our current knowledge on the non-unitarity of the neutrino mixing matrix. However, it is worth mentioning that this constraint can also be regarded as independent and complementary to the bounds in Table \[tab:bounds\].
![DUNE sensitivity to $\alpha_{21}$ for arbitrary values of the Dirac CP phase. The expected sensitivity at $90\%$ of C.L. is indicated by the horizontal line, while the three vertical lines correspond to the benchmark points considered in Fig. \[fig:CPsens\]. []{data-label="fig:a21sensitiv"}](fig_a21-chi2_profile.pdf){width="60.00000%"}
Conclusion and discussion {#sec:conclusions}
=========================
We have reviewed the existing limits on non-unitarity parameters, from weak universality considerations as well as from neutrino oscillation data. We have discussed the model-independent character of the latter constraints, since they only rely upon direct information from the neutrino sector, in contrast with the ones derived form charged lepton processes. We have developed in detail the formalism for neutrino propagation in matter in the presence of non-unitary neutrino mixing. In contrast to the standard unitary case, the neutral current potential contributes to the neutrino Hamiltonian in matter. Here we have focused our analysis on the case of the long-baseline neutrino experiment DUNE. First we have analyzed how the sensitivity to CP violation in DUNE can be affected by the presence of non-unitarity. We have found that DUNE’s potential to probe CP violation is somewhat weakened, although not as much as one might have expected, see Fig. \[fig:CPsens\]. The reason for this, apart from the high statistics, is mainly the fact that the DUNE experiment is characterized by a relatively wide beam, compared with current experiments. This nice feature partly mitigates the ambiguities stressed in [@Miranda:2016wdr]. Moreover, we have investigated how DUNE can probe neutrino properties beyond standard oscillations, such as the parameters characterizing non-unitarity, see Fig. \[fig:a21sensitiv\]. In this respect DUNE is not expected to perform better than previous short baseline oscillation searches at NOMAD, CHORUS and NuTeV. This discouraging result is not surprising, as the sensitivity to non-unitarity comes mainly from probing the “zero-distance effect” and hence involves “near” detection. This could be improved within a setup of the type suggested in Ref. [@Ge:2016xya]. Before closing, we mention two other recent related analyses. In Ref. [@Blennow:2016jkn] the authors considered the effect of extra neutrino states in neutrino oscillations, focusing on the differences and similarities between the case in which these neutrinos are kinematically accessible (sterile neutrinos) or not (non-unitary mixing matrix)[^6]. They choose the stronger model-dependent bounds, for which the effect of non-unitarity on the neutrino signal in DUNE is very small and hence the loss in CP sensitivity. Finally, Ref. [@Pas:2016qbg] suggested a novel possibility of discriminating sterile neutrinos and unitarity violation through CP violation.
This work was funded by the Spanish grants FPA2014-58183-P, Multidark CSD2009-00064, SEV-2014-0398 (MINECO) and PROMETEOII/2014/084 (Generalitat Valenciana). MT is also supported by the grant GV2016-142 (Ayuda para Proyectos Emergentes, Generalitat Valenciana) and by a Ramón y Cajal contract (MINECO). OGM is supported by the CONACyT grant 166639 (Mexico). DVF thanks the URA fellowship that allowed him to visit the theory division at Fermilab where some part of this project was carried on. DVF has been supported by the U.S. Department Of Energy under the DE-SC0013632 and DE-SC0009973 contracts.
[92]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix[URL ]{}[URL ]{}\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
() (), [[`1601.05022`]{}]{}.
() (), [[`1610.04802`]{}]{}.
(), ****, ().
(), ****, (), [[`1203.1669`]{}]{}.
(), ****, (), [[`1204.0626`]{}]{}.
(), ****, (), [[`1201.1386`]{}]{}.
(), ****, (), [[`1207.6632`]{}]{}.
(), ****, (), [[`1502.01550`]{}]{}.
() (), [[`1512.06148`]{}]{}.
, ****, (), [[`astro-ph/0207403`]{}]{}.
, ****, (), [[`1303.2272`]{}]{}.
, ****, (), [[`1004.2318`]{}]{}.
, , , ****, (), [[`astro-ph/9604061`]{}]{}.
, ****, (), [[`hep-ph/9602307`]{}]{}.
, ****, (), [[`astro-ph/0304462`]{}]{}.
, ****, ().
, ****, (), [[`hep-ph/0209094`]{}]{}.
, ****, ().
, ****, (), [[`hep-ph/0207326 v3 KamLAND-updated version`]{}]{}.
, ****, (), [[`hep-ph/0311014`]{}]{}.
, , , , ****, (), [[`hep-ph/0406066`]{}]{}.
, , , ****, (), [[`hep-ph/0002147`]{}]{}.
, , , , ****, (), [[`hep-ph/0105159`]{}]{}.
, , , ****, (), [[`hep-ph/0406280`]{}]{}.
, ****, (), [[`1511.05562`]{}]{}.
, ****, (), [[`1601.03736`]{}]{}.
, ****, (), [[`0802.1434`]{}]{}.
, , , , , ****, (), , [[`1503.08879`]{}]{}.
, , , ****, (), [[`1604.05690`]{}]{}.
, ****, ().
, ****, (), [[`1602.00864`]{}]{}.
, , , ****, (), [[`1306.4669`]{}]{}.
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, (), [[`hep-ph/9509255`]{}]{}.
, , , ****, (), [[`hep-ph/0506296`]{}]{}.
, ****, (), [[`1108.3484`]{}]{}.
(), ****, (), [[`1506.05845`]{}]{}.
(), ****, (), [[`hep-ex/0306037`]{}]{}.
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, (), [[`hep-ph/9402224`]{}]{}.
, , , , , ****, (), [[`1211.3052`]{}]{}.
, , , , ****, (), [[`0901.3589`]{}]{}.
, , , , ****, (), [[`1311.2830`]{}]{}.
, , , ****, (), [[`1406.6978`]{}]{}.
, ****, (), [[`1407.6607`]{}]{}.
(), ****, ().
, ****, (), [[`0707.3439`]{}]{}.
, , , , ****, ().
, ****, ().
, ****, (), [[`1310.7922`]{}]{}.
(), ****, (), [[`hep-ex/0509008`]{}]{}.
, , , ****, ().
, ****, (), [[`1508.05095`]{}]{}.
(), ****, (), [[`hep-ex/0106102`]{}]{}.
(), ****, ().
(), ****, (), [[`0710.3361`]{}]{}.
(), ****, (), [[`hep-ex/0203018`]{}]{}.
, , , (), [[`1308.5700`]{}]{}.
, ****, ().
, ****, ().
, , , , ****, ().
(), ****, ().
(), ****, ().
(), ****, (), [[`1506.06020`]{}]{}.
(), [[`1502.00477`]{}]{}.
, ****, (), [[`1504.04855`]{}]{}.
, , , (), [[`1609.09069`]{}]{}.
, , , , , ****, (), [[`hep-ph/0607020`]{}]{}.
, , , ****, (), [[`1605.08774`]{}]{}.
, ****, ().
, , , ****, (), [[`hep-ph/0111224`]{}]{}.
, , , ****, (), [[`hep-ph/0202048`]{}]{}.
, , , , (), [[`1609.08637`]{}]{}.
, , , , ****, (), [[`1605.01670`]{}]{}.
, , , ****, (), [[`hep-ph/0407333`]{}]{}.
, , , , , ****, (), [[`hep-ph/0701187`]{}]{}.
() (), [[`1606.09550`]{}]{}.
, , , ****, (), [[`1405.7540`]{}]{}.
, ****, (), [[`1607.02500`]{}]{}.
, , , ****, (), [[`1609.08623`]{}]{}.
(), [[`1611.08450`]{}]{}.
[^1]: While direct search bounds for such charged mediators are rather stringent, one can still find “fine-tuned” funnels in parameter space which allow the situation envisaged here.
[^2]: Such model-independent studies of the charged current weak interactions were given in [@Mursula:1984zb; @Fetscher:2008zz].
[^3]: See Refs. [@Antusch:2006vwa; @Fernandez-Martinez:2016lgt], where a different form for the non-unitary neutrino mixing matrix is used.
[^4]: We will assume a non-polarized neutral medium in the calculation of the effective matter potential
[^5]: Similar results have been obtained in Ref.[@Blennow:2016jkn]
[^6]: The comparison between these two scenarios has also been explored in Ref. [@Fong:2016yyh].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove the multiplicity formula for the automorphic discrete spectrum of the metaplectic group ${\mathrm{Mp}}_4$ of rank $2$.'
address:
- 'Department of Mathematics, National University of Singapore, 10 Lower Kent Ridge Road, Singapore 119076'
- 'Department of Mathematics, Kyoto University, Kitashirakawa Oiwake-cho, Sakyo-ku, Kyoto 606-8502, Japan'
author:
- Wee Teck Gan
- Atsushi Ichino
title: 'The automorphic discrete spectrum of ${\mathrm{Mp}}_4$'
---
Introduction
============
In our previous paper [@gi-mp], we studied the automorphic discrete spectrum $L^2_{\mathrm{disc}}({\mathrm{Mp}}_{2n})$ of the metaplectic group ${\mathrm{Mp}}_{2n}$, which is a nonlinear double cover of the symplectic group ${\mathrm{Sp}}_{2n}$ of rank $n$. In particular, we proved a decomposition $$L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_{2n}) = \bigoplus_\phi L^2_\phi({\mathrm{Mp}}_{2n}),$$ where $\phi$ runs over elliptic $A$-parameters for ${\mathrm{Mp}}_{2n}$ and $L^2_\phi({\mathrm{Mp}}_{2n})$ is the near equivalence class determined by $\phi$ (which depends on an auxiliary choice of an additive character). In addition, when $\phi$ is tempered, we proved a further decomposition $$\label{eq:main-intro}
L^2_\phi({\mathrm{Mp}}_{2n}) \cong \bigoplus_\pi m_\pi \pi,$$ where $\pi$ runs over representations in the global $A$-packet associated to $\phi$ and $m_\pi$ is the nonnegative integer predicted by the analog of Arthur’s conjecture [@g-iccm §5.6]. When $n=1$, such a result was first established by Waldspurger [@w1; @w2] as a representation-theoretic reformulation of the Shimura correspondence.
The purpose of this paper is to prove when $n=2$ and $\phi$ is nontempered (see Theorem \[t:main\]), which completes the analysis of $L^2_{\mathrm{disc}}({\mathrm{Mp}}_4)$. We also tabulate the representations in the local $A$-packets for ${\mathrm{Mp}}_4$ explicitly (see Appendix \[a:A-packets\]), which may be useful for arithmetic applications. The case $n=2$ is especially interesting because, similarly to the case $n=1$, it involves the two groups ${\mathrm{SO}}_5$ and ${\mathrm{Sp}}_4$ of the same type $B_2 = C_2$. In particular, we hope that our representation-theoretic formulation will elucidate Ibukiyama’s conjecture [@ibukiyama] on integral and half-integral weight Siegel modular forms of genus $2$.
Acknowledgments {#acknowledgments .unnumbered}
---------------
The first-named author is partially supported by an MOE Tier one grant R-146-000-228-114. The second-named author is partially supported by JSPS KAKENHI Grant Number 19H01781.
Notation {#notation .unnumbered}
--------
If $F$ is a local field of characteristic zero, we fix a nontrivial additive character $\psi$ of $F$. For $a \in F^\times$, we define another nontrivial additive character $\psi_a$ of $F$ by $\psi_a(x) = \psi(ax)$. We also define a quadratic character $\chi_a$ of $F^\times$ by $\chi_a(x) = (a,x)_F$, where $(\cdot,\cdot)_F$ is the quadratic Hilbert symbol of $F$. Let $W_F$ be the Weil group of $F$ and put $$L_F =
\begin{cases}
W_F \times {\mathrm{SL}}_2({\mathbb{C}}) & \text{if $F$ is nonarchimedean;} \\
W_F & \text{if $F$ is archimedean.}
\end{cases}$$ For any linear algebraic group $G$ over $F$, we identify $G$ with its group of $F$-rational points.
If $F$ is a number field with adèle ring ${\mathbb{A}}$, we fix a nontrivial additive character $\psi$ of ${\mathbb{A}}/F$. For $a \in F^\times$, we define another nontrivial additive character $\psi_a$ of ${\mathbb{A}}/F$ by $\psi_a(x) = \psi(ax)$. We also define a quadratic automorphic character $\chi_a$ of ${\mathbb{A}}^\times$ by $\chi_a(x) = \prod_v (a,x_v)_{F_v}$, where $v$ runs over places of $F$. We denote by $\zeta^S(s)$ the partial zeta function of $F$, where $S$ is a finite set of places of $F$. Put $$\mu_{2,{\mathbb{A}}} = \Big\{ (\epsilon_v) \in \prod_v \{ \pm 1 \} \, \Big| \, \text{$\epsilon_v = 1$ for almost all $v$} \Big\}.$$
For any representation $\pi$, we denote by $\pi^\vee$ its contragredient representation. For any abelian locally compact group $S$, we denote by $\widehat{S}$ the group of continuous characters of $S$. For any positive integer $d$, we denote by $S_d$ the unique $d$-dimensional irreducible representation of ${\mathrm{SL}}_2({\mathbb{C}})$. We write $$1^d = \underbrace{1,1,\dots,1}_d.$$ For any $a \in \frac{1}{2} {\mathbb{Z}}$ with $a>0$, we denote by $\mathcal{D}_a$ the $2$-dimensional irreducible representation of $W_{\mathbb{R}}$ induced from the character $z \mapsto (z/\bar{z})^a$ of $W_{\mathbb{C}}= {\mathbb{C}}^\times$.
The multiplicity formula
========================
In this section, we state the multiplicity formula for the automorphic discrete spectrum of ${\mathrm{Mp}}_4$, which is the main result of this paper.
Elliptic $A$-parameters
-----------------------
Let $F$ be a number field with adèle ring ${\mathbb{A}}$ and fix a nontrivial additive character $\psi$ of ${\mathbb{A}}/F$. Recall from [@gi-mp] that an elliptic $A$-parameter for ${\mathrm{Mp}}_4$ is a formal unordered finite direct sum $$\phi = \bigoplus_i \phi_i \boxtimes S_{d_i},$$ where
- $\phi_i$ is an irreducible self-dual cuspidal automorphic representation of ${\mathrm{GL}}_{n_i}({\mathbb{A}})$;
- $S_{d_i}$ is the unique $d_i$-dimensional irreducible representation of ${\mathrm{SL}}_2({\mathbb{C}})$;
- if $d_i$ is odd, then $\phi_i$ is symplectic, i.e. the exterior square $L$-function $L(s, \phi_i, \wedge^2)$ has a pole at $s=1$;
- if $d_i$ is even, then $\phi_i$ is orthogonal, i.e. the symmetric square $L$-function $L(s, \phi_i, {\mathrm{Sym}}^2)$ has a pole at $s=1$;
- if $(\phi_i, d_i) = (\phi_j, d_j)$, then $i=j$;
- $\sum_i n_i d_i = 4$.
Let $S_\phi$ be the global component group of $\phi$, which is defined formally as a free ${\mathbb{Z}}/2{\mathbb{Z}}$-module $$S_\phi = \bigoplus_i ({\mathbb{Z}}/2{\mathbb{Z}}) a_i$$ with a basis $\{a_i\}$, where $a_i$ corresponds to $\phi_i \boxtimes S_{d_i}$. Recall also that Arthur [@arthur (1.5.6)] associated to $\phi$ (which can be regarded as an elliptic $A$-parameter for ${\mathrm{SO}}_5$) a character $\epsilon_\phi$ of $S_\phi$, which plays an important role in the multiplicity formula for the automorphic discrete spectrum of ${\mathrm{SO}}_5$. We define another character $\tilde{\epsilon}_\phi$ of $S_\phi$, which plays a similar role for ${\mathrm{Mp}}_4$, by $$\tilde{\epsilon}_\phi(a_i) = \epsilon_\phi(a_i) \times
\begin{cases}
\epsilon(\frac{1}{2}, \phi_i) & \text{if $\phi_i$ is symplectic;} \\
1 & \text{if $\phi_i$ is orthogonal,} \\
\end{cases}$$ where $\epsilon(\frac{1}{2}, \phi_i) = \pm 1$ is the root number of $\phi_i$.
We can enumerate the elliptic $A$-parameters $\phi$ for ${\mathrm{Mp}}_4$ with associated character $\tilde{\epsilon}_\phi$ as follows.
- We say that $\phi$ is *tempered* if $d_i=1$ for all $i$. In this case, we have $\tilde{\epsilon}_\phi(a_i) = \epsilon(\frac{1}{2}, \phi_i)$ for all $i$.
- We say that $\phi$ is of *Saito–Kurokawa type* if $$\phi = (\rho \boxtimes S_1) \oplus (\chi \boxtimes S_2)$$ for some irreducible cuspidal automorphic representation $\rho$ of ${\mathrm{GL}}_2({\mathbb{A}})$ with trivial central character and some quadratic automorphic character $\chi$ of ${\mathbb{A}}^\times$. In this case, we have $S_\phi \cong ({\mathbb{Z}}/2{\mathbb{Z}})^2$ and $$\tilde{\epsilon}_\phi(a_1) = \epsilon(\tfrac{1}{2}, \rho) \cdot \epsilon(\tfrac{1}{2}, \rho \times \chi), \qquad
\tilde{\epsilon}_\phi(a_2) = \epsilon(\tfrac{1}{2}, \rho \times \chi),$$ where $a_1$ and $a_2$ correspond to $\rho \boxtimes S_1$ and $\chi \boxtimes S_2$, respectively.
- We say that $\phi$ is of *Howe–Piatetski-Shapiro type* if $$\phi = (\chi_1 \boxtimes S_2) \oplus (\chi_2 \boxtimes S_2)$$ for some distinct quadratic automorphic characters $\chi_1$ and $\chi_2$ of ${\mathbb{A}}^\times$. In this case, we have $S_\phi \cong ({\mathbb{Z}}/2{\mathbb{Z}})^2$ and $\tilde{\epsilon}_\phi = 1$.
- We say that $\phi$ is of *Soudry type* if $$\phi = \rho \boxtimes S_2$$ for some irreducible dihedral cuspidal automorphic representation $\rho$ of ${\mathrm{GL}}_2({\mathbb{A}})$ with nontrivial quadratic central character. In this case, we have $S_\phi \cong {\mathbb{Z}}/2{\mathbb{Z}}$ and $\tilde{\epsilon}_\phi = 1$.
- We say that $\phi$ is *principal* if $$\phi = \chi \boxtimes S_4$$ for some quadratic automorphic character $\chi$ of ${\mathbb{A}}^\times$. In this case, we have $S_\phi \cong {\mathbb{Z}}/2{\mathbb{Z}}$ and $\tilde{\epsilon}_\phi = 1$.
Local $A$-packets
-----------------
Let $\phi = \bigoplus_i \phi_i \boxtimes S_{d_i}$ be an elliptic $A$-parameter for ${\mathrm{Mp}}_4$. For each place $v$ of $F$, the localization $\phi_v = \bigoplus_i \phi_{i,v} \boxtimes S_{d_i}$ of $\phi$ at $v$ gives rise to a local $A$-parameter $$\phi_v : L_{F_v} \times {\mathrm{SL}}_2({\mathbb{C}}) \longrightarrow {\mathrm{Sp}}_4({\mathbb{C}})$$ via the local Langlands correspondence [@langlands; @harris-taylor; @henniart; @scholze]. We define the associated $L$-parameter $\varphi_{\phi_v} : L_{F_v} \rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$ by $$\varphi_{\phi_v}(w) = \phi_v \!
\left( w,
\begin{pmatrix}
|w|_v^{\frac{1}{2}} & \\
& |w|_v^{-\frac{1}{2}}
\end{pmatrix}
\right).$$ We denote by $S_{\phi_v}$ and $S_{\varphi_{\phi_v}}$ the component groups of the centralizers of $\phi_v$ and $\varphi_{\phi_v}$ in ${\mathrm{Sp}}_4({\mathbb{C}})$, respectively. Then we have a canonical map $S_\phi \rightarrow S_{\phi_v}$ and a natural surjection $S_{\phi_v} \rightarrow S_{\varphi_{\phi_v}}$.
Let ${\mathrm{Mp}}_4(F_v)$ be the metaplectic double cover of ${\mathrm{Sp}}_4(F_v)$. We will assign to $\phi_v$ a finite set (which depends on the nontrivial additive character $\psi_v$ of $F_v$) $$\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4) = \{ \pi_{\eta_v} \, | \, \eta_v \in \widehat{S}_{\phi_v} \}$$ of semisimple genuine representations of ${\mathrm{Mp}}_4(F_v)$ of finite length indexed by characters of $S_{\phi_v}$. If $\phi$ is tempered, then $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ is defined as the $L$-packet associated to $\phi_v$ (relative to $\psi_v$) defined via the local Shimura correspondence [@ab1; @ab2; @gs]. If $\phi$ is nontempered, then $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ will be defined in §§\[ss:local-princ\], \[ss:local-sk\], \[ss:local-howe-ps\], \[ss:local-soudry\] below. Moreover, we will show that the following properties hold:
- $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ is multiplicity-free;
- $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ contains the $L$-packet $\Pi_{\varphi_{\phi_v}, \psi_v}({\mathrm{Mp}}_4)$ associated to $\varphi_{\phi_v}$ (relative to $\psi_v$);
- the diagram $$\begin{tikzcd}
\Pi_{\phi_v, \psi_v}({\mathrm{Mp}}_4) \arrow[r] & \widehat{S}_{\phi_v} \\
\Pi_{\varphi_{\phi_v}, \psi_v}({\mathrm{Mp}}_4) \arrow[r] \arrow[u,hook] & \widehat{S}_{\varphi_{\phi_v}} \arrow[u,hook]
\end{tikzcd}$$ commutes.
Statement of the main theorem
-----------------------------
Let ${\mathrm{Mp}}_4({\mathbb{A}})$ be the metaplectic double cover of ${\mathrm{Sp}}_4({\mathbb{A}})$. We denote by $$L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_4)$$ the genuine part of the discrete spectrum of the unitary representation $L^2({\mathrm{Sp}}_4(F) \backslash {\mathrm{Mp}}_4({\mathbb{A}}))$ of ${\mathrm{Mp}}_4({\mathbb{A}})$, where we regard ${\mathrm{Sp}}_4(F)$ as a subgroup of ${\mathrm{Mp}}_4({\mathbb{A}})$ via the canonical splitting. In our previous paper [@gi-mp], we proved a decomposition $$L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_4) = \bigoplus_{\phi} L^2_{\phi,\psi}({\mathrm{Mp}}_4),$$ where $\phi$ runs over elliptic $A$-parameters for ${\mathrm{Mp}}_4$ and $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$ is the near equivalence class of irreducible representations $\pi$ in $L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_4)$ such that the $L$-parameter of $\pi_v$ (relative to $\psi_v$) is $\varphi_{\phi_v}$ for almost all places $v$ of $F$. Thus, to understand the spectral decomposition of $L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_4)$, it remains to describe the multiplicity of any irreducible representation of ${\mathrm{Mp}}_4({\mathbb{A}})$ in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$.
Consider the compact group $S_{\phi, {\mathbb{A}}} = \prod_v S_{\phi_v}$ and the group $\widehat{S}_{\phi,{\mathbb{A}}} = \bigoplus_v \widehat{S}_{\phi_v}$ of continuous characters of $S_{\phi, {\mathbb{A}}}$. For any $\eta = \bigotimes_v \eta_v \in \widehat{S}_{\phi, {\mathbb{A}}}$, we may form a semisimple genuine representation $$\pi_\eta = \bigotimes_v \pi_{\eta_v}$$ of ${\mathrm{Mp}}_4({\mathbb{A}})$. Let $\Delta^* \eta = \eta \circ \Delta \in \widehat{S}_\phi$ be the pullback of $\eta$ under the diagonal map $\Delta : S_\phi \rightarrow S_{\phi, {\mathbb{A}}}$.
We now state our main result:
\[t:main\] For any elliptic $A$-parameter $\phi$ for ${\mathrm{Mp}}_4$, we have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\eta \in \widehat{S}_{\phi, {\mathbb{A}}}} m_\eta \pi_\eta,$$ where $$m_\eta =
\begin{cases}
1 & \text{if $\Delta^* \eta = \tilde{\epsilon}_\phi$;} \\
0 & \text{otherwise.}
\end{cases}$$ In particular, $L^2_{{\mathrm{disc}}}({\mathrm{Mp}}_4)$ is multiplicity-free.
This theorem was already proved in our previous paper [@gi-mp] when $\phi$ is tempered and follows from Propositions \[p:princ\], \[p:sk\], \[p:howe-ps\], \[p:soudry\] below when $\phi$ is nontempered.
Metaplectic and orthogonal groups
=================================
In this section, we introduce notation for the groups and representations which appear in this paper. Let $F$ be either a local field of characteristic zero or a number field with adèle ring ${\mathbb{A}}$.
Metaplectic groups {#ss:mp}
------------------
Let $W_n$ be a $2n$-dimensional vector space over $F$ equipped with a nondegenerate antisymmetric bilinear form $\langle \cdot, \cdot \rangle_{W_n}$. We call such $W_n$ a symplectic space over $F$ and denote by ${\mathrm{Sp}}(W_n)$ the associated symplectic group. Choose a basis $w_1,\dots,w_n,w_1^*,\dots,w_n^*$ of $W_n$ such that $$\langle w_i, w_j \rangle_{W_n} = \langle w_i^*, w_j^* \rangle_{W_n} = 0, \qquad \langle w_i, w_j^* \rangle_{W_n} = \delta_{ij}.$$ Given a sequence $\mathbf{k} = (k_1, \dots, k_m)$ of positive integers such that $k_1 + \dots + k_m \le n$, we write $$W_n = X_1 \oplus \cdots \oplus X_m \oplus X_1^* \oplus \cdots \oplus X_m^* \oplus W_{n_0}$$ with $$\begin{aligned}
X_i & = {\operatorname{Span}}(w_{k'_{i-1} + 1}, \dots, w_{k'_i}), \\
X_i^* & = {\operatorname{Span}}(w_{k'_{i-1} + 1}^*, \dots, w_{k'_i}^*), \\
W_{n_0} & = {\operatorname{Span}}(w_{k'_m+1}, \dots, w_n, w_{k'_m+1}^*, \dots, w_n^*),\end{aligned}$$ where $k'_i = k_1 + \dots + k_i$ and $n_0 = n - k_1 - \dots - k_m$. Let $P_{\mathbf{k}} = M_{\mathbf{k}} N_{\mathbf{k}}$ be the parabolic subgroup of ${\mathrm{Sp}}(W_n)$ stabilizing the flag $$X_1 \subset X_1 \oplus X_2 \subset \cdots \subset X_1 \oplus \cdots \oplus X_m,$$ where $$M_{\mathbf{k}} \cong {\mathrm{GL}}(X_1) \times \dots \times {\mathrm{GL}}(X_m) \times {\mathrm{Sp}}(W_{n_0})$$ is the Levi component of $P_{\mathbf{k}}$ stabilizing the flag $$X_1^* \subset X_1^* \oplus X_2^* \subset \cdots \subset X_1^* \oplus \cdots \oplus X_m^*$$ and $N_{\mathbf{k}}$ is the unipotent radical of $P_{\mathbf{k}}$. We also write $P_k = P_{(k)}$ for $1 \le k \le n$ and $B = P_{(1^n)}$, which are a maximal parabolic subgroup and a Borel subgroup of ${\mathrm{Sp}}(W_n)$, respectively.
Suppose that $F$ is local. We denote by ${\mathrm{Mp}}(W_n)$ the metaplectic double cover of ${\mathrm{Sp}}(W_n)$: $$1 \longrightarrow \{ \pm 1 \} \longrightarrow {\mathrm{Mp}}(W_n) \longrightarrow {\mathrm{Sp}}(W_n) \longrightarrow 1.$$ This cover splits over $N_{\mathbf{k}}$ uniquely. Let $\widetilde{P}_{\mathbf{k}}$ and $\widetilde{M}_{\mathbf{k}}$ be the preimages of $P_{\mathbf{k}}$ and $M_{\mathbf{k}}$ in ${\mathrm{Mp}}(W_n)$, respectively, so that $\widetilde{P}_{\mathbf{k}} = \widetilde{M}_{\mathbf{k}} N_{\mathbf{k}}$ and $$\widetilde{M}_{\mathbf{k}} \cong \widetilde{{\mathrm{GL}}}(X_1) \times_{\{ \pm 1 \} } \dots \times_{\{ \pm 1 \} } \widetilde{{\mathrm{GL}}}(X_m) \times_{\{ \pm 1 \} } {\mathrm{Mp}}(W_{n_0}).$$ Here $\widetilde{{\mathrm{GL}}}(X_i)$ is the double cover of ${\mathrm{GL}}(X_i)$ given in [@gi1 §2.5]. Suppose that $F$ is global. We denote by ${\mathrm{Mp}}(W_n)({\mathbb{A}})$ the metaplectic double cover of ${\mathrm{Sp}}(W_n)({\mathbb{A}})$: $$1 \longrightarrow \{ \pm 1 \} \longrightarrow {\mathrm{Mp}}(W_n)({\mathbb{A}}) \longrightarrow {\mathrm{Sp}}(W_n)({\mathbb{A}}) \longrightarrow 1.$$ This cover splits over ${\mathrm{Sp}}(W_n)(F)$ uniquely.
Orthogonal groups {#ss:so}
-----------------
Let $V_n$ be a $(2n+1)$-dimensional vector space over $F$ equipped with a nondegenerate symmetric bilinear form $\langle \cdot, \cdot \rangle_{V_n}$. We call such $V_n$ a quadratic space over $F$ and denote by ${\mathrm{O}}(V_n)$ the associated orthogonal group. Let ${\mathrm{SO}}(V_n)$ be the identity component of ${\mathrm{O}}(V_n)$, so that $${\mathrm{O}}(V_n) = {\mathrm{SO}}(V_n) \times \{ \pm 1\}.$$ Let $V_{\mathrm{an}}$ be an anisotropic kernel of $V_n$ and $r$ the Witt index of $V_n$. Choose a basis $v_1,\dots,v_r,v_1^*,\dots,v_r^*$ of the orthogonal complement of $V_{\mathrm{an}}$ in $V_n$ such that $$\langle v_i, v_j \rangle_{V_n} = \langle v_i^*, v_j^* \rangle_{V_n} = 0, \qquad \langle v_i, v_j^* \rangle_{V_n} = \delta_{ij}.$$ Given a sequence $\mathbf{k} = (k_1, \dots, k_m)$ of positive integers such that $k_1 + \dots + k_m \le r$, we write $$V_n = Y_1 \oplus \cdots \oplus Y_m \oplus Y_1^* \oplus \cdots \oplus Y_m^* \oplus V_{n_0}$$ with $$\begin{aligned}
Y_i & = {\operatorname{Span}}(v_{k'_{i-1} + 1}, \dots, v_{k'_i}), \\
Y_i^* & = {\operatorname{Span}}(v_{k'_{i-1} + 1}^*, \dots, v_{k'_i}^*), \\
V_{n_0} & = {\operatorname{Span}}(v_{k'_m+1}, \dots, v_r, v_{k'_m+1}^*, \dots, v_r^*) \oplus V_{\mathrm{an}},\end{aligned}$$ where $k'_i = k_1 + \dots + k_i$ and $n_0 = n - k_1 - \dots - k_m$. Let $Q_{\mathbf{k}} = L_{\mathbf{k}} U_{\mathbf{k}}$ be the parabolic subgroup of ${\mathrm{SO}}(V_n)$ stabilizing the flag $$Y_1 \subset Y_1 \oplus Y_2 \subset \cdots \subset Y_1 \oplus \cdots \oplus Y_m,$$ where $$L_{\mathbf{k}} \cong {\mathrm{GL}}(Y_1) \times \dots \times {\mathrm{GL}}(Y_m) \times {\mathrm{SO}}(V_{n_0})$$ is the Levi component of $Q_{\mathbf{k}}$ stabilizing the flag $$Y_1^* \subset Y_1^* \oplus Y_2^* \subset \cdots \subset Y_1^* \oplus \cdots \oplus Y_m^*$$ and $U_{\mathbf{k}}$ is the unipotent radical of $Q_{\mathbf{k}}$. We also write $Q_k = Q_{(k)}$ for $1 \le k \le r$, which is a maximal parabolic subgroup of ${\mathrm{SO}}(V_n)$.
We denote by $V_n^+$ the unique (up to isometry) $(2n+1)$-dimensional quadratic space over $F$ with Witt index $n$ and trivial discriminant. Note that ${\mathrm{SO}}(V_n^+)$ is split over $F$. If $F$ is local, then the isometry classes of $(2n+1)$-dimensional quadratic spaces over $F$ with trivial discriminant are classified as follows.
- When $n=0$ or $F = {\mathbb{C}}$, there is a unique such quadratic space $V_n^+$.
- When $n>0$ and $F$ is nonarchimedean, or $n=1$ and $F={\mathbb{R}}$, there are precisely two such quadratic spaces $V_n^+$ and $V_n^-$. Here $V_n^-$ is the $(2n+1)$-dimensional quadratic space over $F$ with Witt index $n-1$ and trivial discriminant.
- When $F = {\mathbb{R}}$, there are precisely $n+1$ such quadratic spaces $V_{p,q}$, where $p$ and $q$ are nonnegative integers such that $p+q=2n+1$ and $q \equiv n \bmod 2$. Here $V_{p,q}$ is the quadratic space over ${\mathbb{R}}$ of signature $(p,q)$. Note that $V_n^+ = V_{n+1,n}$ and $V_1^- = V_{0,3}$.
Parabolic induction
-------------------
Suppose that $F$ is local and fix a nontrivial additive character $\psi$ of $F$. Let $\widetilde{P}_{\mathbf{k}} = \widetilde{M}_{\mathbf{k}} N_{\mathbf{k}}$ be the parabolic subgroup of ${\mathrm{Mp}}(W_n)$ as in §\[ss:mp\]. Then any irreducible genuine representation of $\widetilde{M}_{\mathbf{k}}$ is of the form $$(\tau_1 \otimes \chi_\psi) \boxtimes \dots \boxtimes (\tau_m \otimes \chi_\psi) \boxtimes \pi,$$ where $\tau_i$ is an irreducible representation of ${\mathrm{GL}}(X_i)$, $\chi_\psi$ is the genuine character of $\widetilde{{\mathrm{GL}}}(X_i)$ defined in terms of the Weil index associated to $\psi$ (see [@gi1 §2.6]), and $\pi$ is an irreducible genuine representation of ${\mathrm{Mp}}(W_{n_0})$. We write $$I_{P_{\mathbf{k}}, \psi}(\tau_1, \dots, \tau_m, \pi) =
{\operatorname{Ind}}^{{\mathrm{Mp}}(W_n)}_{\widetilde{P}_{\mathbf{k}}}((\tau_1 \otimes \chi_\psi) \boxtimes \dots \boxtimes (\tau_m \otimes \chi_\psi) \boxtimes \pi)$$ for the associated parabolically induced representation. Note that $I_{P_{\mathbf{k}}, \psi}(\tau_1, \dots, \tau_m, \pi)^\vee = I_{P_{\mathbf{k}}, \psi^{-1}}(\tau_1^\vee, \dots, \tau_m^\vee, \pi^\vee)$ and $$I_{P_{\mathbf{k}}, \psi_a}(\tau_1, \dots, \tau_m, \pi) = I_{P_{\mathbf{k}}, \psi}(\tau_1 \otimes (\chi_a \circ \det), \dots, \tau_m \otimes (\chi_a \circ \det), \pi)$$ for $a \in F^\times$. If $I_{P_{\mathbf{k}}, \psi}(\tau_1, \dots, \tau_m, \pi)$ is a standard module, then we denote by $J_{P_{\mathbf{k}}, \psi}(\tau_1, \dots, \tau_m, \pi)$ its unique irreducible quotient.
Similarly, for the parabolic subgroup $Q_{\mathbf{k}}$ of ${\mathrm{SO}}(V_n)$ as in §\[ss:so\], we write $$I_{Q_{\mathbf{k}}}(\tau_1, \dots, \tau_m, \sigma) =
{\operatorname{Ind}}^{{\mathrm{SO}}(V_n)}_{Q_{\mathbf{k}}}(\tau_1 \boxtimes \dots \boxtimes \tau_m \boxtimes \sigma)$$ for the parabolically induced representation associated to irreducible representations $\tau_i$ and $\sigma$ of ${\mathrm{GL}}(Y_i)$ and ${\mathrm{SO}}(V_{n_0})$, respectively. If $I_{Q_{\mathbf{k}}}(\tau_1, \dots, \tau_m, \sigma)$ is a standard module, then we denote by $J_{Q_{\mathbf{k}}}(\tau_1, \dots, \tau_m, \sigma)$ its unique irreducible quotient.
Local theta lifts
=================
In this section, we introduce local theta lifts and recall some of their basic properties.
Notation
--------
Let $F$ be a local field of characteristic zero and fix a nontrivial additive character $\psi$ of $F$. Let $W_n$ be a $2n$-dimensional symplectic space over $F$ and $V_m$ a $(2m+1)$-dimensional quadratic space over $F$ with trivial discriminant. We denote by $$\omega_{W_n,V_m,\psi}$$ the Weil representation of ${\mathrm{Mp}}(W_n) \times {\mathrm{O}}(V_m)$ with respect to $\psi$. Note that $\omega_{W_n,V_m,\psi}$ depends only on the $(F^\times)^2$-orbit of $\psi$.
For any irreducible genuine representation $\pi$ of ${\mathrm{Mp}}(W_n)$, the maximal $\pi$-isotypic quotient of $\omega_{W_n,V_m,\psi}$ is of the form $$\pi \boxtimes \Theta_{W_n,V_m,\psi}(\pi)$$ for some smooth representation $\Theta_{W_n,V_m,\psi}(\pi)$ of ${\mathrm{O}}(V_m)$ of finite length. We denote by $$\theta_{W_n,V_m,\psi}(\pi)$$ the maximal semisimple quotient of $\Theta_{W_n,V_m,\psi}(\pi)$. Then, by the Howe duality [@howe; @w3; @gt2], $\theta_{W_n,V_m,\psi}(\pi)$ is either zero or irreducible. We also regard $\Theta_{W_n,V_m,\psi}(\pi)$ and $\theta_{W_n,V_m,\psi}(\pi)$ as representations of ${\mathrm{SO}}(V_m)$ via restriction. Note that if $\theta_{W_n,V_m,\psi}(\pi)$ is nonzero, then it remains irreducible as a representation of ${\mathrm{SO}}(V_m)$.
Similarly, for any irreducible representation $\sigma$ of ${\mathrm{O}}(V_m)$, we may define a smooth representation $\Theta_{W_n,V_m,\psi}(\sigma)$ of ${\mathrm{Mp}}(W_n)$ of finite length and its maximal semisimple quotient $\theta_{W_n,V_m,\psi}(\sigma)$, which is either zero or irreducible. Let $\sigma_0$ be an irreducible representation of ${\mathrm{SO}}(V_m)$. For $\epsilon = \pm 1$, let $\sigma^\epsilon$ be the extension of $\sigma_0$ to ${\mathrm{O}}(V_m)$ such that $-1 \in {\mathrm{O}}(V_m)$ acts as the scalar $\epsilon$, which we call the $\epsilon$-extension of $\sigma_0$. If $m \ge n$, then by the conservation relation [@sz], there is at most one $\epsilon$ such that $\theta_{W_n,V_m,\psi}(\sigma^\epsilon)$ is nonzero. We write $\theta_{W_n,V_m,\psi}(\sigma_0) = \theta_{W_n,V_m,\psi}(\sigma^\epsilon)$ if such $\epsilon$ exists, and interpret $\theta_{W_n,V_m,\psi}(\sigma_0)$ as zero otherwise.
Elementary Weil representations {#ss:e-weil}
-------------------------------
Consider the Weil representation $\omega_{W_n,V_0^+,\psi}$ of ${\mathrm{Mp}}(W_n) \times {\mathrm{O}}(V_0^+)$ with respect to $\psi$. We may decompose it as $$\omega_{W_n, V_0^+, \psi} = (\omega_{W_n, \psi}^+ \boxtimes 1) \oplus (\omega_{W_n, \psi}^- \boxtimes \det),$$ where $\omega_{W_n, \psi}^+ = \Theta_{W_n,V_0^+,\psi}(1)$ and $\omega_{W_n, \psi}^- = \Theta_{W_n,V_0^+,\psi}(\det)$ are the big theta lifts of the trivial and nontrivial characters of ${\mathrm{O}}(V_0^+) = \{ \pm 1\}$, respectively. Then both $\omega_{W_n, \psi}^+$ and $\omega_{W_n, \psi}^-$ are irreducible, and we call them the even and odd elementary Weil representations of ${\mathrm{Mp}}(W_n)$ with respect to $\psi$, respectively.
\[l:elem-weil\] We have $$\begin{aligned}
\omega_{W_n, \psi}^+ & = J_{B, \psi}(|\cdot|^{n-\frac{1}{2}}, |\cdot|^{n-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}), \\
\omega_{W_n, \psi}^- & = J_{P, \psi}(|\cdot|^{n-\frac{1}{2}}, |\cdot|^{n-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \omega_{W_1, \psi}^-),\end{aligned}$$ where $P = P_{(1^{n-1})}$.
We use the Schrödinger model of the Weil representation $\omega_{W_n, V_0^+, \psi}$ on the space of Schwartz functions on $F^n$. Then $\omega_{W_n, \psi}^+$ and $\omega_{W_n, \psi}^-$ are realized on the subspaces of even and odd functions, respectively. We may define nonzero ${\mathrm{Mp}}(W_n)$-equivariant maps $$\lambda^+ : \omega_{W_n, \psi}^+ \longrightarrow I_{P_n, \psi}(|\det|^{-\frac{n}{2}}), \qquad
\lambda^- : \omega_{W_n, \psi}^- \longrightarrow I_{P_{n-1}, \psi}(|\det|^{-\frac{n+1}{2}}, \omega_{W_1, \psi}^-)$$ by $$\lambda^+(f)(g) = [\omega^+_{W_n,\psi}(g)f](0), \qquad
\lambda^-(f)(g) = [\omega^-_{W_n,\psi}(g)f](\underbrace{0,0,\dots,0}_{n-1}, \, \cdot \,).$$ Since $\omega_{W_n, \psi}^\pm$ is irreducible and $\omega_{W_1, \psi}^-$ is tempered, the images of $\lambda^+$ and $\lambda^-$ must be the unique irreducible subrepresentations of $$I_{P_n, \psi}(|\det|^{-\frac{n}{2}}) \subset I_{B, \psi}(|\cdot|^{-n+\frac{1}{2}}, |\cdot|^{-n+\frac{3}{2}}, \dots, |\cdot|^{-\frac{1}{2}})$$ and $$I_{P_{n-1}, \psi}(|\det|^{-\frac{n+1}{2}}, \omega_{W_1, \psi}^-) \subset I_{P, \psi}(|\cdot|^{-n+\frac{1}{2}}, |\cdot|^{-n+\frac{3}{2}}, \dots, |\cdot|^{-\frac{3}{2}}, \omega_{W_1, \psi}^-),$$ respectively. The lemma now follows from this and the fact that $(\omega_{W_n, \psi}^\pm)^\vee = \omega_{W_n, \psi^{-1}}^\pm$.
Global theta lifts
==================
In this section, we introduce global theta lifts and recall some of their basic properties.
Notation
--------
Let $F$ be a number field and fix a nontrivial additive character $\psi$ of ${\mathbb{A}}/F$. Let $W_n$ be a $2n$-dimensional symplectic space over $F$ and $V_m$ a $(2m+1)$-dimensional quadratic space over $F$ with trivial discriminant. We denote by $$\omega_{W_n,V_m,\psi} = \bigotimes_v \omega_{W_{n,v},V_{m,v},\psi_v}$$ the Weil representation of ${\mathrm{Mp}}(W_n)({\mathbb{A}}) \times {\mathrm{O}}(V_m)({\mathbb{A}})$ with respect to $\psi$, which is equipped with a natural equivariant map $\phi \mapsto \theta(\phi)$ from $\omega_{W_n,V_m,\psi}$ to the space of left ${\mathrm{Sp}}(W_n)(F) \times {\mathrm{O}}(V_m)(F)$-invariant smooth functions on ${\mathrm{Mp}}(W_n)({\mathbb{A}}) \times {\mathrm{O}}(V_m)({\mathbb{A}})$ of moderate growth.
For any irreducible genuine cuspidal automorphic representation $\pi$ of ${\mathrm{Mp}}(W_n)({\mathbb{A}})$, we define an automorphic representation $\Theta_{W_n,V_m,\psi}(\pi)$ of ${\mathrm{O}}(V_m)({\mathbb{A}})$ as the space spanned by all automorphic forms of the form $$\theta(f,\varphi)(h) = \int_{{\mathrm{Sp}}(W_n)(F) \backslash {\mathrm{Mp}}(W_n)({\mathbb{A}})} \theta(f)(g,h) \overline{\varphi(g)} \, dg$$ for $f \in \omega_{W_n,V_m,\psi}$ and $\varphi \in \pi$. If $\Theta_{W_n,V_m,\psi}(\pi)$ is nonzero and is contained in the space of square-integrable automorphic forms on ${\mathrm{O}}(V_m)({\mathbb{A}})$, then by [@kr Corollary 7.1.3], $\Theta_{W_n,V_m,\psi}(\pi)$ is irreducible and is isomorphic to $\bigotimes_v \theta_{W_{n,v},V_{m,v},\psi_v}(\pi_v)$.
Similarly, for any irreducible cuspidal automorphic representation $\sigma$ of ${\mathrm{O}}(V_m)({\mathbb{A}})$, we may define an automorphic representation $\Theta_{W_n,V_m,\psi}(\sigma)$ of ${\mathrm{Mp}}(W_n)({\mathbb{A}})$.
Tower property
--------------
Given an irreducible cuspidal automorphic representation $\sigma$ of ${\mathrm{O}}(V_m)({\mathbb{A}})$, we consider a family of the global theta lifts $\Theta_{W_n,V_m,\psi}(\sigma)$ to ${\mathrm{Mp}}(W_n)({\mathbb{A}})$ when $W_n$ varies. Let $n_0$ be the smallest nonnegative integer such that $\Theta_{W_{n_0},V_m,\psi}(\sigma)$ is nonzero. Then the tower property [@rallis] says that
- such $n_0$ exists;
- $\Theta_{W_{n_0},V_m,\psi}(\sigma)$ is contained in $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}(W_n))$;
- $\Theta_{W_n,V_m,\psi}(\sigma)$ is nonzero but is not contained in $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}(W_n))$ for all $n > n_0$,
where $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}(W_n))$ denotes the space of genuine cusp forms on ${\mathrm{Mp}}(W_n)({\mathbb{A}})$. Moreover, we have the following refinement by M[œ]{}glin [@moeglin-theta] and Jiang–Soudry [@js Theorem 3.6] (note that there is a typo in [@js Theorem 3.6]: $n+a+1$ should be $2n+a+1$).
\[p:tower\] Assume that $n>n_0$ and $n+n_0>2m$. Then $\Theta_{W_n,V_m,\psi}(\sigma)$ is spanned by residues of Eisenstein series $$\operatorname*{Res}_{s=s_0} E(s,\Phi),$$ where $s_0 = \frac{1}{2}(n+n_0-2m)$ and $\Phi$ runs over sections of $${\operatorname{Ind}}^{{\mathrm{Mp}}(W_n)({\mathbb{A}})}_{\widetilde{P}_{n-n_0}({\mathbb{A}})}(\chi_\psi |\det|^s \boxtimes \Theta_{W_{n_0},V_m,\psi}(\sigma)).$$ In particular, $\Theta_{W_n,V_m,\psi}(\sigma)$ is orthogonal to $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}(W_n))$.
Similarly, given an irreducible genuine cuspidal automorphic representation $\pi$ of ${\mathrm{Mp}}(W_n)({\mathbb{A}})$, an analogous result holds for a family of the global theta lifts $\Theta_{W_n,V_m,\psi}(\pi)$ to ${\mathrm{O}}(V_m)({\mathbb{A}})$ when $V_m$ varies in a fixed Witt tower.
Nonvanishing
------------
We now discuss the nonvanishing of global theta lifts. For this, it suffices to compute the Petersson inner products of the global theta lifts. This is achieved by the Rallis inner product formula [@kr; @gqt; @yamana], which expresses the inner products in terms of special values of automorphic $L$-functions and which implies the following.
\[p:nonvanish1\] Let $\sigma$ be an irreducible cuspidal automorphic representation of ${\mathrm{O}}(V_m)({\mathbb{A}})$ and $L(s, \sigma)$ its standard $L$-function. Assume that $m \le n \le 2m$ and that $\Theta_{W_{n-1}, V_m, \psi}(\sigma)$ is zero. Then $\Theta_{W_n, V_m, \psi}(\sigma)$ is nonzero if and only if
- $\Theta_{W_{n,v}, V_{m,v}, \psi_v}(\sigma_v)$ is nonzero for all $v$; and
- $L(s,\sigma)$ is holomorphic and nonzero at $s=n-m+\frac{1}{2}$.
An analogous result also holds for global theta lifts in the other direction, but in this paper, we only need the following weaker result, which is a consequence of the regularized Siegel–Weil formula [@kr; @ichino; @wu] (see also the argument in the proof of [@kr Theorem 7.2.5]).
\[p:nonvanish2\] Let $\pi$ be an irreducible genuine cuspidal automorphic representation of ${\mathrm{Mp}}(W_n)({\mathbb{A}})$ and $L_\psi^S(s, \pi)$ its partial standard $L$-function relative to $\psi$, where $S$ is a sufficiently large finite set of places of $F$.
1. Assume that $m < n$ and that $L_\psi^S(s, \pi)$ has a pole at $s=n-m+\frac{1}{2}$. Then there exits a $(2m+1)$-dimensional quadratic space $V_m$ over $F$ with trivial discriminant such that $\Theta_{W_n,V_m,\psi}(\pi)$ is nonzero.
2. Assume that $m=n$ and that $L_\psi^S(s, \pi)$ is holomorphic and nonzero at $s=\frac{1}{2}$. Then there exits a $(2n+1)$-dimensional quadratic space $V_n$ over $F$ with trivial discriminant such that $\Theta_{W_n,V_n,\psi}(\pi)$ is nonzero.
The residual spectrum of ${\mathrm{Mp}}_4$ {#s:residual}
==========================================
Recall the decomposition $$L^2_{\mathrm{disc}}({\mathrm{Mp}}_4) = L^2_{{\mathrm{cusp}}}({\mathrm{Mp}}_4) \oplus L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$$ of the discrete spectrum into the cuspidal and residual spectra, and the further decomposition $$L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4) = L^2_{P_1}({\mathrm{Mp}}_4) \oplus L^2_{P_2}({\mathrm{Mp}}_4) \oplus L^2_B({\mathrm{Mp}}_4)$$ according to cuspidal supports. In this section, we review the result of Gao [@gao] which describes the structure of $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$.
Notation
--------
Let $F$ be a number field with adèle ring ${\mathbb{A}}$ and fix a nontrivial additive character $\psi$ of ${\mathbb{A}}/F$. As in §\[ss:mp\], let $P_{\mathbf{k}}$ be the parabolic subgroup of ${\mathrm{Sp}}_{2n}$ associated to a sequence $\mathbf{k} = (k_1, \dots, k_m)$ of positive integers such that $k_1 + \dots + k_m \le n$, so that its Levi component is isomorphic to ${\mathrm{GL}}_{k_1} \times \dots \times {\mathrm{GL}}_{k_m} \times {\mathrm{Sp}}_{2n_0}$ with $n_0 = n - k_1 - \dots - k_m$. For any irreducible representation $\tau_i$ of ${\mathrm{GL}}_{k_i}({\mathbb{A}})$ and any irreducible genuine representation $\pi$ of ${\mathrm{Mp}}_{2n_0}({\mathbb{A}})$ such that $I_{P_{\mathbf{k}}, \psi_v}(\tau_{1,v}, \dots, \tau_{m,v}, \pi_v)$ is a standard module for all $v$, we set $$J_{P_{\mathbf{k}}, \psi}(\tau_1, \dots, \tau_m, \pi) =
\bigotimes_v J_{P_{\mathbf{k}}, \psi_v}(\tau_{1,v}, \dots, \tau_{m,v}, \pi_v).$$
Structure of $L^2_{P_1}({\mathrm{Mp}}_4)$
-----------------------------------------
Recall that $P_1 = P_{(1)}$ is the maximal parabolic subgroup of ${\mathrm{Sp}}_4$ with Levi component ${\mathrm{GL}}_1 \times {\mathrm{Sp}}_2$. Recall also that, if $\pi$ is an irreducible genuine cuspidal automorphic representation of ${\mathrm{Mp}}_2({\mathbb{A}})$, then the associated global $A$-parameter $\phi$ is of the form
- $\phi = \rho \boxtimes S_1$ for some irreducible cuspidal automorphic representation $\rho$ of ${\mathrm{GL}}_2({\mathbb{A}})$ with trivial central character; or
- $\phi = \chi \boxtimes S_2$ for some quadratic automorphic character $\chi = \chi_a$ of ${\mathbb{A}}^\times$ with $a \in F^\times$, in which case there is a nonempty finite set $S(\pi)$ of places of $F$ of even cardinality such that $$\pi_v \cong
\begin{cases}
J_{B,\psi_v}(\chi_v|\cdot|_v^{\frac{1}{2}}) & \text{if $v \notin S(\pi)$;} \\
\omega_{W_{1,v}, \psi_{a,v}}^- & \text{if $v \in S(\pi)$.}
\end{cases}$$
By [@gao Theorem 4.5], we have $$L^2_{P_1}({\mathrm{Mp}}_4) = L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{pr}} \oplus L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{SK}} \oplus L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{HPS}}$$ with $$\begin{aligned}
L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{pr}} & \cong \bigoplus_{\chi} \bigoplus_{\pi} J_{P_1,\psi}(\chi|\cdot|^{\frac{3}{2}}, \pi), \label{P1-pr} \\
L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{SK}} & \cong \bigoplus_{\chi, \rho} \bigoplus_{\pi} J_{P_1,\psi}(\chi|\cdot|^{\frac{1}{2}}, \pi), \label{P1-SK} \\
L^2_{P_1}({\mathrm{Mp}}_4)_{\mathrm{HPS}} & \cong \bigoplus_{\chi_1, \chi_2} \bigoplus_{\pi} J_{P_1,\psi}(\chi_1|\cdot|^{\frac{1}{2}}, \pi), \label{P1-HPS}\end{aligned}$$ where
- in , $\chi$ runs over quadratic automorphic characters of ${\mathbb{A}}^\times$, and $\pi$ runs over irreducible genuine cuspidal automorphic representations of ${\mathrm{Mp}}_2({\mathbb{A}})$ with $A$-parameter $\chi \boxtimes S_2$;
- in , $\chi, \rho$ run over pairs of a quadratic automorphic character of ${\mathbb{A}}^\times$ and an irreducible cuspidal automorphic representation of ${\mathrm{GL}}_2({\mathbb{A}})$ with trivial central character such that $L(\frac{1}{2}, \rho \times \chi) \ne 0$, and $\pi$ runs over irreducible genuine cuspidal automorphic representations of ${\mathrm{Mp}}_2({\mathbb{A}})$ with $A$-parameter $\rho \boxtimes S_1$;
- in , $\chi_1, \chi_2$ run over ordered pairs of distinct quadratic automorphic characters of ${\mathbb{A}}^\times$, and $\pi$ runs over irreducible genuine cuspidal automorphic representations of ${\mathrm{Mp}}_2({\mathbb{A}})$ with $A$-parameter $\chi_2 \boxtimes S_2$ such that $\chi_{1,v} \ne \chi_{2,v}$ for all $v \in S(\pi)$.
Note that
- $J_{P_1,\psi}(\chi|\cdot|^{\frac{3}{2}}, \pi)$ in belongs to the near equivalence class determined by the principal $A$-parameter $\chi \boxtimes S_4$;
- $J_{P_1,\psi}(\chi|\cdot|^{\frac{1}{2}}, \pi)$ in belongs to the near equivalence class determined by the $A$-parameter $(\rho \boxtimes S_1) \oplus (\chi \boxtimes S_2)$ of Saito–Kurokawa type;
- $J_{P_1,\psi}(\chi_1|\cdot|^{\frac{1}{2}}, \pi)$ in belongs to the near equivalence class determined by the $A$-parameter $(\chi_1 \boxtimes S_2) \oplus (\chi_2 \boxtimes S_2)$ of Howe–Piatetski-Shapiro type.
Structure of $L^2_{P_2}({\mathrm{Mp}}_4)$
-----------------------------------------
Recall that $P_2 = P_{(2)}$ is the maximal parabolic subgroup of ${\mathrm{Sp}}_4$ with Levi component ${\mathrm{GL}}_2$. By [@gao Theorem 3.4], we have $$L^2_{P_2}({\mathrm{Mp}}_4) \cong \bigoplus_\rho J_{P_2, \psi}(\rho \otimes |\det|^{\frac{1}{2}}),$$ where $\rho$ runs over irreducible dihedral cuspidal automorphic representations of ${\mathrm{GL}}_2({\mathbb{A}})$ with nontrivial quadratic central characters. Note that $J_{P_2, \psi}(\rho \otimes |\det|^{\frac{1}{2}})$ belongs to the near equivalence class determined by the $A$-parameter $\rho \boxtimes S_2$ of Soudry type.
Structure of $L^2_B({\mathrm{Mp}}_4)$
-------------------------------------
Recall that $B = P_{(1,1)}$ is the Borel subgroup of ${\mathrm{Sp}}_4$ with Levi component ${\mathrm{GL}}_1 \times {\mathrm{GL}}_1$. By [@gao Theorem 5.10], we have $$L^2_B({\mathrm{Mp}}_4) = L^2_B({\mathrm{Mp}}_4)_{\mathrm{pr}} \oplus L^2_B({\mathrm{Mp}}_4)_{\mathrm{HPS}}$$ with $$\begin{aligned}
L^2_B({\mathrm{Mp}}_4)_{\mathrm{pr}} & \cong \bigoplus_{\chi} J_{B,\psi}(\chi|\cdot|^{\frac{3}{2}}, \chi|\cdot|^{\frac{1}{2}}), \label{B-pr} \\
L^2_B({\mathrm{Mp}}_4)_{\mathrm{HPS}} & \cong \bigoplus_{\chi_1,\chi_2} J_{B,\psi}(\chi_1|\cdot|^{\frac{1}{2}}, \chi_2|\cdot|^{\frac{1}{2}}), \label{B-HPS}\end{aligned}$$ where
- in , $\chi$ runs over quadratic automorphic characters of ${\mathbb{A}}^\times$;
- in , $\chi_1, \chi_2$ run over unordered pairs of distinct quadratic automorphic characters of ${\mathbb{A}}^\times$.
Note that
- $J_{B,\psi}(\chi|\cdot|^{\frac{3}{2}}, \chi|\cdot|^{\frac{1}{2}})$ in belongs to the near equivalence class determined by the principal $A$-parameter $\chi \boxtimes S_4$;
- $J_{B,\psi}(\chi_1|\cdot|^{\frac{1}{2}}, \chi_2|\cdot|^{\frac{1}{2}})$ in belongs to the near equivalence class determined by the $A$-parameter $(\chi_1 \boxtimes S_2) \oplus (\chi_2 \boxtimes S_2)$ of Howe–Piatetski-Shapiro type.
Principal $A$-packets
=====================
In this section, we construct the $A$-packet associated to a principal $A$-parameter $$\phi = \chi_a \boxtimes S_4$$ with $a \in F^\times$. This is the most degenerate nontempered $A$-packet and is given by elementary Weil representations.
Local $A$-packets {#ss:local-princ}
-----------------
Suppose that $F$ is local and $\phi = \chi_a \boxtimes S_4$ is a local $A$-parameter with $a \in F^\times$. Then we have $S_{\phi} \cong {\mathbb{Z}}/2{\mathbb{Z}}$, so that we may identify $\widehat{S}_\phi$ with $\mu_2$. We define the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ by $$\Pi_{\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^\epsilon \, | \, \epsilon \in \mu_2 \}, \qquad
\pi^\epsilon = \omega_{W_2, \psi_a}^\epsilon.$$ More explicitly, we have $$\pi^+ = J_{B, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \chi_a|\cdot|^{\frac{1}{2}}), \qquad
\pi^- = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \omega_{W_1,\psi_a}^-)$$ by Lemma \[l:elem-weil\]. Hence $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ is multiplicity-free and we have $\Pi_{\varphi_\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^+ \}$ as required, where $\varphi_\phi$ is the $L$-parameter associated to $\phi$ and $\Pi_{\varphi_\phi,\psi}({\mathrm{Mp}}_4)$ is its $L$-packet.
Structure of $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$
-----------------------------------------------
Suppose that $F$ is global. For $\epsilon \in \mu_{2,{\mathbb{A}}}$, put $$\pi^\epsilon = \bigotimes_v \pi_v^{\epsilon_v},$$ where $\pi_v^{\epsilon_v}$ is the representation in the local $A$-packet $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ defined above.
\[p:princ\] We have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\epsilon} \pi^\epsilon,$$ where $\epsilon$ runs over elements in $\mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$.
Proof of Proposition \[p:princ\]
--------------------------------
For $\epsilon \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$, we may realize the elementary Weil representation $\pi^\epsilon$ on the space ${\mathcal{V}}^\epsilon$ of the associated theta functions on ${\mathrm{Mp}}_4({\mathbb{A}})$. Indeed, by the tower property and Proposition \[p:tower\], ${\mathcal{V}}^\epsilon$ is nonzero and is contained in $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$. Moreover, by the description of $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$ in §\[s:residual\] (see in particular and ), we have $$\label{eq:res-pr}
L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cap L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4) = \bigoplus_\epsilon {\mathcal{V}}^\epsilon,$$ where $\epsilon$ runs over elements in $\mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$.
It remains to show that the orthogonal complement of $\bigoplus_\epsilon {\mathcal{V}}^\epsilon$ in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$ is zero. Suppose on the contrary that there exists an irreducible genuine automorphic representation $\pi$ of ${\mathrm{Mp}}_4({\mathbb{A}})$ occurring in this orthogonal complement. Then $\pi$ is cuspidal by and we have $$L^S_{\psi_a}(s, \pi) = \zeta^S(s+\tfrac{3}{2}) \cdot \zeta^S(s+\tfrac{1}{2}) \cdot \zeta^S(s-\tfrac{1}{2}) \cdot \zeta^S(s-\tfrac{3}{2}),$$ where $S$ is a sufficiently large finite set of places of $F$. Since $L^S_{\psi_a}(s, \pi)$ has a pole at $s=\frac{5}{2}$, it follows from Proposition \[p:nonvanish2\] that the global theta lift $\Theta_{W_2,V_0^+,\psi_a}(\pi)$ to ${\mathrm{O}}(V_0^+)({\mathbb{A}})$ is nonzero. Hence, by the adjunction formula, $\pi$ is not orthogonal to ${\mathcal{V}}^\epsilon$ for some $\epsilon \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$, which is a contradiction. This completes the proof of Proposition \[p:princ\].
$A$-packets of Saito–Kurokawa type {#s:sk}
==================================
In this section, we construct the $A$-packet associated to an $A$-parameter of Saito–Kurokawa type $$\phi = (\rho \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$$ with an irreducible cuspidal automorphic representation $\rho$ of ${\mathrm{GL}}_2({\mathbb{A}})$ with trivial central character and $a \in F^\times$. Since the associated character $\tilde{\epsilon}_\phi$ is nontrivial, this is the most interesting nontempered $A$-packet from the viewpoint of the multiplicity formula.
Local $A$-packets {#ss:local-sk}
-----------------
Suppose that $F$ is local and $\phi = (\rho \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ is a local $A$-parameter with a $2$-dimensional symplectic almost tempered representation $\rho$ of $L_F$ and $a \in F^\times$. Here we say that a representation $\rho$ of $L_F$ is *almost tempered* if, for any irreducible summand $\rho_0$ of $\rho$, the image of $W_F$ under $\rho_0 |\cdot|^{-s_0}$ is bounded for some $s_0 \in {\mathbb{R}}$ with $|s_0| < \frac{1}{2}$. Then we have $$S_\phi \cong
\begin{cases}
({\mathbb{Z}}/2{\mathbb{Z}})^2 & \text{if $\rho$ is irreducible;} \\
({\mathbb{Z}}/2{\mathbb{Z}})^2 / ({\mathbb{Z}}/2{\mathbb{Z}}\oplus \{ 0 \}) & \text{if $\rho$ is reducible,}
\end{cases}$$ so that we may identify $\widehat{S}_\phi$ with $$\begin{cases}
\mu_2 \times \mu_2 & \text{if $\rho$ is irreducible;} \\
\{ 1 \} \times \mu_2 & \text{if $\rho$ is reducible.}
\end{cases}$$
We construct the associated $A$-packet by using theta lifts from ${\mathrm{O}}(V_1)$, where $V_1$ is a $3$-dimensional quadratic space over $F$ with trivial discriminant. As explained in §\[ss:so\], up to isometry, there are precisely two such quadratic spaces $V_1^+$ and $V_1^-$ when $F$ is nonarchimedean or $F = {\mathbb{R}}$, and there is a unique such quadratic space $V_1^+$ when $F = {\mathbb{C}}$. For $\epsilon \in \mu_2$, we may identify ${\mathrm{SO}}(V_1^\epsilon)$ with $$(B^\epsilon)^\times / F^\times$$ for some quaternion algebra $B^\epsilon$ over $F$. Let $\sigma_0^\epsilon$ be the irreducible representation of $(B^\epsilon)^\times$ with $L$-parameter $\rho \otimes \chi_a$. Here $\sigma_0^-$ does not exist if $\rho$ is reducible, in which case we interpret $\sigma_0^-$ as zero. Since $\rho \otimes \chi_a$ is symplectic, we may regard $\sigma_0^\epsilon$ as a representation of ${\mathrm{SO}}(V_1^\epsilon)$. For $\epsilon, \epsilon' \in \mu_2$, let $\sigma^{\epsilon,\epsilon'}$ be the $\epsilon'$-extension of $\sigma_0^\epsilon$ to ${\mathrm{O}}(V_1^\epsilon)$. Let $\epsilon_1, \epsilon_2 \in \mu_2$ be such that $$\label{eq:e1e2-SK}
\epsilon = \epsilon_1 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \epsilon(\tfrac{1}{2},\rho \times \chi_a) \cdot \chi_a(-1), \qquad
\epsilon' = \epsilon_1 \cdot \epsilon_2 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \chi_a(-1).$$ Note that $\epsilon = \epsilon_1$ if $\rho$ is reducible.
1. \[item:SK-i\] The theta lift $\theta_{W_1,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$ to ${\mathrm{Mp}}_2(F)$ is nonzero if and only if one of the following holds:
- $\rho$ is irreducible and $\epsilon_2 = 1$;
- $\rho$ is reducible and $\epsilon_1 = \epsilon_2 = 1$.
2. \[item:SK-ii\] The theta lift $\theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$ to ${\mathrm{Mp}}_4(F)$ is nonzero if and only if one of the following holds:
- $\rho$ is irreducible and
- if $F$ is nonarchimedean and $\rho = \chi_a \boxtimes S_2$, or $F = {\mathbb{R}}$ and $\rho = \mathcal{D}_{\frac{1}{2}}$, then $$(\epsilon_1,\epsilon_2) \ne
\begin{cases}
(1,-1) & \text{if $\chi_a \ne 1$;} \\
(-1,-1) & \text{if $\chi_a = 1$;}
\end{cases}$$
- otherwise, $\epsilon_1, \epsilon_2$ are arbitrary;
- $\rho$ is reducible and $\epsilon_1 = 1$.
The assertions and follow from and Lemma \[l:o3mp4-nonzero\] below, respectively.
We now define the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ by $$\Pi_{\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^{\epsilon_1,\epsilon_2} \, | \, \epsilon_1,\epsilon_2 \in \mu_2 \}, \qquad
\pi^{\epsilon_1,\epsilon_2} = \theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'}),$$ where $\epsilon, \epsilon' \in \mu_2$ are as in . Then, by Lemma \[l:o3mp4-mult-free\] below, $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ is multiplicity-free. Moreover, since $$\pi^{\epsilon_1,+} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \theta_{W_1,V_1^\epsilon, \psi_a}(\sigma^{\epsilon,\epsilon'}))$$ by Lemma \[l:ind-princ-1\] below, we have $\Pi_{\varphi_\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^{+,+}, \pi^{-,+} \}$ if $\rho$ is irreducible and $\Pi_{\varphi_\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^{+,+} \}$ if $\rho$ is reducible as required, where $\varphi_\phi$ is the $L$-parameter associated to $\phi$ and $\Pi_{\varphi_\phi,\psi}({\mathrm{Mp}}_4)$ is its $L$-packet. In Appendix \[a:A-packets\] below, we will describe $\pi^{\epsilon_1,\epsilon_2}$ explicitly.
Structure of $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$
-----------------------------------------------
Suppose that $F$ is global. For $\epsilon_1, \epsilon_2 \in \mu_{2,{\mathbb{A}}}$, put $$\pi^{\epsilon_1,\epsilon_2} = \bigotimes_v \pi_v^{\epsilon_{1,v},\epsilon_{2,v}},$$ where $\pi_v^{\epsilon_{1,v},\epsilon_{2,v}}$ is the representation in the local $A$-packet $\Pi_{\phi_v,\psi_v}({\mathrm{Mp}}_4)$ defined above.
\[p:sk\] We have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\epsilon_1,\epsilon_2} \pi^{\epsilon_1,\epsilon_2},$$ where $\epsilon_1,\epsilon_2$ run over elements in $\mu_{2,{\mathbb{A}}}$ such that $$\label{eq:epsilon-sk}
\prod_v \epsilon_{1,v} = \epsilon(\tfrac{1}{2}, \rho) \cdot \epsilon(\tfrac{1}{2}, \rho \times \chi_a), \qquad
\prod_v \epsilon_{2,v} = \epsilon(\tfrac{1}{2}, \rho \times \chi_a).$$
Proof of Proposition \[p:sk\]
-----------------------------
For $\epsilon_1, \epsilon_2 \in \mu_{2,{\mathbb{A}}}$ satisfying , we define $\epsilon, \epsilon' \in \mu_{2,{\mathbb{A}}}$ by $$\epsilon_v = \epsilon_{1,v} \cdot \epsilon(\tfrac{1}{2},\rho_v) \cdot \epsilon(\tfrac{1}{2},\rho_v \times \chi_{a,v}) \cdot \chi_{a,v}(-1), \qquad
\epsilon'_v = \epsilon_{1,v} \cdot \epsilon_{2,v} \cdot \epsilon(\tfrac{1}{2},\rho_v) \cdot \chi_{a,v}(-1),$$ so that $\prod_v \epsilon_v = \prod_v \epsilon_v' = 1$. Put $$\sigma^{\epsilon, \epsilon'} = \bigotimes_v \sigma_v^{\epsilon_v, \epsilon'_v},$$ where $\sigma_v^{\epsilon_v, \epsilon'_v}$ is the representation of ${\mathrm{O}}(V_{1,v}^{\epsilon_v})$ defined above with $L$-parameter $\rho_v \otimes \chi_{a,v}$. If $\sigma^{\epsilon, \epsilon'}$ is nonzero, then $\sigma^{\epsilon, \epsilon'}$ is an irreducible cuspidal automorphic representation of ${\mathrm{O}}(V_1^{\epsilon})({\mathbb{A}})$, where $V_1^{\epsilon}$ is the $3$-dimensional quadratic space over $F$ with trivial discriminant such that $V_1^\epsilon \otimes_F F_v = V_{1,v}^{\epsilon_v}$ for all $v$.
We may realize $\pi^{\epsilon_1, \epsilon_2}$ on the global theta lift $${\mathcal{V}}^{\epsilon_1,\epsilon_2} = \Theta_{W_2,V_1^\epsilon,\psi_a}(\sigma^{\epsilon,\epsilon'}).$$ Indeed, by the tower property and Propositions \[p:tower\], \[p:nonvanish1\], we have
- if the abstract representation $\pi^{\epsilon_1,\epsilon_2}$ is nonzero, then the global theta lift ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is nonzero;
- if $\epsilon_{2,v} = 1$ for all $v$ and $L(\frac{1}{2}, \rho \times \chi_a)$ is nonzero, then ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is contained in $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$;
- otherwise, ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is contained in $L^2_{{\mathrm{cusp}}}({\mathrm{Mp}}_4)$.
Moreover, by the description of $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$ in §\[s:residual\] (see in particular ), we have $$\label{eq:res-SK}
L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cap L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4) \subset \bigoplus_{\epsilon_1,\epsilon_2} {\mathcal{V}}^{\epsilon_1,\epsilon_2},$$ where $\epsilon_1,\epsilon_2$ run over elements in $\mu_{2,{\mathbb{A}}}$ satisfying .
It remains to show that the orthogonal complement of $\bigoplus_{\epsilon_1,\epsilon_2} {\mathcal{V}}^{\epsilon_1,\epsilon_2}$ in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$ is zero. Suppose on the contrary that there exists an irreducible genuine automorphic representation $\pi$ of ${\mathrm{Mp}}_4({\mathbb{A}})$ occurring in this orthogonal complement. Then $\pi$ is cuspidal by and we have $$L^S_{\psi_a}(s,\pi) = L^S(s,\rho \times \chi_a) \cdot \zeta^S(s+\tfrac{1}{2}) \cdot \zeta^S(s-\tfrac{1}{2}),$$ where $S$ is a sufficiently large finite set of places of $F$. Since $L^S_{\psi_a}(s,\pi)$ has a pole at $s = \frac{3}{2}$, it follows from Proposition \[p:nonvanish2\] that the global theta lift $\Theta_{W_2,V_1^\epsilon, \psi_a}(\pi)$ to ${\mathrm{O}}(V_1^\epsilon)({\mathbb{A}})$ is nonzero for some $\epsilon \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$. Moreover, since $\theta_{W_{2,v},V_{1,v}^{\epsilon_v}, \psi_{a,v}}(\pi_v) \cong \sigma_v^{+,+}$ for almost all $v$, we have $\Theta_{W_2,V_1^\epsilon, \psi_a}(\pi) = \sigma^{\epsilon,\epsilon'}$ for some $\epsilon' \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v' = 1$ by the strong multiplicity one theorem. Hence, by the adjunction formula, $\pi$ is not orthogonal to ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ for some $\epsilon_1,\epsilon_2 \in \mu_{2,{\mathbb{A}}}$ satisfying , which is a contradiction. This completes the proof of Proposition \[p:sk\].
$A$-packets of Howe–Piatetski-Shapiro type {#s:howe-ps}
==========================================
In this section, we construct the $A$-packet associated to an $A$-parameter of Howe–Piatetski-Shapiro type $$\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$$ with $a, b \in F^\times$ such that $\chi_a \ne \chi_b$.
Local $A$-packets {#ss:local-howe-ps}
-----------------
Suppose that $F$ is local and $\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ is a local $A$-parameter with $a,b \in F^\times$. Then we have $$S_\phi \cong
\begin{cases}
({\mathbb{Z}}/2{\mathbb{Z}})^2 & \text{if $\chi_a \ne \chi_b$;} \\
({\mathbb{Z}}/2{\mathbb{Z}})^2 / \Delta {\mathbb{Z}}/2{\mathbb{Z}}& \text{if $\chi_a = \chi_b$,}
\end{cases}$$ so that we may identify $\widehat{S}_\phi$ with $$\begin{cases}
\mu_2 \times \mu_2 & \text{if $\chi_a \ne \chi_b$;} \\
\Delta \mu_2 & \text{if $\chi_a = \chi_b$.}
\end{cases}$$
We construct the associated $A$-packet by using theta lifts from ${\mathrm{O}}(V_1)$ as in §\[ss:local-sk\], but with the $L$-parameter $\rho \boxtimes S_1$ replaced by the $A$-parameter $\chi_b \boxtimes S_2$. For $\epsilon \in \mu_2$, let $\sigma_0^\epsilon$ be the irreducible representation of $(B^\epsilon)^\times$ with $A$-parameter $\chi_{ab} \boxtimes S_2$, i.e. $$\sigma_0^\epsilon = \chi_{ab} \circ \mathrm{N}_{B^\epsilon},$$ where $\mathrm{N}_{B^\epsilon}$ is the reduced norm on $B^\epsilon$. Here $\sigma_0^-$ does not exist if $F = {\mathbb{C}}$, in which case we interpret $\sigma_0^-$ as zero. We may regard $\sigma_0^\epsilon$ as a representation of ${\mathrm{SO}}(V_1^\epsilon)$. For $\epsilon, \epsilon' \in \mu_2$, let $\sigma^{\epsilon,\epsilon'}$ be the $\epsilon'$-extension of $\sigma_0^\epsilon$ to ${\mathrm{O}}(V_1^\epsilon)$. Let $\epsilon_1, \epsilon_2 \in \mu_2$ be such that $$\label{eq:e1e2-HPS}
\epsilon = \epsilon_2, \qquad
\epsilon' = \epsilon_1 \cdot \epsilon_2 \cdot \chi_{ab}(-1).$$
1. \[item:HPS-i\] The theta lift $\theta_{W_1,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$ to ${\mathrm{Mp}}_2(F)$ is nonzero if and only if one of the following holds:
- $F \ne {\mathbb{C}}$, $\chi_a \ne \chi_b$, and $\epsilon_1 = 1$;
- $F \ne {\mathbb{C}}$, $\chi_a = \chi_b$, and $\epsilon_1 = \epsilon_2$;
- $F = {\mathbb{C}}$ (so that $\chi_a = \chi_b$) and $\epsilon_1 = \epsilon_2 = 1$.
2. \[item:HPS-ii\] The theta lift $\theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$ to ${\mathrm{Mp}}_4(F)$ is nonzero if and only if one of the following holds:
- $F$ is nonarchimedean, $\chi_a \ne \chi_b$, and $\epsilon_1, \epsilon_2$ are arbitrary;
- $F = {\mathbb{R}}$, $\chi_a \ne \chi_b$, and $(\epsilon_1, \epsilon_2) \ne (-1,-1)$;
- $F \ne {\mathbb{C}}$, $\chi_a = \chi_b$, and $\epsilon_1 = \epsilon_2$;
- $F = {\mathbb{C}}$ (so that $\chi_a = \chi_b$) and $\epsilon_1 = \epsilon_2 = 1$.
The assertions and follow from and Lemma \[l:o3mp4-nonzero\] below, respectively.
We now define the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ by $$\Pi_{\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi^{\epsilon_1,\epsilon_2} \, | \, \epsilon_1,\epsilon_2 \in \mu_2 \}, \qquad
\pi^{\epsilon_1,\epsilon_2} = \theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'}),$$ where $\epsilon, \epsilon' \in \mu_2$ are as in . Then, by Lemma \[l:o3mp4-mult-free\] below, $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ is multiplicity-free. Moreover, since $$\pi^{+,+} = J_{B, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \chi_b |\cdot|^{\frac{1}{2}})$$ by Lemma \[l:howe-ps-local+\] below, we have $\Pi_{\varphi_\phi, \psi}({\mathrm{Mp}}_4) = \{ \pi^{+,+} \}$ as required, where $\varphi_\phi$ is the $L$-parameter associated to $\phi$ and $\Pi_{\varphi_\phi, \psi}({\mathrm{Mp}}_4)$ is its $L$-packet. In Appendix \[a:A-packets\] below, we will describe $\pi^{\epsilon_1,\epsilon_2}$ explicitly. We only remark that if $\chi_a \ne \chi_b$, then it follows from Lemmas \[l:howe-ps-local+\] and \[l:howe-ps-local-\] below that $$\pi^{+,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_b}^-), \qquad
\pi^{-,+} = J_{P_1, \psi}(\chi_b |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_a}^-),$$ which we will use later.
Structure of $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$
-----------------------------------------------
Suppose that $F$ is global. For $\epsilon_1, \epsilon_2 \in \mu_{2,{\mathbb{A}}}$, put $$\pi^{\epsilon_1,\epsilon_2} = \bigotimes_v \pi_v^{\epsilon_{1,v},\epsilon_{2,v}},$$ where $\pi_v^{\epsilon_{1,v},\epsilon_{2,v}}$ is the representation in the local $A$-packet $\Pi_{\phi_v, \psi_v}({\mathrm{Mp}}_4)$ defined above.
\[p:howe-ps\] We have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\epsilon_1,\epsilon_2} \pi^{\epsilon_1,\epsilon_2},$$ where $\epsilon_1,\epsilon_2$ run over elements in $\mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_{1,v} = \prod_v \epsilon_{2,v} = 1$.
Proof of Proposition \[p:howe-ps\]
----------------------------------
The proof is similar to that of Proposition \[p:sk\]. For $\epsilon_1, \epsilon_2 \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_{1,v} = \prod_v \epsilon_{2,v} = 1$, we define $\epsilon, \epsilon' \in \mu_{2,{\mathbb{A}}}$ by $$\epsilon_v = \epsilon_{2,v}, \qquad
\epsilon'_v = \epsilon_{1,v} \cdot \epsilon_{2,v} \cdot \chi_{ab,v}(-1),$$ so that $\prod_v \epsilon_v = \prod_v \epsilon_v' = 1$. Put $$\sigma^{\epsilon, \epsilon'} = \bigotimes_v \sigma_v^{\epsilon_v, \epsilon'_v},$$ where $\sigma_v^{\epsilon_v, \epsilon'_v}$ is the representation of ${\mathrm{O}}(V_{1,v}^{\epsilon_v})$ defined above with $A$-parameter $\chi_{ab,v} \boxtimes S_2$. If $\sigma^{\epsilon, \epsilon'}$ is nonzero, then $\sigma^{\epsilon, \epsilon'}$ is a $1$-dimensional automorphic representation of ${\mathrm{O}}(V_1^{\epsilon})({\mathbb{A}})$, where $V_1^{\epsilon}$ is the $3$-dimensional quadratic space over $F$ with trivial discriminant such that $V_1^\epsilon \otimes_F F_v = V_{1,v}^{\epsilon_v}$ for all $v$.
Suppose that $\epsilon_{2,v} \ne 1$ for some $v$, so that $V_1^\epsilon$ is anisotropic and $\sigma^{\epsilon, \epsilon'}$ is cuspidal. Then we may realize $\pi^{\epsilon_1, \epsilon_2}$ on the global theta lift $${\mathcal{V}}^{\epsilon_1,\epsilon_2} = \Theta_{W_2,V_1^\epsilon,\psi_a}(\sigma^{\epsilon,\epsilon'}).$$ Indeed, by the tower property and Propositions \[p:tower\], \[p:nonvanish1\], we have:
- if the abstract representation $\pi^{\epsilon_1,\epsilon_2}$ is nonzero, then the global theta lift ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is nonzero;
- if $$\epsilon_{1,v} =
\begin{cases}
1 & \text{if $\chi_{a,v} \ne \chi_{b,v}$;} \\
\epsilon_{2,v} & \text{if $\chi_{a,v} = \chi_{b,v}$}
\end{cases}$$ for all $v$, then ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is contained in $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$;
- otherwise, ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ is contained in $L^2_{{\mathrm{cusp}}}({\mathrm{Mp}}_4)$.
Moreover, by the description of $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$ in §\[s:residual\] (see in particular and ), we may realize $\pi^{\epsilon_1,\epsilon_2}$ on a subspace ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ of $L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4)$ even if $\epsilon_{2,v} = 1$ for all $v$, and we have $$\label{eq:res-HPS}
L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cap L^2_{{\mathrm{res}}}({\mathrm{Mp}}_4) \subset \bigoplus_{\epsilon_1,\epsilon_2} {\mathcal{V}}^{\epsilon_1,\epsilon_2},$$ where $\epsilon_1,\epsilon_2$ run over elements in $\mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_{1,v} = \prod_v \epsilon_{2,v} = 1$.
It remains to show that the orthogonal complement of $\bigoplus_{\epsilon_1,\epsilon_2} {\mathcal{V}}^{\epsilon_1,\epsilon_2}$ in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$ is zero. Suppose on the contrary that there exists an irreducible genuine automorphic representation $\pi$ of ${\mathrm{Mp}}_4({\mathbb{A}})$ occurring in this orthogonal complement. Then $\pi$ is cuspidal by and we have $$L^S_{\psi_a}(s,\pi) = \zeta^S(s+\tfrac{1}{2}) \cdot \zeta^S(s-\tfrac{1}{2}) \cdot L^S(s+\tfrac{1}{2},\chi_{ab}) \cdot L^S(s-\tfrac{1}{2},\chi_{ab}),$$ where $S$ is a sufficiently large finite set of places of $F$. Since $L^S_{\psi_a}(s,\pi)$ has a pole at $s = \frac{3}{2}$, it follows from Proposition \[p:nonvanish2\] that the global theta lift $\Theta_{W_2,V_1^\epsilon, \psi_a}(\pi)$ to ${\mathrm{O}}(V_1^\epsilon)({\mathbb{A}})$ is nonzero for some $\epsilon \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v = 1$. Moreover, since $\theta_{W_{2,v},V_{1,v}^{\epsilon_v}, \psi_{a,v}}(\pi_v) \cong \sigma_v^{+,+}$ for almost all $v$, we have $\Theta_{W_2,V_1^\epsilon, \psi_a}(\pi) = \sigma^{\epsilon,\epsilon'}$ for some $\epsilon' \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_v' = 1$ by the strong multiplicity one theorem. Hence, if $\epsilon_v \ne 1$ for some $v$, then by the adjunction formula, $\pi$ is not orthogonal to ${\mathcal{V}}^{\epsilon_1,\epsilon_2}$ for some $\epsilon_1,\epsilon_2 \in \mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_{1,v} = \prod_v \epsilon_{2,v} = 1$ and $\epsilon_{2,v} \ne 1$ for some $v$, which is a contradiction. This forces $\epsilon_v = 1$ for all $v$, i.e. $V_1^\epsilon = V_1^+$. Since $\sigma^{\epsilon,\epsilon'}$ is not cuspidal, the global theta lift $\Theta_{W_2,V_0^+, \psi_a}(\pi)$ to ${\mathrm{O}}(V_0^+)({\mathbb{A}})$ is nonzero by the tower property, so that $$\pi_v \cong J_{B, \psi_v}(\chi_{a,v} |\cdot|_v^{\frac{3}{2}}, \chi_{a,v} |\cdot|_v^{\frac{1}{2}})$$ for almost all $v$. This contradicts the assumption that $\pi$ occurs in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$ and completes the proof of Proposition \[p:howe-ps\].
$A$-packets of Soudry type {#s:soudry}
==========================
In this section, we construct the $A$-packet associated to an $A$-parameter of Soudry type $$\phi = \rho \boxtimes S_2$$ with an irreducible dihedral cuspidal automorphic representation $\rho$ of ${\mathrm{GL}}_2({\mathbb{A}})$ with nontrivial quadratic central character. This is the most troublesome nontempered $A$-packet in the sense that unlike the other nontempered $A$-packets, it cannot be constructed by using theta lifts from smaller orthogonal groups. We will deal with it in the same way as [@gi-mp], where we dealt with the tempered $A$-packets by using theta lifts to much larger orthogonal groups.
Local $A$-packets {#ss:local-soudry}
-----------------
Suppose that $F$ is local and $\phi = \rho \boxtimes S_2$ is a local $A$-parameter with a $2$-dimensional orthogonal tempered representation $\rho$ of $L_F$. Note that such $\rho$ can be regarded as a representation of $W_F$. We construct the associated $A$-packet $$\Pi_{\phi, \psi}({\mathrm{Mp}}_4) = \{ \pi_\eta \, | \, \eta \in \widehat{S}_\phi \}$$ in a nonuniform way depending on whether $\rho$ is irreducible or not.
### The reducible case
Suppose that $\rho$ is reducible. Then $\rho$ is of the form either $$\rho = \chi \oplus \chi^{-1}$$ for some unitary character $\chi$ of $F^\times$ such that $\chi^2 \ne 1$, or $$\rho = \chi_a \oplus \chi_b$$ for some $a, b \in F^\times$. In the former case, $S_\phi$ is trivial and we define the unique representation in $\Pi_{\phi, \psi}({\mathrm{Mp}}_4)$ as the unique irreducible genuine representation of ${\mathrm{Mp}}_4(F)$ with $L$-parameter $\varphi_\phi$ associated to $\phi$. In the latter case, $\phi$ is of Howe–Piatetski-Shapiro type and we have already constructed $\Pi_{\phi, \psi}({\mathrm{Mp}}_4)$ in §\[ss:local-howe-ps\].
### The irreducible case
Suppose that $\rho$ is irreducible. Then either $F$ is nonarchimedean or $F={\mathbb{R}}$. Also, we have $S_\phi \cong {\mathbb{Z}}/2{\mathbb{Z}}$, so that we may identify $\widehat{S}_\phi$ with $\mu_2$. We define the $A$-packet $\Pi_{\phi, \psi}({\mathrm{Mp}}_4)$ by $$\Pi_{\phi, \psi}({\mathrm{Mp}}_4) = \{ \pi^\epsilon \, | \, \epsilon \in \mu_2 \}$$ with the representation $\pi^\epsilon$ given as follows. Put $$\pi^+ = J_{P_2,\psi}(\tau \otimes |\det|^{\frac{1}{2}}),$$ where $\tau$ is the irreducible square-integrable representation of ${\mathrm{GL}}_2(F)$ with $L$-parameter $\rho$. When $F$ is nonarchimedean, we consider a symplectic representation $\varphi = \rho \boxtimes S_2$ of $L_F = W_F \times {\mathrm{SL}}_2({\mathbb{C}})$ and put $$\pi^- = \pi^-_\varphi,$$ where $\pi^-_\varphi$ is the irreducible genuine square-integrable representation of ${\mathrm{Mp}}_4(F)$ with $L$-parameter $\varphi$ (relative to $\psi$) associated to the nontrivial character of $S_\varphi \cong {\mathbb{Z}}/ 2{\mathbb{Z}}$. When $F = {\mathbb{R}}$, we write $\rho = \mathcal{D}_\kappa$ with some positive integer $\kappa$ and put $$\pi^- = \pi_\Lambda \oplus \pi_\Lambda^\vee,$$ where $\pi_\Lambda$ is the genuine discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $\widetilde{{\mathrm{U}}}(2)$-type $\Lambda = (\kappa+\frac{3}{2}, \kappa+\frac{3}{2})$. This is the only instance of a reducible representation of ${\mathrm{Mp}}_4(F)$ in an $A$-packet.
Structure of $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$
-----------------------------------------------
Suppose that $F$ is global. For $\eta \in \widehat{S}_{\phi,{\mathbb{A}}}$, put $$\pi_\eta = \bigotimes_v \pi_{\eta_v},$$ where $\pi_{\eta_v}$ is the representation in the local $A$-packet $\Pi_{\phi_v, \psi_v}({\mathrm{Mp}}_4)$ defined above.
\[p:soudry\] We have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\eta} \pi_\eta,$$ where $\eta$ runs over elements in $\widehat{S}_{\phi,{\mathbb{A}}}$ such that $\Delta^* \eta = 1$.
The multiplicity preservation
-----------------------------
Since we have constructed the local $A$-packet $\Pi_{\phi_v, \psi_v}({\mathrm{Mp}}_4)$ in a nonuniform way, it is somewhat tricky to construct the near equivalence class $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$. We follow [@gi-mp] and appeal to some result of J.-S. Li [@li97] on theta lifts in the stable range. Namely, we fix an integer $r>3$ (in fact, we will take $r=4$ later) and consider (abstract) theta lifts from ${\mathrm{Mp}}_4({\mathbb{A}})$ to ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$, where ${\mathrm{SO}}_{2r+5} = {\mathrm{SO}}(V_{r+2}^+)$ denotes the split odd special orthogonal group of rank $r+2$. For any irreducible genuine unitary representation $\pi$ of ${\mathrm{Mp}}_4({\mathbb{A}})$, we may define an irreducible unitary representation $\theta_\psi(\pi)$ of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$ by $$\theta_\psi(\pi) = \bigotimes_v \theta_{W_{2,v}, V_{r+2,v}^+, \psi_v}(\pi_v).$$ Let $L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ be the near equivalence class in the automorphic discrete spectrum of ${\mathrm{SO}}_{2r+5}$ determined by the elliptic $A$-parameter $$\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r}).$$
\[p:mult-preservation\] Writing $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_\pi m_{\phi,\psi}(\pi) \pi,$$ where $\pi$ runs over irreducible genuine unitary representations of ${\mathrm{Mp}}_4({\mathbb{A}})$ and $m_{\phi,\psi}(\pi)$ is the multiplicity of $\pi$ in $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$, we have $$L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5}) \cong \bigoplus_\pi m_{\phi,\psi}(\pi) \theta_\psi(\pi).$$
The proof is similar to that of [@gi-mp Corollary 4.2]. Put $$\begin{aligned}
m(\pi) & = \dim_{\mathbb{C}}{\operatorname{Hom}}_{{\mathrm{Mp}}_4({\mathbb{A}})}(\pi, \mathcal{A}({\mathrm{Mp}}_4)), \\
m_{\mathrm{disc}}(\pi) & = \dim_{\mathbb{C}}{\operatorname{Hom}}_{{\mathrm{Mp}}_4({\mathbb{A}})}(\pi, \mathcal{A}^2({\mathrm{Mp}}_4)),\end{aligned}$$ where $\mathcal{A}({\mathrm{Mp}}_4)$ is the space of genuine automorphic forms on ${\mathrm{Mp}}_4({\mathbb{A}})$ and $\mathcal{A}^2({\mathrm{Mp}}_4)$ is the subspace of square-integrable automorphic forms. Similarly, for any irreducible unitary representation $\sigma$ of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$, we may define its multiplicities $m(\sigma)$ and $m_{\mathrm{disc}}(\sigma)$. Then the result of J.-S. Li [@li97] says that $$\label{eq:JSLi}
m_{\mathrm{disc}}(\pi) \le m_{\mathrm{disc}}(\theta_\psi(\pi)) \le m(\theta_\psi(\pi)) \le m(\pi).$$ This and the Howe duality imply that there is an embedding $$\bigoplus_\pi m_{\phi,\psi}(\pi) \theta_\psi(\pi) {\lhook\joinrel\longrightarrow}L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5}).$$ To show that this embedding is an isomorphism, it suffices to prove the following:
- $m_{\mathrm{disc}}(\pi) = m_{\mathrm{disc}}(\theta_\psi(\pi))$ for any irreducible genuine unitary representation $\pi$ of ${\mathrm{Mp}}_4({\mathbb{A}})$ such that the $L$-parameter of $\pi_v$ (relative to $\psi_v$) is $\varphi_{\phi_v}$ for almost all $v$;
- any irreducible summand of $L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ is isomorphic to $\theta_\psi(\pi)$ for some irreducible summand $\pi$ of $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$.
The first assertion follows from and Lemma \[l:soudry-mult\] below. The second assertion follows from the first one as in the proof of [@gi-mp Corollary 4.2].
\[l:soudry-mult\] Let $\pi$ be an irreducible genuine unitary representation of ${\mathrm{Mp}}_4({\mathbb{A}})$ such that the $L$-parameter of $\pi_v$ (relative to $\psi_v$) is $\varphi_{\phi_v}$ for almost all $v$. Then we have $$m_{\mathrm{disc}}(\pi) = m(\pi).$$
We may assume that $m(\pi)>0$. For any realization of $\pi$ on a subspace ${\mathcal{V}}$ of $\mathcal{A}({\mathrm{Mp}}_4)$, we need to show that ${\mathcal{V}}$ is contained in $\mathcal{A}^2({\mathrm{Mp}}_4)$. We may further assume that ${\mathcal{V}}$ is not contained in $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}_4)$. Since the weak lift of $\pi$ to ${\mathrm{GL}}_4({\mathbb{A}})$ is $$(\rho \otimes |\det|^{\frac{1}{2}}) \boxplus (\rho \otimes |\det|^{-\frac{1}{2}}),$$ where $\boxplus$ denotes the isobaric sum, it follows from the argument in the proof of [@gi-mp Proposition 4.1] that the cuspidal support of ${\mathcal{V}}$ is on $P_2$ and is of the form $$\rho \otimes |\det|^{\pm \frac{1}{2}}.$$ In particular, $\pi$ is a subrepresentation of $I_{P_2,\psi}(\rho \otimes |\det|^{\pm \frac{1}{2}})$. On the other hand, we have $\pi_v \cong J_{P_2,\psi_v}(\rho_v \otimes |\det|_v^{\frac{1}{2}})$ for almost all $v$, which does not occur as a subrepresentation of $I_{P_2,\psi_v}(\rho_v \otimes |\det|_v^{\frac{1}{2}})$. Hence the cuspidal support of ${\mathcal{V}}$ must be $\rho \otimes |\det|^{-\frac{1}{2}}$, so that ${\mathcal{V}}$ is contained in $\mathcal{A}^2({\mathrm{Mp}}_4)$ by the square-integrability criterion. This completes the proof.
\[r:howe-ps\] In fact, Proposition \[p:mult-preservation\] holds for other elliptic $A$-parameters for ${\mathrm{Mp}}_4$. We give some details here in the case of $A$-parameters of Howe–Piatetski-Shapiro type, which we will use later. Suppose that $\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ with $a, b \in F^\times$ such that $\chi_a \ne \chi_b$. Let $\pi$ be an irreducible genuine unitary representation of ${\mathrm{Mp}}_4({\mathbb{A}})$ such that the $L$-parameter of $\pi_v$ (relative to $\psi_v$) is $\varphi_{\phi_v}$ for almost all $v$, so that its weak lift to ${\mathrm{GL}}_4({\mathbb{A}})$ is $$\chi_a |\cdot|^{\frac{1}{2}} \boxplus \chi_b |\cdot|^{\frac{1}{2}} \boxplus \chi_a |\cdot|^{-\frac{1}{2}} \boxplus \chi_b |\cdot|^{-\frac{1}{2}}.$$ Then, for any realization of $\pi$ on a subspace ${\mathcal{V}}$ of $\mathcal{A}({\mathrm{Mp}}_4)$ which is not contained in $\mathcal{A}_{{\mathrm{cusp}}}({\mathrm{Mp}}_4)$, it follows from the same argument that the cuspidal support of ${\mathcal{V}}$ is on $B$ and is of the form $\chi_a |\cdot|^{-\frac{1}{2}} \otimes \chi_b |\cdot|^{-\frac{1}{2}}$. This implies that $m_{\mathrm{disc}}(\pi) = m(\pi)$, from which we can deduce Proposition \[p:mult-preservation\] for $\phi$.
Proof of Proposition \[p:soudry\]
---------------------------------
We first apply Arthur’s multiplicity formula to the near equivalence class $L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$. For the elliptic $A$-parameter $\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r})$, we have $$S_{\theta(\phi)} \cong S_\phi \oplus {\mathbb{Z}}/2{\mathbb{Z}},$$ where the extra factor ${\mathbb{Z}}/2{\mathbb{Z}}$ corresponds to $1 \boxtimes S_{2r}$. Similarly, for each place $v$ of $F$ and the local $A$-parameter $\theta(\phi_v) = \phi_v \oplus (1 \boxtimes S_{2r})$, we have $$S_{\theta(\phi_v)} \cong S_{\phi_v} \oplus {\mathbb{Z}}/2{\mathbb{Z}}.$$ Put $$\overline{S}_{\theta(\phi_v)} = S_{\theta(\phi_v)} / \langle z_{\theta(\phi_v)} \rangle,$$ where $z_{\theta(\phi_v)}$ is the image of $-1 \in {\mathrm{Sp}}_{2r+4}({\mathbb{C}})$ in $S_{\theta(\phi_v)}$. Then we have a natural isomorphism $$\iota_v : S_{\phi_v} {\lhook\joinrel\longrightarrow}S_{\theta(\phi_v)} {\relbar\joinrel\twoheadrightarrow}\overline{S}_{\theta(\phi_v)}.$$ Let $$\Pi_{\theta(\phi_v)}({\mathrm{SO}}_{2r+5}) = \{ \sigma_{\xi_v} \, | \, \xi_v \in \widehat{\overline{S}}_{\theta(\phi_v)} \}$$ be the local $A$-packet defined by Arthur [@arthur] consisting of some semisimple representations of ${\mathrm{SO}}_{2r+5}(F_v)$ of finite length. For $\xi \in \widehat{\overline{S}}_{\theta(\phi), {\mathbb{A}}}$, put $$\sigma_\xi = \bigotimes_v \sigma_{\xi_v}.$$ Then, noting that the character $\epsilon_{\theta(\phi)}$ associated to $\theta(\phi)$ is trivial, we deduce from Arthur’s multiplicity formula [@arthur Theorem 1.5.2] that $$\label{eq:mult-so}
L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5}) \cong \bigoplus_\xi \sigma_\xi,$$ where $\xi$ runs over elements in $\widehat{\overline{S}}_{\theta(\phi), {\mathbb{A}}}$ such that $\Delta^* \xi = 1$.
We now use Proposition \[p:mult-preservation\] to transfer the structure of $L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ to $L^2_{\phi,\psi}({\mathrm{Mp}}_4)$. For $\eta_v \in \widehat{S}_{\phi_v}$ and $\xi_v = \eta_v \circ \iota_v^{-1} \in \widehat{\overline{S}}_{\theta(\phi_v)}$, write $$\sigma_{\xi_v} = \bigoplus_i m_{\xi_v,i} \sigma_{\xi_v,i}$$ with some positive integers $m_{\xi_v,i}$ and some pairwise distinct irreducible representations $\sigma_{\xi_v,i}$ of ${\mathrm{SO}}_{2r+5}(F_v)$, and put $$\label{eq:tilde_pi_eta}
\tilde{\pi}_{\eta_v} = \bigoplus_i m_{\xi_v,i} \theta_{W_{2,v}, V_{r+2,v}^+,\psi_v}(\sigma_{\xi_v,i}).$$ For $\eta \in \widehat{S}_{\phi,{\mathbb{A}}}$, put $$\tilde{\pi}_\eta = \bigotimes_v \tilde{\pi}_{\eta_v}.$$ Then, by Proposition \[p:mult-preservation\] and , we have $$L^2_{\phi,\psi}({\mathrm{Mp}}_4) \cong \bigoplus_{\eta} \tilde{\pi}_\eta,$$ where $\eta$ runs over elements in $\widehat{S}_{\phi,{\mathbb{A}}}$ such that $\Delta^* \eta = 1$. Thus, to finish the proof of Proposition \[p:soudry\], we may take $r=4$ and need to prove the following.
\[p:soudry-local\] Assume that $r=4$. Then we have $$\tilde{\pi}_{\eta_v} = \pi_{\eta_v}$$ for $\eta_v \in \widehat{S}_{\phi_v}$. Here $\tilde{\pi}_{\eta_v}$ is defined by , whereas $\pi_{\eta_v}$ is defined in §\[ss:local-soudry\].
This proposition will be proved in the next section. In fact, it holds for any $r>3$ when either $F$ is nonarchimedean or $F={\mathbb{C}}$.
We recall the analog of here in the case of $A$-parameters of Howe–Piatetski-Shapiro type, which we will use later. Suppose that $\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ with $a, b \in F^\times$ such that $\chi_a \ne \chi_b$ and put $\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r})$. For each place $v$ of $F$, let $$\Pi_{\theta(\phi_v)}({\mathrm{SO}}_{2r+5}) = \{ \sigma_v^{\epsilon_{1,v},\epsilon_{2,v}} \, | \, \epsilon_{1,v}, \epsilon_{2,v} \in \mu_2 \}$$ be the local $A$-packet defined by Arthur [@arthur] consisting of some semisimple representations of ${\mathrm{SO}}_{2r+5}(F_v)$ of finite length, where we identify $\widehat{\overline{S}}_{\theta(\phi_v)}$ with $$\begin{cases}
\mu_2 \times \mu_2 & \text{if $\chi_{a,v} \ne \chi_{b,v}$;} \\
\Delta \mu_2 & \text{if $\chi_{a,v} = \chi_{b,v}$,}
\end{cases}$$ and interpret $\sigma_v^{\epsilon_{1,v},\epsilon_{2,v}}$ as zero if $\chi_{a,v} = \chi_{b,v}$ and $\epsilon_{1,v} \ne \epsilon_{2,v}$. For $\epsilon_1, \epsilon_2 \in \mu_{2,{\mathbb{A}}}$, put $$\sigma^{\epsilon_1,\epsilon_2} = \bigotimes_v \sigma_v^{\epsilon_{1,v},\epsilon_{2,v}}.$$ Then, noting that the character $\epsilon_{\theta(\phi)}$ associated to $\theta(\phi)$ is trivial, we deduce from Arthur’s multiplicity formula [@arthur Theorem 1.5.2] that $$\label{eq:mult-so2}
L^2_{\theta(\phi)}({\mathrm{SO}}_{2r+5}) \cong \bigoplus_{\epsilon_1,\epsilon_2} \sigma^{\epsilon_1,\epsilon_2},$$ where $\epsilon_1,\epsilon_2$ run over elements in $\mu_{2,{\mathbb{A}}}$ such that $\prod_v \epsilon_{1,v} = \prod_v \epsilon_{2,v} = 1$.
Proof of Proposition \[p:soudry-local\] {#s:pf-soudry}
=======================================
Suppose that $F$ is local and consider local $A$-parameters $$\phi = \rho \boxtimes S_2, \qquad
\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r})$$ with a $2$-dimensional orthogonal tempered representation $\rho$ of $L_F$ and an integer $r>3$. Let $$\Pi_{\phi,\psi}({\mathrm{Mp}}_4) = \{ \pi_\eta \, | \, \eta \in \widehat{S}_\phi \}, \qquad
\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}) = \{ \sigma_\xi \, | \, \xi \in \widehat{\overline{S}}_{\theta(\phi)} \}$$ be the associated $A$-packets. Note that $\pi_\eta$ is irreducible unless $F = {\mathbb{R}}$, $\rho$ is irreducible, and $\eta$ is nontrivial. We denote by $$\theta_\psi = \theta_{W_2,V_{r+2}^+,\psi}$$ the theta lift from ${\mathrm{Mp}}_4 = {\mathrm{Mp}}(W_2)$ to ${\mathrm{SO}}_{2r+5} = {\mathrm{SO}}(V_{r+2}^+)$. For any semisimple genuine representation $\pi$ of ${\mathrm{Mp}}_4(F)$ of finite length, write $\pi = \bigoplus_i m_i \pi_i$ with some positive integers $m_i$ and some pairwise distinct irreducible genuine representations $\pi_i$ of ${\mathrm{Mp}}_4(F)$, and put $$\theta_\psi(\pi) = \bigoplus_i m_i \theta_\psi(\pi_i).$$ Then we need to show that $$\theta_\psi(\pi_\eta) = \sigma_{\xi}$$ for $\eta \in \widehat{S}_\phi$ and $\xi = \eta \circ \iota^{-1} \in \widehat{\overline{S}}_{\theta(\phi)}$ (under the assumption that $r=4$ when $F={\mathbb{R}}$), where $\iota : S_{\phi} \hookrightarrow S_{\theta(\phi)} \twoheadrightarrow \overline{S}_{\theta(\phi)}$ is the natural isomorphism.
We will consider the various cases in turn.
The reducible case I
--------------------
Suppose that $\rho = \chi \oplus \chi^{-1}$ for some unitary character $\chi$ of $F^\times$ such that $\chi^2 \ne 1$. We denote by $\pi_\phi$ and $\sigma_{\theta(\phi)}$ the unique representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ and $\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$, respectively. By definition, we have $$\pi_\phi = J_{B, \psi}(\chi |\cdot|^{\frac{1}{2}}, \chi^{-1} |\cdot|^{\frac{1}{2}}).$$ Also, by the inductive property of $A$-packets [@moeglin11 §4.2], [@mr-c §6], [@mr3 §5], we have $$\sigma_{\theta(\phi)} = I_{Q_2}(\chi \circ \det, 1),$$ where the right-hand side is irreducible. In fact, the right-hand side is a quotient of $$I_Q(\chi |\cdot|^{\frac{1}{2}}, \chi |\cdot|^{-\frac{1}{2}}, |\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}})
\cong
I_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, \chi |\cdot|^{\frac{1}{2}}, \chi^{-1} |\cdot|^{\frac{1}{2}}),$$ where $Q = Q_{(1^{r+2})}$, so that $\sigma_{\theta(\phi)}$ is the unique irreducible representation of ${\mathrm{SO}}_{2r+5}(F)$ with $L$-parameter $\varphi_{\theta(\phi)}$. Hence, by Lemma \[l:ind-princ-2\] below, we have $$\theta_\psi(\pi_\phi) = \sigma_{\theta(\phi)}.$$
The reducible case II
---------------------
Suppose that $\rho = \chi_a \oplus \chi_b$ for some $a, b \in F^\times$. For $\epsilon_1, \epsilon_2 \in \mu_2$, we denote by $\pi^{\epsilon_1,\epsilon_2}$ and $\sigma^{\epsilon_1,\epsilon_2}$ the representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ and $\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ associated to $(\epsilon_1,\epsilon_2)$, respectively, where we identify $\widehat{S}_\phi \cong \widehat{\overline{S}}_{\theta(\phi)}$ with a subgroup of $\mu_2 \times \mu_2$. By definition, we have $$\pi^{+,+} = \pi_\phi,$$ where $\pi_\phi$ is the unique irreducible genuine representation of ${\mathrm{Mp}}_4(F)$ with $L$-parameter $\varphi_\phi$ (relative to $\psi$). Also, by [@arthur Proposition 7.4.1], we have $$\label{eq:sigma++}
\sigma^{+,+} \supset \sigma_{\theta(\phi)},$$ where $\sigma_{\theta(\phi)}$ is the unique irreducible representation of ${\mathrm{SO}}_{2r+5}(F)$ with $L$-parameter $\varphi_{\theta(\phi)}$. Moreover, if $F$ is nonarchimedean and $\chi_a \ne \chi_b$, then the $A$-parameter $\theta(\phi)$ is elementary in the sense of M[œ]{}glin [@moeglin06] and we can describe $\sigma^{\epsilon_1,\epsilon_2}$ explicitly. For example, if further $\chi_a, \chi_b, 1$ are pairwise distinct, then we have $$\begin{aligned}
\sigma^{+,+} & = J_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, \chi_a |\cdot|^{\frac{1}{2}}, \chi_b |\cdot|^{\frac{1}{2}}), \\
\sigma^{+,-} & = J_{Q'}(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \chi_a |\cdot|^{\frac{1}{2}}, \sigma^-_{\varphi_b}), \\
\sigma^{-,+} & = J_{Q'}(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \chi_b |\cdot|^{\frac{1}{2}}, \sigma^-_{\varphi_a})\end{aligned}$$ by her explicit construction of $A$-packets (see also [@xu §6]). Here $Q = Q_{(1^{r+2})}$, $Q' = Q_{(1^r)}$, and for any $c \in F^\times$ such that $\chi_c \ne 1$, $\sigma^-_{\varphi_c}$ is the irreducible square-integrable representation of ${\mathrm{SO}}_5(F)$ with $L$-parameter $\varphi_c = (\chi_c \boxtimes S_2) \oplus (1 \boxtimes S_2)$ associated to the nontrivial character of $\overline{S}_{\varphi_c} \cong {\mathbb{Z}}/2{\mathbb{Z}}$.
\[l:mult-free\] The $A$-packet $\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ is multiplicity-free.
In a more general context, the assertion was completely proved by M[œ]{}glin [@moeglin11] when $F$ is nonarchimedean and by M[œ]{}glin–Renard [@mr-c] when $F = {\mathbb{C}}$, and was largely proved by M[œ]{}glin–Renard [@mr1; @mr2] when $F = {\mathbb{R}}$. Here we give a proof based on theta lifts.
Fix the following data:
- a number field ${\mathbb{F}}$ with adèle ring ${\mathbb{A}}$ and distinct places $v_0,v_1$ such that ${\mathbb{F}}_{v_0} = {\mathbb{F}}_{v_1} = F$;
- a nontrivial additive character $\varPsi$ of ${\mathbb{A}}/{\mathbb{F}}$ such that $\varPsi_{v_0}, \varPsi_{v_1}$ belong to the $(F^\times)^2$-orbit of $\psi$;
- $\alpha, \beta \in {\mathbb{F}}^\times$ such that $\chi_{\alpha,v_0} = \chi_{\alpha,v_1} = \chi_a$, $\chi_{\beta,v_0} = \chi_{\beta,v_1} = \chi_b$, and $\chi_\alpha \ne \chi_\beta$.
Put $$\varPhi = (\chi_\alpha \boxtimes S_2) \oplus (\chi_\beta \boxtimes S_2),$$ so that $\varPhi_{v_0} = \varPhi_{v_1} = \phi$. Then, by , there is an embedding $$\Big( \bigoplus_{\epsilon_1, \epsilon_2 \in \mu_2} ( \sigma^{\epsilon_1,\epsilon_2} \otimes \sigma^{\epsilon_1,\epsilon_2} ) \Big)
\otimes \Big( \bigotimes_{v \ne v_0, v_1} \sigma_{\theta(\varPhi_v)} \Big)
{\lhook\joinrel\longrightarrow}L^2_{\theta(\varPhi)}({\mathrm{SO}}_{2r+5}),$$ where $\theta(\varPhi) = \varPhi \oplus (1 \boxtimes S_{2r})$ and $\sigma_{\theta(\varPhi_v)}$ is the unique irreducible representation of ${\mathrm{SO}}_{2r+5}({\mathbb{F}}_v)$ with $L$-parameter $\varphi_{\theta(\varPhi_v)}$. Hence, for any irreducible representation $\sigma$ of ${\mathrm{SO}}_{2r+5}(F)$, we have $$\sum_{\epsilon_1, \epsilon_2 \in \mu_2} m^{\epsilon_1,\epsilon_2}(\sigma)^2 \le m_{\mathrm{disc}}(\varSigma),$$ where $$m^{\epsilon_1,\epsilon_2}(\sigma) = \dim_{\mathbb{C}}{\operatorname{Hom}}_{{\mathrm{SO}}_{2r+5}(F)}(\sigma,\sigma^{\epsilon_1,\epsilon_2})$$ and $\varSigma$ is the irreducible representation of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$ such that
- $\varSigma_{v_0} = \varSigma_{v_1} = \sigma$;
- $\varSigma_v = \sigma_{\theta(\varPhi_v)}$ for all $v \ne v_0,v_1$.
On the other hand, by Proposition \[p:howe-ps\] and the analog of Proposition \[p:mult-preservation\] for $A$-parameters of Howe–Piatetski-Shapiro type (see Remark \[r:howe-ps\]), we have $m_{\mathrm{disc}}(\varSigma) \le 1$, so that $$\sum_{\epsilon_1, \epsilon_2 \in \mu_2} m^{\epsilon_1,\epsilon_2}(\sigma)^2 \le 1.$$ This implies the assertion.
We denote by ${\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}))$ the set of irreducible representations of ${\mathrm{SO}}_{2r+5}(F)$ occurring in $\sigma^{\epsilon_1,\epsilon_2}$ for some $\epsilon_1,\epsilon_2 \in \mu_2$.
\[l:bij\] The theta lift induces a bijection $$\theta_\psi : \Pi_{\phi,\psi}({\mathrm{Mp}}_4) \longrightarrow {\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})).$$
Let ${\mathbb{F}}, v_0, v_1, \varPhi, \varPsi$ be as in the proof of Lemma \[l:mult-free\]. For a given $\pi \in \Pi_{\phi,\psi}({\mathrm{Mp}}_4)$, let $\varPi$ be the irreducible genuine representation of ${\mathrm{Mp}}_4({\mathbb{A}})$ such that
- $\varPi_{v_0} = \varPi_{v_1} = \pi$;
- $\varPi_v = \pi_{\varPhi_v}$ for all $v \ne v_0,v_1$,
where $\pi_{\varPhi_v}$ is the unique irreducible genuine representation of ${\mathrm{Mp}}_4({\mathbb{F}}_v)$ with $L$-parameter $\varphi_{\varPhi_v}$ (relative to $\varPsi_v$). Then $\varPi$ occurs in $L^2_{\varPhi,\varPsi}({\mathrm{Mp}}_4)$ by Proposition \[p:howe-ps\], so that $\theta_\varPsi(\varPi)$ occurs in $L^2_{\theta(\varPhi)}({\mathrm{SO}}_{2r+5})$ by Remark \[r:howe-ps\]. In particular, we have $$\theta_\psi(\pi) = \theta_\varPsi(\varPi)_{v_0} \in {\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})).$$ This defines a map $\theta_\psi:\Pi_{\phi,\psi}({\mathrm{Mp}}_4) \rightarrow {\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}))$, which is injective by the Howe duality.
Conversely, for a given $\sigma \in {\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}))$, let $\varSigma$ be the irreducible representation of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$ such that
- $\varSigma_{v_0} = \varSigma_{v_1} = \sigma$;
- $\varSigma_v = \sigma_{\theta(\varPhi_v)}$ for all $v \ne v_0,v_1$.
Then $\varSigma$ occurs in $L^2_{\theta(\varPhi)}({\mathrm{SO}}_{2r+5})$ by , so that $\theta_\varPsi(\varSigma)$ occurs in $L^2_{\varPhi,\varPsi}({\mathrm{Mp}}_4)$ by Remark \[r:howe-ps\]. In particular, we have $$\theta_\psi(\sigma) = \theta_\varPsi(\varSigma)_{v_0} \in \Pi_{\phi,\psi}({\mathrm{Mp}}_4).$$ This implies that the map $\theta_\psi: \Pi_{\phi,\psi}({\mathrm{Mp}}_4) \rightarrow {\operatorname{JH}}(\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}))$ is surjective.
\[l:irred\] For any $\epsilon_1, \epsilon_2 \in \mu_2$, $\sigma^{\epsilon_1,\epsilon_2}$ is either zero or irreducible.
Suppose on the contrary that $\sigma^{\epsilon_1,\epsilon_2}$ is nonzero and reducible. Then, by Lemma \[l:mult-free\], we have $\sigma^{\epsilon_1,\epsilon_2} \supset \sigma' \oplus \sigma''$ for some inequivalent irreducible representations $\sigma',\sigma''$ of ${\mathrm{SO}}_{2r+5}(F)$. Let ${\mathbb{F}}, v_0, v_1, \varPhi, \varPsi$ be as in the proof of Lemma \[l:mult-free\]. Let $\varSigma$ be the irreducible representation of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$ such that
- $\varSigma_{v_0} = \sigma'$;
- $\varSigma_{v_1} = \sigma''$;
- $\varSigma_v = \sigma_{\theta(\varPhi_v)}$ for all $v \ne v_0,v_1$.
Then $\varSigma$ occurs in $L^2_{\theta(\varPhi)}({\mathrm{SO}}_{2r+5})$ by , so that $\varPi = \theta_\varPsi(\varSigma)$ occurs in $L^2_{\varPhi,\varPsi}({\mathrm{Mp}}_4)$ by Remark \[r:howe-ps\]. On the other hand, by Lemma \[l:bij\], we have $$\varPi_{v_0} = \theta_\psi(\sigma') = \pi^{\epsilon_1',\epsilon_2'}, \qquad
\varPi_{v_1} = \theta_\psi(\sigma'') = \pi^{\epsilon_1'',\epsilon_2''}$$ for some $\epsilon_1', \epsilon_2', \epsilon_1'', \epsilon_2'' \in \mu_2$ such that $(\epsilon_1',\epsilon_2') \ne (\epsilon_1'',\epsilon_2'')$. Also, by Lemma \[l:ind-princ-2\] below, we have $$\varPi_v = \theta_{\varPsi_v}(\sigma_{\theta(\varPhi_v)}) = \pi_{\varPhi_v}$$ for all $v \ne v_0,v_1$. Hence it follows from Proposition \[p:howe-ps\] that $\varPi$ does not occur in $L^2_{\varPhi,\varPsi}({\mathrm{Mp}}_4)$. This is a contradiction and completes the proof.
By Lemmas \[l:bij\] and \[l:irred\], the theta lift induces a bijection $$\theta_\psi : \Pi_{\phi,\psi}({\mathrm{Mp}}_4) \longrightarrow \Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5}).$$ We need to show that $$\theta_\psi(\pi^{\epsilon_1,\epsilon_2}) = \sigma^{\epsilon_1,\epsilon_2}$$ for $\epsilon_1,\epsilon_2 \in \mu_2$.
By Lemma \[l:ind-princ-2\] below, we have $$\theta_\psi(\pi^{+,+}) = \sigma^{+,+},$$ noting that $\pi^{+,+} = \pi_\phi$ and $\sigma^{+,+} = \sigma_{\theta(\phi)}$ (where the latter follows from and Lemma \[l:irred\]). This implies that $\theta_\psi(\pi^{-,-}) = \sigma^{-,-}$ when $\chi_a = \chi_b$. Also, we will show in §\[sss:howe-ps-nonarch\] below that $$\label{eq:howe-ps-nonarch}
\theta_\psi(\pi^{+,-}) = \sigma^{+,-}, \qquad
\theta_\psi(\pi^{-,+}) = \sigma^{-,+}$$ when $F$ is nonarchimedean and $\chi_a,\chi_b,1$ are pairwise distinct. This implies that $\theta_\psi(\pi^{-,-}) = \sigma^{-,-}$ in this case.
Finally, suppose that $F$ is arbitrary and that $\chi_a \ne \chi_b$. Fix the following data:
- a number field ${\mathbb{F}}$ with adèle ring ${\mathbb{A}}$ and distinct places $v_0,v_1$ such that ${\mathbb{F}}_{v_0} = F$ and ${\mathbb{F}}_{v_1}$ is nonarchimedean;
- a nontrivial additive character $\varPsi$ of ${\mathbb{A}}/{\mathbb{F}}$ such that $\varPsi_{v_0}$ belongs to the $(F^\times)^2$-orbit of $\psi$;
- $\alpha, \beta \in {\mathbb{F}}^\times$ such that $\chi_{\alpha,v_0} = \chi_a$, $\chi_{\beta,v_0} = \chi_b$, and $\chi_{\alpha,v_1}, \chi_{\beta,v_1}, 1$ are pairwise distinct.
Put $$\varPhi = (\chi_\alpha \boxtimes S_2) \oplus (\chi_\beta \boxtimes S_2),$$ so that $\varPhi_{v_0} = \phi$. For given $\epsilon_1, \epsilon_2 \in \mu_2$, let $\varSigma$ be the irreducible representation of ${\mathrm{SO}}_{2r+5}({\mathbb{A}})$ such that
- $\varSigma_{v_0} = \sigma^{\epsilon_1,\epsilon_2}$;
- $\varSigma_{v_1} = \sigma_{v_1}^{\epsilon_1,\epsilon_2}$;
- $\varSigma_v = \sigma_{\theta(\varPhi_v)}$ for all $v \ne v_0,v_1$.
Then $\varSigma$ occurs in $L^2_{\theta(\varPhi)}({\mathrm{SO}}_{2r+5})$ by , so that $\varPi = \theta_\varPsi(\varSigma)$ occurs in $L^2_{\varPhi,\varPsi}({\mathrm{Mp}}_4)$ by Remark \[r:howe-ps\]. On the other hand, we have $$\varPi_{v_0} = \theta_\psi(\sigma^{\epsilon_1,\epsilon_2}) = \pi^{\epsilon_1',\epsilon_2'}$$ for some $\epsilon_1',\epsilon_2' \in \mu_2$. Since we have already shown that $$\varPi_{v_1} = \theta_{\varPsi_{v_1}}(\sigma_{v_1}^{\epsilon_1,\epsilon_2}) = \pi_{v_1}^{\epsilon_1,\epsilon_2}$$ and $$\varPi_v = \theta_{\varPsi_v}(\sigma_{\theta(\varPhi_v)}) = \pi_{\varPhi_v}$$ for all $v \ne v_0,v_1$, it follows from Proposition \[p:howe-ps\] that $(\epsilon_1',\epsilon_2') = (\epsilon_1,\epsilon_2)$.
This completes the proof of Proposition \[p:soudry-local\] when $\rho$ is reducible.
The irreducible case
--------------------
Suppose that $\rho$ is irreducible. Then either $F$ is nonarchimedean or $F = {\mathbb{R}}$. For $\epsilon \in \mu_2$, we denote by $\pi^\epsilon$ and $\sigma^\epsilon$ the representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ and $\Pi_{\theta(\phi)}({\mathrm{SO}}_{2r+5})$ associated to $\epsilon$, respectively, where we identify $\widehat{S}_\phi \cong \widehat{\overline{S}}_{\theta(\phi)}$ with $\mu_2$.
### The nonarchimedean case
Suppose that $F$ is nonarchimedean. By definition, we have $$\pi^+ = J_{P_2,\psi}(\tau \otimes |\det|^{\frac{1}{2}}), \qquad
\pi^- = \pi^-_\varphi,$$ where $\tau$ is the irreducible supercuspidal representation of ${\mathrm{GL}}_2(F)$ with $L$-parameter $\rho$ and $\pi^-_\varphi$ is the irreducible genuine square-integrable representation of ${\mathrm{Mp}}_4(F)$ with $L$-parameter $\varphi = \rho \boxtimes S_2$ (relative to $\psi$) associated to the nontrivial character of $S_\varphi \cong {\mathbb{Z}}/2{\mathbb{Z}}$. Moreover, the $A$-parameter $\theta(\phi)$ is elementary in the sense of M[œ]{}glin [@moeglin06] and we have $$\begin{aligned}
\sigma^+ & = J_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, \tau \otimes |\det|^{\frac{1}{2}}), \\
\sigma^- & = J_{Q'}(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \sigma^-_{\varphi'})\end{aligned}$$ by her explicit construction of $A$-packets (see also [@xu §6]). Here $Q = Q_{(1^r,2)}$, $Q' = Q_{(1^{r-1})}$, and $\sigma^-_{\varphi'}$ is the irreducible square-integrable representation of ${\mathrm{SO}}_7(F)$ with $L$-parameter $\varphi' = (\rho \boxtimes S_2) \oplus (1 \boxtimes S_2)$ associated to the nontrivial character of $\overline{S}_{\varphi'} \cong {\mathbb{Z}}/2{\mathbb{Z}}$. By Lemma \[l:ind-princ-2\] below, we have $$\theta_\psi(\pi^+) = \sigma^+.$$ Also, by [@atobe-gan Theorem 1.4], we have $$\theta_\psi(\pi^-) = \sigma^-.$$
### The real case
Suppose that $F = {\mathbb{R}}$. When $r=4$, we will show in §\[sss:soudry-real\] below that $$\label{eq:soudry-real}
\theta_\psi(\pi^\epsilon) = \sigma^\epsilon$$ for $\epsilon \in \mu_2$.
This completes the proof of Proposition \[p:soudry-local\] when $\rho$ is irreducible.
Local theta lifts from ${\mathrm{O}}_3$ to ${\mathrm{Mp}}_4$
============================================================
In this appendix, we prove some properties of the local theta lift used in §§\[s:sk\] and \[s:howe-ps\]. Let $F$ be a local field of characteristic zero.
Properties of the theta lift
----------------------------
We consider the theta lift from ${\mathrm{O}}(V_1^\epsilon)$ to ${\mathrm{Mp}}(W_2)$.
\[l:o3mp4-nonzero\] Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(V_1^\epsilon)$. Then we have $$\theta_{W_2,V_1^\epsilon,\psi}(\sigma) \ne 0
\, \Longleftrightarrow \, \sigma \ne \det,$$ where $\det$ is the nontrivial character of ${\mathrm{O}}(V_1^\epsilon)$ trivial on ${\mathrm{SO}}(V_1^\epsilon)$.
Put $$\begin{aligned}
n^+(\sigma) & = \min \{ n \ge 0 \, | \, \theta_{W_n,V_1^\epsilon,\psi}(\sigma) \ne 0 \}, \\
n^-(\sigma) & = \min \{ n \ge 0 \, | \, \theta_{W_n,V_1^\epsilon,\psi}(\sigma \otimes \det) \ne 0 \}.\end{aligned}$$ Since $n^-(\sigma) = 0$ if and only if $\sigma = \det$, the assertion follows from the conservation relation $$n^+(\sigma) + n^-(\sigma) = 3$$ proved by Sun–Zhu [@sz].
\[l:o3mp4-mult-free\] Let $\sigma_1$ and $\sigma_2$ be irreducible representations of ${\mathrm{O}}(V_1^{\epsilon_1})$ and ${\mathrm{O}}(V_1^{\epsilon_2})$, respectively. Assume that $\theta_{W_2,V_1^{\epsilon_1},\psi}(\sigma_1)$ and $\theta_{W_2,V_1^{\epsilon_2},\psi}(\sigma_2)$ are nonzero. Then we have $$\theta_{W_2,V_1^{\epsilon_1},\psi}(\sigma_1) = \theta_{W_2,V_1^{\epsilon_2},\psi}(\sigma_2)
\, \Longleftrightarrow \, \epsilon_1 = \epsilon_2, \, \sigma_1 = \sigma_2.$$
By the Howe duality, it suffices to show that if $\theta_{W_2,V_1^{\epsilon_1},\psi}(\sigma_1) = \theta_{W_2,V_1^{\epsilon_2},\psi}(\sigma_2)$, then $\epsilon_1 = \epsilon_2$. We may assume that $F \ne {\mathbb{C}}$. For any $\epsilon \in \mu_2$ and any irreducible genuine representation $\pi$ of ${\mathrm{Mp}}(W_2)$, put $$r^\epsilon(\pi) = \min \{ r \ge 0 \, | \, \theta_{W_2,V_r,\psi}(\pi) \ne 0, \, \epsilon(V_r) = \epsilon \},$$ where $\epsilon(V_r)$ denotes the Hasse–Witt invariant of a $(2r+1)$-dimensional quadratic space $V_r$ over $F$ with trivial discriminant. By the conservation relation $$r^+(\pi) + r^-(\pi) =5$$ proved by Sun–Zhu [@sz], we have $r^\epsilon(\pi) \ge 3$ for some $\epsilon \in \mu_2$. This implies the assertion.
\[l:ind-princ-1\] Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(V_1^\epsilon)$. Assume that either $\sigma$ is tempered, or $\epsilon = 1$ and $\sigma|_{{\mathrm{SO}}(V_1^\epsilon)} = I_{Q_1}(\chi |\cdot|^s)$ for some unitary character $\chi$ of $F^\times$ and some $s \in {\mathbb{R}}$ with $|s|<\frac{1}{2}$. If $$\label{eq:root}
\sigma(-1) = \epsilon \cdot \epsilon(\tfrac{1}{2}, \sigma),$$ then we have $$\theta_{W_2,V_1^\epsilon,\psi}(\sigma) = J_{P_1, \psi}(|\cdot|^{\frac{1}{2}}, \theta_{W_1,V_1^\epsilon,\psi}(\sigma)),$$ where $\epsilon(s, \sigma)$ is the standard $\epsilon$-factor of $\sigma$ relative to $\psi$ (whose value at $s = \frac{1}{2}$ does not depend on $\psi$).
Recall that $$\label{eq:dichotomy}
\theta_{W_1,V_1^\epsilon,\psi}(\sigma) \ne 0 \, \Longleftrightarrow \,
\sigma(-1) = \epsilon \cdot \epsilon(\tfrac{1}{2}, \sigma)$$ by the result of Waldspurger [@w1; @w2] (see also [@gs §5]). If we put $\pi = \theta_{W_1,V_1^\epsilon,\psi}(\sigma)$, then by assumption, either $\pi$ is tempered, or $\pi = I_{B,\psi}(\chi |\cdot|^s)$ for some unitary character $\chi$ of $F^\times$ and some $s \in {\mathbb{R}}$ with $|s| < \frac{1}{2}$. When $F$ is nonarchimedean, the assertion follows from [@gt1 Proposition 3.2] (which is stated for $\pi$ tempered but continues to hold for $\pi = I_{B,\psi}(\chi |\cdot|^s)$ as above). When $F = {\mathbb{R}}$, the assertion will be proved in §\[sss:ind-princ-1-real\] below. When $F = {\mathbb{C}}$, the assertion follows from [@ab1 Theorem 2.8].
Recall that ${\mathrm{SO}}(V_1^\epsilon) \cong (B^\epsilon)^\times/F^\times$ for some quaternion algebra $B^\epsilon$ over $F$. For any quadratic character $\chi$ of $F^\times$, we may regard $\chi \circ \mathrm{N}_{B^\epsilon}$ as a representation of ${\mathrm{SO}}(V_1^\epsilon)$, where $\mathrm{N}_{B^\epsilon}$ is the reduced norm on $B^\epsilon$.
\[l:howe-ps-local+\] Let $\sigma$ be a $1$-dimensional representation of ${\mathrm{O}}(V_1^+)$ such that $\sigma|_{{\mathrm{SO}}(V_1^+)} = \chi \circ \mathrm{N}_{B^+}$ for some quadratic character $\chi$ of $F^\times$.
1. \[item:howe-ps-1\] If $\sigma(-1) = \chi(-1)$, then we have $$\theta_{W_2, V_1^+, \psi}(\sigma) = J_{B, \psi}(\chi |\cdot|^{\frac{1}{2}}, |\cdot|^{\frac{1}{2}}).$$
2. \[item:howe-ps-2\] If $\sigma(-1) = -\chi(-1)$ and $\chi \ne 1$, then we have $$\theta_{W_2, V_1^+, \psi}(\sigma) = J_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi}^-).$$
When $F$ is nonarchimedean or $F = {\mathbb{R}}$, the assertion will be proved in §§\[sss:howe-ps-local+-nonarch\] and \[sss:howe-ps-local+-real\] below. When $F = {\mathbb{C}}$, the assertion follows from [@ab1 Theorem 2.8].
\[l:howe-ps-local-\] Let $\sigma$ be a $1$-dimensional representation of ${\mathrm{O}}(V_1^-)$ such that $\sigma|_{{\mathrm{SO}}(V_1^-)} = \chi \circ \mathrm{N}_{B^-}$ for some quadratic character $\chi = \chi_a$ of $F^\times$ with $a \in F^\times$. If $\sigma(-1) = -\chi(-1)$ and $\chi \ne 1$, then we have $$\theta_{W_2, V_1^-, \psi}(\sigma) = J_{P_1, \psi}(|\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_a}^-).$$
Since $\epsilon(\tfrac{1}{2}, \sigma) = \chi(-1)$ and $\theta_{W_1,V_1^-,\psi}(\sigma) = \omega_{W_1,\psi_a}^-$, the assertion follows from Lemma \[l:ind-princ-1\].
The nonarchimedean case
-----------------------
Suppose that $F$ is nonarchimedean. Let $$W_2 = X_1 \oplus X_1^* \oplus W_1, \qquad
V_1^+ = Y_1 \oplus Y_1^* \oplus V_0^+$$ be the decompositions as in §§\[ss:mp\] and \[ss:so\] with $$\begin{aligned}
X_1 & = {\operatorname{Span}}(w_1), &
Y_1 & = {\operatorname{Span}}(v_1), \\
X_1^* & = {\operatorname{Span}}(w_1^*), &
Y_1^* & = {\operatorname{Span}}(v_1^*). \end{aligned}$$ Recall that $P_1$ is the maximal parabolic subgroup of ${\mathrm{Sp}}(W_2)$ stabilizing $X_1$. Let $B_1$ be the Borel subgroup of ${\mathrm{O}}(V_1^+)$ stabilizing $Y_1$.
### Proof of Lemma \[l:howe-ps-local+\] {#sss:howe-ps-local+-nonarch}
Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(V_1^+)$. Assume that $\sigma \ne \det$, so that $\theta_{W_2,V_1^+,\psi}(\sigma)$ is nonzero by Lemma \[l:o3mp4-nonzero\]. We will compute $\theta_{W_2,V_1^+,\psi}(\sigma)$ under the further assumption that there is an embedding $$\sigma {\lhook\joinrel\longrightarrow}{\operatorname{Ind}}^{{\mathrm{O}}(V_1^+)}_{B_1}(\chi_s \boxtimes \sigma_0),$$ where $\chi_s = \chi |\cdot|^s$ for some unitary character $\chi$ of ${\mathrm{GL}}(Y_1) \cong F^\times$ and some $s \in {\mathbb{R}}$, and $\sigma_0$ is a character of ${\mathrm{O}}(V_0^+) = \{ \pm 1\}$. Note that if $\chi = 1$ and $s = -\frac{1}{2}$, then since $\sigma \ne \det$, we have $\sigma(-1) = 1$. Without loss of generality, we may assume that $s \ne \frac{3}{2}$. Put $\epsilon_0 = \sigma_0(-1) = \sigma(-1) \cdot \chi(-1)$.
We have $$\begin{aligned}
\Theta_{W_2,V_1^+,\psi}(\sigma)^*
& = {\operatorname{Hom}}_{{\mathrm{O}}(V_1^+)}(\omega_{W_2,V_1^+,\psi}, \sigma) \\
& \subset {\operatorname{Hom}}_{{\mathrm{O}}(V_1^+)}(\omega_{W_2,V_1^+,\psi}, {\operatorname{Ind}}^{{\mathrm{O}}(V_1^+)}_{B_1}(\chi_s \boxtimes \sigma_0)) \\
& = {\operatorname{Hom}}_{{\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+)}(R_{B_1}(\omega_{W_2,V_1^+,\psi}), \chi_s \boxtimes \sigma_0), \end{aligned}$$ where $*$ denotes the linear dual and $R_{B_1}$ denotes the normalized Jacquet functor with respect to $B_1$. By the result of Kudla [@kudla], $R_{B_1}(\omega_{W_2,V_1^+,\psi})$ has a filtration $$R_{B_1}(\omega_{W_2,V_1^+,\psi}) = R^0 \supset R^1 \supset \{ 0 \}$$ of ${\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+) \times {\mathrm{Mp}}(W_2)$-modules such that $$\begin{aligned}
R^0/R^1 & \cong |{\det}_{Y_1}|^{\frac{3}{2}} \boxtimes \omega_{W_2,V_0^+,\psi}, \\
R^1 & \cong {\operatorname{Ind}}^{{\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+) \times {\mathrm{Mp}}(W_2)}_{{\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+) \times \widetilde{P}_1}(\mathcal{S}(\operatorname{Isom}(X_1,Y_1)) \boxtimes \omega_{W_1,V_0^+,\psi}).\end{aligned}$$ Here ${\mathrm{GL}}(Y_1) \times \widetilde{{\mathrm{GL}}}(X_1)$ acts on $\mathcal{S}(\operatorname{Isom}(X_1, Y_1))$ by $$[(a, \tilde{b}) \cdot f](g) = \chi_\psi(\tilde{b}) \cdot f(a^{-1} \circ g \circ b)$$ for $a \in {\mathrm{GL}}(Y_1)$, $\tilde{b} \in \widetilde{{\mathrm{GL}}}(X_1)$ with projection $b \in {\mathrm{GL}}(X_1)$, and $f \in \mathcal{S}(\operatorname{Isom}(X_1, Y_1))$. Since $s \ne \frac{3}{2}$, the action of ${\mathrm{GL}}(Y_1)$ shows that $${\operatorname{Hom}}_{{\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+)}(R^0/R^1, \chi_s \boxtimes \sigma_0) = \{ 0 \},$$ so that $$\begin{aligned}
\Theta_{W_2,V_1^+,\psi}(\sigma)^*
& \subset {\operatorname{Hom}}_{{\mathrm{GL}}(Y_1) \times {\mathrm{O}}(V_0^+)}(R^1, \chi_s \boxtimes \sigma_0) \\
& = I_{P_1, \psi}(\chi_s^{-1}, \Theta_{W_1, V_0^+, \psi}(\sigma_0))^*.\end{aligned}$$ Thus, noting that $\Theta_{W_1, V_0^+, \psi}(\sigma_0) = \omega_{W_1, \psi}^{\epsilon_0}$, we obtain a surjection $$\label{eq:surj-o3mp4}
I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^{\epsilon_0})
{\relbar\joinrel\twoheadrightarrow}\Theta_{W_2,V_1^+,\psi}(\sigma).$$ In particular, when $\chi^2 = 1$ and $s = -\frac{1}{2}$, this proves Lemma \[l:howe-ps-local+\] in the nonarchimedean case.
### More properties of the theta lift
We prove more properties of the local theta lift, which are not used in the proof of the main theorem but will be necessary when we describe local $A$-packets explicitly in Appendix \[a:A-packets\] below.
\[l:more-theta-nonarch-1\] Let $\sigma$ be an irreducible supercuspidal representation of ${\mathrm{O}}(V_1^\epsilon)$. Assume that $\sigma \ne \det$ when $\epsilon = -1$. If $\sigma(-1) = - \epsilon \cdot \epsilon(\frac{1}{2}, \sigma)$, then we have $$\theta_{W_2, V_1^\epsilon, \psi}(\sigma) = \theta_{W_2, V_2^\epsilon, \psi}({\mathrm{St}}^\epsilon(1,\sigma_0)),$$ where ${\mathrm{St}}^\epsilon(1,\sigma_0)$ is the irreducible square-integrable representation of ${\mathrm{SO}}(V_2^\epsilon)$ contained in $I_{Q_1}(|\cdot|^{\frac{1}{2}}, \sigma_0)$ with $\sigma_0 = \sigma|_{{\mathrm{SO}}(V_1^\epsilon)}$ (see Lemmas \[l:nsc-so+q1\] and \[l:nsc-so-q1\] below).
Put $\pi = \theta_{W_2, V_1^{\epsilon}, \psi}(\sigma)$. By Lemma \[l:o3mp4-nonzero\] and , $\pi$ is nonzero and supercuspidal. Hence, by the local Shimura correspondence [@gs], $\theta_{W_2, V_2^\epsilon, \psi}(\pi)$ is nonzero and square-integrable.
We now consider the normalized Jacquet module of $\omega_{W_2, V_2^\epsilon, \psi}$ with respect to the maximal parabolic subgroup $Q_1$ of ${\mathrm{SO}}(V_2^\epsilon)$. By the result of Kudla [@kudla], there is an $F^\times \times {\mathrm{SO}}(V_1^\epsilon) \times {\mathrm{Mp}}(W_2)$-equivariant surjection $$R_{Q_1}(\omega_{W_2, V_2^\epsilon, \psi}) {\relbar\joinrel\twoheadrightarrow}|\cdot|^{\frac{1}{2}} \boxtimes \omega_{W_2, V_1^\epsilon, \psi}.$$ Since there is a surjection $\omega_{W_2, V_1^\epsilon, \psi} \twoheadrightarrow \sigma_0 \boxtimes \pi$, this gives rise via Frobenius reciprocity to a nonzero ${\mathrm{SO}}(V_2^\epsilon) \times {\mathrm{Mp}}(W_2)$-equivariant map $$\omega_{W_2, V_2^\epsilon, \psi} \longrightarrow I_{Q_1}(|\cdot|^{\frac{1}{2}}, \sigma_0) \boxtimes \pi.$$ Hence $\theta_{W_2, V_2^\epsilon, \psi}(\pi)$ is a subquotient of $I_{Q_1}(|\cdot|^{\frac{1}{2}}, \sigma_0)$. Since $\theta_{W_2, V_2^\epsilon, \psi}(\pi)$ is square-integrable, $\theta_{W_2, V_2^\epsilon, \psi}(\pi)$ must be ${\mathrm{St}}^\epsilon(1, \sigma_0)$. This completes the proof.
\[l:more-theta-nonarch-2\] Let $\sigma$ be an irreducible square-integrable representation of ${\mathrm{O}}(V_1^+)$ such that $\sigma|_{{\mathrm{SO}}(V_1^+)}$ is the unique irreducible subrepresentation of $I_{Q_1}(\chi |\cdot|^{\frac{1}{2}})$ for some quadratic character $\chi$ of $F^\times$.
1. If $\sigma(-1) = -\chi(-1)$ and $\chi \ne 1$, then we have $$\theta_{W_2, V_1^+, \psi}(\sigma) = \widetilde{{\mathrm{St}}}_\psi(\chi, \omega_{W_1, \psi}^-),$$ where $\widetilde{{\mathrm{St}}}_\psi(\chi, \omega_{W_1, \psi}^-)$ is the irreducible genuine square-integrable representation of ${\mathrm{Mp}}(W_2)$ contained in $I_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \omega_{W_1, \psi}^-)$ (see Lemma \[l:nsc-mp-p1\] below).
2. If $\sigma(-1) = 1$ and $\chi = 1$, then we have $$\theta_{W_2, V_1^+, \psi}(\sigma) = \pi_{{\mathrm{gen}}, \psi}({\mathrm{st}}),$$ where $\pi_{{\mathrm{gen}}, \psi}({\mathrm{st}})$ is the irreducible genuine tempered representation of ${\mathrm{Mp}}(W_2)$ contained in $I_{P_1, \psi}(|\cdot|^{\frac{1}{2}}, \omega_{W_1, \psi}^+)$ (see Lemma \[l:nsc-mp-p1\] below).
Note that $$\epsilon(\tfrac{1}{2}, \sigma) =
\begin{cases}
\chi(-1) & \text{if $\chi \ne 1$;} \\
-1 & \text{if $\chi = 1$,}
\end{cases}$$ so that $\theta_{W_1, V_1^+, \psi}(\sigma)$ is zero by . The assertion follows from .
\[l:more-theta-nonarch-3\] Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(V_1^+)$ such that $\sigma|_{{\mathrm{SO}}(V_1^+)} = I_{Q_1}(\chi |\cdot|^s)$ for some unitary character $\chi$ of $F^\times$ and some $s \in {\mathbb{R}}$ with $0 \le s <\frac{1}{2}$. If $\sigma(-1) = - \chi(-1)$, then we have $$\theta_{W_2, V_1^+, \psi}(\sigma) = J_{P_1,\psi}(\chi |\cdot|^s, \omega^-_{W_1,\psi}).$$
The assertion follows from .
The real case
-------------
Suppose that $F = {\mathbb{R}}$. We work in the category of $({\mathfrak{g}},K)$-modules.
### Notation {#sss:notation-real}
Let ${\mathrm{O}}(n)$ be the maximal compact subgroup of ${\mathrm{GL}}_n({\mathbb{R}})$ defined by $${\mathrm{O}}(n) = \{ g \in {\mathrm{GL}}_n({\mathbb{R}}) \, | \, {}^t g^{-1} = g \}.$$ Put $n_0 = [\frac{n}{2}]$. As in [@adams p. 18], we parametrize the irreducible representations of ${\mathrm{O}}(n)$ by highest weights $$(a_1,\dots,a_{n_0};\epsilon)$$ with $a_i \in {\mathbb{Z}}$ such that $a_1 \ge \dots \ge a_{n_0} \ge 0$ and $\epsilon = \pm 1$. Note that the two representations associated to $(a_1,\dots,a_{n_0};1)$ and $(a_1,\dots,a_{n_0};-1)$ are isomorphic if and only if $n$ is even and $a_{n_0}> 0$.
Let $V_{p,q}$ be the quadratic space over ${\mathbb{R}}$ of signature $(p,q)$ with $p+q$ odd. We realize the orthogonal group ${\mathrm{O}}(p,q) = {\mathrm{O}}(V_{p,q})$ as $${\mathrm{O}}(p,q) = \left\{ g \in {\mathrm{GL}}_{p+q}({\mathbb{R}}) \, \left| \,
{}^t g
\begin{pmatrix}
\mathbf{1}_p & \\
& -\mathbf{1}_q
\end{pmatrix}
g =
\begin{pmatrix}
\mathbf{1}_p & \\
& -\mathbf{1}_q
\end{pmatrix}
\right. \right\}.$$ Let $K \cong {\mathrm{O}}(p) \times {\mathrm{O}}(q)$ be the maximal compact subgroup of ${\mathrm{O}}(p,q)$ defined by $$K = \{ g \in {\mathrm{O}}(p,q) \, | \, {}^t g^{-1} = g \}.$$ Put $p_0 = [\frac{p}{2}]$ and $q_0 = [\frac{q}{2}]$. As above, we parametrize the irreducible representations of $K$ by highest weights $$(a_1,\dots,a_{p_0};\epsilon) \boxtimes (b_1,\dots,b_{q_0};\delta)$$ with $a_i, b_j \in {\mathbb{Z}}$ such that $a_1 \ge \dots \ge a_{p_0} \ge 0$, $b_1 \ge \dots \ge b_{q_0} \ge 0$ and $\epsilon, \delta = \pm 1$.
Let $W_n$ be the $2n$-dimensional symplectic space over ${\mathbb{R}}$. We realize the symplectic group ${\mathrm{Sp}}_{2n}({\mathbb{R}}) = {\mathrm{Sp}}(W_n)$ as $${\mathrm{Sp}}_{2n}({\mathbb{R}}) = \left\{ g \in {\mathrm{GL}}_{2n}({\mathbb{R}}) \, \left| \,
{}^t g
\begin{pmatrix}
& \mathbf{1}_n \\
-\mathbf{1}_n &
\end{pmatrix}
g =
\begin{pmatrix}
& \mathbf{1}_n \\
-\mathbf{1}_n &
\end{pmatrix}
\right. \right\}.$$ Let $\overline{K'} \cong {\mathrm{U}}(n)$ be the maximal compact subgroup of ${\mathrm{Sp}}_{2n}({\mathbb{R}})$ defined by $$\overline{K'} = \{ g \in {\mathrm{Sp}}_{2n}({\mathbb{R}}) \, | \, {}^t g^{-1} = g \}$$ and $K'$ the preimage of $\overline{K'}$ in ${\mathrm{Mp}}_{2n}({\mathbb{R}})$. Following the normalization in [@adams p. 19], we parametrize the irreducible genuine representations of $K'$ by highest weights $$(a_1, \dots, a_n)$$ with $a_i \in {\mathbb{Z}}+ \frac{1}{2}$ such that $a_1 \ge \dots \ge a_n$. Note that this parametrization depends on $\psi$.
For any $a \in \frac{1}{2} {\mathbb{Z}}$ with $a>0$, we denote by $D_a$ the relative discrete series representation of ${\mathrm{GL}}_2({\mathbb{R}})$ of weight $2a + 1$ with central character trivial on ${\mathbb{R}}_+^\times$, so that $\mathcal{D}_a$ is the $L$-parameter of $D_a$. For any $a \in \frac{1}{2} {\mathbb{Z}}\smallsetminus {\mathbb{Z}}$, we denote by $\widetilde{D}_{a,\psi}$ the genuine discrete series representation of ${\mathrm{Mp}}_2({\mathbb{R}})$ of weight (relative to the parametrization depending on $\psi$) $$\begin{cases}
a+1 & \text{if $a>0$;} \\
a-1 & \text{if $a<0$.}
\end{cases}$$
### Theta correspondence over ${\mathbb{R}}$
We recall some basic properties of the theta correspondence over ${\mathbb{R}}$ (see [@howe; @kashiwara-vergne; @adams]).
Let $\mathcal{P}$ be the Fock model of the Weil representation $\omega_{W_n, V_{p,q}, \psi}$. Namely, $\mathcal{P} = \bigoplus_{d=0}^\infty \mathcal{P}_d$ is the space of polynomials in $n(p+q)$ variables and $\mathcal{P}_d$ is the subspace of homogeneous polynomials of degree $d$, which is invariant under the action of $K \times K'$. For any irreducible representation $\mu$ of $K$ occurring in $\mathcal{P}$, we define $\deg \mu$ as the smallest nonnegative integer $d$ such that the $\mu$-isotypic component of $\mathcal{P}_d$ is nonzero. If $$\mu = (a_1,\dots,a_k,0,\dots,0;\epsilon) \boxtimes (b_1,\dots,b_l,0,\dots,0;\delta)$$ with $a_k, b_l > 0$, then we have $$\deg \mu = \sum_{i=1}^k a_i + \sum_{j=1}^l b_j + k' + l',$$ where $$\label{eq:k'l'}
k' =
\begin{cases}
0 & \text{if $\epsilon = 1$;} \\
p-2k & \text{if $\epsilon = -1$,}
\end{cases}
\qquad
l' =
\begin{cases}
0 & \text{if $\delta = 1$;} \\
q-2l & \text{if $\delta = -1$.}
\end{cases}$$ In particular, $\deg \mu$ does not depend on $n$. Similarly, for any irreducible genuine representation $\mu'$ of $K'$ occurring in $\mathcal{P}$, we may define $\deg \mu'$. If $$\mu' = (a_1,\dots,a_n) + \frac{p-q}{2} \cdot (1,\dots,1)$$ with $a_i \in {\mathbb{Z}}$, then we have $$\deg \mu' = \sum_{i=1}^n |a_i|.$$ In particular, $\deg \mu'$ depends only on $p-q$.
Let $\mathcal{H}$ be the space of joint harmonics, which is a $K \times K'$-invariant subspace of $\mathcal{P}$. Then $\mathcal{H}$ is multiplicity-free as a representation of $K \times K'$ and induces a correspondence between representations of $K$ and $K'$ given as follows. Let $\mu$ be an irreducible representation of $K$ and $\mu'$ an irreducible genuine representation of $K'$. Then $\mu$ and $\mu'$ correspond if and only if $\mu$ and $\mu'$ are of the form $$\mu = (a_1,\dots,a_k,0,\dots,0;\epsilon) \boxtimes (b_1,\dots,b_l,0,\dots,0;\delta)$$ and $$\mu' = (a_1,\dots,a_k, \underbrace{1, \dots, 1}_{k'}, 0,\dots,0, \underbrace{-1,\dots,-1}_{l'}, -b_l,\dots, -b_1) + \frac{p-q}{2} \cdot (1,\dots,1)$$ with $a_k, b_l > 0$ and $k+k'+l+l' \le n$, where $k',l'$ are as in . Note that $\mu$ and $\mu'$ determine each other.
Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(p,q)$ such that the theta lift $\pi = \theta_{W_n,V_{p,q}, \psi}(\sigma)$ to ${\mathrm{Mp}}_{2n}({\mathbb{R}})$ is nonzero. Let $\mu$ be a $K$-type of $\sigma$, i.e. an irreducible representation of $K$ occurring in $\sigma|_K$. We say that $\mu$ is of *minimal degree* in $\sigma$ if $\deg \mu$ is minimal among all $K$-types of $\sigma$. In this case, $\mu$ occurs in $\mathcal{H}$. Let $\mu'$ be the irreducible genuine representation of $K'$ corresponding to $\mu$. Then $\mu'$ is a $K'$-type of $\pi$ and is of minimal degree in $\pi$. An analogous result also holds when we switch the roles of $\sigma$ and $\pi$.
We also need the notion of *lowest $K$-types* introduced by Vogan [@vogan]. In particular, we will use the following properties:
- any irreducible tempered representation with real infinitesimal character is uniquely determined by its unique lowest $K$-type (see also §\[sss:temp-mp4-real\] below);
- any lowest $K$-type of a standard module occurs with multiplicity one;
- the set of lowest $K$-types of a standard module agrees with that of its unique irreducible quotient.
These two notions of $K$-types are closely related as follows. Assume for simplicity that $p+q=2n+1$. Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(p,q)$ such that the theta lift $\pi = \theta_{W_n,V_{p,q}, \psi}(\sigma)$ to ${\mathrm{Mp}}_{2n}({\mathbb{R}})$ is nonzero. Then, by [@ab2 Corollary 5.2], we have:
- if $\mu$ is a lowest $K$-type of $\sigma$, then $\mu$ is of minimal degree in $\sigma$;
- if $\mu'$ is a lowest $K'$-type of $\pi$, then $\mu'$ is of minimal degree in $\pi$.
### Proof of Lemma \[l:ind-princ-1\] {#sss:ind-princ-1-real}
Suppose first that $\sigma$ is a principal series representation of ${\mathrm{O}}(2,1)$. More generally, let $\sigma$ be an irreducible representation of ${\mathrm{O}}(2,1)$ such that there is a surjection $${\operatorname{Ind}}^{{\mathrm{O}}(2,1)}_{B_1}(\chi_s \boxtimes \sigma_0) {\relbar\joinrel\twoheadrightarrow}\sigma,$$ where $B_1$ is the Borel subgroup of ${\mathrm{O}}(2,1)$, $\chi_s = \chi |\cdot|^s$ for some unitary character $\chi$ of ${\mathbb{R}}^\times$ and some $s \in {\mathbb{R}}$ with $s \ge 0$, and $\sigma_0$ is a character of ${\mathrm{O}}(1)$. Put $$\delta_0 = \chi(-1), \qquad
\epsilon_0 = \sigma_0(-1).$$ Then is equivalent to $$\epsilon_0 = 1,$$ in which case we have $\theta_{W_1,V_{2,1},\psi}(\sigma) = J_{B,\psi}(\chi_s)$. Let $\mu_0$ be a lowest $(K \cap T_1)$-type of $\chi_s \boxtimes \sigma_0$, where $T_1$ is the Levi component of $B_1$. Let $\mu$ be a lowest $K$-type of $\sigma$, so that $\mu$ occurs in ${\operatorname{Ind}}^{{\mathrm{O}}(2,1)}_{B_1}(\chi_s \boxtimes \sigma_0)$ with multiplicity one. As we will explicate below, we assume the following conditions:
- $\mu_0$ is of minimal degree in $\chi_s \boxtimes \sigma_0$;
- $\mu$ is of minimal degree in ${\operatorname{Ind}}^{{\mathrm{O}}(2,1)}_{B_1}(\chi_s \boxtimes \sigma_0)$;
- $\deg \mu = \deg \mu_0$;
- the restriction of $\mu$ to $K \cap T_1$ contains $\mu_0$.
Then, by the induction principle of Adams–Barbasch (see [@ab1 Proposition 3.25], [@ab2 Theorem 8.7]), $\theta_{W_2,V_{2,1}, \psi}(\sigma)$ is a subquotient of $$I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^{\epsilon_0})$$ containing $\mu'$, where $\mu'$ is the $K'$-type corresponding to $\mu$.
Assume that $(\delta_0, \epsilon_0) \ne (1,-1)$. We may take $\mu_0$ and $\mu$ given by $$\mu_0 = (\enspace; \delta_0) \boxtimes (\enspace; \epsilon_0), \qquad
\mu = (0; \epsilon_0) \boxtimes (\enspace; \delta_0 \epsilon_0),$$ which satisfy the conditions above and $$\deg \mu = \deg \mu_0 =
\begin{cases}
0 & \text{if $(\delta_0, \epsilon_0) = (1,1)$;} \\
1 & \text{if $(\delta_0, \epsilon_0) = (-1,1)$;} \\
2 & \text{if $(\delta_0, \epsilon_0) = (-1,-1)$.}
\end{cases}$$ Then we have $$\mu' =
\begin{cases}
(\frac{1}{2}, \frac{1}{2}) & \text{if $(\delta_0, \epsilon_0) = (1,1)$;} \\
(\frac{1}{2}, -\frac{1}{2}) & \text{if $(\delta_0, \epsilon_0) = (-1,1)$;} \\
(\frac{3}{2}, \frac{3}{2}) & \text{if $(\delta_0, \epsilon_0) = (-1,-1)$.}
\end{cases}$$ If $\epsilon_0 = 1$, then it follows from [@ab2 Proposition 6.10] that $\mu'$ is a lowest $K'$-type of the principal series representation $$I_{B,\psi}(\chi_s^{-1}, |\cdot|^{\frac{1}{2}})$$ of ${\mathrm{Mp}}_4({\mathbb{R}})$, which has $I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^+)$ as a quotient. Hence, if $\epsilon_0 = 1$ and $0 \le s \le \frac{1}{2}$, then we have $$\label{eq:ind-princ-1-real}
\theta_{W_2,V_{2,1},\psi}(\sigma) = J_{B,\psi}(|\cdot|^{\frac{1}{2}}, \chi_s).$$
Suppose next that $\sigma$ is a discrete series representation of ${\mathrm{O}}(p,q)$ with $(p,q) = (2,1)$ or $(0,3)$. We write the $L$-parameter of $\sigma|_{{\mathrm{SO}}(p,q)}$ as $\mathcal{D}_{\kappa-\frac{1}{2}}$ with some positive integer $\kappa$. Then, by , we have $$\sigma(-1) =
\begin{cases}
(-1)^\kappa & \text{if $(p,q) = (2,1)$;} \\
(-1)^{\kappa-1} & \text{if $(p,q) = (0,3)$.}
\end{cases}$$ Let $\mu$ be the lowest $K$-type of $\sigma$ given by $$\mu =
\begin{cases}
(\kappa;1) \boxtimes (\enspace;1) & \text{if $(p,q) = (2,1)$;} \\
(\kappa-1;1) & \text{if $(p,q) = (0,3)$.}
\end{cases}$$ Put $\pi_0 = \theta_{W_1, V_{p,q}, \psi}(\sigma)$, so that $\pi_0 = \widetilde{D}_{\lambda,\psi}$ with $$\lambda =
\begin{cases}
\kappa-\frac{1}{2} & \text{if $(p,q) = (2,1)$;} \\
-\kappa+\frac{1}{2} & \text{if $(p,q) = (0,3)$.}
\end{cases}$$ Since $\mu$ is of minimal degree in $\sigma$, it follows from the induction principle [@ab2 Theorem 8.4] that $\theta_{W_2,V_{p,q},\psi}(\sigma)$ is a subquotient of $$I_{P_1, \psi}(|\cdot|^{-\frac{1}{2}}, \pi_0)$$ containing $\mu'$, where $\mu'$ is the $K'$-type corresponding to $\mu$ and is given by $$\mu' =
\begin{cases}
(\kappa+\frac{1}{2}, \frac{1}{2}) & \text{if $(p,q) = (2,1)$;} \\
(-\frac{3}{2}, -\kappa-\frac{1}{2}) & \text{if $(p,q) = (0,3)$.}
\end{cases}$$ By [@ab2 Proposition 6.10], $\mu'$ is a lowest $K'$-type of $I_{P_1, \psi}(|\cdot|^{-\frac{1}{2}}, \pi_0)$, so that $$\theta_{W_2,V_{p,q},\psi}(\sigma) = J_{P_1, \psi}(|\cdot|^{\frac{1}{2}}, \pi_0).$$ This completes the proof of Lemma \[l:ind-princ-1\] in the real case.
### Proof of Lemma \[l:howe-ps-local+\] {#sss:howe-ps-local+-real}
We retain the notation of §\[sss:ind-princ-1-real\], so that $\sigma$ is a quotient of ${\operatorname{Ind}}^{{\mathrm{O}}(2,1)}_{B_1}(\chi_s \boxtimes \sigma_0)$. Then follows from . To prove , we may assume that $(\delta_0, \epsilon_0) = (-1, -1)$. Since $\omega_{W_1,\psi}^- = \widetilde{D}_{\frac{1}{2},\psi}$, it follows from [@ab2 Proposition 6.10] that $\mu'$ is a lowest $K'$-type of $$I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^-).$$ Hence, if $(\delta_0, \epsilon_0) = (-1, -1)$ and $s \ge 0$, then we have $$\theta_{W_2, V_{2,1}, \psi}(\sigma) = J_{P_1, \psi}(\chi_s, \omega_{W_1,\psi}^-).$$ This completes the proof of Lemma \[l:howe-ps-local+\] in the real case.
### More properties of the theta lift
We prove more properties of the local theta lift, which are not used in the proof of the main theorem but will be necessary when we describe local $A$-packets explicitly in Appendix \[a:A-packets\] below.
\[l:more-theta-real-1\] Let $\sigma$ be a discrete series representation of ${\mathrm{O}}(p,q)$ with $(p,q) = (2,1)$ or $(0,3)$ and with lowest $K$-type $\mu$ given by $$\mu =
\begin{cases}
(\kappa; 1) \boxtimes (\enspace;-1) & \text{if $(p,q) = (2,1)$;} \\
(\kappa-1; -1) & \text{if $(p,q) = (0,3)$}
\end{cases}$$ for some positive integer $\kappa$. Assume that $\sigma \ne \det$ (i.e. $\kappa > 1$) when $(p,q) = (0,3)$.
1. If $(p,q) = (2,1)$, then $\theta_{W_2, V_{2,1}, \psi}(\sigma)$ is the genuine (limit of) discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $K'$-type $(\kappa+\frac{1}{2}, -\frac{1}{2})$ (relative to $\psi$).
2. If $(p,q) = (0,3)$, then $\theta_{W_2, V_{0,3}, \psi}(\sigma)$ is the genuine discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $K'$-type $(-\frac{5}{2}, -\kappa-\frac{1}{2})$ (relative to $\psi$).
Note that $$\sigma(-1) =
\begin{cases}
(-1)^{\kappa-1} & \text{if $(p,q) = (2,1)$;} \\
(-1)^\kappa & \text{if $(p,q) = (0,3)$}
\end{cases}$$ and $\epsilon(\frac{1}{2}, \sigma) = (-1)^\kappa$, so that $\theta_{W_1, V_{p,q}, \psi}(\sigma)$ is zero by . Put $\pi = \theta_{W_2, V_{p,q}, \psi}(\sigma)$, which is nonzero by Lemma \[l:o3mp4-nonzero\]. Let $\mu'$ be the irreducible genuine representation of $K'$ corresponding to $\mu$: $$\mu' =
\begin{cases}
(\kappa+\frac{1}{2}, -\frac{1}{2}) & \text{if $(p,q) = (2,1)$;} \\
(-\frac{5}{2}, -\kappa-\frac{1}{2}) & \text{if $(p,q) = (0,3)$.}
\end{cases}$$ Since $\mu$ is the unique $K$-type of minimal degree in $\sigma$, $\mu'$ is the unique $K'$-type of minimal degree in $\pi$. Hence we deduce that $\mu'$ is the unique lowest $K'$-type of $\pi$, noting that $\theta_{W_2, V_{3,2}, \psi}(\pi)$ is nonzero.
Since $\pi$ has real infinitesimal character by [@przebinda], it remains to show that $\pi$ is tempered (see also §\[sss:temp-mp4-real\] below). For any irreducible genuine nontempered representation $\pi'$ of ${\mathrm{Mp}}_4({\mathbb{R}})$, it follows from [@ab2 Proposition 6.10] that $\pi'$ has lowest $K'$-types given as follows.
- Suppose that $\pi' = J_{P_1, \psi}(\chi |\cdot|^s, \widetilde{D}_{a,\psi})$ for some unitary character $\chi$ of ${\mathbb{R}}^\times$, some $s \in {\mathbb{R}}$ with $s>0$, and some $a \in \frac{1}{2} {\mathbb{Z}}\smallsetminus {\mathbb{Z}}$. Then $\pi'$ has a unique lowest $K'$-type $$\begin{cases}
(a+1,\frac{1}{2}) & \text{if $a>0$ and $\chi(-1) = 1$;} \\
(a+1,\frac{3}{2}) & \text{if $a>0$ and $\chi(-1) = -1$;} \\
(-\frac{3}{2},a-1) & \text{if $a<0$ and $\chi(-1) = 1$;} \\
(-\frac{1}{2},a-1) & \text{if $a<0$ and $\chi(-1) = -1$.}
\end{cases}$$
- Suppose that $\pi' = J_{P_2, \psi}(D_a \otimes |\det|^s)$ for some $a \in \frac{1}{2} {\mathbb{Z}}$ with $a>0$ and some $s \in {\mathbb{C}}$ with ${\operatorname{Re}}s > 0$. Then $\pi'$ has a unique lowest $K'$-type $$(a+\tfrac{1}{2}, -a-\tfrac{1}{2})$$ if $a \in {\mathbb{Z}}$ and two lowest $K'$-types $$(a+1,-a), \qquad
(a,-a-1)$$ if $a \notin {\mathbb{Z}}$.
- Suppose that $\pi' = J_{B, \psi}(\chi_1 |\cdot|^{s_1}, \chi_2 |\cdot|^{s_2})$ for some unitary characters $\chi_1, \chi_2$ of ${\mathbb{R}}^\times$ and some $s_1, s_2 \in {\mathbb{R}}$ with $s_1 > 0$ and $s_1 \ge s_2 \ge 0$. Put $\epsilon_1 = \chi_1(-1)$ and $\epsilon_2 = \chi_2(-1)$. Then $\pi'$ has a unique lowest $K'$-type $$\begin{cases}
(\frac{1}{2}, \frac{1}{2}) & \text{if $(\epsilon_1, \epsilon_2) = (1,1)$;} \\
(\frac{1}{2}, -\frac{1}{2}) & \text{if $(\epsilon_1, \epsilon_2) = (1,-1), (-1,1)$;} \\
(-\frac{1}{2}, -\frac{1}{2}) & \text{if $(\epsilon_1, \epsilon_2) = (-1,-1)$.}
\end{cases}$$
In particular, the set of lowest $K'$-types of $\pi'$ does not agree with that of $\pi$, which is a singleton $\{ \mu' \}$. Hence $\pi$ is tempered. This completes the proof.
\[l:more-theta-real-2\] Let $\sigma$ be an irreducible representation of ${\mathrm{O}}(2,1)$ such that $\sigma|_{{\mathrm{SO}}(2,1)} = I_{Q_1}(\chi |\cdot|^s)$ for some unitary character $\chi$ of ${\mathbb{R}}^\times$ and some $s \in {\mathbb{R}}$ with $0 \le s <\frac{1}{2}$. If $\sigma(-1) = -\chi(-1)$, then we have $$\theta_{W_2, V_{2,1}, \psi}(\sigma) = J_{P_1,\psi}(\chi |\cdot|^s, \omega^-_{W_1,\psi}).$$
By assumption, we have $\sigma = {\operatorname{Ind}}^{{\mathrm{O}}(2,1)}_{B_1}(\chi_s \boxtimes \sigma_0)$, where $B_1$ is the Borel subgroup of ${\mathrm{O}}(2,1)$, $\chi_s = \chi |\cdot|^s$, and $\sigma_0$ is the nontrivial character of ${\mathrm{O}}(1)$. Since $\theta_{W_1, V_{1,0}, \psi}(\sigma_0) = \omega_{W_1,\psi}^-$, it follows from the induction principle [@ab2 Theorem 8.4] that $\theta_{W_2, V_{2,1}, \psi}(\sigma)$ is a subquotient of $$I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^-).$$ On the other hand, by [@gi-real Proposition 2.3], $I_{P_1, \psi}(\chi_s^{-1}, \omega_{W_1, \psi}^-)$ is irreducible. This implies the assertion.
Local theta lifts from ${\mathrm{Mp}}_4$ to ${\mathrm{SO}}_{2r+5}$ with $r>3$ {#a:mp4so2r+5}
=============================================================================
In this appendix, we prove some properties of the local theta lift used in §§\[s:soudry\] and \[s:pf-soudry\]. Let $F$ be a local field of characteristic zero.
Properties of the theta lift
----------------------------
We consider the theta lift from ${\mathrm{Mp}}(W_2)$ to ${\mathrm{SO}}(V_{r+2}^+)$ with $r>3$.
\[l:ind-princ-2\] Let $\rho$ be a $2$-dimensional orthogonal tempered representation of $L_F$. Put $\phi = \rho \boxtimes S_2$ and $\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r})$. Let $\pi_\phi$ be the unique irreducible genuine representation of ${\mathrm{Mp}}(W_2)$ with $L$-parameter $\varphi_\phi$ (relative to $\psi$) and $\sigma_{\theta(\phi)}$ the unique irreducible representation of ${\mathrm{SO}}(V_{r+2}^+)$ with $L$-parameter $\varphi_{\theta(\phi)}$. Then we have $$\theta_{W_2,V_{r+2}^+,\psi}(\pi_\phi) = \sigma_{\theta(\phi)}.$$
When $F$ is nonarchimedean or $F = {\mathbb{R}}$, the assertion will be proved in §§\[sss:ind-princ-2-nonarch\] and \[sss:ind-princ-2-real\] below. When $F = {\mathbb{C}}$, the assertion follows from [@ab1 Theorem 2.8].
We also need the properties when $F$ is nonarchimedean and when $F = {\mathbb{R}}$, which will be recalled and proved below.
The nonarchimedean case
-----------------------
Suppose that $F$ is nonarchimedean. Let $$W_2 = X_1 \oplus X_1^* \oplus W_1 = X_2 \oplus X_2^*, \qquad
V_{r+2}^+ = Y_1 \oplus Y_1^* \oplus V_{r+1}^+ = Y_2 \oplus Y_2^* \oplus V_r^+$$ be the decompositions as in §§\[ss:mp\] and \[ss:so\] with $$\begin{aligned}
X_1 & = {\operatorname{Span}}(w_1), &
Y_1 & = {\operatorname{Span}}(v_1), \\
X_1^* & = {\operatorname{Span}}(w_1^*), &
Y_1^* & = {\operatorname{Span}}(v_1^*), \\
X_2 & = {\operatorname{Span}}(w_1, w_2), &
Y_2 & = {\operatorname{Span}}(v_1, v_2), \\
X_2^* & = {\operatorname{Span}}(w_1^*, w_2^*), &
Y_2^* & = {\operatorname{Span}}(v_1^*, v_2^*). \end{aligned}$$ We write $$X_2 = X_1 \oplus X_1'$$ with $X_1' = {\operatorname{Span}}(w_2)$. Recall that $P_i$ and $Q_i$ are the maximal parabolic subgroups of ${\mathrm{Sp}}(W_2)$ and ${\mathrm{SO}}(V_{r+2}^+)$ stabilizing $X_i$ and $Y_i$, respectively. Let $\mathcal{B}$ be the Borel subgroup of ${\mathrm{GL}}(X_2)$ stabilizing $X_1$.
### Proof of Lemma \[l:ind-princ-2\] {#sss:ind-princ-2-nonarch}
Let $\tau$ be the irreducible self-dual tempered representation of ${\mathrm{GL}}(X_2) \cong {\mathrm{GL}}_2(F)$ with $L$-parameter $\rho$. Then we have either
- $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}(X_2)}_{\mathcal{B}}(\chi \boxtimes \chi^{-1})$ for some unitary character $\chi$ of $F^\times$ such that $\chi^2 \ne 1$; or
- $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}(X_2)}_{\mathcal{B}}(\chi_a \boxtimes \chi_b)$ for some $a, b \in F^\times$; or
- $\tau$ is supercuspidal.
Let $\pi$ be an irreducible genuine representation of ${\mathrm{Mp}}(W_2)$. We will compute $\theta_{W_2,V_{r+2}^+,\psi}(\pi)$ under the assumption that there is an embedding $$\pi {\lhook\joinrel\longrightarrow}I_{P_2, \psi}(\tau_s),$$ where $\tau_s = \tau \otimes |\det|^s$ for some $s \in {\mathbb{R}}$ with $s \le 0$.
We have $$\begin{aligned}
\Theta_{W_2,V_{r+2}^+,\psi}(\pi)^*
& = {\operatorname{Hom}}_{{\mathrm{Mp}}(W_2)}(\omega_{W_2,V_{r+2}^+,\psi}, \pi) \\
& \subset {\operatorname{Hom}}_{{\mathrm{Mp}}(W_2)}(\omega_{W_2,V_{r+2}^+,\psi}, I_{P_2,\psi}(\tau_s)) \\
& = {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_2)}(R_{P_2}(\omega_{W_2,V_{r+2}^+,\psi}), \tau_s \otimes \chi_\psi),\end{aligned}$$ where $*$ denotes the linear dual and $R_{P_2}$ denotes the normalized Jacquet functor with respect to $P_2$. By the result of Kudla [@kudla], $R_{P_2}(\omega_{W_2,V_{r+2}^+,\psi})$ has a filtration $$R_{P_2}(\omega_{W_2,V_{r+2}^+,\psi}) = R^0 \supset R^1 \supset R^2 \supset \{ 0 \}$$ of $\widetilde{{\mathrm{GL}}}(X_2) \times {\mathrm{SO}}(V_{r+2}^+)$-modules such that $$\begin{aligned}
R^0/R^1 & \cong \chi_\psi |{\det}_{X_2}|^{r+1}, \\
R^1/R^2 & \cong {\operatorname{Ind}}^{\widetilde{{\mathrm{GL}}}(X_2) \times {\mathrm{SO}}(V_{r+2}^+)}_{\widetilde{\mathcal{B}} \times Q_1}(\chi_\psi |{\det}_{X_1}|^{r+\frac{1}{2}} \boxtimes \mathcal{S}(\operatorname{Isom}(Y_1,X_1'))), \\
R^2 & \cong {\operatorname{Ind}}^{\widetilde{{\mathrm{GL}}}(X_2) \times {\mathrm{SO}}(V_{r+2}^+)}_{\widetilde{{\mathrm{GL}}}(X_2) \times Q_2}(\mathcal{S}(\operatorname{Isom}(Y_2,X_2))).\end{aligned}$$ Here $\widetilde{{\mathrm{GL}}}(X_1') \times {\mathrm{GL}}(Y_1)$ acts on $\mathcal{S}(\operatorname{Isom}(Y_1, X_1'))$ by $$[(\tilde{a}, b) \cdot f](g) = \chi_\psi(\tilde{a}) \cdot f(a^{-1} \circ g \circ b)$$ for $\tilde{a} \in \widetilde{{\mathrm{GL}}}(X_1')$ with projection $a \in {\mathrm{GL}}(X_1')$, $b \in {\mathrm{GL}}(Y_1)$, and $f \in \mathcal{S}(\operatorname{Isom}(Y_1, X_1'))$, and $\widetilde{{\mathrm{GL}}}(X_2) \times {\mathrm{GL}}(Y_2)$ acts on $\mathcal{S}(\operatorname{Isom}(Y_2, X_2))$ similarly. Since $s \le 0$, the actions of $\widetilde{{\mathrm{GL}}}(X_2)$ and $\widetilde{{\mathrm{GL}}}(X_1)$ show that $${\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_2)}(R^0/R^1, \tau_s \otimes \chi_\psi) = \{ 0 \}$$ and $$\begin{aligned}
& {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_2)}(R^1/R^2, \tau_s \otimes \chi_\psi) \\
& = {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_1) \times \widetilde{{\mathrm{GL}}}(X_1')}({\operatorname{Ind}}^{\widetilde{{\mathrm{GL}}}(X_1) \times \widetilde{{\mathrm{GL}}}(X_1') \times {\mathrm{SO}}(V_{r+2}^+)}_{\widetilde{{\mathrm{GL}}}(X_1) \times \widetilde{{\mathrm{GL}}}(X_1') \times Q_1}(\chi_\psi |{\det}_{X_1}|^{r+\frac{1}{2}} \boxtimes \mathcal{S}(\operatorname{Isom}(Y_1,X_1'))), R_{\overline{\mathcal{B}}}(\tau_s \otimes \chi_\psi)) \\
& = \{ 0 \},\end{aligned}$$ respectively, where $\overline{\mathcal{B}}$ is the Borel subgroup of ${\mathrm{GL}}(X_2)$ opposite to $\mathcal{B}$, so that $$\begin{aligned}
\Theta_{W_2,V_{r+2}^+,\psi}(\pi)^*
& \subset {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_2)}(R^2, \tau_s \otimes \chi_\psi) \\
& = I_{Q_2}(\tau_s^\vee, 1)^*.\end{aligned}$$ Thus, we obtain a surjection $$I_{Q_2}(\tau_s^\vee, 1)
{\relbar\joinrel\twoheadrightarrow}\Theta_{W_2,V_{r+2}^+,\psi}(\pi).$$ When $s = -\frac{1}{2}$, the left-hand side is a quotient of $$\label{l:ind-princ-2-pf}
I_{Q_{(2,r)}}(\tau \otimes |\det|^{\frac{1}{2}}, |\det|^{\frac{r}{2}}),$$ where $Q_{(i,j)}$ is the parabolic subgroup of ${\mathrm{SO}}(V_{r+2}^+)$ with Levi component ${\mathrm{GL}}_i \times {\mathrm{GL}}_j \times {\mathrm{SO}}(V_{r-i-j+2}^+)$. By [@zelevinsky Theorem 4.2], the representation $${\operatorname{Ind}}^{{\mathrm{GL}}_{r+2}(F)}_{\mathcal{Q}_{(2,r)}}((\tau \otimes |\det|^{\frac{1}{2}}) \boxtimes |\det|^{\frac{r}{2}})$$ is irreducible and is isomorphic to $${\operatorname{Ind}}^{{\mathrm{GL}}_{r+2}(F)}_{\mathcal{Q}_{(r,2)}}(|\det|^{\frac{r}{2}} \boxtimes (\tau \otimes |\det|^{\frac{1}{2}})),$$ where $\mathcal{Q}_{(i,j)}$ is the maximal parabolic subgroup of ${\mathrm{GL}}_{i+j}$ with Levi component ${\mathrm{GL}}_i \times {\mathrm{GL}}_j$. Hence is isomorphic to $$I_{Q_{(r,2)}}(|\det|^{\frac{r}{2}}, \tau \otimes |\det|^{\frac{1}{2}}),$$ which is a quotient of $$I_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, \tau \otimes |\det|^{\frac{1}{2}}),$$ where $Q = Q_{(1^r,2)}$. This proves Lemma \[l:ind-princ-2\] in the nonarchimedean case.
### Proof of {#sss:howe-ps-nonarch}
Let $a, b \in F^\times$ be such that $\chi_a, \chi_b, 1$ are pairwise distinct. Then says that $$\theta_{W_2,V_{r+2}^+,\psi}(J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_b}^-)) = J_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \chi_a |\cdot|^{\frac{1}{2}}, \sigma^-_{\varphi_b}),$$ where $Q = Q_{(1^r)}$ and $\sigma^-_{\varphi_b}$ is the irreducible square-integrable representation of ${\mathrm{SO}}(V_2^+)$ with $L$-parameter $\varphi_b = (\chi_b \boxtimes S_2) \oplus (1 \boxtimes S_2)$ associated to the nontrivial character of $\overline{S}_{\varphi_b} \cong {\mathbb{Z}}/ 2 {\mathbb{Z}}$.
Put $\pi_0 = \omega_{W_1,\psi_b}^-$ to ease notation. Let $\pi$ be an irreducible genuine representation of ${\mathrm{Mp}}(W_2)$. We will compute $\theta_{W_2,V_{r+2}^+,\psi}(\pi)$ under the assumption that there is an embedding $$\pi {\lhook\joinrel\longrightarrow}I_{P_1, \psi}(\chi_s, \pi_0),$$ where $\chi_s = \chi |\cdot|^s$ for some unitary character $\chi$ of ${\mathrm{GL}}(X_1) \cong F^\times$ and some $s \in {\mathbb{R}}$ with $s \le 0$.
We have $$\begin{aligned}
\Theta_{W_2,V_{r+2}^+,\psi}(\pi)^*
& = {\operatorname{Hom}}_{{\mathrm{Mp}}(W_2)}(\omega_{W_2,V_{r+2}^+,\psi}, \pi) \\
& \subset {\operatorname{Hom}}_{{\mathrm{Mp}}(W_2)}(\omega_{W_2,V_{r+2}^+,\psi}, I_{P_1, \psi}(\chi_s, \pi_0)) \\
& = {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1)}(R_{P_1}(\omega_{W_2,V_{r+2}^+,\psi}), \chi_s \chi_\psi \boxtimes \pi_0), \end{aligned}$$ where $*$ denotes the linear dual and $R_{P_1}$ denotes the normalized Jacquet functor with respect to $P_1$. By the result of Kudla [@kudla], $R_{P_1}(\omega_{W_2,V_{r+2}^+,\psi})$ has a filtration $$R_{P_1}(\omega_{W_2,V_{r+2}^+,\psi}) = R^0 \supset R^1 \supset \{ 0 \}$$ of $\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1) \times {\mathrm{SO}}(V_{r+2}^+)$-modules such that $$\begin{aligned}
R^0/R^1 & \cong \chi_\psi |{\det}_{X_1}|^{r+\frac{1}{2}} \boxtimes \omega_{W_1,V_{r+2}^+,\psi}, \\
R^1 & \cong {\operatorname{Ind}}^{\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1) \times {\mathrm{SO}}(V_{r+2}^+)}_{\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1) \times Q_1}(\mathcal{S}(\operatorname{Isom}(Y_1,X_1)) \boxtimes \omega_{W_1,V_{r+1}^+,\psi}).\end{aligned}$$ Here $\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{GL}}(Y_1)$ acts on $\mathcal{S}(\operatorname{Isom}(Y_1, X_1))$ by $$[(\tilde{a}, b) \cdot f](g) = \chi_\psi(\tilde{a}) \cdot f(a^{-1} \circ g \circ b)$$ for $\tilde{a} \in \widetilde{{\mathrm{GL}}}(X_1)$ with projection $a \in {\mathrm{GL}}(X_1)$, $b \in {\mathrm{GL}}(Y_1)$, and $f \in \mathcal{S}(\operatorname{Isom}(Y_1, X_1))$. Since $s \le 0$, the action of $\widetilde{{\mathrm{GL}}}(X_1)$ shows that $${\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1)}(R^0/R^1, \chi_s \chi_\psi \boxtimes \pi_0) = \{ 0 \},$$ so that $$\begin{aligned}
\Theta_{W_2,V_{r+2}^+,\psi}(\pi)^*
& \subset {\operatorname{Hom}}_{\widetilde{{\mathrm{GL}}}(X_1) \times {\mathrm{Mp}}(W_1)}(R^1, \chi_s \chi_\psi \boxtimes \pi_0) \\
& = I_{Q_1}(\chi_s^{-1}, \Theta_{W_1,V_{r+1}^+,\psi}(\pi_0))^*.\end{aligned}$$ Thus, we obtain a surjection $$\label{eq:howe-ps-nonarch-pf}
I_{Q_1}(\chi_s^{-1}, \Theta_{W_1,V_{r+1}^+,\psi}(\pi_0))
{\relbar\joinrel\twoheadrightarrow}\Theta_{W_2,V_{r+2}^+,\psi}(\pi).$$ On the other hand, since $\pi_0$ is supercuspidal, $\Theta_{W_1,V_{r+1}^+,\psi}(\pi_0)$ is irreducible. Hence, by [@atobe-gan Theorem 1.4], we have $$\Theta_{W_1,V_{r+1}^+,\psi}(\pi_0) = J_{Q'}(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \sigma^-_{\varphi_b}),$$ where $Q' = Q_{(1^{r-1})}$. In particular, when $\chi = \chi_a$ and $s = -\frac{1}{2}$, the left-hand side of is a quotient of $$I_Q(\chi_a |\cdot|^{\frac{1}{2}}, |\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \sigma^-_{\varphi_b})
\cong
I_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{3}{2}}, \chi_a |\cdot|^{\frac{1}{2}}, \sigma^-_{\varphi_b}).$$ This proves .
The real case
-------------
Suppose that $F = {\mathbb{R}}$. We work in the category of $({\mathfrak{g}},K)$-modules. We retain the notation of §\[sss:notation-real\].
### Proof of Lemma \[l:ind-princ-2\] {#sss:ind-princ-2-real}
Put $(p,q) = (r+3,r+2)$. Let $\tau$ be the irreducible self-dual tempered representation of ${\mathrm{GL}}_2({\mathbb{R}})$ with $L$-parameter $\rho$. Then we have either
- $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ for some unitary characters $\chi_1, \chi_2$ of ${\mathbb{R}}^\times$ such that $\chi_1 \chi_2 = 1$ or $\{\chi_1, \chi_2\} = \{ 1, \mathrm{sgn} \}$, where $\mathcal{B}$ is the Borel subgroup of ${\mathrm{GL}}_2$; or
- $\tau = D_\kappa$ for some positive integer $\kappa$.
Put $$\epsilon_1 = \chi_1(-1), \qquad
\epsilon_2 = \chi_2(-1)$$ in the former case. Let $\pi$ be an irreducible genuine representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ such that there is a surjection $$I_{P_2, \psi}(\tau_s) {\relbar\joinrel\twoheadrightarrow}\pi,$$ where $\tau_s = \tau \otimes |\det|^s$ for some $s \in {\mathbb{R}}$ with $s \ge 0$. Let $\mu'_0$ be a lowest $(\overline{K'} \cap M_2)$-type of $\tau_s$, where $M_2$ is the Levi component of $P_2$. Let $\mu'$ be a lowest $K'$-type of $\pi$, so that $\mu'$ occurs in $I_{P_2, \psi}(\tau_s)$ with multiplicity one. As we will explicate below, we assume the following conditions:
- $\mu'_0$ is of minimal degree in $\tau_s$;
- $\mu'$ is of minimal degree in $I_{P_2, \psi}(\tau_s)$;
- $\deg \mu' = \deg \mu'_0$;
- the restriction of $\mu'$ to $K' \cap \widetilde{M}_2$ contains $\mu'_0 \otimes \chi_\psi$.
Then, by the induction principle of Adams–Barbasch (see [@ab1 Proposition 3.25], [@ab2 Theorem 8.7]), $\theta_{W_2,V_{p,q}, \psi}(\pi)$ is a subquotient of $$I_{Q_2}(\tau_s^\vee, 1)$$ containing $\mu$, where $\mu$ is the $K$-type corresponding to $\mu'$.
By [@ab2 Proposition 6.10], we may take $\mu'_0$ and $\mu'$ given by $$\mu'_0 =
\begin{cases}
(0;1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,1)$;} \\
(1;1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,-1), (-1,1)$;} \\
(0;-1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (-1,-1)$;} \\
(2\kappa+1;1) & \text{if $\tau = D_\kappa$}
\end{cases}$$ and $$\mu' =
\begin{cases}
(\frac{1}{2}, \frac{1}{2}) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,1)$;} \\
(\frac{1}{2}, -\frac{1}{2}) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,-1), (-1,1)$;} \\
(-\frac{1}{2}, -\frac{1}{2}) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (-1,-1)$;} \\
(\kappa+\frac{1}{2}, -\kappa-\frac{1}{2}) & \text{if $\tau = D_\kappa$,}
\end{cases}$$ which satisfy the conditions above (since $\theta_{W_2,V_{3,2}, \psi}(\pi)$ is nonzero) and $$\deg \mu' = \deg \mu'_0 =
\begin{cases}
0 & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,1)$;} \\
1 & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,-1), (-1,1)$;} \\
2 & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (-1,-1)$;} \\
2\kappa+1 & \text{if $\tau = D_\kappa$.}
\end{cases}$$ Then we have $$\mu =
\begin{cases}
(0,\dots,0;1) \boxtimes (0,\dots,0;1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,1)$;} \\
(0,\dots,0;1) \boxtimes (1,0,\dots,0;1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (1,-1), (-1,1)$;} \\
(0,\dots,0;1) \boxtimes (1,1,0,\dots,0;1) & \text{if $\tau = {\operatorname{Ind}}^{{\mathrm{GL}}_2({\mathbb{R}})}_{\mathcal{B}}(\chi_1 \boxtimes \chi_2)$ and $(\epsilon_1, \epsilon_2) = (-1,-1)$;} \\
(\kappa,0,\dots,0;1) \boxtimes (\kappa+1,0,\dots,0;1) & \text{if $\tau = D_\kappa$.}
\end{cases}$$ If we put $Q' = Q_{(2,1^r)}$, then it follows from [@ab2 Proposition 6.10] that $\mu$ is a lowest $K$-type of $$I_{Q'}(\tau_s^\vee, |\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}),$$ which has $I_{Q_2}(\tau_s^\vee, 1)$ as a quotient. Hence, if $0 \le s \le \frac{1}{2}$, then we have $$\theta_{W_2,V_{p,q},\psi}(\pi) = J_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, \tau_s),$$ where $Q = Q_{(1^r,2)}$. This completes the proof of Lemma \[l:ind-princ-2\] in the real case.
### Some $A$-packets
Let $G$ be the split odd special orthogonal group of rank $r+2$, so that $G = {\mathrm{SO}}(p,q)$ with $(p,q) = (r+3,r+2)$. Let $\theta$ be the Cartan involution of $G$ defined by $\theta(g) = {}^t g^{-1}$. Let ${\mathfrak{g}}$ be the complexified Lie algebra of $G$.
We consider local $A$-parameters $$\phi = \mathcal{D}_{\kappa} \boxtimes S_2, \qquad
\theta(\phi) = \phi \oplus (1 \boxtimes S_{2r})$$ with a positive integer $\kappa$. For $\epsilon \in \mu_2$, let $\sigma^\epsilon$ be the representation in the $A$-packet $\Pi_{\theta(\phi)}(G)$ associated to $\epsilon$. Then, by [@mr1; @mr2], we have $$\sigma^+ = A_{{\mathfrak{q}}_1}(\lambda_1), \qquad
\sigma^- = A_{{\mathfrak{q}}_0}(\lambda_0) \oplus A_{{\mathfrak{q}}_2}(\lambda_2),$$ where ${\mathfrak{q}}_i$ is a $\theta$-stable parabolic subalgebra of ${\mathfrak{g}}$ whose normalizer $L_i$ in $G$ satisfies $$L_i \cong {\mathrm{U}}(i,2-i) \times {\mathrm{SO}}(p-2i, q-4+2i),$$ $\lambda_i$ is the $1$-dimensional representation of $L_i$ given by $$\lambda_i = {\det}^{\kappa-r-1} \boxtimes 1,$$ and $A_{{\mathfrak{q}}_i}(\lambda_i)$ is the cohomologically induced representation defined by [@knappvogan (5.6)]. Note that $\lambda_i$ is in the weakly fair range, so that $A_{{\mathfrak{q}}_i}(\lambda_i)$ is a (possibly zero, possibly reducible) unitary representation of $G$ of finite length. If $\kappa > r$, then $\lambda_i$ is in the good range, so that $A_{{\mathfrak{q}}_i}(\lambda_i)$ is nonzero and irreducible. Moreover, by [@knappvogan Theorem 11.216], we have $$A_{{\mathfrak{q}}_1}(\lambda_1) \cong J_Q(|\cdot|^{r-\frac{1}{2}}, |\cdot|^{r-\frac{3}{2}}, \dots, |\cdot|^{\frac{1}{2}}, D_\kappa \otimes |\det|^{\frac{1}{2}})$$ for $\kappa > r$, where $Q = Q_{(1^r,2)}$.
\[l:atlas\] Assume that $r=4$. Then $A_{{\mathfrak{q}}_i}(\lambda_i)$ is nonzero and irreducible. Moreover, we have $$A_{{\mathfrak{q}}_1}(\lambda_1) \cong J_Q(|\cdot|^{\frac{7}{2}}, |\cdot|^{\frac{5}{2}}, |\cdot|^{\frac{3}{2}}, |\cdot|^{\frac{1}{2}}, D_\kappa \otimes |\det|^{\frac{1}{2}}),$$ where $Q = Q_{(1^4,2)}$.
It remains to prove the assertion for $\kappa \le 4$, but this can be checked by the Atlas software [@atlas]. For example, if $i=1$ and $\kappa=2$, then we have:
atlas> set G=SO(7,6)
Variable G: RealForm
atlas> Aq_reducible(KGB(G,14),[2,0,0,2,0,0],[1,0,0,1,0,0])
Value:
1*parameter(x=1953,lambda=[11,7,9,1,3,1]/2,nu=[7,1,5,-1,3,1]/2) [44]
atlas> set P=Parabolic:([4],KGB(G,2336))
Variable P: ([int],KGBElt)
atlas> set p=parameter(KGB(Levi(P),0),[0,0,0,0,5,-3]/2,[7,5,3,1,1,1]/2)
Variable p: Param
atlas> finalize(real_induce_standard(p,G))
Value:
1*parameter(x=1953,lambda=[11,7,9,1,3,1]/2,nu=[7,1,5,-1,3,1]/2) [44]
### Proof of {#sss:soudry-real}
For $\epsilon \in \mu_2$, let $\pi^\epsilon$ be the representation in the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4)$ associated to $\epsilon$. By definition, we have $$\pi^+ = J_{P_2,\psi}(D_\kappa \otimes |\det|^{\frac{1}{2}}), \qquad
\pi^- = \pi_\Lambda \oplus \pi_\Lambda^\vee,$$ where $\pi_\Lambda$ is the genuine discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $K'$-type $\Lambda = (\kappa+\frac{3}{2}, \kappa+\frac{3}{2})$ (relative to the parametrization depending on $\psi$). Note that $\pi^-$ does not depend on $\psi$. Then says that $$\theta_{W_2, V_{p,q}, \psi}(\pi^+) = A_{{\mathfrak{q}}_1}(\lambda_1), \qquad
\theta_{W_2, V_{p,q}, \psi}(\pi^-) = A_{{\mathfrak{q}}_0}(\lambda_0) \oplus A_{{\mathfrak{q}}_2}(\lambda_2)$$ when $(p,q) = (7,6)$.
The first assertion follows from Lemmas \[l:ind-princ-2\] and \[l:atlas\]. To prove the second assertion, we consider the theta lift $\theta_{W_2, V_{p,q},\psi}(\pi_\Lambda)$ when $4 \le p \le q+5$. By [@ab2 Theorem 3.3], we have $$\pi_{\Lambda} = \theta_{W_2, V_{5,0}, \psi}(\sigma_\Lambda),$$ where $\sigma_\Lambda$ is the irreducible representation of ${\mathrm{O}}(5,0)$ with highest weight $(\kappa-1,\kappa-1;1)$. Hence it follows from [@lmt Theorem 1.4] combined with induction in stages (see [@knappvogan Corollary 11.86]) that $\theta_{W_2, V_{p,q}, \psi}(\pi_\Lambda)$ is a subquotient of $A_{\mathfrak{q}}(\lambda)$, where ${\mathfrak{q}}$ is a $\theta$-stable parabolic subalgebra of $\mathfrak{so}(p,q)$ whose normalizer $L$ in ${\mathrm{SO}}(p,q)$ satisfies $$L \cong {\mathrm{U}}(2, 0) \times {\mathrm{SO}}(p-4,q)$$ and $\lambda$ is the $1$-dimensional representation of $L$ given by $${\det}^{\kappa-\frac{p+q-3}{2}} \boxtimes 1.$$ In particular, when $\{p,q\} = \{6,7\}$, we have $$\theta_{W_2, V_{p,q},\psi}(\pi_\Lambda) = A_{\mathfrak{q}}(\lambda)$$ by Lemma \[l:atlas\]. On the other hand, by [@ab2 Lemma 1.5], we have $$\theta_{W_2, V_{p,q}, \psi}(\pi_\Lambda^\vee) = \theta_{W_2, V_{q,p}, \psi}(\pi_\Lambda).$$ This completes the proof of .
Local $A$-packets for ${\mathrm{Mp}}_4$ {#a:A-packets}
=======================================
In this appendix, we describe local $A$-packets for ${\mathrm{Mp}}_4$ explicitly. Let $F$ be a local field of characteristic zero.
The nonarchimedean case
-----------------------
Suppose that $F$ is nonarchimedean. Let ${\mathrm{st}}$ be the Steinberg representation of ${\mathrm{GL}}_2(F)$ and put ${\mathrm{st}}_\chi = {\mathrm{st}}\otimes (\chi \circ \det)$ for any quadratic character $\chi$ of $F^\times$. We may regard ${\mathrm{st}}_\chi$ as a representation of ${\mathrm{SO}}(V_1^+)$. Let $\widetilde{{\mathrm{st}}}_{\chi, \psi}$ be the irreducible genuine square-integrable representation of ${\mathrm{Mp}}(W_1)$ contained in $I_{B,\psi}(\chi |\cdot|^{\frac{1}{2}})$.
### Elliptic tempered representations of ${\mathrm{Mp}}(W_2)$
We first introduce some notation. For this, we need to enumerate the elliptic tempered representations of ${\mathrm{Mp}}(W_2)$ which are not supercuspidal.
The following lemmas summarize the result of Hanzer–Matić [@hanzer-matic] on the composition series of parabolically induced representations of ${\mathrm{Mp}}(W_2)$.
\[l:nsc-mp-p1\] Let $\chi$ be a unitary character of $F^\times$. Let $s \in {\mathbb{R}}$ with $s \ge 0$. Let $\pi$ be either an irreducible genuine square-integrable representation of ${\mathrm{Mp}}(W_1)$ or an even elementary Weil representation of ${\mathrm{Mp}}(W_1)$. Then $I_{P_1, \psi}(\chi |\cdot|^s, \pi)$ is reducible if and only if one of the following holds:
1. \[nsc-mp-p1i\] $\chi^2 = 1$ (so that $\chi = \chi_a$ for some $a \in F^\times$), $s = \frac{1}{2}$, and $\pi$ is supercuspidal but $\pi \ne \omega_{W_1,\psi_a}^-$, in which case we have $$0 \longrightarrow \widetilde{{\mathrm{St}}}_\psi(\chi, \pi) \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \pi) \longrightarrow J_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \pi) \longrightarrow 0,$$ where $\widetilde{{\mathrm{St}}}_\psi(\chi, \pi)$ is an irreducible square-integrable representation;
2. $\chi^2 = 1$, $s = \frac{1}{2}$, and $\pi = \widetilde{{\mathrm{st}}}_{\mu, \psi}$ for some quadratic character $\mu$ of $F^\times$, in which case we have $$0 \longrightarrow \Pi \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \widetilde{{\mathrm{st}}}_{\mu, \psi}) \longrightarrow J_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \widetilde{{\mathrm{st}}}_{\mu, \psi}) \longrightarrow 0$$ with $$\Pi =
\begin{cases}
\widetilde{{\mathrm{St}}}_\psi(\chi, \widetilde{{\mathrm{st}}}_{\mu, \psi}) & \text{if $\chi \ne \mu$;} \\
\pi_{{\mathrm{gen}}, \psi}({\mathrm{st}}_\chi) & \text{if $\chi = \mu \ne 1$;} \\
\pi_{\mathrm{ng}, \psi}({\mathrm{st}}_\chi) & \text{if $\chi = \mu = 1$,}
\end{cases}$$ where $\widetilde{{\mathrm{St}}}_\psi(\chi, \widetilde{{\mathrm{st}}}_{\mu, \psi})$ is an irreducible square-integrable representation (which is isomorphic to $\widetilde{{\mathrm{St}}}_\psi(\mu, \widetilde{{\mathrm{st}}}_{\chi, \psi})$), and $\pi_{{\mathrm{gen}}, \psi}({\mathrm{st}}_\chi)$ and $\pi_{\mathrm{ng}, \psi}({\mathrm{st}}_\chi)$ are the representations as in Lemma \[l:nsc-mp-p2\] below;
3. \[nsc-mp-p1iii\] $\chi^2 = 1$, $s = \frac{1}{2}$, and $\pi = \omega_{W_1,\psi_b}^+$ for some $b \in F^\times$, in which case we have $$0 \longrightarrow \Pi \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_b}^+) \longrightarrow J_{B, \psi}(\chi |\cdot|^{\frac{1}{2}}, \chi_b |\cdot|^{\frac{1}{2}}) \longrightarrow 0$$ with $$\Pi =
\begin{cases}
J_{P_1, \psi}(\chi_b |\cdot|^{\frac{1}{2}}, \widetilde{{\mathrm{st}}}_{\chi,\psi}) & \text{if $\chi \ne \chi_b$;} \\
\pi_{\mathrm{ng}, \psi}({\mathrm{st}}_\chi) & \text{if $\chi = \chi_b \ne 1$;} \\
\pi_{{\mathrm{gen}}, \psi}({\mathrm{st}}_\chi) & \text{if $\chi = \chi_b = 1$;}
\end{cases}$$
4. \[nsc-mp-p1iv\] $\chi^2 = 1$, $s = \frac{3}{2}$, and $\pi = \widetilde{{\mathrm{st}}}_{\chi, \psi}$, in which case we have $$0 \longrightarrow \widetilde{{\mathrm{St}}}^+_{\chi, \psi} \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{3}{2}}, \widetilde{{\mathrm{st}}}_{\chi, \psi}) \longrightarrow J_{P_1, \psi}(\chi |\cdot|^{\frac{3}{2}}, \widetilde{{\mathrm{st}}}_{\chi, \psi}) \longrightarrow 0,$$ where $\widetilde{{\mathrm{St}}}^+_{\chi, \psi}$ is an irreducible square-integrable representation;
5. $\chi^2 = 1$ (so that $\chi = \chi_a$ for some $a \in F^\times$), $s = \frac{3}{2}$, and $\pi = \omega_{W_1,\psi_a}^-$, in which case we have $$0 \longrightarrow \widetilde{{\mathrm{St}}}^-_{\chi, \psi} \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{3}{2}}, \omega_{W_1,\psi_a}^-) \longrightarrow \omega_{W_2,\psi_a}^- \longrightarrow 0,$$ where $\widetilde{{\mathrm{St}}}^-_{\chi, \psi}$ is an irreducible square-integrable representation;
6. $\chi^2 = 1$ (so that $\chi = \chi_a$ for some $a \in F^\times$), $s = \frac{3}{2}$, and $\pi = \omega_{W_1,\psi_a}^+$, in which case we have $$0 \longrightarrow J_{P_2, \psi}({\mathrm{st}}_\chi \otimes |\det|) \longrightarrow I_{P_1, \psi}(\chi |\cdot|^{\frac{3}{2}}, \omega_{W_1, \psi_a}^+) \longrightarrow \omega_{W_2, \psi_a}^+ \longrightarrow 0.$$
\[l:nsc-mp-p2\] Let $\tau$ be an irreducible unitary square-integrable representation of ${\mathrm{GL}}_2(F)$ with central character $\omega_\tau$. Let $s \in {\mathbb{R}}$ with $s \ge 0$. Then $I_{P_2, \psi}(\tau \otimes |\det|^s)$ is reducible if and only if one of the following holds:
1. \[nsc-mp-p2i\] $\omega_\tau = 1$ and $s=0$, in which case we have $$I_{P_2, \psi}(\tau) = \pi_{{\mathrm{gen}}, \psi}(\tau) \oplus \pi_{\mathrm{ng}, \psi}(\tau),$$ where $\pi_{{\mathrm{gen}}, \psi}(\tau)$ is an irreducible $\psi$-generic tempered representation and $\pi_{\mathrm{ng}, \psi}(\tau)$ is an irreducible non-$\psi$-generic tempered representation;
2. $\tau$ is self-dual and supercuspidal, $\omega_\tau \ne 1$, and $s = \frac{1}{2}$, in which case we have $$0 \longrightarrow \widetilde{{\mathrm{St}}}_\psi(\tau) \longrightarrow I_{P_2, \psi}(\tau \otimes |\det|^{\frac{1}{2}}) \longrightarrow J_{P_2, \psi}(\tau \otimes |\det|^{\frac{1}{2}}) \longrightarrow 0,$$ where $\widetilde{{\mathrm{St}}}_\psi(\tau)$ is an irreducible square-integrable representation;
3. $\tau = {\mathrm{st}}_\chi$ for some quadratic character $\chi$ of $F^\times$ and $s=1$, in which case we have $$0 \longrightarrow \widetilde{{\mathrm{St}}}^+_{\chi,\psi} \longrightarrow I_{P_2, \psi}({\mathrm{st}}_\chi \otimes |\det|) \longrightarrow J_{P_2, \psi}({\mathrm{st}}_\chi \otimes |\det|) \longrightarrow 0,$$ where $\widetilde{{\mathrm{St}}}^+_{\chi,\psi}$ is the representation as in Lemma \[l:nsc-mp-p1\].
The representations described in the lemmas above exhaust all irreducible genuine elliptic tempered representations of ${\mathrm{Mp}}(W_2)$ which are not supercuspidal.
### Elliptic tempered representations of ${\mathrm{SO}}(V_2^\epsilon)$
We also need to enumerate the elliptic tempered representations of ${\mathrm{SO}}(V_2^\epsilon)$ which are not supercuspidal. We write $I_{Q_i}({\cdots}) = I^\epsilon_{Q_i}({\cdots})$ and $J_{Q_i}({\cdots}) = J^\epsilon_{Q_i}({\cdots})$ to indicate that they are representations of ${\mathrm{SO}}(V_2^\epsilon)$. Let ${\mathrm{St}}^\epsilon$ be the Steinberg representation of ${\mathrm{SO}}(V_2^\epsilon)$ and put ${\mathrm{St}}^\epsilon_\chi = {\mathrm{St}}^\epsilon \otimes (\chi \circ \nu)$ for any quadratic character $\chi$ of $F^\times$, where $\nu$ denotes the spinor norm.
The following lemmas summarize the result of Sally–Tadić [@sally-tadic] (see also [@matic]) on the composition series of parabolically induced representations of ${\mathrm{SO}}(V_2^+)$.
\[l:nsc-so+q1\] Let $\chi$ be a unitary character of $F^\times$. Let $s \in {\mathbb{R}}$ with $s \ge 0$. Let $\sigma$ be an irreducible square-integrable representation of ${\mathrm{SO}}(V_1^+)$. Then $I^+_{Q_1}(\chi |\cdot|^s, \sigma)$ is reducible if and only if one of the following holds:
1. \[nsc-so+q1i\] $\chi^2 = 1$, $s = \frac{1}{2}$, and $\sigma$ is supercuspidal, in which case we have $$0 \longrightarrow {\mathrm{St}}^+(\chi, \sigma) \longrightarrow I^+_{Q_1}(\chi |\cdot|^{\frac{1}{2}}, \sigma) \longrightarrow J^+_{Q_1}(\chi |\cdot|^{\frac{1}{2}}, \sigma) \longrightarrow 0,$$ where ${\mathrm{St}}^+(\chi, \sigma)$ is an irreducible square-integrable representation;
2. $\chi^2 = 1$, $s = \frac{1}{2}$, and $\sigma = {\mathrm{st}}_\mu$ for some quadratic character $\mu$ of $F^\times$, in which case we have $$0 \longrightarrow \Sigma \longrightarrow I^+_{Q_1}(\chi |\cdot|^{\frac{1}{2}}, {\mathrm{st}}_\mu) \longrightarrow J^+_{Q_1}(\chi |\cdot|^{\frac{1}{2}}, {\mathrm{st}}_\mu) \longrightarrow 0$$ with $$\Sigma =
\begin{cases}
{\mathrm{St}}^+(\chi, {\mathrm{st}}_\mu) & \text{if $\chi \ne \mu$;} \\
\sigma_{\mathrm{gen}}({\mathrm{st}}_\chi) & \text{if $\chi = \mu$,}
\end{cases}$$ where ${\mathrm{St}}^+(\chi, {\mathrm{st}}_\mu)$ is an irreducible square-integrable representation (which is isomorphic to ${\mathrm{St}}^+(\mu, {\mathrm{st}}_\chi)$) and $\sigma_{\mathrm{gen}}({\mathrm{st}}_\chi)$ is the representation as in Lemma \[l:nsc-so+q2\] below;
3. $\chi^2 = 1$, $s = \frac{3}{2}$, and $\sigma = {\mathrm{st}}_\chi$, in which case we have $$0 \longrightarrow {\mathrm{St}}^+_\chi \longrightarrow I^+_{Q_1}(\chi |\cdot|^{\frac{3}{2}}, {\mathrm{st}}_\chi) \longrightarrow J^+_{Q_1}(\chi |\cdot|^{\frac{3}{2}}, {\mathrm{st}}_\chi) \longrightarrow 0.$$
\[l:nsc-so+q2\] Let $\tau$ be an irreducible unitary square-integrable representation of ${\mathrm{GL}}_2(F)$ with central character $\omega_\tau$. Let $s \in {\mathbb{R}}$ with $s \ge 0$. Then $I^+_{Q_2}(\tau \otimes |\det|^s)$ is reducible if and only if one of the following holds:
1. \[nsc-so+q2i\] $\omega_\tau = 1$ and $s=0$, in which case we have $$I^+_{Q_2}(\tau) = \sigma_{\mathrm{gen}}(\tau) \oplus \sigma_{\mathrm{ng}}(\tau),$$ where $\sigma_{\mathrm{gen}}(\tau)$ is an irreducible generic tempered representation and $\sigma_{\mathrm{ng}}(\tau)$ is an irreducible nongeneric tempered representation;
2. $\tau$ is self-dual and supercuspidal, $\omega_\tau \ne 1$, and $s=\frac{1}{2}$, in which case we have $$0 \longrightarrow {\mathrm{St}}^+(\tau) \longrightarrow I^+_{Q_2}(\tau \otimes |\det|^{\frac{1}{2}}) \longrightarrow J^+_{Q_2}(\tau \otimes |\det|^{\frac{1}{2}}) \longrightarrow 0,$$ where ${\mathrm{St}}^+(\tau)$ is an irreducible square-integrable representation;
3. $\tau = {\mathrm{st}}_\chi$ for some quadratic character $\chi$ of $F^\times$ and $s=1$, in which case we have $$0 \longrightarrow {\mathrm{St}}^+_\chi \longrightarrow I^+_{Q_2}({\mathrm{st}}_\chi \otimes |\det|) \longrightarrow J^+_{Q_2}({\mathrm{st}}_\chi \otimes |\det|) \longrightarrow 0.$$
We can easily describe the composition series of parabolically induced representations of ${\mathrm{SO}}(V_2^-)$ as follows.
\[l:nsc-so-q1\] Let $\chi$ be a unitary character of $F^\times$. Let $s \in {\mathbb{R}}$ with $s \ge 0$. Let $\sigma$ be an irreducible representation of ${\mathrm{SO}}(V_1^-)$. Then $I^-_{Q_1}(\chi |\cdot|^s, \sigma)$ is reducible if and only if one of the following holds:
1. \[nsc-so-q1i\] $\chi^2 = 1$, $s=\frac{1}{2}$, and $\sigma \ne \chi \circ \nu$, in which case we have $$0 \longrightarrow {\mathrm{St}}^-(\chi, \sigma) \longrightarrow I^-_{Q_1}(\chi |\cdot|^{\frac{1}{2}}, \sigma) \longrightarrow J^-_{Q_1}(\chi|\cdot|^{\frac{1}{2}}, \sigma) \longrightarrow 0,$$ where ${\mathrm{St}}^-(\chi, \sigma)$ is an irreducible square-integrable representation;
2. $\chi^2 = 1$, $s=\frac{3}{2}$, and $\sigma = \chi \circ \nu$, in which case we have $$0 \longrightarrow {\mathrm{St}}^-_\chi \longrightarrow I^-_{Q_1}(\chi |\cdot|^{\frac{3}{2}}, \chi \circ \nu) \longrightarrow J^-_{Q_1}(\chi |\cdot|^{\frac{3}{2}}, \chi \circ \nu) \longrightarrow 0.$$
The representations described in the lemmas above exhaust all irreducible elliptic tempered representations of ${\mathrm{SO}}(V_2^\epsilon)$ which are not supercuspidal.
### Local Shimura correspondence
The local Shimura correspondence [@gs] is a bijection between representations of ${\mathrm{Mp}}(W_2)$ and ${\mathrm{SO}}(V_2^\epsilon)$ defined via theta lifts, which restricts to a bijection $$\Pi_{\varphi, \psi}({\mathrm{Mp}}(W_2)) \, \longleftrightarrow \,
\Pi_\varphi({\mathrm{SO}}(V_2^+)) \sqcup \Pi_\varphi({\mathrm{SO}}(V_2^-))$$ for any $L$-parameter $\varphi: W_F \times {\mathrm{SL}}_2({\mathbb{C}}) \rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$. The following table describes this bijection explicitly when any irreducible summand of $\varphi$ is symplectic.
$\varphi:$ $L$-parameter $\Pi_{\varphi, \psi}({\mathrm{Mp}}(W_2))$ $\Pi_{\varphi}({\mathrm{SO}}(V_2^\epsilon))$
------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------
$\varphi = \varrho \boxtimes S_1$ $\pi^{\epsilon}:$ supercuspidal $\sigma^{\epsilon}:$ supercuspidal
$\varphi = \rho \boxtimes S_2$ $\pi^+ = \widetilde{{\mathrm{St}}}_\psi(\tau)$ $\sigma^+ = {\mathrm{St}}^+(\tau)$
$\pi^-:$ supercuspidal $\sigma^-:$ supercuspidal
$\varphi = \chi_a \boxtimes S_4$ $\pi^+ = \widetilde{{\mathrm{St}}}^+_{\chi_a, \psi}$ $\sigma^+ = {\mathrm{St}}^+_{\chi_a}$
$\pi^- = \widetilde{{\mathrm{St}}}^-_{\chi_a, \psi}$ $\sigma^- = {\mathrm{St}}^-_{\chi_a}$
$\varphi = 1 \boxtimes S_4$ $\pi^+ = \widetilde{{\mathrm{St}}}^-_{1,\psi}$ $\sigma^+ = {\mathrm{St}}_1^+$
$\pi^- = \widetilde{{\mathrm{St}}}^+_{1,\psi}$ $\sigma^- = {\mathrm{St}}_1^-$
$\varphi = (\rho_1 \boxtimes S_1) \oplus (\rho_2 \boxtimes S_1)$ $\pi^{\epsilon_1, \epsilon_2}:$ supercuspidal $\sigma^{\epsilon_1, \epsilon_2}:$ supercuspidal
$\varphi = (\rho_0 \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ $\pi^{+,+} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \pi_0^+)$ $\sigma^{+,+} = {\mathrm{St}}^+(\chi_a,\sigma_0^+)$
$\pi^{+,-} = \theta_{W_2, V_1^{\varepsilon_a}, \psi_a}((\sigma_0^{\varepsilon_a} \otimes \nu_a)^{-\varepsilon_0 \chi_a(-1)})$ $\sigma^{+,-} = \theta_{W_1, V_2^-, \psi_a}(\pi^{\varepsilon_a}_0) \otimes \nu_a$
$\pi^{-,+} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \pi_0^-)$ $\sigma^{-,+} = {\mathrm{St}}^-(\chi_a,\sigma_0^-)$
$\pi^{-,-} = \theta_{W_2, V_1^{-\varepsilon_a}, \psi_a}((\sigma_0^{-\varepsilon_a} \otimes \nu_a)^{\varepsilon_0 \chi_a(-1)})$ $\sigma^{-,-} = \theta_{W_1, V_2^+, \psi_a}(\pi^{-\varepsilon_a}_0) \otimes \nu_a$
$\varphi = (\rho_0 \boxtimes S_1) \oplus (1 \boxtimes S_2)$ $\pi^{+,+} = \theta_{W_2, V_1^+, \psi}(\sigma_0^{+, -\varepsilon_0})$ $\sigma^{+,+} = {\mathrm{St}}^+(1,\sigma_0^+)$
$\pi^{+,-} = \widetilde{{\mathrm{St}}}_\psi(1, \pi_0^+)$ $\sigma^{+,-} = \theta_{W_1, V_2^-, \psi}(\pi_0^+)$
$\pi^{-,+} = \theta_{W_2, V_1^-, \psi}(\sigma_0^{-, \varepsilon_0})$ $\sigma^{-,+} = {\mathrm{St}}^-(1,\sigma_0^-)$
$\pi^{-,-} = \widetilde{{\mathrm{St}}}_\psi(1, \pi_0^-)$ $\sigma^{-,-} = \theta_{W_1, V_2^+, \psi}(\pi_0^-)$
$\varphi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ $\pi^{+,+} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \widetilde{{\mathrm{st}}}_{\chi_b,\psi})$ $\sigma^{+,+} = {\mathrm{St}}^+(\chi_a, {\mathrm{st}}_{\chi_b})$
$\pi^{+,-} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \omega_{W_1,\psi_b}^-)$ $\sigma^{+,-} = {\mathrm{St}}^-(\chi_a, \nu_b)$
$\pi^{-,+} = \widetilde{{\mathrm{St}}}_\psi(\chi_b, \omega_{W_1,\psi_a}^-)$ $\sigma^{-,+} = {\mathrm{St}}^-(\chi_b, \nu_a)$
$\pi^{-,-} = \theta_{W_2, V_1^-, \psi_b}(\nu_{ab}^{\chi_{ab}(-1)})$ $\sigma^{-,-} = \theta_{W_1, V_2^+, \psi_b}(\omega_{W_1, \psi_a}^-) \otimes \nu_b$
$\varphi = (\chi_a \boxtimes S_2) \oplus (1 \boxtimes S_2)$ $\pi^{+,+} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \omega_{W_1,\psi}^-)$ $\sigma^{+,+} = {\mathrm{St}}^+(\chi_a, {\mathrm{st}}_1)$
$\pi^{+,-} = \widetilde{{\mathrm{St}}}_\psi(\chi_a, \widetilde{{\mathrm{st}}}_{1,\psi})$ $\sigma^{+,-} = {\mathrm{St}}^-(\chi_a, \nu_1)$
$\pi^{-,+} = \theta_{W_2, V_1^-, \psi}(\nu_a^{\chi_a(-1)})$ $\sigma^{-,+} = {\mathrm{St}}^-(1, \nu_a)$
$\pi^{-,-} = \widetilde{{\mathrm{St}}}_\psi(1, \omega_{W_1,\psi_a}^-)$ $\sigma^{-,-} = \theta_{W_1, V_2^+, \psi}(\omega_{W_1, \psi_a}^-)$
$\varphi = \varphi_0 \oplus \varphi_0$ $\pi^{+,+} = \pi_{{\mathrm{gen}}, \psi}(\tau_0)$ $\sigma^{+,+} = \sigma_{\mathrm{gen}}(\tau_0)$
$\pi^{-,-} = \pi_{\mathrm{ng}, \psi}(\tau_0)$ $\sigma^{-,-} = \sigma_{\mathrm{ng}}(\tau_0)$
- $a, b \in F^\times$ such that $\chi_a, \chi_b, 1$ are pairwise distinct
- $\varrho:$ $4$-dimensional irreducible symplectic representation of $W_F$
- $\rho_0, \rho_1, \rho_2:$ $2$-dimensional irreducible symplectic representation of $W_F$ such that $\rho_1 \ne \rho_2$
- $\rho:$ $2$-dimensional irreducible orthogonal representation of $W_F$
- $\varphi_0:$ $2$-dimensional irreducible symplectic representation of $L_F$
- $\Pi_{\rho_0,\psi}({\mathrm{Mp}}(W_1)) = \{ \pi_0^\epsilon \, | \, \epsilon \in \mu_2 \}$
- $\Pi_{\rho_0}({\mathrm{SO}}(V_1^\epsilon)) = \{ \sigma_0^\epsilon \}$
- $\Pi_\rho({\mathrm{GL}}_2(F)) = \{ \tau \}$
- $\Pi_{\varphi_0}({\mathrm{GL}}_2(F)) = \{ \tau_0 \}$
- $\nu_a = \chi_a \circ \nu$ with $\nu:$ spinor norm
- $\nu_a^\epsilon:$ $\epsilon$-extension of $\nu_a$
- $\varepsilon_0 = \epsilon(\frac{1}{2}, \rho_0)$
- $\varepsilon_a = \epsilon(\frac{1}{2}, \rho_0) \cdot \epsilon(\frac{1}{2}, \rho_0 \times \chi_a) \cdot \chi_a(-1)$
### Nontempered $A$-packets
For any $A$-parameter $\phi : L_F \times {\mathrm{SL}}_2({\mathbb{C}}) \rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$, we have described most of the representations in the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}(W_2))$ explicitly in the body of this paper. The following lemmas determine the remaining representations.
Suppose that $\phi = (\rho \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ with a $2$-dimensional symplectic almost tempered representation $\rho$ of $L_F$ and $a \in F^\times$.
1. If either
- $\rho = \rho_0 \boxtimes S_1$ for some $2$-dimensional irreducible symplectic representation $\rho_0$ of $W_F$; or
- $\rho = \chi_b \boxtimes S_2$ for some $b \in F^\times$ such that $\chi_a \ne \chi_b$,
then for any $\epsilon_1 \in \mu_2$, we have $$\pi^{\epsilon_1, -} = \pi^{\epsilon_1, -}_\varphi,$$ where $\pi^{\epsilon_1,-}_\varphi$ is the irreducible genuine square-integrable representation of ${\mathrm{Mp}}(W_2)$ with $L$-parameter $\varphi = \rho \oplus (\chi_a \boxtimes S_2)$ (relative to $\psi$) associated to $(\epsilon_1,-1) \in \widehat{S}_\varphi = \mu_2 \times \mu_2$.
2. If $\rho = \chi_a \boxtimes S_2$, then for any $\epsilon_1 \in \mu_2$, we have $$\pi^{\epsilon_1, -} =
\begin{cases}
0 & \text{if $\chi_a \ne 1$ and $\epsilon_1 = 1$;} \\
\pi_\varphi^{-,-} & \text{if $\chi_a \ne 1$ and $\epsilon_1 = -1$;} \\
\pi_\varphi^{+,+} & \text{if $\chi_a = 1$ and $\epsilon_1 = 1$;} \\
0 & \text{if $\chi_a = 1$ and $\epsilon_1 = -1$,}
\end{cases}$$ where $\pi^{\epsilon_1,\epsilon_1}_\varphi$ is the irreducible genuine tempered representation of ${\mathrm{Mp}}(W_2)$ with $L$-parameter $\varphi = (\chi_a \boxtimes S_2) \oplus (\chi_a \boxtimes S_2)$ (relative to $\psi$) associated to $(\epsilon_1,\epsilon_1) \in \widehat{S}_\varphi = \Delta \mu_2$.
3. If $\rho = (\chi |\cdot|^s \boxtimes S_1) \oplus (\chi^{-1} |\cdot|^{-s} \boxtimes S_1)$ for some unitary character $\chi$ of $F^\times$ and some $s \in {\mathbb{R}}$ with $0 \le s < \frac{1}{2}$, then we have $$\pi^{+,-} = J_{P_1,\psi}(\chi |\cdot|^s, \omega^-_{W_1,\psi_a}).$$
For $\epsilon_1 \in \mu_2$, put $\epsilon = \epsilon_1 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \epsilon(\tfrac{1}{2},\rho \times \chi_a) \cdot \chi_a(-1)$ and $\epsilon' = - \epsilon_1 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \chi_a(-1)$. Let $\sigma_0^\epsilon$ be the irreducible representation of $(B^\epsilon)^\times$ with $L$-parameter $\rho \otimes \chi_a$ and $\sigma^{\epsilon,\epsilon'}$ the $\epsilon'$-extension of $\sigma_0^\epsilon$ to ${\mathrm{O}}(V_1^\epsilon)$. Since $\pi^{\epsilon_1, -} = \theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$, the lemma follows from Lemmas \[l:o3mp4-nonzero\], \[l:more-theta-nonarch-1\], \[l:more-theta-nonarch-2\], and \[l:more-theta-nonarch-3\].
Suppose that $\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ with $a,b \in F^\times$.
1. If $\chi_a \ne \chi_b$, then we have $$\pi^{-,-} = \pi^{-,-}_\varphi,$$ where $\pi^{-,-}_\varphi$ is the irreducible genuine square-integrable representation of ${\mathrm{Mp}}(W_2)$ with $L$-parameter $\varphi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ (relative to $\psi$) associated to $(-1,-1) \in \widehat{S}_\varphi = \mu_2 \times \mu_2$.
2. If $\chi_a = \chi_b$, then we have $$\pi^{-,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \widetilde{{\mathrm{st}}}_{\chi_a, \psi}).$$
Put $\epsilon' = \chi_{ab}(-1)$. Let $\sigma^{-, \epsilon'}$ be the $\epsilon'$-extension of $\chi_{ab} \circ \mathrm{N}_{B^-}$ to ${\mathrm{O}}(V_1^-)$. Since $\pi^{-,-} = \theta_{W_2, V_1^-, \psi_a}(\sigma^{-, \epsilon'})$, the lemma follows from Lemmas \[l:ind-princ-1\] and \[l:more-theta-nonarch-1\].
The following table describes the representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}(W_2))$ when $\phi$ is nontempered and any irreducible summand of $\phi$ is symplectic.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\phi:$ $A$-parameter $\Pi_{\phi, \psi}({\mathrm{Mp}}(W_2))$
------------------------------------------------------------------------------------------------------ -----------------------------------------------------------------------------------------------------
$\phi = \chi_a \boxtimes S_4$ $\pi^+ = J_{B, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \chi_a|\cdot|^{\frac{1}{2}})$
$\pi^- = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \omega_{W_1,\psi_a}^-)$
$\phi = (\rho_0 \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ with $\rho_0 \ne \chi_a \boxtimes S_2$ $\pi^{\epsilon_1,+} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \pi_0^{\epsilon_1})$
$\pi^{\epsilon_1, -} = \pi^{\epsilon_1, -}_\varphi$
$\phi = (\rho_0 \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ with $\rho_0 = \chi_a \boxtimes S_2$ $\pi^{\epsilon_1,+} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \pi_0^{\epsilon_1})$
$\pi^{\epsilon_1, -} =
\begin{cases}
0 & \text{if $\chi_a \ne 1$ and $\epsilon_1 = 1$} \\
\pi_\varphi^{-,-} & \text{if $\chi_a \ne 1$ and $\epsilon_1 = -1$} \\
\pi_\varphi^{+,+} & \text{if $\chi_a = 1$ and $\epsilon_1 = 1$} \\
0 & \text{if $\chi_a = 1$ and $\epsilon_1 = -1$}
\end{cases}
$
$\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ $\pi^{+,+} = J_{B, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \chi_b |\cdot|^{\frac{1}{2}})$
$\pi^{+,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_b}^-)$
$\pi^{-,+} = J_{P_1, \psi}(\chi_b |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_a}^-)$
$\pi^{-,-} = \pi^{-,-}_\varphi$
$\phi = (\chi_a \boxtimes S_2) \oplus (\chi_a \boxtimes S_2)$ $\pi^{+,+} = J_{B, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \chi_a |\cdot|^{\frac{1}{2}})$
$\pi^{-,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \widetilde{{\mathrm{st}}}_{\chi_a, \psi})$
$\phi = \rho \boxtimes S_2$ $\pi^+ = J_{P_2,\psi}(\tau \otimes |\det|^{\frac{1}{2}})$
$\pi^- = \pi^-_\varphi$
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- $a, b \in F^\times$ such that $\chi_a \ne \chi_b$
- $\rho_0:$ $2$-dimensional irreducible symplectic representation of $L_F$
- $\rho:$ $2$-dimensional irreducible orthogonal representation of $L_F$
- $\varphi = \phi \circ \Delta:$ $4$-dimensional symplectic representation of $L_F$
- $\Delta : W_F \times {\mathrm{SL}}_2({\mathbb{C}}) \rightarrow W_F \times {\mathrm{SL}}_2({\mathbb{C}}) \times {\mathrm{SL}}_2({\mathbb{C}}) :$ diagonal map
- $\Pi_{\rho_0, \psi}({\mathrm{Mp}}(W_1)) = \{ \pi_0^\epsilon \, | \, \epsilon \in \mu_2 \}$
- $\Pi_\rho({\mathrm{GL}}_2(F)) = \{ \tau \}$
The real case
-------------
Suppose that $F = {\mathbb{R}}$. We retain the notation of §\[sss:notation-real\].
### (Limit of) discrete series representations of ${\mathrm{Mp}}_4({\mathbb{R}})$ {#sss:temp-mp4-real}
Recall from [@vogan00 Theorem 8.1] that there is a bijection $$\begin{tikzcd}
\{ \text{irreducible genuine tempered representations of ${\mathrm{Mp}}_4({\mathbb{R}})$ with real infinitesimal character} \} \arrow[d] \\
\{ \text{irreducible genuine representations of $K'$} \} \arrow[u]
\end{tikzcd}$$ sending $\pi$ in the first set (see [@vogan00 Definition 8.5] for the precise definition) to the unique lowest $K'$-type of $\pi$, where $K'$ is the maximal compact subgroup of ${\mathrm{Mp}}_4({\mathbb{R}})$. In particular, any genuine (limit of) discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ (which always has real infinitesimal character) is uniquely determined by its lowest $K'$-type.
We now describe the $L$-packet $\Pi_{\varphi,\psi}({\mathrm{Mp}}_4({\mathbb{R}}))$ explicitly for any $L$-parameter $\varphi : W_{\mathbb{R}}\rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$ when any irreducible summand of $\varphi$ is symplectic. We may write $$\varphi = \mathcal{D}_a \oplus \mathcal{D}_b$$ with some $a, b \in \frac{1}{2} + {\mathbb{Z}}$ such that $a \ge b > 0$ and $$\Pi_{\varphi, \psi}({\mathrm{Mp}}_4({\mathbb{R}})) = \{ \pi^{\epsilon_1,\epsilon_2} \, | \, \epsilon_1, \epsilon_2 \in \mu_2 \},$$ where we interpret $\pi^{\epsilon_1,\epsilon_2}$ as zero if $a = b$ and $\epsilon_1 \ne \epsilon_2$. If $a > b$, then by [@luo §2.2], $\pi^{\epsilon_1,\epsilon_2}$ is the genuine discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $K'$-type $\Lambda^{\epsilon_1,\epsilon_2}$ (relative to $\psi$) given by $$\begin{aligned}
\Lambda^{+,+} & = (a+1, -b), \\
\Lambda^{+,-} & = (a+1,b+2), \\
\Lambda^{-,+} & = (-b-2,-a-1), \\
\Lambda^{-,-} & = (b, -a-1).\end{aligned}$$ If $a=b$, then $\pi^{\epsilon_1,\epsilon_2}$ is the genuine limit of discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with lowest $K'$-type $\Lambda^{\epsilon_1, \epsilon_2}$ (relative to $\psi$) given by $$\begin{aligned}
\Lambda^{+,+} & = (a+1, -a), \\
\Lambda^{-,-} & = (a, -a-1), \end{aligned}$$ so that $$I_{P_2, \psi}(D_a) = \pi^{+,+} \oplus \pi^{-,-}.$$ The representations described above exhaust all genuine (limit of) discrete series representations of ${\mathrm{Mp}}_4({\mathbb{R}})$.
### Nontempered $A$-packets
For any $A$-parameter $\phi : W_{\mathbb{R}}\times {\mathrm{SL}}_2({\mathbb{C}}) \rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$, we have described most of the representations in the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4({\mathbb{R}}))$ explicitly in the body of this paper. The following lemmas determine the remaining representations.
Suppose that $\phi = (\rho \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ with a $2$-dimensional symplectic almost tempered representation $\rho$ of $W_{\mathbb{R}}$ and $a \in {\mathbb{R}}^\times$.
1. If $\rho = \mathcal{D}_{\kappa - \frac{1}{2}}$ for some positive integer $\kappa$, then for any $\epsilon_1 \in \mu_2$, we have $$\pi^{\epsilon_1, -} =
\begin{cases}
\pi_\varphi^{\epsilon_1, \chi_a(-1)} & \text{if $\kappa>1$ or $\epsilon_1 = \chi_a(-1)$;} \\
0 & \text{if $\kappa=1$ and $\epsilon_1 = -\chi_a(-1)$,}
\end{cases}$$ where $\pi_\varphi^{\epsilon_1, \chi_a(-1)}$ is the genuine (limit of) discrete series representation of ${\mathrm{Mp}}_4({\mathbb{R}})$ with $L$-parameter $\varphi = \mathcal{D}_{\kappa - \frac{1}{2}} \oplus \mathcal{D}_{\frac{1}{2}}$ (relative to $\psi$) associated to $(\epsilon_1, \chi_a(-1)) \in \widehat{S}_\varphi \subset \mu_2 \times \mu_2$.
2. If $\rho = \chi |\cdot|^s \oplus \chi^{-1} |\cdot|^{-s}$ for some unitary character $\chi$ of ${\mathbb{R}}^\times$ and some $s \in {\mathbb{R}}$ with $0 \le s < \frac{1}{2}$, then we have $$\pi^{+,-} = J_{P_1,\psi}(\chi |\cdot|^s, \omega^-_{W_1,\psi_a}).$$
For $\epsilon_1 \in \mu_2$, put $\epsilon = \epsilon_1 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \epsilon(\tfrac{1}{2},\rho \times \chi_a) \cdot \chi_a(-1)$ and $\epsilon' = - \epsilon_1 \cdot \epsilon(\tfrac{1}{2},\rho) \cdot \chi_a(-1)$. Let $\sigma_0^\epsilon$ be the irreducible representation of $(B^\epsilon)^\times$ with $L$-parameter $\rho \otimes \chi_a$ and $\sigma^{\epsilon,\epsilon'}$ the $\epsilon'$-extension of $\sigma_0^\epsilon$ to ${\mathrm{O}}(V_1^\epsilon)$. Since $\pi^{\epsilon_1, -} = \theta_{W_2,V_1^{\epsilon}, \psi_a}(\sigma^{\epsilon, \epsilon'})$, the lemma follows from Lemmas \[l:o3mp4-nonzero\], \[l:more-theta-real-1\], and \[l:more-theta-real-2\].
Suppose that $\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ with $a,b \in {\mathbb{R}}^\times$.
1. If $\chi_a \ne \chi_b$, then we have $$\pi^{-,-} = 0.$$
2. If $\chi_a = \chi_b$, then we have $$\pi^{-,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, (\omega^-_{W_1,\psi_a})^\vee).$$
Put $\epsilon' = \chi_{ab}(-1)$. Let $\sigma^{-, \epsilon'}$ be the $\epsilon'$-extension of $\chi_{ab} \circ \mathrm{N}_{B^-}$ to ${\mathrm{O}}(V_1^-)$. Since $\pi^{-,-} = \theta_{W_2, V_1^-, \psi_a}(\sigma^{-, \epsilon'})$, the lemma follows from Lemmas \[l:o3mp4-nonzero\] and \[l:ind-princ-1\].
The following table describes the representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}_4({\mathbb{R}}))$ when $\phi$ is nontempered and any irreducible summand of $\phi$ is symplectic.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\phi:$ $A$-parameter $\Pi_{\phi, \psi}({\mathrm{Mp}}_4({\mathbb{R}}))$
------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------
$\phi = \chi_a \boxtimes S_4$ $\pi^+ = J_{B, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \chi_a|\cdot|^{\frac{1}{2}})$
$\pi^- = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{3}{2}}, \omega_{W_1,\psi_a}^-)$
$\phi = (\mathcal{D}_{\kappa - \frac{1}{2}} \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$ $\pi^{\epsilon_1,+} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \widetilde{D}_{\lambda,\psi})$
$\pi^{\epsilon_1, -} =
\begin{cases}
\pi_\varphi^{\epsilon_1, \chi_a(-1)} & \text{if $\kappa>1$ or $\epsilon_1 = \chi_a(-1)$} \\
0 & \text{if $\kappa=1$ and $\epsilon_1 = -\chi_a(-1)$}
\end{cases}
$
$\phi = (\chi_a \boxtimes S_2) \oplus (\chi_b \boxtimes S_2)$ $\pi^{+,+} = J_{B, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \chi_b |\cdot|^{\frac{1}{2}})$
$\pi^{+,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_b}^-)$
$\pi^{-,+} = J_{P_1, \psi}(\chi_b |\cdot|^{\frac{1}{2}}, \omega_{W_1,\psi_a}^-)$
$\pi^{-,-} = 0$
$\phi = (\chi_a \boxtimes S_2) \oplus (\chi_a \boxtimes S_2)$ $\pi^{+,+} = J_{B, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, \chi_a |\cdot|^{\frac{1}{2}})$
$\pi^{-,-} = J_{P_1, \psi}(\chi_a |\cdot|^{\frac{1}{2}}, (\omega^-_{W_1,\psi_a})^\vee)$
$\phi = \mathcal{D}_\kappa \boxtimes S_2$ $\pi^+ = J_{P_2,\psi}(D_\kappa \otimes |\det|^{\frac{1}{2}})$
$\pi^- = \pi_{\varphi}^{+,-} \oplus \pi_{\varphi}^{-,+}$
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- $a, b \in {\mathbb{R}}^\times$ such that $\chi_a \ne \chi_b$
- $\kappa:$ positive integer
- $\lambda =
\begin{cases}
\kappa-\frac{1}{2} & \text{if $\epsilon_1 = 1$} \\
-\kappa+\frac{1}{2} & \text{if $\epsilon_1 = -1$}
\end{cases}
$
- $\varphi =
\begin{cases}
\mathcal{D}_{\kappa - \frac{1}{2}} \oplus \mathcal{D}_{\frac{1}{2}}
& \text{if $\phi = (\mathcal{D}_{\kappa - \frac{1}{2}} \boxtimes S_1) \oplus (\chi_a \boxtimes S_2)$} \\
\mathcal{D}_{\kappa + \frac{1}{2}} \oplus \mathcal{D}_{\kappa - \frac{1}{2}}
& \text{if $\phi = \mathcal{D}_\kappa \boxtimes S_2$}
\end{cases}
$
The complex case
----------------
Suppose that $F = {\mathbb{C}}$. For any $A$-parameter $\phi : W_{\mathbb{C}}\times {\mathrm{SL}}_2({\mathbb{C}}) \rightarrow {\mathrm{Sp}}_4({\mathbb{C}})$, we have described the representations in the $A$-packet $\Pi_{\phi,\psi}({\mathrm{Mp}}_4({\mathbb{C}}))$ explicitly in the body of this paper. The following table describes the representations in $\Pi_{\phi,\psi}({\mathrm{Mp}}_4({\mathbb{C}}))$ when $\phi$ is nontempered and any irreducible summand of $\phi$ is symplectic.
$\phi:$ $A$-parameter $\Pi_{\phi, \psi}({\mathrm{Mp}}_4({\mathbb{C}}))$
----------------------------------------------------- -------------------------------------------------------------------------
$\phi = 1 \boxtimes S_4$ $\pi^+ = J_{B, \psi}(|\cdot|^{\frac{3}{2}}, |\cdot|^{\frac{1}{2}})$
$\pi^- = J_{P_1, \psi}(|\cdot|^{\frac{3}{2}}, \omega_{W_1,\psi}^-)$
$\phi = (1 \boxtimes S_2) \oplus (1 \boxtimes S_2)$ $\pi^{+,+} = J_{B, \psi}(|\cdot|^{\frac{1}{2}}, |\cdot|^{\frac{1}{2}})$
$\pi^{-,-} = 0$
[99]{}
J. Adams, *The theta correspondence over $\mathbb{R}$.* Harmonic analysis, group representations, automorphic forms and invariant theory, 1–39, Lect. Notes Ser. Inst. Math. Sci. Natl. Univ. Singap. **12**, World Sci. Publ., Hackensack, NJ, 2007.
J. Adams and D. Barbasch, *Reductive dual pair correspondence for complex groups.* J. Funct. Anal. **132** (1995), no. 1, 1–42.
J. Adams and D. Barbasch, *Genuine representations of the metaplectic group.* Compos. Math. **113** (1998), no. 1, 23–66.
J. Arthur, *The endoscopic classification of representations: orthogonal and symplectic groups.* American Mathematical Society Colloquium Publications **61**, American Mathematical Society, Providence, RI, 2013.
H. Atobe and W. T. Gan, *Local theta correspondence of tempered representations and Langlands parameters.* Invent. Math. **210** (2017), no. 2, 341–415.
W. T. Gan, *A Langlands program for covering groups?* Proceedings of the Sixth International Congress of Chinese Mathematicians. Vol. I, 57–78, Adv. Lect. Math. **36**, Int. Press, Somerville, MA, 2017.
W. T. Gan and A. Ichino, *Formal degrees and local theta correspondence.* Invent. Math. **195** (2014), no. 3, 509–672.
W. T. Gan and A. Ichino, *The Shimura–Waldspurger correspondence for ${\mathrm{Mp}}_{2n}$.* Ann. of Math. (2) **188** (2018), no. 3, 965–1016.
W. T. Gan and A. Ichino, *On the irreducibility of some induced representations of real reductive Lie groups.* Tunis. J. Math. **1** (2019), no. 1, 73–107.
W. T. Gan, Y. Qiu, and S. Takeda, *The regularized Siegel–Weil formula (the second term identity) and the Rallis inner product formula.* Invent. Math. **198** (2014), no. 3, 739–831.
W. T. Gan and G. Savin, *Representations of metaplectic groups I: epsilon dichotomy and local Langlands correspondence.* Compos. Math. **148** (2012), no. 6, 1655–1694.
W. T. Gan and S. Takeda, *On the Howe duality conjecture in classical theta correspondence.* Advances in the theory of automorphic forms and their $L$-functions, 105–117, Contemp. Math. **664**, Amer. Math. Soc., Providence, RI, 2016.
W. T. Gan and S. Takeda, *A proof of the Howe duality conjecture.* J. Amer. Math. Soc. **29** (2016), no. 2, 473–493.
F. Gao, *The residual spectrum of $\mathrm{Mp}_4(\mathbf{A}_k)$.* Trans. Amer. Math. Soc. **366** (2014), no. 11, 6151–6182.
M. Hanzer and I. Matić, *The unitary dual of $p$-adic $\widetilde{\mathrm{Sp}(2)}$.* Pacific J. Math. **248** (2010), no. 1, 107–137.
M. Harris and R. Taylor, *The geometry and cohomology of some simple Shimura varieties.* Annals of Mathematics Studies **151**, Princeton University Press, Princeton, NJ, 2001.
G. Henniart, *Une preuve simple des conjectures de Langlands pour $\mathrm{GL}(n)$ sur un corps $p$-adique.* Invent. Math. **139** (2000), no. 2, 439–455.
R. Howe, *Transcending classical invariant theory.* J. Amer. Math. Soc. **2** (1989), no. 3, 535–552.
T. Ibukiyama, *A conjecture on a Shimura type correspondence for Siegel modular forms, and Harder’s conjecture on congruences.* Modular forms on Schiermonnikoog, 107–144, Cambridge Univ. Press, Cambridge, 2008.
A. Ichino, *On the regularized Siegel-Weil formula.* J. Reine Angew. Math. **539** (2001), 201–234.
D. Jiang and D. Soudry, *On the genericity of cuspidal automorphic forms of $\mathrm{SO}(2n+1)$. II.* Compos. Math. **143** (2007), no. 3, 721–748.
M. Kashiwara and M. Vergne, *On the Segal-Shale-Weil representations and harmonic polynomials.* Invent. Math. **44** (1978), no. 1, 1–47.
A. W. Knapp and D. A. Vogan, Jr., *Cohomological induction and unitary representations.* Princeton Mathematical Series **45**, Princeton University Press, Princeton, NJ, 1995.
S. S. Kudla, *On the local theta-correspondence.* Invent. Math. **83** (1986), no. 2, 229–255.
S. S. Kudla and S. Rallis, *A regularized Siegel–Weil formula: the first term identity.* Ann. of Math. (2) **140** (1994), no. 1, 1–80.
R. P. Langlands, *On the classification of irreducible representations of real algebraic groups.* Representation theory and harmonic analysis on semisimple Lie groups, 101–170, Math. Surveys Monogr. **31**, Amer. Math. Soc., Providence, RI, 1989.
J.-S. Li, *Automorphic forms with degenerate Fourier coefficients.* Amer. J. Math. **119** (1997), no. 3, 523–578.
H. Y. Loke, J.-J. Ma, and U-L. Tang, *Transfers of $K$-types on local theta lifts of characters and unitary lowest weight modules.* Israel J. Math. **201** (2014), no. 1, 1–24.
C. Luo, *Endoscopic character identities for metaplectic groups.* arXiv:1801.10302
I. Matić, *The unitary dual of $p$-adic $\mathrm{SO}(5)$.* Proc. Amer. Math. Soc. **138** (2010), no. 2, 759–767.
C. M[œ]{}glin, *Quelques propriétés de base des séries théta.* J. Lie Theory **7** (1997), no. 2, 231–238.
C. M[œ]{}glin, *Sur certains paquets d’Arthur et involution d’Aubert–Schneider–Stuhler généralisée.* Represent. Theory **10** (2006), 86–129.
C. M[œ]{}glin, *Multiplicité $1$ dans les paquets d’Arthur aux places $p$-adiques.* On certain $L$-functions, 333–374, Clay Math. Proc. **13**, Amer. Math. Soc., Providence, RI, 2011.
C. M[œ]{}glin and D. Renard, *Paquets d’Arthur des groupes classiques complexes.* Around Langlands correspondences, 203–256, Contemp. Math. **691**, Amer. Math. Soc., Providence, RI, 2017.
C. M[œ]{}glin and D. Renard, *Sur les paquets d’Arthur des groupes unitaires et quelques conséquences pour les groupes classiques.* Pacific J. Math. **299** (2019), no. 1, 53–88.
C. M[œ]{}glin and D. Renard, *Sur les paquets d’Arthur des groupes classiques réels.* arXiv:1703.07226
C. M[œ]{}glin and D. Renard, *Sur les paquets d’Arthur aux places réelles, translation.* arXiv:1704.05096
T. Przebinda, *The duality correspondence of infinitesimal characters.* Colloq. Math. **70** (1996), no. 1, 93–102.
S. Rallis, *On the Howe duality conjecture.* Compos. Math. **51** (1984), no. 3, 333–399.
P. J. Sally, Jr. and M. Tadić, *Induced representations and classifications for $\mathrm{GSp}(2,F)$ and $\mathrm{Sp}(2,F)$.* Mém. Soc. Math. France (N.S.) **52** (1993), 75–133.
P. Scholze, *The local Langlands correspondence for $\mathrm{GL}_n$ over $p$-adic fields.* Invent. Math. **192** (2013), no. 3, 663–715.
B. Sun and C.-B. Zhu, *Conservation relations for local theta correspondence.* J. Amer. Math. Soc. **28** (2015), no. 4, 939–983.
D. A. Vogan, Jr., *Representations of real reductive Lie groups.* Progress in Mathematics **15**, Birkhäuser, Boston, MA, 1981.
D. A. Vogan, Jr., *The method of coadjoint orbits for real reductive groups.* Representation theory of Lie groups, 179–238, IAS/Park City Math. Ser. **8**, Amer. Math. Soc., Providence, RI, 2000.
J.-L. Waldspurger, *Correspondance de Shimura.* J. Math. Pures Appl. (9) **59** (1980), no. 1, 1–132.
J.-L. Waldspurger, *Correspondances de Shimura et quaternions.* Forum Math. **3** (1991), no. 3, 219–307.
J.-L. Waldspurger, *Démonstration d’une conjecture de dualité de Howe dans le cas $p$-adique, $p \neq 2$.* Festschrift in honor of I. I. Piatetski-Shapiro on the occasion of his sixtieth birthday, Part I, 267–324, Israel Math. Conf. Proc. **2**, Weizmann, Jerusalem, 1990.
C. Wu, *A critical case of Rallis inner product formula.* Sci. China Math. **60** (2017), no. 2, 201–222
B. Xu, *On M[œ]{}glin’s parametrization of Arthur packets for $p$-adic quasisplit $Sp(N)$ and $SO(N)$.* Canad. J. Math. **69** (2017), no. 4, 890–960.
S. Yamana, *$L$-functions and theta correspondence for classical groups.* Invent. Math. **196** (2014), no. 3, 651–732.
A. V. Zelevinsky, *Induced representations of reductive $\mathfrak{p}$-adic groups. II. On irreducible representations of $\mathrm{GL}(n)$.* Ann. Sci. École Norm. Sup. (4) **13** (1980), no. 2, 165–210.
*Atlas of Lie Groups and Representations.* <http://www.liegroups.org/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Within the framework of ImQMD05, we study several isospin sensitive observables, such as DR(n/p) ratios, isospin transport ratio (isospin diffusion), yield ratios for LCPs between the projectile region and mid-rapidity region for the reaction systems Ni+Ni, Zn+Zn, Sn+Sn at low-intermediate energies. Our results show that those observables are sensitive to the density dependence of symmetry energy, and also depend on the cluster formation mechanism. By comparing these calculations to the data, the information of the symmetry energy and reaction mechanism is obtained.'
address:
- '$^1$ China Institute of Atomic Energy, P.O. Box 275 (10), Beijing 102413, P.R. China'
- '$^2$ College of Physics and Technology,Guangxi Normal University, Guilin 541004,P.R. China'
- '$^3$ INFN, Laboratori Nazionali del Sud, Catania, Italy'
- '$^4$ National Superconducting Cyclotron Laboratory and Joint Institute of Nuclear Astrophysics, Michigan State University, East Lansing, MI 48824, USA'
author:
- 'Yingxun Zhang$^1$, Zhuxia Li$^1$, Chengshuang Zhou$^{1,2}$, Jixian Chen$^{1,2}$, M.Colonna$^3$, P.Danielewicz$^4$ and M.B.Tsang$^4$'
title: Probing the symmetry energy with isospin ratio from nucleons to fragments
---
Introduction
============
The nuclear symmetry energy plays an important role in the properties of nuclei and neutron stars [@Latti2001; @Latti2004; @Stein2005; @Anna2006; @Yakov2004]. To a good approximantion, it can be written as $$E_{sym}=S(\rho)\delta^2.$$ where $\delta=(\rho_n-\rho_p)/(\rho_n+\rho_p)$, is the isospin asymmetry; $\rho_{n}$, $\rho_{p}$, are the neutron, proton densities, and $S(\rho)$ describes the density dependence of the symmetry energy. Theoretical predictions for $S(\rho)$ from microscopic nucleon-nucleon interactions show large uncertainties, especially in the region of suprasaturation density [@Brown91; @BALi08]. Constraining the density dependence of the symmetry energy has become one of the main goals in nuclear physics and has stimulated many theoretical and experimental studies [@BALi08; @Danie02; @Fuch06; @Garg04; @HSXu00; @Tsang01; @Shett04; @Tsang04; @LWCh04; @qfli05; @qfli06; @TXLiu07; @BALi05; @Fami06; @BALi97; @BALi06; @BALi00; @BALi04; @Yong06; @Tsang09; @Gior10; @Napo10; @zhang05; @zhang08]. Heavy Ion Collisions (HIC) with asymmetric nuclei provide a unique opportunity for laboratory studies of the density dependence of the symmetry energy because a large range of densities can be momentarily achieved during HICs. In theoretical studies with transport models, the isospin ratio observables which are constructed from the isospin contents of emitted nucleons or fragments, such as Y(n)/Y(p) and DR(n/p) for emitted nucleons[@Fami06; @BALi97; @zhang08], isospin transport ratios $R_i$ constructed from the isospin asymmetry of projectile residues (or emitted resource)[@Tsang04; @LWCh04; @Tsang09; @zhang12], and $R^{mid}_{yield}$, constructed from the yields of LCP between the mid-rapidity and projectile region[@Kohley], have been prove to be primarily sensitive to the density dependence of the symmetry energy. By comparing the theoretical predictions to the experimental data, the sought-after constraints can be obtained.
One frequently utilized transport models to describe the heavy ion collisions is the Boltzmann-Uehling-Uhlenbeck (BUU) equation, which provides an approximate Wigner transform of the one-body density matrix as its solution[@Bertsch88]. $$\begin{aligned}
\frac{\partial f}{\partial t}+v \cdot \nabla_{\mathbf{r}}f-\nabla_{\mathbf{r}}U \cdot \nabla_{\mathbf{p}}f && =
-\frac{1}{(2\pi)^6}\int d^3p_2d^3p_{2'}d\Omega\frac{d\sigma}{d\Omega}v_{12}\nonumber\\
&&\times \{ [ff_2(1-f_{1'})(1-f_{2'})]-[f_{1'}f_{2'}(1-f)(1-f_{2})]\nonumber\\
&&\times (2\pi)^3\delta^{3}(\mathbf{p}+\mathbf{p}_2-\mathbf{p}_{1'}-\mathbf{p}_{2'})\} \label{buueq}\end{aligned}$$ The l.h.s. of this equation is the total differential of $f$ with respect to the time assuming a potential $U$. Usually a Skyrme-parametrization of the real part of the G-matrix or Skyrme-like energy density functional are employed as the nucleonic potential which describe the influence of the different isospin asymmetric nuclear equation of state (asy-EOS). Stochastic extensions of these mean-field based approaches have been introduced (see [@Chomaz04] and references therein). For instance, in the so-called Stochastic Mean Field (SMF) [@Baran02] model fluctuations are injected in coordinate space by agitating the spacial density profile. The r.h.s. of Eq. (\[buueq\]) contains a Boltzmann collision integral which describes the influence of binary hard-core collisions and are realized by the test particles.
Another frequently utilized approaches, known as the Molecular Dynamics Model (QMD)that represent the individual nucleons as Gaussian “wave-packet” with mean values that move in according the Ehrenfest theorem; i.e. Hamilton’s equations[@Aiche87]. $$\begin{aligned}
\dot{\mathbf{r}_i}=\frac{\mathbf{p}_i}{m}+\nabla_{\mathbf{p}_i}\sum_{j}\langle V_{ij} \rangle=\nabla_{\mathbf{p}_i}\sum_{j}\langle H \rangle\\
\dot{\mathbf{p}_i}=-\nabla_{\mathbf{r}_i}\sum_{j}\langle V_{ij} \rangle=-\nabla_{\mathbf{r}_i}\sum_{j}\langle H \rangle\end{aligned}$$ The expectation of the total Hamiltonian $<H>$ is obtained from the real part of the G-matrix or the Skyrme energy density functional, and it describes the influence of the different asy-EOS. The collision part in the QMD models are handled as same as the way in BUU type models but it is for nucleons rather than the test particles.
In this work, we choose to simulate nuclear collisions with the code ImQMD05 developed at the China Institute of Atomic Energy (CIAE), details of this code are described in Ref. [@zhang05; @zhang08; @zhang06; @zhang07], for studying several isospin ratio observables, such as DR(n/p), isospin transport ratios $R_i$, $R^{mid}_{yield}$ ratios for the yields of LCP between the mid-rapidity and projectile region and their relations to the fragmentation mechanism. For brevity, we limit our discussion here to the parameterization of the symmetry energy used in our calculations, which is of the form $$S(\rho)=\frac{1}{3}\frac{\hbar^2}{2m}\rho^{2/3}_{0}(\frac{3\pi^2}{2}\frac{\rho}{\rho_{0}})^{2/3}+\frac{C_{s}}{2}(\frac{\rho}{\rho_{0}})^{\gamma_{i}}.\label{srho}$$ where $m$ is the nucleon mass and the symmetry coefficient $C_s=35.19MeV$. Using this particular parameterization, the symmetry energy at subsaturation densities increases with decreasing $\gamma_i$, while the opposite is true for supranormal densities. In general, the EoS is labeled as stiff-asy for $\gamma_i>1$, and as soft-asy for $\gamma_i<1$. Finally, we also give a brief discussion the recent comparisons between the ImQMD05 and SMF calculations for further understanding the theoretical issue in the describing the reaction mechanism at low-intermediate energy heavy ion collisions.
isospin ratio from nucleon to fragments
=======================================
Isospin ratios which are constructed from the isospin contents of fragments, such as $R(n/p)=Y(n)/Y(p)$ (or named as n/p ratio), DR(n/p) from neutron-rich and neutron-poor systems, isospin transport ratios $R_i$ and $R^{mid}_{yield}=2 Y_{LCP}$($y^0<0.5$)/$Y_{LCP}$(0.5$<y^0<$1.5), are sensitive to the density dependence of symmetry energy. In this section, we will check their sensitivities to the density dependence of symmetry energy and try to get the information of symmetry energy from them.
n/p ratio and DR(n/p) ratio
---------------------------
The neutron to proton ratio $R_{n/p}=Y(n)/Y(p)$ of pre-equilibrium emitted neutron over proton spectra was considered as a sensitive observable to the density dependence of symmetry energy[@BALi97], because it has a straightforward link to the symmetry energy. In order to reduce the sensitivity to uncertainties in the neutron detection efficiencies and sensitivity to relative uncertainties in energy calibrations of neutrons and protons, the double ratio $$DR(n/p)=R_{n/p}(A)/R_{n/p}(B)$$ had been measured by Famiano and compared with the transport model prediction[@Fami06; @BALi97].
We performed calculations of collisions at an impact parameter of $b=2 fm$ at an incident energy of $50 MeV$ per nucleon for two systems: $A=^{124}Sn+^{124}Sn$ and $B=^{112}Sn+^{112}Sn$ with ImQMD05 to study the $DR(n/p)$ ratio for emitted nucleons[@zhang08]. The shaded regions in the left panel of Fig.\[fig-dr\] show the range, determined by uncertainties in the simulations, of predicted double ratios $DR(n/p)=R_{n/p}(A)/
R_{n/p}(B)$ of the nucleons emitted between $70^{\circ}$ and $110^{\circ}$ in the center of mass frame as a function of the center of mass nucleon energy, for $\gamma_{i}=0.5$ and $2.0$. The double ratios $DR(n/p)$ are higher for the EOS with the weaker symmetry energy density dependence $\gamma_{i}=0.5$ than that for $\gamma_{i}=2.0$ because the nucleons mainly emit from the lower density region at intermediate energy HICs. Compare to the data on $DR(n/p)$ for emitted nucleons(solid stars), the general trend of data $DR(n/p)$ are qualitatively reproduced and the data seem to be closer to the calculation employing the EOS with $\gamma_{i} =0.5$. The right panel of fig.\[fig-dr\] show the coalescence-invariant double ratio. The coalescence-invariant double ratios are constructed by including all neutrons and protons emitted at a given velocity, regardless of whether they are emitted free or within a cluster. The data are shown as open stars and the calculation results are shown as shaded region. Here, the measurement and simulation results illustrate that the fragments with $Z\ge2$ mainly contribute to the low energy spectra and do not affect the high-energy $DR(n/p)$ data very much.
 (Left) DR(n/p) ratios for emitted free nucleons and (Right) coalescent-invariant $DR(n/p)$ from the ImQMD simulations are plotted as shadow region.](fig-dr.eps){width="40.00000%"}
In order to constrain the range of $\gamma_{i}$ from the $DR(n/p)$ data that had been published, a series calculations for two systems, $A=^{124}Sn+^{124}Sn$ and $B=^{112}Sn+^{112}Sn$, have been performed by varying $\gamma_{i}=0.35, 0.5, 0.75, 1.0$ and $2.0$[@Tsang09]. Since the emitted nucleons are mainly from the subnormal densities at this energies, the n/p ratios of emitted nucleons are associated with the values of symmetry energy at subnormal density. Therefor, the $DR(n/p)$ ratio should increase with decreasing $\gamma_{i}$. However, in the limit of very small $\gamma_{i}\ll0.35$, the finite system completely disintegrates and the $DR(n/p)$ ratio decrease and approaches the limit of reaction system, $(N/Z)_{124}/(N/Z)_{112}=1.2$. As a consequence of these two competing effects, the double ratio values peak around the $\gamma_{i}=0.7$. Despite the large experiment uncertainties for higher energy data, those comparisons definitely rule out very soft ($\gamma_{i}=0.35$) and very stiff ($\gamma_{i}=2.0$) density dependence of symmetry energy. The $\chi^2$ analysis suggest that within a $2\sigma$ uncertainty, parameters of $\gamma_{i}$ fall in the range of $0.4\le\gamma_{i}\le1.05$ for the $C_{s}=35.2MeV$.
isospin transport ratio
-----------------------
When the projectile and target nuclei come into contact, there can be exchange of nucleons between them. If the neutron to proton ratios of the projectile and target differ greatly, the net nucleon flux can cause a diffusion of the asymmetry $\delta$ reducing the difference between the asymmetries of two nuclei. This isospin diffusion process, which depends on the magnitude of the symmetry energy, affects the isospin asymmetry of the projectile and target residues in peripheral HICs. The isospin transport ratio $R_i$ has been introduced [@Tsang04] to quantify the isospin diffusion effects, $$R_i=\frac{2X-X_{aa}-X_{bb}}{X_{aa}-X_{bb}}, \label{Ridef}$$ where X is an isospin observable and the subscripts $a$ and $b$ represent the neutron rich and neutron-poor nuclei. In this work, we use $a$ and $b$ to denote the projectile (first index) and target (second index) combination. where $\mathrm{a=^{124}Sn}$, and $\mathrm{b=^{112}Sn}$. We obtain the value of $R_{i}$ by comparing three reaction systems, $\mathrm{a+a}$, $\mathrm{b+b}$ and $\mathrm{a+b}$ (or $\mathrm{b+a}$). Construction of the transport ratio minimize the influence of other effects besides isospin diffusion effects on the fragment yields, such as preequilibrium emission and secondary decay, by rescaling the observable X for the asymmetric a+b system by its values for the neutron-rich and neutron-deficient symmetric systems, which do not experience isospin diffusion. Based on Eq. (\[Ridef\]), one expects $R_i=\pm1$ in the absence of isospin diffusion and $R_i\sim0$ if isospin equilibrium is achieved. Eq. (\[Ridef\]) also dictates that two different observables, $\mathrm{X}$, will give the same results if they are linearly related. In one experiment, $\mathrm{X}$ was taken as the isoscaling parameter, $\alpha$, obtained from the yield of the light particles near the projectile rapidity[@Tsang01], to measure the isospin diffusion ability in heavy ion collisions. In transport models [@Tsang04; @LWCh04], the isospin asymmetry $\delta$ of the projectile residues (emitting source) has been used to compute $R_i(\delta)$ because it is linearly related to the isoscaling parameters $\alpha$[@TXLiu07; @Tsang01; @Ono03].
We analyze the amount of isospin diffusion with ImQMD05 by constructing a tracer from the isospin asymmetry of all emitted nucleons (N) and fragments (frag), including the heavy residue if it exists, with velocity cut $v^{N,frag}_z>0.5v^{c.m.}_{beam}$ (nearly identical results are obtained with higher velocity cut $v^{N,frag}_z> 0.7v^{c.m.}_{beam}$). This represents the full projectile-like emitting source, and should be comparable to what has been measured in experiments. Fig.2 shows the results of isospin transport ratios $R_i(X=\delta_{N,frag})$ (upright triangles) as a function of the impact parameter for a soft symmetry case ($\gamma_i=0.5$, open symbols) and a stiff symmetry case ($\gamma_i=2.0$, closed symbols). $R_i$ obtained with soft-symmetry case is smaller than those obtained with stiff-symmetry potential case. This is consistent with the expectation that higher symmetry energy at subnormal density leads to larger isospin diffusion effects (smaller $R_i$ values).
{width="40.00000%"}
$R_i$ depends weakly on impact parameter over a range extending from central ($b=3fm$) to mid peripheral($b=8fm$) collisions. Interestingly, the isospin equilibrium and global thermal equilibrium are not reached even for central collisions. Our results show, that neither the effective interaction is sufficiently strong nor the collisions are sufficiently frequent (most of them are Pauli suppressed) to mix the projectile and target nucleons completely. These two effects prevent the combined system from attaining isospin equilibrium even in central collisions. With impact parameter increasing for $b>5fm$, the overlap region and thus the number of nucleons transferred from projectile and target decreases, causing the $R_i$ values to increase.
In peripheral collisions, most often, a large residue remains. If it decouples from the full emitting source before it equilibrates, it may experience a different amount of diffusion than the full emitting source examined by $X=\delta_{N,frag}$. To examine this, we constructed a tracer using the isospin asymmetry of the heaviest fragments with charge $Z_{max}>20$ in the projectile region. This tracer is mainly relevant to peripheral collisions as the central collisions are dominated by multifragmentation and very few large projectile fragments survive. The dependence of $R_i(X=\delta_{Z_{max}>20})$ for impact parameter $b\geq5fm$ is shown as open and closed circles in Fig. 2. The isospin transport ratios constructed from the different isospin tracers have different values especially in the case of $\gamma_i=0.5$. Stronger isospin equilibration (smaller $R_i$ values) is observed in the isospin transport ratios $R_i(X=\delta_{N,frag})$ constructed from nucleons and fragments than $R_i(X=\delta_{Z_{max}>20})$ constructed from the heaviest fragments with $Z_{max} > 20$. Since isospin diffusion mainly occurs through the low-density neck region, and the system breaks up before isospin equilibrium, the asymmetry of the projectile and target residues do not achieve equilibrium and, larger $R_i(X=\delta_{Z_{max}\ge20})$ values result. In contrast, there is more mixing of nucleons from the target and projectile in the neck region due to the isospin diffusion. Consequently, rupture of the neutron-rich neck is predicted to result in the production of neutron-rich fragments at mid rapidity.
Since fragments are formed at all rapidities, we can examine the rapidity dependence of $R_i$ to obtain more information about the reaction dynamics. Fig. 3 shows $R_i$ as a function of the scaled rapidity $y/y_{beam}$. The symbols in the leftmost panel are experimental data obtained in Ref. [@TXLiu07] for three centrality gates. This transport ratio was generated using the isospin tracer $X=ln(Y(^7Li)/Y(^7Be))$ where $Y(^7Li)/Y(^7Be)$ is the yield ratio of the mirror nuclei, $^7Li$ and $^7Be$ [@TXLiu07]. As expected the values of $R_i$ obtained from peripheral collisions (solid stars) are larger than those obtained in central collisions (open stars). For comparison, the ImQMD05 calculations of $R_i(X=\delta_{N,frag})$ are plotted as lines in the middle and right panels for a range of impact parameters. The middle panel contains the results from the soft symmetry potential ($\gamma_i=0.5$) while the right panel shows the results from the stiff symmetry potential ($\gamma_i=2.0$). The impact parameter trends and magnitude of the data are more similar to the results of the calculations from soft symmetry potentials ($\gamma_i=0.5$) for peripheral collisions.
We have performed the $\chi^2$ analysis for both observables, $R_i$ and $R_i(y)$, for constraining the density dependence of symmetry energy. Using the same 2$\sigma$ criterion, the analysis brackets the regions $0.45\le\gamma_i\le0.95$ is obtained. It is consistent with previous analysis on DR(n/p). However, the experimental trend of $R_i$ gated on the most central collisions (open stars) is not reproduced by the calculations. The experimental data indicate more equilibration for central collisions near mid rapidity while the transport model indicates more transparency. The equilibration in the E/A = 50MeV data may be the result of the impact parameter determination from charged particle multiplicity wherein the most central collisions are assumed to be the ones with highest charge particle multiplicity. For the most central events, a gate on the highest multiplicity, may select events in which more nucleon-nucleon collisions occur rather than a strict selection on the most central impact parameters.
![(Color online) (Left panel) Experimental $R_i$ as a function of rapidity for three centrality gates \[16\]. (Middle panel) The calculated results of $R_i(X=\delta_{N,frag})$ as a function of rapidity for $b= 2, 4, 6, 8 fm$ for $\gamma_i=0.5$ and (Right panel) $\gamma_i=2.0$.](fig8_Ri_rap_b.ps){width="45.00000%"}
$R^{mid}_{yield}$ ratios for light charged particles
-----------------------------------------------------
The yield ratios, $R^{mid}_{yield}$, for LCPs between the projectile region and mid-rapidity region, are defined as $$R^{mid}_{yield}=\frac{2\cdot Yield(0.0\leq Y_r \leq0.5)}{Yield(0.5\leq Y_r \leq1.5)},$$ where $Y_r=\frac{Y_{c.m.}}{Y^{c.m.}_{proj}}$ is the reduced rapidity. It reflect the isospin migration ability and have been measured by *Kohley.et.al.* for $\mathrm{^{70}Zn+^{70}Zn}$, $\mathrm{^{64}Zn+^{64}Zn}$, $\mathrm{^{64}Ni+^{64}Ni}$ at the beam energy of 35 MeV/nucleon for middle peripheral collisions [@Kohley]. The data show a clear preference for emission around the mid-rapidity region for more neutron-rich LCPs resulting from the isospin migration mechanism through the neck region between the projectile and target[@Baran04; @zhang05]. Theoretical study by SMF model[@Rizzo08] demonstrates that $R^{mid}_{yield}$ is sensitive to the density dependence of symmetry energy. The experimental trends are reproduced by SMF model. However, there are largest discrepancies on the reduced rapidity distribution for the yields of proton and $^3He$, and also on their values of $R^{mid}_{yield}$ as mentioned in Reference[@Kohley]. The discrepancies were thought to be related to the statistical decay of QP at later stages of the reaction [@Kohley]. In the points of reaction dynamics, the different fragmentation mechanism in the transport models simulations also lead to different behaviors of the rapidity distribution for LCPs besides the effects from secondary decay. Thus, it is instructive to study the rapidity distribution of LCP with ImQMD05.
In Fig.\[ref-Fig.3\](a), we present the multiplicity distribution for fragments with $Z\geq 3$ for $\mathrm{^{70}Zn+^{70}Zn}$ at $E_{beam}$=35 MeV/nucleon and impact parameter b=4fm and $\gamma_i=2.0$. We find that **half** of events belongs to multi-fragmentation process which are defined by multiplicity for fragments with charge $Z>3$, i.e.,$M(Z\geq 3)>3$. The rest are the binary ($M(Z\geq 3)=2$) and ternary ($M(Z\geq 3)=3$) fragmentation events. It suggests that the binary, ternary fragmentation and multi-fragmentation coexist around 35MeV/nucleon. In Fig.\[ref-Fig.3\] (b) and (c), we plot the reduced rapidity distribution for the yields of $^3He$ and $^6He$ obtained with three kinds of fragmentation process, binary (square symbols), ternary (circle symbols) and multi-fragmentation (triangle symbols) which are selected by $M(Z\geq 3)=2, 3$ and $>3$. The yields of $^3He$ and $^6He$ in Fig.\[ref-Fig.3\] are normalized to per event. It is clear that the binary events produce more $^3He$ and $^6He$ at mid-rapidity relative to that produce in multi-fragmentation events. For $\gamma_i=2.0$ case, the yield of $^3He$ at $Y_r=0$ obtained with binary fragmentation events is 35% larger than that with multi-fragmentation events. For neutron-rich LCP, for example, the yield of $^6He$ at $Y_r=0$ obtained in binary fragmentation events is 70% larger than that in multi-fragmentation events due to isospin migration.
 (a) The multiplicity distribution for fragments with $Z\geq 3$ ($M(Z\geq 3)$). (b) is the reduced rapidity ($Y_r$) distribution for the yield of $^3He$ with binary (square symbols), ternary (circle symbols) and multi-fragmentation (triangle symbols) process. (c) is for $^6He$. All of those results are for $\mathrm{^{70}Zn+^{70}Zn}$ at E=35 MeV/u for b=4fm and $\gamma_i=2.0$.](pm-the6.eps)
Fig.\[ref-Fig.4\] shows the calculated results for the rapidity distribution of light charged particles *p*, *d*, *t*, $^{3}He$, $^{4}He$ and $^{6}He$ for $^{64}Ni+^{64}Ni$ at b=4fm with 100,000 events. The distributions are normalized with the yield at $Y_r=0$ for comparing with data in Ref [@Kohley] to understand the fragmentation mechanism, because the fragmentation mechanism mainly determine the shape of the rapidity distribution for LCP. The open circles are for $\gamma_i=0.5$ and solid symbols are for $\gamma_i=2.0$. The data are taken from the Ref [@Kohley] and plotted as stars. Both our calculations and data show that the width of distribution decreases with the mass of LCPs increasing. For the rapidity distributions of $^3H$ and $^3He$, the width of distribution for $^3H$ is smaller than that for $^3He$ due to the isospin migration. By comparing the simulated results to the data, the ImQMD05 calculations with stiffer symmetry energy can well reproduce the data at forward rapidity region ($Y_r>0$) for all *p*, *d*, *t*, $^3He$, $^4He$ and $^6He$. For the backward rapidity region ($Y_r<0$), there are obvious differences between the results from ImQMD05 calculations and the data because the efficiency for detection of LCPs at the backward [@Kohley] are not included in this ImQMD05 calculations.
In order to constrain the symmetry energy by the rapidity distribution of LCPs, we further calculate $R_{yield}^{mid}$ in a series of $\gamma_i=0.5, 0.75, 1.0, 2.0$. In Fig.\[ref-Fig.5\], we present the results of $R^{mid}_{yield}$ as a function of the AZ of emitted particles for three reaction systems $\mathrm{^{64}Zn+^{64}Zn}$, $\mathrm{^{64}Ni+^{64}Ni}$ and $\mathrm{^{70}Zn+^{70}Zn}$ at b=4fm. The open symbols are the results for $\gamma_i=0.5, 0.75, 1.0$ and $2.0$. The solid stars are the data from [@Kohley]. Since the isospin migration occurs in the neutron-rich neck region, the $R^{mid}_{yield}$ shows an increasing trend with the values of isospin asymmetry of LCPs increasing for the same element. The more neutron-rich the LCPs is, the larger the $R^{mid}_{yield}$ is. Furthermore, the calculated results show the values of $R^{mid}_{yield}$ for neutron-rich isotopes are sensitive to the density dependence of symmetry energy. The calculations with stiffer symmetry energy predict larger values of $R^{mid}_{yield}$ due to the stronger isospin migration effects. This conclusion is as same as the results obtained with SMF model [@Kohley]. As shown in Fig.\[ref-Fig.5\], the ImQMD05 calculations with stiffer symmetry energy ($\gamma_i\ge0.75$) can reasonably reproduce the data of $R^{mid}_{yield}$ as a function of AZ for $\mathrm{^{64}Zn+^{64}Zn}$. But our calculations underestimate the $R^{mid}_{yield}$ values of neutron-rich light charged particles, such as $^6He$, for the neutron-rich reaction systems $\mathrm{^{64}Ni+^{64}Ni}$ and $\mathrm{^{70}Zn+^{70}Zn}$. It could come from the lacking of fine structure effects of neutron-rich elements (such as neutron-skin, stability of lighter neutron-rich elements), and the impact parameter smearing effects in the transport model simulations. Even though our calculations can reproduce the $R^{mid}_{yield}$ data for $^{64}Zn+^{64}Zn$, the definitely constraints on the symmetry energy with the data of $R^{mid}_{yield}$ can not be obtained before we fix the problems on the theoretical predictions of $R^{mid}_{yield}$ for neutron-rich reaction system.
$R_{yield}^{mid}$ values as a function of the charge times mass (ZA) for *p* (ZA=1), *d* (ZA=2), *t* (ZA=3), $^3He$ (ZA=6), $^4He$ (ZA=8), $^6He$ (ZA=12). The open symbols are the results obtained with ImQMD05 for $\gamma_i=0.5, 0.75, 1.0$ and $2.0$. The solid stars are the data from [@Kohley].](Fig.5.EPS)
remarks on the comparisons between ImQMD05 and SMF
--------------------------------------------------
Up to date, there are several constraints on the symmetry energy with isospin sensitive observables, DR(n/p) raitos, isospin transport ratios $R_i$ and $R^{mid}_{yield}$ ratios, by adopting different type of transport models, such as QMD type and BUU type[@Tsang09; @LWChen05; @Rizzo08] . There are overlap between the results of symmetry energy from different approaches, but they are different in detail. Thus, further understanding the issues in the QMD type and BUU type would be crucial in theoretical studies for improving the constraints on the symmetry energy.
At the code level, both BUU and QMD models propagate particles classically under the influence of a mean field potential, which is calculated self-consistently the positions and momenta of the particles, and allow scattering by nucleon-nucleon collisions due to the residual interaction. The Pauli principle in both approaches is enforced by application of Pauli blocking factors. These similarities in implementation have lead to similarities in predictions for many collision observables [@Aich89].
There are also significant differences in these approaches. In the BUU equations, each nucleon is represented by 200-1000 test particels that generate the mean field and suffer the collisions. In QMD, there is one test particle per nucleon. A-body correlations and cluster formation are not native to the original BUU approach; which is supposed to provide the Wigner transform of the one body density matrix. On the other hand, many-body correlations and fluctuations can arise from the A-body dynamics of QMD approach. Such A-body correlations are suppressed in BUU approach, but correlations can arise in both approaches from the amplification of mean field instabilities in spinodal region [@Chomaz04]. Collision algorithms in the QMD approach modify the momenta of individual nucleons, while in BUU approach, only the momenta of test particles are modified. Depending on the details of the in-medium cross sections that are implemented, the blocking of collisions can also be more restrictive for QMD than for BUU, leading to fewer collisions and therefore a greater transparency.
Fragments can be formed in QMD approaches due to the A-body correlations and these correlations are mapped onto the asymptotic final fragments by a minimum spanning tree algorithm. Serval different methods have been developed to allow BUU codes to calculate cluster production. In the Stochastic Mean Field (SMF) approach (one of the BUU type model), the time evolution of the one-body phase-space distribution $f$ is governed by the nuclear mean-field, two-body scattering, and a fluctuating (stochastic) term which causes the fragmentation [@Baran02; @Baran04; @Rizzo08; @Colonna98]. Since there are typically more than 100 test particles per nucleon, collision induced fluctuations are smaller in BUU than in QMD possibly suppressing the fragment formation rates.
As an example, we present the results of average charge number for $Z\ge3$ as a function of rapidity obtained with ImQMD05 and SMF models for $^{124}Sn+^{124}Sn$ at $E_{beam}=50AMeV$ and b=6, 8fm in Fig.\[ref-zhangsmf\]. The solid lines are the results from ImQMD05, and dashed lines are from SMF calculations[@zhmari]. Squares are for b=6fm, and circles are for b=8fm. The results obtained with QMD simulations show that the average charge number of $Z\ge3$ increase with rapidity increasing before $y/y^{c.m.}_{beam}\sim1$ in the forward region in the QMD simulations, and two peaks appear around the projectile and target rapidity region, respectively. In the SMF simulations, the peak appears at mid-rapidity besides two peaks around projectile and target rapidity. It clearly shown that the strict Pauli blocking in the QMD models simulations leads to a greater transparency than that in the SMF type simulations. The larger fluctuation in the QMD models lead to more fragments and light particles emitted than that from SMF predictions. As results, the fragments distributed over the whole rapidity region and heavier fragment is, larger velocity is. In SMF calculations, the fragmentation process is less effective due to the reduced amplitude of fluctuations and many-body correlations. This enhances the appearance of binary and ternary processes in semi-peripheral heavy ion reactions, according to the prominent role of the mean-field dynamics. As a consequence, the intermediate mass fragments tend to distribute at mid-rapidity. It also lead that the the average charge number of $Z\ge3$ have a peak around mid-rapidity and narrower rapidity distribution of lighter clusters than the results from QMD models.
 The average charge number for $Z\ge3$ as a function of rapidity for $^{124}Sn+^{124}Sn$ at b=6,8fm with $\gamma_i=2.0$. The solid lines are the results from ImQMD05, and dashed lines are from SMF.](smf-imqmd.eps){width="35.00000%"}
Summary
=======
In summary, we have investigated the influences of the density dependence of the symmetry energy on several different isospin ratio observables, such as DR(n/p) ratio, isospin transport ratios $R_i$, the rapidity dependence of isospin transport ratio $R_i(y)$ and $R^{mid}_{yield}$ raitos (the yield ratios of LCP between the mid-rapidity and projectile region) with ImQMD05. The study shows that these isospin ratio observables are sensitive to the density dependence of symmetry energy. This conclusion is similar to conclusions reached using BUU approaches in the range of symmetry energies studied here. By comparing the calculated results to data, the very soft ($\gamma_i=0.35$) and very stiff symmetry energy ($\gamma_i=2.0$) are ruled out.
Cluster formation is important for intermediate energy heavy ion collisions, and it modifies the spectral double ratios at $E_{c.m.}<
40 MeV$. We also tested different tracers by constructing corresponding isospin transport ratios for them using different symmetry energies. For weakly density dependent symmetry energies (small $\gamma_i$) with large symmetry energies at sub-saturation densities, the values of $R_i$ for the residue tracer $X=\delta_{Z_{max}>20}$ are larger than those extracted from the entire emitting source, i.e., $X=\delta_{N,frag}$. The difference between these two tracers can be examined experimentally as a new probe of the symmetry energy and reaction mechanism.
By studying reaction systems $\mathrm{^{64}Zn+^{64}Zn}$, $\mathrm{^{64}Ni+^{64}Ni}$ and $\mathrm{^{70}Zn+^{70}Zn}$ at the beam energy of 35 MeV per nucleon and b=4fm within the framework of ImQMD05, we find that half of events belongs to the multi-fragmentation mechanism, and half of them is of binary and ternary fragmentation events. The binary and ternary events produce more light charged particles at middle rapidity, and the multi-fragmentation events broaden the reduced rapidity distribution for the yields of LCPs. Both the data and our calculations illustrate that the reaction systems seems more transparency and more fragments, light particles emitted. As results, the data of the reduced rapidity distribution for the yields of LCPs and $R_{yield}^{mid}$ as a function of AZ for $^{64}Ni+^{64}Ni$ can be well reproduced by the ImQMD05 calculations. For neutron rich reaction systems $\mathrm{^{64}Ni+^{64}Ni}$ and $\mathrm{^{70}Zn+^{70}Zn}$, our calculations underestimate the $R^{mid}_{yield}$ values of neutron-rich light charged particles, such as $^6He$, it could be cause by the lacking of fine structure effects for lighter elements in the transport models, and the impact parameter smearing effects.
Acknowledgments
===============
This work has been supported by the Chinese National Science Foundation under Grants 11075215, 10979023, 10875031, 11005022,11005155, 10235030, and the national basic research program of China No. 2007CB209900. We wish to acknowledge the support of the National Science Foundation Grants No. PHY-0606007.
References {#references .unnumbered}
==========
[9]{}
J. Lattimer and M. Prakash, Ap. J. 550, 426 (2001). J. Lattimer and M. Prakash, Science 304, 536 (2001). A. Steiner et al., Phys. Rep. 411, 325 (2005). A. L. Watts and T. E. Strohmayer, Ap. J. 637, L117 (2006). D. G. Yakovlev and C. J. Pethick, Annu. Rev. Astron. Astrophys. 42, 169 (2004). B. A. Brown, Phys. Rev. C 43, R1513 (1991). B. Li et al., Phys. Rep 464, 113 (2008). P. Danielewicz, R. Lacey, and W. G.Lynch, Science 298, 1592 (2006). C. Fuch and H. Wolter, Eur. Phys. J. A 30, 5 (2006). U. Garg, Nucl. Phys. A 731, 3 (2004). H. S. Xu et al., Phys. Rev. Lett. 85, 716 (2000). M. B. Tsang et al., Phys. Rev. Lett. 86, 5023 (2001). D. V. Shetty et al., Phys. Rev. C 70, 011601 (2004). M. B. Tsang et al., Phys. Rev. Lett. 92, 062701 (2004). L.-W. Chen, C. M. Ko, and B.-A. Li, Phys. Rev. Lett. 94, 032701 (2005). Q. Li et al., Phys. Rev. C 72, 034613 (2005). Q. Li et al., Phys. Rev. C 73, 051601 (2006). T. X. Liu et al., Phys. Rev. C 76, 034603 (2007). B.-A. Li and L.-W. Chen, Phys. Rev. C 72, 064611 (2005). M. A. Famiano et al., Phys. Rev. Lett. 97, 052701 (2006). B. Li, C. Ko, and Z. Ren, Phys. Rev. Lett. 78, 1644 (1997). B. Li, L. W. Chen, G. C. Yong, and W. Zuo, Phys. Lett. B 634, 378 (2006). B. Li, Phys. Rev. Lett. 85, 4221 (2000). B. Li, Nucl. Phys. A 734, 539c (2004). G. Yong, B. Li, and L. Chen, Phys. Rev. C 73, 034603 (2006). M. B. Tsang, Y. Zhang, P. Danielewicz, M. Famiano, Z. Li, W. G. Lynch, and A. W. Steiner, Phys. Rev. Lett. 102, 122701 (2009). B. Giordano et al., Phys. Rev. C 81, 044611 (2010). P. Napolitani et al., Phys. Rev. C. 81, 044619 (2010). Y. Zhang and Z. Li, Phys. Rev. C 71, 024604 (2005). Y. Zhang, P. Danielewicz, M. Famiano, Z. Li, W. G. Lynch, and M. B. Tsang, Phys. Lett. B 664, 145 (2008). Yingxun Zhang, D.D.S.Coupland, P.Danielewicz, et.al., Phys.Rev.C85, 024602(2012). Z. Kohley, L.W. May, S. Wuenschel, et.al., Phys.Rev.C83, 044601(2011).
G.F. Bertsch, S. Das Gupta, Phys.Rep. 4, 189(1988). P.Chomaz, M.Colonna, J.Randrup, Phys.Rep. 389, 263 (2004). V. Baran et. al, Nucl. Phys. A 703, 603 (2002). J. Aichelin, A. Rosenhauer, G. Peilert, H. Stocker, and W. Greiner, Phys. Rev. Lett. 58, 1926 (1987). Y. Zhang and Z. Li, Phys. Rev. C 74, 014602 (2006). Y. Zhang, Z. Li, and P. Danielewicz, Phys. Rev. C 75, 034615 (2007). A. Ono, P. Danielewicz, W. A. Friedman, W. G. Lynch, and M. B. Tsang, Phys. Rev. C 68, 051601(R) (2003).
V. Baran et al, Nucl. Phys. A 730, 329 (2004) J. Rizzo et al, Nucl. Phys. A 806, 79 (2008). in priviate communication
L. W. Chen, C. M. Ko, B. A. Li, Phys. Rev. C 72, 064309 (2005). M. Colonna et al., Nucl. Phys. A 642, 449 (1998). J.Aichelin, C.Hartnack, A. Bohnet, Zhuxia Li, G.Peilert, H.Stocker and W. Greiner, Phys.Lett. B 224, 34(1989).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We model how the mid-infrared colors of Young Stellar Objects (YSOs) vary with stellar temperature. The spectral energy distribution (SED) of each object has contributions from thermal emission of circumstellar dust, from direct stellar photospheric emission, and from scattered stellar emission. We first isolate the effects of stellar contributions (direct+scattered) to the SED using homologous “Class I” models: the distribution of circumstellar matter is chosen to scale with stellar temperature $T_\star$ such that the shape of the thermal contribution to the SED remains constant. The relative contribution of stellar direct and scattered light varies with $T_\star$, changing the $1-10 \mu$m (mid-infrared; MIR) colors. Stellar light contributes more to the MIR emission of YSOs with lower temperature stars ($T_\star \sim$ 4000 K) because the emission peak wavelength of the star is closer to that of the thermal radiation. In YSOs with hotter central stars, since the peak of the stellar and thermal spectra are more separated in wavelength, the $1-10 \mu$m spectrum is closer to a pure thermal spectrum and the objects are redder.
Next we consider realistic Class 0, I, and II source models and find that the other dominant effect of varying stellar temperature on YSO SEDs is that of the inner disk wall: In high-$T_\star$ models, the dust destruction radius is much further out with a consequently larger inner disk wall that contributes relatively more to the $2-10\mu$m flux. This effect partially offsets that of the stellar contribution leading to varying behaviors of the $2-10\mu$m flux: In Class 0 sources, the trend is for higher $T_\star$ models to have redder colors. In Class I sources, the trend applies with some exceptions. In Class II sources, $2-10\mu$m colors become redder going from $T_\star=4000$ to 8000 due to decreasing stellar contribution at $T_\star = 8000 K$, and then become blue again from 8000 to 31500 K due to increasing inner disk wall contribution. Near edge-on inclinations, the color behavior is completely different.
Our modeled MIR protostellar colors have implications for interpretations of [*Spitzer* ]{} IRAC observations of star formation regions: It is commonly assumed that the slope of the SED at $1-10\mu$m is directly related to evolutionary state. We show that inclination effects, aperture size, scattered light, and stellar temperature cause a broad spread in the colors of a source at a single evolutionary state. Color-magnitude diagrams can help sort out these effects by separating sources with different $T_\star$ based on their different brightness (for sources at the same distance).
author:
- 'Barbara. A. Whitney, Rémy Indebetouw, J. E. Bjorkman, & Kenneth Wood,'
title: '2-D Models of Protostars: III. Effects of Stellar Temperature'
---
Introduction
============
With the launch and successful operation of the [*Spitzer Space Telescope*]{} [@werner04], there is a wealth of mid-infrared data being collected on star formation regions near and far [e.g. @allen04; @megeath04; @reach04; @whitney04]. The IRAC camera [@fazio04] provides unprecedented sensitivity and mapping speed using four filters centered on 3.6, 4.5, 5.8, and 8 $\mu$m, and the MIPS camera [@rieke04] is functioning particularly well in the 24$\mu$m band [e.g. @muzerolle04]. Furthermore, many current ground-based facilities and recent space based facilities (e.g. ISO) are optimized to collect data in the near/mid-IR. Thus, in many cases, scientific analysis of star forming clusters is based only on broad-band colors in the 1-30$\mu$m range.
Traditionally the slope of the SED in this region, parameterized by the spectral index $\alpha$ ($=d \log \lambda F_\lambda / d \log \lambda$) is used to classify evolutionary state [@lada87; @lada99]. A sequence of evolutionary classes is fairly well-defined for protostars of moderate mass ($M<1-2M_\sun$). Sources for which $\alpha$ is positive are thought to have large (thousands of AU) infalling envelopes and are classified as Class 0 to Class I. Class 0 sources are very young ($10^4$ yrs) protostars (André et al. 1993) with high infall rates and very collimated outflows. Class I sources are older ($10^5$ yrs) with lower infall rates and larger bipolar cavities carved by meandering jets and molecular outflows (Gómez et al. 1997; Richer et al. 2000; Reipurth et al. 2000). Sources with $\alpha<0$ are identified as Class II sources, pre-main sequence stars surrounded by flared accretion disks. Obviously there are intermediate stages of evolution between Class I and II and it is commonly thought that the spectral index continuum is also a measure of the evolutionary continuum between Class 0 and II [@kenyon95].
Contrasting with the well-defined class system for low-mass stars, the evolution of high-mass protostars is poorly understood. Disk accretion is more difficult to model theoretically in the case of a high-mass protostar because of strong radiation pressure from the central source, high accretion rates, and large disk masses leading to instabilities. Some theorists have suggested protostellar coalescence as an alternative formation mechanism [@bonnell], but others have succeeded in producing viable accretion models [@behrend; @maeder]. The large volume of mid-infrared data currently being collected in protostars of all masses makes it important to understand the infrared properties of high-mass protostars. In particular, does the slope of the SED relate to evolutionary state, as for low-mass stars, and what other effects are present that should be taken into account when analyzing the data?
This paper is part of a series on using 2-D radiative transfer models to interpret data of Young Stellar Objects (YSOs). In previous papers (Whitney et al. 2003a,b; Papers I & II), we showed that inclination effects can blur the separation of evolutionary states in mid-IR color-color diagrams of low-mass YSOs. As an example, in edge-on Class I sources, the mid-IR flux is dominated by scattered light due to the large extinction in the disk blocking all stellar and inner disk radiation; this leads to mid-IR colors that are bluer than Class II sources. In contrast, pole-on Class I sources are less red than average due to lower extinctions in their partially-evacuated bipolar cavities. Color-magnitude diagrams and near-IR polarization measurements can help sort out this blending since both flux and polarization vary with inclination (e.g., edge-on Class I and II sources will be blue, faint and have high near-IR polarization).
In this paper, we investigate the variation in mid-IR colors of YSOs due to the temperature ($T_\star$), or mass, of the central star. We assume that disk accretion occurs for stars of any mass, with physical properties that scale with mass and agree with observations. We find that varying the stellar temperature has two competing effects on the MIR colors of YSOs. The first is the relative contribution of stellar scattered+direct flux to the 1-10 $\mu$m flux for sources with differing $T_\star$. We demonstrate this effect by showing a set of homologous “Class I” models in §2.2 in which the thermal spectral shapes are very similar between models with four different stellar temperatures. The disk geometries in these high-$T_\star$ models are unrealistic because they are chosen to scale homologously with the low-$T_\star$ disk. §2.3 shows Class I models with more realistic disk properties in the high-$T_\star$ models. These models illustrate a second effect on the MIR SEDs, that of the increasing inner disk wall contribution from the higher $T_\star$ sources. This partially offsets the reddening effect of the stellar contribution in Class I sources. §2.4 extends the models to Class 0 and II sources and shows that the inner disk wall effect is even more important in Class II sources, not surprisingly. We show color-color diagrams in §2.5 which show overlap between evolutionary states due to inclination and stellar temperature; however the color-magnitude diagrams provide some guidance in separating effects of stellar temperature. A brief summary is presented in §3.
Models
======
Radiative Transfer
------------------
We use a 3-D Monte Carlo Radiative transfer code[^1] described in Paper I, which uses the radiative equilibrium method developed by Bjorkman & Wood (2001). The geometries considered in this paper are 2-D so we use a 2-D grid ($r-\theta$). The model geometries consist of a stellar source, a flared accretion disk, and a rotationally-flattened infalling envelope (Ulrich 1976; Terebey, Shu & Cassen 1984) with partially-evacuated bipolar cavities (Papers I & II). The envelope density decreases via a power-law at large radii ($\rho \propto r^{-1.5}$) and then merges with the ambient density of the surrounding molecular cloud. Luminosity is generated by the central star and accretion in the disk. This radiation is scattered and reprocessed by the surrounding circumstellar dust. We note that we solve for the 3-D temperature in any specified circumstellar geometry; therefore we naturally compute a hot surface on the inner disk wall as shown in Figure \[temp\], as we have in all our previous publications using these models (e.g., Wood et al. 2002a,b, Papers I & II, Rice et al. 2003, Walker et al. 2004). In addition, our cavity dust is hot and has a high emissivity, despite its low density (Paper I, Figure 7). Therefore our Class 0-II models include mid-IR thermal emission from the warm disk and cavity regions, in addition to any envelope component. Our models also accurately compute scattering and polarization using arbitrary scattering phase functions. Our models conserve flux absolutely (that is, to 0%). The primary source of error in our models is photon counting statistics (and, when comparing to data, knowledge of the appropriate input circumstellar geometry and dust properties). Running more photons produces higher signal-to-noise spectra. The models produced for this paper took 3 hours each to run (on 2 GHz PCs running g77). The exiting photons were binned into 10 inclinations and 200 frequencies to produce SEDs.
### Dust Properties and Dust Sublimation Radius
We use similar grain properties as in Paper II (Table 3): a large-grain model for the high-density regions in the disk (Wood et al. 2002b); a medium-sized grain model for the upper layers of the disk (Cotera et al. 2001); and for the envelope and cavity, a grain model that gives an extinction curve typical of molecular clouds with $r_V$, the ratio of total-to-selective extinction, equal to 4.3. The dust sublimation temperature is chosen to be 1600 K.
The disk dust sublimation radius was calculated through iteration by running the code several times and setting the opacity to be zero in grid cells when the temperature rises above $T_{sub}$. In the final iteration, there are no cells with $T > T_{sub}$ (Walker et al. 2004). We determined an empirical formula that fit our range of stellar temperatures: $$R_{sub}/R_\star=(T_{sub}/T_\star)^{-2.085}.$$ This is very similar to the optically [*thick*]{} blackbody radiative equilibrium temperature, $r \propto T^{-2}$, and thus likely has little dependence on the dust opacity law, unlike the optically thin radiative equilibrium limit [@lamers99; @beckwith90]. This behavior is obviously due to the fact that the inner disk wall is opaque at all wavelengths; this is true over a wide range of disk masses (Wood et al. 2002b).
Homologous Models
-----------------
We start by specifying a Class I model for a low-mass (0.5 M$_\sun$) YSO, based on previous observations and models (Kenyon et al. 1993a,b, Whitney et al. 1997, Lucas & Roche 1997, 1998; Padgett et al. 1999). Then we will scale this model homologously for different stellar temperatures. To construct a homologous model, we require that the inner and outer radii be scaled to the dust destruction radius and that the optical depths (at a given inclination angle) be the same for all the models (Ivezić & Elitzur 1997; Carciofi et al. 2004). This will result in a homologous temperature distribution for the circumstellar dust, and the shape of the thermal contribution to the SED will be invariant (only scaled by the increased luminosity). Table 1 shows the model parameters that result. For all the models we choose the inner radius to be the dust destruction radius. For the low-mass model, the outer radius is 3000 AU (chosen to be as small as reasonable since the high temperature model will be huge). The disk radius and envelope centrifugal radius is 300 AU, and the envelope infall rate is $6.7\times 10^{-6}$ M$_\sun$/yr. The model includes a flared disk of mass 0.01 $M_\sun$. The ratio of disk scale height to radius at the disk outer radius is $h/r = 0.12$ (chosen to match the HH30 disk, Burrows et al. 1996; Wood et al. 1998). For simplicity in comparing models, we set the disk accretion rate to 0, so the disk radiates from thermal reprocessing of starlight only. The bipolar cavity has a curved shape ($z = a (x^2+y^2)^b$, where $b=1.5$ and $a$ is set by the opening angle) and an opening angle of 20 degrees at the outer radius. The bipolar cavity is filled with constant-density dust with an optical depth along the polar direction of $\tau_V=5$ at V (0.55 $\mu$m). Because the low-$T_\star$ models have smaller outer envelope radii, they have correspondingly higher density in the bipolar cavities. Based on these parameter choices, the optical depth through the envelope at an inclination of 60 degrees is $\tau_V$=25; and through the disk midplane is $\tau_V$=193,000.
The parameters for the other models are then chosen to give the same inner and outer radii scaled to dust destruction radius, and the same optical depths at 0(cavity), 60(envelope), and 90(disk). Thus the envelope and disk masses and radii grow with the higher $T_\star$ models, as shown in Table 1. We choose stellar parameters appropriate for young stars of age $3 \times 10^5$ yrs [@siess].
Figure 2 shows SEDs for the homologous model. Figure \[homoga\] shows only the thermal spectrum from each model, and does not include the stellar direct or scattered radiation. From this we can see that the shapes of the spectra are similar for all models. The temperature structure and therefore thermal emission is determined only by the luminosity of the incident radiation (which sets the dust destruction radius) and where the radiation is absorbed (determined by geometry and density). The slight differences between the models are due to the fact that the models are not perfectly homologous for two main reasons: 1) The geometries are slightly different between the models because the ratio of the stellar size to dust destruction radius varies: In the high-$T_\star$ models, the star is effectively a point source, and in the low-$T_\star$ models it is larger than the inner disk wall height; and 2) Because the dust scattering albedo is non-zero and varies with wavelength, the total absorbed luminosity varies slightly between the models (i.e., the dust scattering albedo at the wavelengths where most of the stellar flux is emitted is lower in the low-$T_\star$ models than the high-$T_\star$ models so a slightly higher fraction of flux is absorbed in the low $T_\star$ models). However, the differences are slight and they do not detract from the main point, that the thermal spectra are similar between the models.
Figure \[homogb\] shows the total spectrum from each model. Here we see that the stellar contribution (scattered and direct) is greater in the low-$T_\star$ models. Thus, compared to the high-$T_\star$ models, the spectrum is relatively blue. We emphasize that the high-temperature models are redder in the 1-10 $\mu$m region, not because the envelope mass is larger but because the spectrum in this region is a more pure thermal spectrum. Note also that one reason the thermal spectra are similar for all the models is that the dust sublimation temperature $T_{sub}$ sets the cutoff for the maximum temperature of the thermal radiation. Thus the peak of the thermal emission occurs at the same wavelength for all the models. If there were no cutoff with $T_{sub}$, then the high $T_\star$ sources would have a more continuous blend between stellar and thermal spectra.
Class I Models guided by observations
-------------------------------------
The high-$T_\star$ models in Figures 2 are not realistic, given the homologous scaling of the disk parameters from the low-$T_\star$ disk. Here we show more realistic geometries gleaned from previous observations and modeling of high-mass sources [e.g. @alvarez04; @beltran; @beuther; @sandell04; @sandell00; @shepherd]. We keep the envelope optical depths similar between the models to help understand the comparisons between the models better. The model parameters are shown in Table 2. The main difference between these and the homologous model is in the disk parameters. Like the homologous models, the disk inner radii are set to the dust sublimation radius, $R_{sub}$. The outer radii are chosen based on observations cited above. The disk masses are chosen to be 5% of the stellar mass. The disk scale heights are calculated at $R_{sub}$ from the analytic solution of the hydrostatic equation assuming the disk temperature is vertically isothermal. Since the disk temperature is known at $R_{sub}$, it is straightforward to calculate the scale height $h$ at this location (equal to the sound speed divided by the Keplerian velocity; Bjorkman 1997). The gaussian scale height at each radius is then $h=h_{sub} (r/R_{sub})^{1.25}$. Disk accretion is included, but the effect on the SED is minor since the disk accretion luminosity is relatively small for all of the models (Table 2). The outer envelope radii are left large since the hotter stars will heat up surrounding ambient material out to several pc. Note that the envelope masses of the high-$T_\star$ models are large due to the large outer radii. However, the relevant parameter for the radiative transfer is optical depth, which is similar between the models. Therefore the variation of the thermal emission is due to circumstellar geometry.
Figure \[reala\] shows that hotter $T_\star$ models have relatively more $1-10 \mu$m thermal emission than the cooler $T_\star$ models. This is due to the larger inner disk walls in the high-$T_\star$ models as a result of the larger dust destruction radius ($h/r$ at $R_{sub}$ in Table 2). This is similar to the “puffed-up” inner disk region invoked by Natta et al. (2001) to explain the SEDs of Ae/Be stars. The wall intercepts and reprocesses radiation near the dust sublimation temperature $T_{sub}$ with its peak radiation at about 2 $\mu$m. This effect did not appear in the homologous model because $h/r$ at $R_{sub}$ was the same in all the models (Table 1). The increased $1-10 \mu$m emission in the high-$T_\star$ models is counteracted slightly when the stellar emission is included (Figure \[realb\]) but for the models with $T_\star > 8000$K, the colors are bluer towards pole-on inclinations and higher $T_\star$ due to emission from the disk wall. This figure shows that the disk structure and emission properties affect the 1-10$\mu$m spectrum in Class I sources. Note that the SEDs in Figs. \[reala\] and \[realb\] include the flux from the entire envelope which extends to nearly 2 pc in the case of the hot star model (Table 1). Figure \[realc\] shows the results integrated in a 3000 AU radius aperture (1.5at a distance of 2 kpc), more typical of aperture photometry observations. In this case, the hottest $T_\star$ model has less short- and long-wave flux in the smaller aperture giving a more rounded SED shape. It is rather striking that the SEDs of these four Class I models, with nearly identical optical depths in the cavity and envelope, have such different shapes.
Other Evolutionary States: Class 0 and II
-----------------------------------------
To see how stellar temperature affects other evolutionary states, we also show SEDs of Class 0 and II sources for the four stellar temperatures. We keep the stellar parameters the same for the Class 0 and II models even though they would obviously evolve over this time period. However, this allows us to isolate better the differences between the resulting SEDs (e.g., the disk scales heights will be similar if the stellar properties do not change). Figure \[c0\] shows SEDs for Class 0 sources with the model parameters in Table 3. The envelope optical depths are four times higher than the Class I model, and the polar optical depths are two times higher. The disk radii are smaller and disk masses higher (0.1 times the stellar mass), giving larger midplane optical depths. The SEDs are shown integrated in a 3000 AU aperture. Except for the pole-on inclinations, these show a trend for the high-$T_\star$ models to have redder $1-10\mu$m colors. As in the Class I models, the disk walls contribute more radiation at 2-10$\mu$m in the high-$T_\star$ models but this is only apparent towards pole-on inclinations ($i < 45$).
Class II SEDs are shown in Figure \[cii\] with the model parameters in Table 4. The disk parameters are similar to the Class I models except that the masses are lower, since they are more evolved. At a wavelength range of 1-2$\mu$m, there is a tendency for the models to become more red with increasing $T_\star$. At 2 $\mu$m the disk emission kicks in and the larger walls from the high-$T_\star$ models show a blue spectrum towards pole-on inclinations from $2-10 \mu$m. However, the edge-on inclinations are red or flat in the models due to obscuration by the flaring outer regions. Thus when inclination is included, there is again a large spread in colors at $1-10 \mu$m. We note that the disk models do not include the effect of gas opacity inside the dust destruction radius. This should not be a problem in the low mass disks due to low opacities (Lada & Adams 1992) but it may be important in the high-mass disks. In addition, PAH emission is likely an important contributor in the high-$T_\star$ models. This is beyond the scope of this paper and will be explored in the future.
Color-color and Color-magnitude diagrams
----------------------------------------
Figure \[cc\] shows color-color plots in the IRAC bands \[3.6\]-\[4.5\] vs. \[5.8\]-\[8.0\] for the Class I models (§2.3 and Table 2) integrated in a 3000 AU radius aperture. There is a trend for the high-$T_\star$ models to be more red, though there is overlap due to the broad spread in inclination within a model. The color range of these Class I models (with similar optical depths) spans the range of observations in the four star forming clusters presented by @allen04! The grey box in Figure \[cc\] shows the region denoted by Allen et al. as the approximate domain of Class II sources. Some of edge-on Class I models are blueward of this domain (see Figure 3c). Figure \[cm\] shows a color-magnitude diagram for the Class I models. This shows some spread in both magnitude and color with each model but a general trend for the high-$T_\star$ models to be brighter (obviously) and more red. The Class I models in Figure 6 can be compared to those presented by Allen et al. (2004) in their Figure 1. Their models show a trend for higher luminosity sources to be more red in \[5.8\]-\[8.0\] and higher density envelopes (presumably younger sources) to be more red in \[3.6\]-\[4.5\]. It is not clear if they varied stellar temperature in their Class I models, but our models show a similar trend for higher luminosity sources to have redder \[3.6\]-\[4.5\] colors due to stellar temperature effects. The main difference however is that our models in general give bluer colors for the same envelope parameters (density or infall rate) due to our inclusion of partially evacuated bipolar cavities and flared disks in the Class I models (Paper I, Figure 12). Thus we would likely estimate a younger evolutionary state (higher density envelope) for a given set of observational colors.
Finally, we show color-color and color-magnitude plots in Figures \[ccall\] and \[cmall\] adding in Class 0 and Class II sources. There are several interesting things to note in Fig. \[ccall\]: the six reddest sources in \[5.8\]-\[8.0\] are Class II sources (with stellar temperatures 8000 and 15000 K). The eight bluest sources in \[5.8\]-\[8.0\] are Class 0 and I sources (with $T_\star=4000$ K). This is backwards from the common wisdom. The \[3.6\]-\[4.5\] color behaves more as expected with Class 0-I sources most red and Class I-II sources most blue. There is somewhat of a dearth of sources in the “Class II domain” of Allen et al. (2004) (the grey box). This is due to the fact that we computed only one cool-$T_\star$ disk model, whereas most T Tauri stars in a low-mass cloud are likely cool with a range of disk masses. Allen et al.’s disk models all used a stellar temperature of 4000 K and inclinations of 30and 60. Both the hottest and coolest of our disk models fall at the right edge of their Class II domain region. The mid-temperature disks lie just to the right for most inclinations and then to the far right for the partially obscured (near edge-on) sources. Most of the Class 0 sources fall in the color range of 1-2 in both \[5.8\]-\[8.0\] and \[3.6\]-\[4.5\]. The Class I sources, on the other hand, span nearly the entire range of the plot. They suffer the most variation due to inclination, stellar temperature, inner disk wall, and scattered light effects. And Class II and 0 sources can be found at unexpected locations in the color-color plot as well, albeit with lower frequency.
Figure \[cmall\] shows a more systematic behavior in the color-magnitude plot. The hotter $T_\star$ sources tend to be brighter, and within a given flux range, their colors follow an evolutionary sequence with bluer colors being younger. Thus, the color-magnitude plot provides some guidance in separating stellar temperature effects from evolutionary effects.
Conclusions
===========
In this series of papers, we have shown that inclination effects, aperture size, scattered light, and stellar temperature cause a broad spread in the colors of a source at a single evolutionary state. There is systematic behavior, as shown in the color-magnitude diagrams (Fig. 9), but the behavior is not as simple as using the slope of the observed SED (or the IRAC \[3.6\]-\[4.5\] vs. \[5.8\]-\[8.0\] color-color plot) to estimate evolutionary state (Fig. 8) for a given source. There are trends in color space that could be applied in a statistical sense to a cluster. More modeling than the small grid presented here would be useful to provide a statistical guide for interpreting mid-IR color-color plots. Our codes are now publicly available and we are in the process of computing a large grid of models which will also be publicly available.
We note that these problems for interpreting evolutionary state based on SEDs are also reduced if long wavelength ($\lambda > 100 \mu$m) observations are obtained. These are much less sensitive to geometry (and inclination) and thus 1-D or simple disk models do a reasonable job estimating circumstellar mass and hence evolutionary state, assuming younger sources have more massive circumstellar envelopes (e.g., Mueller et al. 2002). Also, if the temperature of the stellar source can be estimated, the problems of interpreting the mid-IR colors in terms of evolutionary state become less severe. The stellar contribution can be estimated at wavelengths shortward of 2$\mu$m where disk emission does not contribute. In our preliminary modeling of the [*Spitzer*]{} IRAC data in the giant H II region RCW 49 (Whitney et al. 2004), we find that we can estimate the stellar temperature by fitting multi-band photometry across the $1-10 \mu$m range (2MASS JHK and the four IRAC bands). This improves our ability to simultaneously estimate the evolutionary state and central source temperature. We will present these results in our next paper.
This work was supported by the NASA LTSA Program (NAG5-8933, BAW); by NASA’s Spitzer Space Telescope Legacy Science Program through Contract Number 1224653 (RI); by the National Science Foundation (AST-0307686, JEB); and through a UK PPARC Advanced Fellowship (KW).
Allen, L. E., et al. 2004, ApJS, 154, 363
Alvarez, C., Hoare, M., & Lucas, P. 2004, , 419, 203
André, P., Ward-Thompson, D., & Barsony, M. 1993, , 406, 122
Beckwith, S. V. W., Sargent, A. I., Chini, R. S., & Gusten, R. 1990, , 99, 924
Behrend, R. & Maeder, A. 2001, , 373, 190
Beltr[' a]{}n, M. T., Cesaroni, R., Neri, R., Codella, C., Furuya, R. S., Testi, L., & Olmi, L. 2004, , 601, L187
Beuther, H., et al. 2004, ArXiv Astrophysics e-prints, astro-ph/0402505
Bjorkman, J. E. 1997, in Stellar Atmospheres: Theory and Observations, ed J.P. De Greve, R. Blomme, and H. Hensberge (New York:Springer), 239
Bjorkman, J. E., & Wood, K. 2001, ApJ, 554, 615
Bonnell, I. A., Bate, M. R., & Zinnecker, H. 1998, , 298, 93
Burrows, C. J., et al. 1996, , 473, 437
Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245
Carciofi,ÊA.ÊC., Bjorkman,ÊJ.ÊE., & Magalh‹es,ÊA.ÊM. 2004, ApJ, 604, 238
Cotera, A., Whitney, B. A., Young, E., Wolff, M. J., Wood, K., Povich, M., Schneider, G., Rieke, M., & Thompson, R. 2001, , 556, 958
Fazio, G. G. et al., 2004, ApJS, 154, 10
Gómez, M., Whitney, B. A., & Kenyon, S. J. 1997, , 114, 265
Indebetouw, R. et al. 2004, ApJ, submitted
Ivezić, Z. & Elitzur,ÊM. 1997, MNRAS, 287, 799
Kenyon, S. J., Calvet, N., & Hartmann, L. 1993a, , 414, 676
Kenyon, S. J., Whitney, B., Gómez, M., & Hartmann, L. 1993b, , 414, 773
Kenyon, S. J. & Hartmann, L. 1995, , 101, 117
Lada, C. J. 1987, in Star Forming Regions, edited by M. Peimbert & J. Jugaka (Dordrecht, Reidel), 1
Lada, C. J. & Adams, F. C. 1992, , 393, 278
Lada, C. J. 1999, in The Origin of Stars and Planetary Systems, edited by C. J. Lada & N. D. Kylafis (Kluwer: Dordrecht), 143
Lamers, H. J. G. L. M. & Cassinelli, J. P. 1999, in Introduction to Stellar Winds, (Cambridge: Cambridge University Press), 163
Lucas, P. W., & Roche, P. F. 1997, , 286, 895
Lucas, P. W., & Roche, P. F. 1998, , 299, 699
Maeder, A. & Behrend, R. 2002, , 281, 75
Megeath, S. T., et al. 2004, ApJS, 154, 367
Mueller,ÊK. E., Shirley,ÊY.ÊL., Evans,ÊN. ÊJ.,ÊII, Jacobson,ÊH.ÊR. 2002, ApJs, 143, 469
Muzerolle, J., et al. 2004, ApJS, 154, 379
Natta,ÊA., Prusti,ÊT., Neri,ÊR., Wooden,ÊD., Grinin,ÊV.ÊP., & Mannings,ÊV. 2001, , 371, 186
Padgett, D. L., Brandner, W., Stapelfeldt, K. R., Strom, S. E., Terebey, S., & Koerner, D. 1999, , 117, 1490
Reach, W. T., et al. 2004, ApJS, 154, 385
Reipurth, B., Yu, K. C., Heathcote, S., Bally, J., & Rodríguez, L. F. 2000, , 120, 1449
Rice, W. K. M., Wood, K., Armitage, P. J., Whitney, B. A., & Bjorkman, J. E. 2003, MNRAS, 342, 79
Richer, J. S., Shepherd, D. S., Cabrit, S., Bachiller, R., & Churchwell, E. 2000, in Protostars and Planets IV, eds. V. Mannings, A. P. Boss, S. R. Russell (Tucson: University of Arizona Press), p. 867
Rieke, G. H. et al. 2004, ApJS, 154, 25
Sandell, G. 2000, , 358, 242
Sandell, G. & Sievers, A. 2004, , 600, 269
Shepherd, D. S., Claussen, M. J., & Kurtz, S. E. 2001, Science, 292, 1513
Siess, L., Dufour, E., & Forestini, M. 2000, , 358, 593
Terebey, S., Shu, F. H., & Cassen, P. 1984, , 286, 52
Ulrich, R. K. 1976, , 210, 377
Walker, C., Wood, K., Lada, C. J., Robitaille, T., Bjorkman, J. E., & Whitney, B. 2004, MNRAS, 351, 607
Werner, M. W. et al. 2004, , 154, 1
Whitney, B. A., Kenyon, S. J., & Gómez, M. 1997, , 485, 703
Whitney, B. A., Wood, K., Bjorkman, J. E., & Wolff, M. J. 2003a, , 591, 1049 (Paper I)
Whitney, B. A., Wood, K., Bjorkman, J. E., & Cohen, M. 2003b, , 598, 1099 (Paper II)
Whitney, B. A., et al. 2004, ApJS, 154, 315
Wood,ÊK., Kenyon,ÊS.ÊJ.; Whitney,ÊB., & Turnbull,ÊM. 1998, ApJ, 497, 404
Wood, K., Wolff, M. J., Bjorkman, J. E., & Whitney, B. 2002a, , 564, 887
Wood, K., Lada, C. J., Bjorkman, J. E., Kenyon, S. J., Whitney, B., & Wolff, M. J. 2002b, , 567, 1183
[lllll]{} Envelope infall rate ($/ 10^{-5} M_\sun/$yr) & 0.67 & 10.5 & 10 & 49\
Envelope mass ($M_\sun$) & 0.12 & 18 & 46 & 2150\
Stellar radius ($R_\sun$) & 4 & 11.6 & 5 & 7.3\
Stellar luminosity ($L_\sun$) & 3.67 & 494 & 1134 & 46900\
Stellar mass ($M_\sun$) & 0.5 & 6.0 & 6.2 & 20\
Envelope & Disk inner radius ($R_\star$) & 6.7 & 29 & 106 & 500\
Envelope & Disk inner radius (AU) & 0.125 & 1.55 & 2.47 & 16.9\
Envelope outer radius (AU) & 3000 & 36900 & 59000 & 404000\
Disk mass ($M_\sun$) & 0.01 & 1.28 & 3.06 & 138\
Disk outer radius (AU) & 300 & 3691 & 5900 & 40400\
$h/r$ at $R_\star$ & 0.011 & 0.0074 & 0.0053 & 0.0036\
$h/r$ at $R_{sub}$ & 0.017 & 0.017 & 0.017 & 0.017\
Cavity opening angle () & 20 & 20 & 20 & 20\
Cavity density ($/ 10^{-20}$ gm cm$^{-3}$) & 37 & 3.0 & 1.9 & 0.28\
$\tau_V$ ($i=0$) & 5 & 5 & 5 & 5\
$\tau_V$ ($i=60$) & 25 & 25 & 25 & 25\
$\tau_V$ ($i=90$, $/ 10^5$) & 1.93 & 1.93 & 1.93 & 1.93\
[lllll]{}
Envelope infall rate ($/ 10^{-5} M_\sun/$yr) & 0.67 & 3.8 & 4.8 & 22\
Envelope mass ($M_\sun$) & 0.12 & 8.5 & 31 & 1430\
Disk mass ($M_\sun$) & 0.02 & 0.30 & 0.31 & 1\
Disk accretion rate ($/ 10^{-8} M_\sun/$yr) & 2.1 & 24 & 24 & 120\
Disk accretion luminosity ($/10^{-4} L_{acc}/L_\star $) & 17 & 1.4 & 0.38 & 0.022\
Disk outer radius (AU) & 300 & 400 & 500 & 500\
$h/r$ at $R_\star$ & 0.025 & 0.018 & 0.016 & 0.016\
$h/r$ at $R_{sub}$ & 0.041 & 0.041 & 0.051 & 0.074\
Cavity density ($/ 10^{-20}$ gm cm$^{-3}$) & 37 & 3.0 & 1.9 & 0.28\
$\tau_V$ ($i=90$, $/ 10^5$) & 1.63 & 1.77 & 0.80 & 0.28\
[lllll]{}
Envelope infall rate ($/ 10^{-5} M_\sun/$yr) & 0.88 & 5.2 & 6.1 & 25\
Envelope mass ($M_\sun$) & 0.15 & 11.5 & 58 & 2700\
Disk mass ($M_\sun$) & 0.05 & 0.60 & 0.62 & 2\
Disk accretion rate ($/ 10^{-8} M_\sun/$yr) & 32 & 250 & 250 & 1400\
Disk accretion luminosity ($/10^{-4} L_{acc}/L_\star $) & 251 & 14 & 4.0 & 0.25\
Disk outer radius (AU) & 50 & 80 & 100 & 100\
$h/r$ at $R_\star$ & 0.025 & 0.018 & 0.016 & 0.016\
$h/r$ at $R_{sub}$ & 0.040 & 0.041 & 0.051 & 0.074\
Cavity density ($/ 10^{-20}$ gm cm$^{-3}$) & 74 & 6.05 & 3.8 & 0.55\
Cavity opening angle () & 10 & 10 & 10 & 10\
$\tau_V$ ($i=0$) & 10 & 10 & 10 & 10\
$\tau_V$ ($i=60$) & 100 & 100 & 100 & 100\
$\tau_V$ ($i=90$, $/ 10^5$) & 25 & 18.6 & 7.9 & 2.87\
[lllll]{}
Disk mass ($M_\sun$) & 0.01 & 0.15 & 0.15 & 1\
Disk accretion rate ($/ 10^{-8} M_\sun/$yr) & 1.1 & 12 & 12 & 59\
Disk accretion luminosity ($/10^{-4} L_{acc}/L_\star $) & 8.3 & 0.68 & 0.19 & 0.001\
Disk outer radius (AU) & 300 & 400 & 500 & 500\
$h/r$ at $R_\star$ & 0.025 & 0.018 & 0.016 & 0.016\
$h/r$ at $R_{sub}$ & 0.041 & 0.041 & 0.051 & 0.074\
$\tau_V$ ($i=90$$/ 10^5$) & 0.82 & 0.89 & 0.39 & 0.14\
[^1]: Source code, instructions for running, and sample plotting tools are available at http://gemelli.spacescience.org/$\sim$bwhitney/codes/codes.html
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The goal of ranking and selection (R&S) procedures is to identify the best stochastic system from among a finite set of competing alternatives. Such procedures require constructing estimates of each system’s performance, which can be obtained simultaneously by running multiple independent replications on a parallel computing platform. However, nontrivial statistical and implementation issues arise when designing R&S procedures for a parallel computing environment. Thus we propose several design principles for parallel R&S procedures that preserve statistical validity and maximize core utilization, especially when large numbers of alternatives or cores are involved. These principles are followed closely by our parallel Good Selection Procedure (GSP), which, under the assumption of normally distributed output, (i) guarantees to select a system in the indifference zone with high probability, (ii) runs efficiently on up to 1,024 parallel cores, and (iii) in an example uses smaller sample sizes compared to existing parallel procedures, particularly for large problems (over $10^6$ alternatives). In our computational study we discuss two methods for implementing GSP on parallel computers, namely the Message-Passing Interface (MPI) and Hadoop MapReduce and show that the latter provides good protection against core failures at the expense of a significant drop in utilization due to periodic unavoidable synchronization.'
author:
- |
Eric C. Ni\
Cornell University\
Ithaca, New York 14853, USA\
`[email protected]`
- |
Dragos F. Ciocan\
INSEAD\
Fontainebleau 77305, FRANCE\
`[email protected]`
- |
Shane G. Henderson\
Cornell University\
Ithaca, New York 14853, USA\
`[email protected]`
- |
Susan R. Hunter\
Purdue University\
West Lafayette, Indiana 47907, USA\
`[email protected]`
bibliography:
- 'smaster.bib'
- 'sim.bib'
- 'Ni\_Sim.bib'
title: Efficient Ranking and Selection in Parallel Computing Environments
---
Introduction
============
The simulation optimization (SO) problem is a nonlinear optimization problem in which the objective function is defined implicitly through a Monte Carlo simulation, and thus can only be observed with error. Such problems are common in a variety of applications including transportation, public health, and supply chain management; for these and other examples, see `SimOpt.org` [@simoptlib]. For overviews of methods to solve the SO problem, see, e.g., [@1994fu; @1998andWSC; @2005fugloaprWSC; @2013pasgho].
We consider the case of SO on finite sets, in which the decision variables can be categorical, integer-ordered and finite, or a finite “grid” constructed from a continuous space. Formally, the SO problem on finite sets can be written as $$\begin{aligned}
\quad \max_{i\in {\mathcal{S}}} \,\, \mu_i=E[X(i;\xi)] \label{eq:problem}\end{aligned}$$ where ${\mathcal{S}}=\{1,\hdots,k\}$ is a finite set of design points or “systems” indexed by $i$, and $\xi$ is a random element used to model the stochastic nature of simulation experiments. (In the remainder of the paper we assume that $\mu_1 \le
\mu_2 \le \cdots \le \mu_k$, so that system $k$ is the best.) The objective function $\mu:{\mathcal{S}}\rightarrow\mathbb{R}$ cannot be computed exactly, but can be estimated using output from a stochastic simulation represented by $X(\cdot;\xi)$. While the feasible space ${\mathcal{S}}$ may have topology, as in the finite but integer-ordered case, we consider only methods to solve the SO problem in that (i) do not exploit such topology or structural properties of the function, and that (ii) apply when the computational budget permits at least *some* simulation of *every* system. Such methods are called *ranking and selection* (R&S) procedures. R&S procedures are frequently used in simulation studies because structural properties, such as convexity, are difficult to verify for simulation models and rarely hold. They can also be used in conjunction with heuristic search procedures in a variety of ways [@Pichitlamken2006; @BNK2003], making them useful even if not all systems can be simulated. See [@KN2006] for an excellent introduction to, and overview of, R&S procedures. We remark here that while R&S problems are closely related to best-arm problems, there are several differences between these bodies of literature. Almost always, the algorithms developed in the best-arm literature assume that only one system is simulated at a time [see, e.g., @Jamieson2014; @bubces12] and that simulation outputs are bounded, or that all variances have a known bound.
R&S procedures are designed to offer one of several types of probabilistic guarantees, and can be Bayesian or frequentist in nature. Bayesian procedures offer guarantees related to a loss function associated with a non-optimal choice; see [@BCS2007] and @Chen2015 [Chapter 3]. Frequentist procedures typically offer one of two statistical guarantees; in defining these guarantees, let $\delta>0$ be a known constant and let $\alpha\in(0,1)$ be a parameter selected by the user. The [*Probability of Correct Selection*]{} (PCS) guarantee is a guarantee that, whenever $\mu_k -\mu_{k-1} \ge \delta$, the probability of selecting the best system $k$ when the procedure terminates is greater than $1-\alpha$. Henceforth, the assumption that $\mu_k -\mu_{k-1} \ge \delta$ will be called the [*PCS assumption*]{}; if $\mu_k - \mu_{k-1} < \delta$ then a PCS guarantee does not hold. In contrast, the [*Probability of Good Selection*]{} (PGS) guarantee is a guarantee that the probability of selecting a system with objective value within $\delta$ of the best is greater than $1-\alpha$. That is, the PGS guarantee implies PGS $= \mathrm{P}[\text{Select a system } K \text{ such that } {\mu_k-
\mu_K\le \delta}] \geq 1-\alpha$. A PGS guarantee makes no assumption about the configuration of the means and is the same as the “probably approximately correct” guarantee in best-arm literature.
Traditionally, R&S procedures were limited to problems with a modest number of systems $k$, say $k \le 100$, due to the need to assume worst-case mean configurations to construct validity proofs. The advent of screening, i.e., discarding clearly inferior alternatives early on [@Nelson2001; @Kim2006; @Hong2006], has allowed R&S to be applied to larger problems, say $k \le 500$. Exploiting parallel computing is a natural next step as argued in, e.g., [@fu02]. By employing parallel cores, simulation output can be generated at a higher rate, and a parallel R&S procedure should complete in a smaller amount of time than its sequential equivalent, and allowing larger problems to be solved.
[@hei88; @glyhei90; @glyhei91] explored the use of parallel computers to construct valid simulation estimators, but R&S procedures that exploit parallel computing have emerged only recently. [@Luo2000] and [@Yoo2009] employ a web-based computing environment and present a parallel procedure under the optimal computing budget allocation (OCBA) framework. (OCBA has impressive empirical performance, but does not offer PCS or PGS guarantees.) [@Chen2005] tests a sequential pairwise hypothesis testing approach on a local network of computers. More recently, [@Luo2013] develop a parallel adaptation of a fully-sequential R&S procedure that provides an asymptotic (as $\delta \to 0$) PCS guarantee. [@Luo2013] is the best known existing method for parallel ranking and selection that provides a form of PCS guarantee on the returned solution.
In this paper, we (i) identify opportunities and challenges that arise from adopting a parallel computing environment to solve large-scale R&S problems, (ii) propose a Good Selection Procedure (GSP) that solves R&S problems on parallel computers, and (iii) implement our procedure in two different parallel computing frameworks. We make the following contributions.
**Theoretical contributions.** We propose a number of design principles that promote efficiency and validity in such an environment, and demonstrate them in a new parallel GSP. GSP showcases the power of these design principles in that it greatly extends the boundary on the size of solvable R&S problems. While the method of [@Luo2013] can solve on the order of $10^4$ systems, one of our implementations of GSP is capable of solving R&S problems with more than $10^6$ systems. Our computational results include such a problem, which we solve in under 6 minutes on $10^3$ cores. Another important theoretical contribution of this paper is the redesigned screening method in GSP which, unlike many fully-sequential procedures [@KimNels01; @Hong2006], does not rely on the PCS assumption. Accordingly, many systems can lie within the indifference-zone, i.e., have an objective function value within $\delta$ of that of System $k$, as will usually be the case when the number of systems is very large. Our procedure then provides the same PGS guarantee as existing indifference-zone procedures like [@Nelson2001] but with far smaller sample sizes. **Practical contributions.** The GSP procedure discussed in this paper is intended for any parallel, shared or non-shared memory platform where cores can communicate with each other. As long as no core fails during execution, it should deliver expected results regardless of the hardware specification. The procedure is also amenable to a range of existing parallel computing frameworks. We offer implementations of GSP based on MPI (Message-Passing Interface) and Hadoop MapReduce, and show how they differ in construction and in performance. The reasons for our choice of implementation frameworks are twofold:
- Both MPI and MapReduce are among the most popular and mature platforms for deploying parallel code, on a wide range of systems ranging from high performance supercomputers to commodity clusters such as Amazon EC2.
- MPI and MapReduce provide points of comparison between two different parallel design philosophies. Broadly speaking, the former enables low level tailoring and optimization in the implementation of a parallel procedure, while the latter is more of a “one-size-fits-all” framework which delegates as much of the implementation complexity as possible to the MapReduce package itself.
As we shall see, MPI is the more efficient of the two, achieving speed and utilization gains of around a factor of magnitude over MapReduce. On the other hand, MapReduce offers acceptable performance for large scale problems, and is more robust to reliability issues that may arise in cloud-computing environments where parallel tasks may fail to complete due to unresponsive cores.
The remainder of the paper is organized as follows. §\[sect:DesignPrinciples\] discusses the design principles followed in creating GSP to promote efficiency and ensure the procedure’s validity. §\[sect:GoodSelectionProcedure\] describes our multi-stage parallel R&S procedure GSP, and establishes the PGS guarantee. Computational studies in §\[sect:computation\] support our assertions on the quality of GSP and its parallel implementations, and point to open-access repositories where the code can be obtained. An appendix contains more proof detail, and further information on the MPI and MapReduce implementations. This paper is a considerable outgrowth of the conference papers [@Ni2013; @Ni2014; @Ni2015].
Design Principles for Parallel R&S Procedures {#sect:DesignPrinciples}
=============================================
R&S procedures are essentially made up of three computational tasks: (1) deciding what simulations to run next, (2) running simulations, and (3) screening (computing statistical estimators and determining which systems are inferior). On a single-core computer, these tasks are repeatedly performed in a certain order until a termination criterion is met. On a parallel platform, multiple cores can simultaneously perform one or several of these tasks.
In this section, we discuss various issues that arise when a R&S procedure is designed for and implemented on parallel platforms to solve large-scale R&S problems. We argue that failing to consider these issues may result in impractically expensive or invalid procedures. We recommend strategies by which these issues can be addressed, and illustrate how we incorporate them in our procedure presented in §\[sect:GoodSelectionProcedure\], which iteratively runs Tasks (1) through (3) in multiple stages. For discussing the design principles for parallel R&S procedures in this section, we consider a parallel computing environment that satisfies the following properties.
\[assum:Independence\] (*Core Independence*) A fixed number of processing units (“cores”) are employed to execute the parallel procedure. Each core is capable of performing its own set of computations without interfering with other cores unless instructed to do so. Each core has its own memory and does not access the memory of other cores.
\[assum:Message-passing\] (*Message-passing*) The cores are capable of communicating through sending and receiving messages of common data types and arbitrary lengths.
\[assum:Reliability\] (*Reliability*) Cores do not “fail” or suddenly become unavailable. Messages are never “lost”.
Many parallel computer platforms satisfy the first two assumptions, but some are subject to the risk of core failure, which may interrupt the computation in various ways. For clarity, we work under the reliability assumption and defer the design of failure-proof procedures to §\[sect:implementation\_hadoop\] where we discuss Hadoop MapReduce. Similar to [@Luo2013] and [@Ni2013], we consider a master-worker framework, using a uniquely executed “master” process (typically run on a dedicated “master” core) to coordinate the parallel procedure, and letting other cores (the “workers”) work according to the master’s instructions. To the extent possible we want to avoid synchronization delays, where one core cannot continue until another core completes its task, as we will see in §\[sect:implementation\_parallel\].
Implications of Random Completion Times {#sect:RandCompTime}
---------------------------------------
Consider the simplest case where only Task (2), running simulations, is run in parallel, and each simulation replication completes in a random amount of time. To construct estimators for a single system simulated by multiple cores, one can either collect a fixed number of replications in a random completion time, or a random number of replications in a fixed completion time [@hei88]. [@hei88] and [@glyhei90; @glyhei91] discuss unbiased estimators of each type. Because a random number of replications collected after a fixed amount of time may not be i.i.d. with the desired distribution upon which much of the screening theory depends [@hei88; @glyhei91; @Ni2013; @Luo2013], we confine our attention to estimators that produce a fixed number of replications in a random completion time. (The cause of this difficulty can be traced to dependence between the estimated objective function and computational time.)
Using estimators that produce a fixed number of replications in a random completion time for parallel R&S places a restriction on the manner in which replications can validly be farmed out to and collected from the workers. Consider the case where more than one core simulates the same system, and replications generated in parallel are aggregated to produce a single estimator. A naïve way is to collect replications from any core following the order in which they are generated, but as demonstrated in @Ni2013 [§3.1], the estimators may be biased, making it hard to establish provable statistical guarantees. In contrast, a valid method is to place the finished replications in a predetermined order and use them as if they are generated following that order, to avoid “re-ordering” of the simulation replications caused by random completion time.
Under this principle, our GSP in §\[sect:GoodSelectionProcedure\] ensures that the simulation results generated in parallel are initiated, collected, assembled and used by the screening routine in an ordered manner. Specifically, in Stage 2 of GSP, when the master instructs a worker to simulate system $i$ for a batch of replications (Step \[step:Stage2Main\]\[step:NHHsim\]), the batch index is also received by the worker. When the batch is completed, its statistics are sent back to the master alongside the batch index (Step \[step:Stage2Main\]\[step:NHHsim\]), which signals its pre-determined position in the assembled batch sequence on the master. This ensures that the batch statistics sent to workers for screening (Step \[step:Stage2Main\]\[step:NHHscreening\]) follow the exact order in which they were initiated, and constructed estimators are unbiased with the correct distribution. [@Luo2013] discuss a similar approach which they refer to as “vector-filling”.
Allocating Tasks to the Master and Workers
------------------------------------------
Previous work on parallel R&S procedures [@Chen2000; @Yoo2009; @Luo2011; @Luo2013] focuses almost exclusively on pushing Task (2), running simulations, to parallel cores. In those procedures, usually the master is solely responsible for Tasks (1) and (3), deciding what simulations to run next and screening, and the workers perform Task (2) in parallel. In this setting, the benefit of using a parallel computing platform is entirely attributed to distributing simulation across parallel cores, hence reducing the total amount of time required by Task (2).
However, the master could potentially become a bottleneck in a number of ways. First, as noted by [@Luo2011], the master can be overwhelmed with messages. Second, for the master to keep track of all simulation results requires a large amount of memory, especially when the number of systems is large [@Luo2013]. Finally, when the number of systems is large and simulation output is generated by many workers concurrently, running Tasks (1) and (3) on the master alone may become relatively slow, resulting in a waste of core hours on workers waiting for the master’s further instructions. Therefore, a truly scalable parallel R&S procedure should allow its users a simple way to control the level of communication, use the memory efficiently, and distribute as many tasks as possible across parallel cores. In addition, it should perform some form of load-balancing to minimize idling on workers.
### Batching to Reduce Communication Load {#sect:batching}
One way to reduce the number of messages handled by the master is to control communication frequency by having the workers run simulation replications in batches and only communicate once after each batch is finished.
Since R&S procedures typically use summary statistics rather than individual observations when screening systems, it may even suffice for the worker to compute and report batch statistics instead of point observations from every single replication. Indeed, a useful property of our statistic for screening systems $i$ and $j$ is that it is updated using only the sample means over the entirety of the most recent batch $r$, instead of requiring the collection of individual replication outcomes. These sample means can be independently computed on the worker(s) running the $r$th batch of systems $i$ and $j$, and the amount of communication needed in reporting them to the master is constant and does not grow with the batch size. The distribution of batches in parallel must be handled with care. Most importantly, since using a random number of replications after a fixed run time may introduce bias (as we have shown in §\[sect:RandCompTime\]), a valid procedure should employ a predetermined and fixed batch size for each system, which may vary across different systems. Batches generated in parallel for the same system should be assembled according to a predetermined order, following the same argument used in §\[sect:RandCompTime\]. Furthermore, if the procedure requires screening upon completion of every batch, then it is necessary to perform screening steps following the assembled order.
### Allocating Simulation Time to Systems {#sect:BatchSizes}
When multiple systems survive a round of screening, R&S procedures need to decide which system(s) to simulate next (possibly on multiple cores), and how many replications to take. While sequential procedures usually sample one replication from the chosen system(s), or multiple replications from a single system, it is natural for a parallel procedure to consider strategies that sample multiple replications from multiple systems. In doing so, the parallel procedure may adopt sampling strategies such that simulation resources are allocated to surviving systems in a most efficient manner.
The best practice in making such allocations depends on the specific screening method. For instance, in [@Hong2006] as well as GSP, screening between systems $i$ and $j$ is based on a scaled Brownian motion $B([\sigma_i^2/n_i+\sigma_{j}^2/n_{j}]^{-1})$ where $B(\cdot)$ denotes a standard Brownian motion (with zero drift and unit volatility), $n_i$ is the sample size and $\sigma_i^2$ is the variance of system $i$. To drive this Brownian motion rapidly with the fewest samples possible, which accelerates screening, [@Hong2006] recommended that the ratio $n_i/\sigma_i$ be kept equal across all surviving systems.
The above recommendation implicitly assumes that simulation completion time is fixed for all systems, and is suboptimal when completion time varies across systems. Suppose all workers are identical, and each replication of system $i$ takes a fixed amount of time $T_i$ to simulate on any worker. We can then formulate the problem of advancing the above Brownian motion as $$\begin{aligned}
\max\,\,\, &[\sigma_i^2/n_i+\sigma_{j}^2/n_{j}]^{-1}\\
\text{s.t. } & n_iT_i+n_{j}T_{j}=T\end{aligned}$$ which yields the optimal computing time allocation $$\begin{aligned}
\frac{n_iT_i}{n_{j}T_{j}}=\frac{\sigma_i\sqrt{T_i}}{\sigma_{j}\sqrt{T_{j}}}.
\label{eq:optimBatchSize}\end{aligned}$$ This result is consistent with a conclusion in [@glywhi92], that when simulation completion time $T_i$ varies, an asymptotic measure of efficiency per replication is inversely proportional to $\sigma_i^2 E[T_i]$. In practice, $T_i$ is unknown and possibly random, so both $E[T_i]$ and $\sigma^2$ need to be estimated in a preliminary stage. Suppose they are estimated by some estimators $\bar{T_i}$ and $S_i^2$. Then we recommend setting the batch size for each system $i$ proportional to $S_i/\sqrt{\bar{T_i}}$ following .
### Distributed screening {#sect:DistributedScreening}
In fully sequential R&S procedures, e.g., [@Kim2006; @Hong2006], each screening step typically involves doing a fixed amount of calculation between every pair of systems to decide if one system is better than another with a certain degree of statistical confidence. The amount of work is proportional to the number of pairs of systems, which is $O(k^2)$.
In the serial R&S literature, the computational cost of screening is assumed to be negligible compared to that of simulation because the number of systems $k$ is usually quite small and each simulation replication may take orders of magnitude longer than $O(k^2)$ screening operations required in each iteration. Under this assumption, it is tempting to simply have the master handle all screening after the workers complete a simulation batch. This approach can easily be implemented and proven to be statistically valid. However, it may become computationally inefficient because all workers stay idle while the master screens, so a total amount of $O(ck^2)$ processing time is wasted, where $c$ is the number of workers. For a large problem with a million systems solved on a thousand cores, the wasted processing time per round of screening can easily amount to thousands of core hours, reducing the benefits from a parallel implementation dramatically. Moreover, if the procedure requires computing and storing in memory some quantities for each system pair (for instance, the variance of differences between systems), then the total amount of $O(k^2)$ memory may easily exceed the limit for a single core.
It is therefore worth considering strategies that distribute screening among workers. A natural strategy is to assign roughly $k/c$ systems to each worker, and let it screen among those systems only, as illustrated in Figure \[fig:screenOnWorkers\]. By doing so, each worker screens $k/c$ systems, occupying only $O(k^2/c^2)$ memory, and performing $O(k^2/c^2)$ work in parallel. Hence the wall-clock time for each round of screening is reduced by a factor of $c^2$.
Under the distributed screening scheme, not all pairs of systems are compared, so fewer systems may get eliminated. The reduction in effectiveness of screening can be compensated by sharing some good systems across workers. In Figure \[fig:screenOnWorkers\], for example, each core shares its own (estimated) best system with other cores, and each system is screened against other systems on the same core, as well as $O(c)$ good systems from other cores. This greatly improves the chance that each system is screened against a good one, despite the extra work to share those good systems. As illustrated in Figure \[fig:screenOnWorkers\], the additional number of pairs that need to be screened on each core is only $O(k)$ when the best system on each core is shared. Alternatively, the procedure may also choose to share only a smaller number $c'\ll c$ of good systems, so that the communication workload associated with this sharing does not increase as the number of workers increases.
The statistical validity of some screening-based R&S procedures (e.g. @KimNels01 [@Hong2006; @Luo2013]) requires screening to be performed once every replication (or batch of replications) is simulated. This implies that, when the identity of the estimated-best system(s) changes, the master has to communicate all previous replication results of the new estimated-best system(s) to the workers, so that they can perform all of the screening steps up to the current replication to ensure validity of the screening. (If screening on a strict subsequence of replications, it may be sufficient to communicate summary statistics.) Such “catch-up” screening was used, for instance, in [@Pichitlamken2006], in a different context. In §\[sect:GoodSelectionProcedure\], we employ a probabilistic bound that removes the need for catch-up screening in GSP.
Besides core hours, distributing screening across workers also saves memory space on the master. In our implementation of GSP, the master keeps a complete copy of batch statistics only for a small number of systems that are estimated to be the best. For a system that is not among the best, the master acts as an intermediary, keeping statistics for only the most recent batches that have not been collected by a worker. Whenever some batch statistics are sent to a worker (for screening in Steps \[step:Stage1Main\]\[step:NHHStage1Screening\] or \[step:Stage2Main\]\[step:NHHscreening\] of GSP), they can be deleted on the master. This helps to even out memory usage across cores, making the procedure capable of solving larger problems without the need to use slower forms of storage.
Random Number Stream Management
-------------------------------
The validity and performance of simulation experiments and simulation optimization procedures relies substantially on the quality and efficiency of (pseudo) random number generators. For a discussion of random number generators and their desirable properties, see [@LEcuyer2006].
To avoid unnecessary synchronization, each core may run its own random number generator independently of other cores. Some strategies for generating independent random numbers in parallel have been proposed in the literature. [@Mascagni2000] consider a class of random number generators which are parametrized so that each valid parametrization is assigned to one core. [@Karl2014] adopt [@LEcuyer2002]’s `RngStream` package, which supports streams and substreams, and demonstrated a way to distribute `RngStream` objects across parallel cores.
Both methods set up parallel random number generation in such a way that once initialized, each core will be able to generate a unique, statistically independent stream of pseudo random numbers, which we denote as $U_w$, for each $w=1,2,\ldots, c $. If a core has to switch between systems to simulate, one can partition $U_w$ into substreams $\{U_w^i:i=1,2,\ldots, k\}$, simulating system $i$ using $U_w^i$ only. It follows that for any system $i$, $U_w^i$ for different $w$ are independent as they are substreams of independent $U_w$’s, so simulation replicates generated in parallel with $\{U_w^i:w=1,2,\ldots,c\}$ are also i.i.d. Moreover, if it is desirable to separate sources of randomness in a simulation, it may help to further divide $U_w^i$ into subsubstreams, each used by a single source of randomness.
In practice, one does not need to pre-compute and store all random numbers in a (sub)stream, as long as jumping ahead to the next (sub)stream and switching between different (sub)streams are fast. Such operations are easily achievable in constant computational cost; see [@LEcuyer2002] for an example.
Although our procedure does not support the use of common random numbers (CRN), it is worth noting that the above framework easily extends to accommodate CRN as follows. Begin by having one identical stream $U_0$ set up on all cores and partitioning it into substreams $\{U_0(\ell): 1\le \ell\le L\}$ for sufficiently large $L$. Let the master keep variables $\{\ell_i:i=1,2,\ldots,k\}$ which count the total number of replications already generated for system $i$ over all workers. Each time the master initiates a new replication of system $i$ on a worker, it instructs the worker to simulate system $i$ using substream $\{U_0(\ell_i+1)\}$ and adds 1 to $\ell_i$. This ensures that for any $\ell>0$, the $\ell$th replication of every system is generated by the same substream $\{U_0(\ell)\}$.
The Parallel Good Selection Procedure {#sect:GoodSelectionProcedure}
=====================================
In this section, we provide a R&S procedure GSP that incorporates the design principles from §\[sect:DesignPrinciples\], and is implementable on a wide spectrum of parallel platforms. Our procedure applies to the general case in which the system mean and variance are both unknown and need to be estimated (an earlier version of the procedure under the known variance case is discussed in @Ni2014), and does not permit the use of common random numbers. We prove that the procedure offers a PGS guarantee for normally distributed observations.
The Setup
---------
GSP consists of four broad stages. In an optional Stage 0, workers run $n_0$ simulation replications for each system in parallel to estimate completion times, which are subsequently used to try to balance the workload. As discussed in §\[sect:BatchSizes\], Stage 0 samples are then dropped and not used to form estimators of $\mu_i$’s due to the potential correlation between simulation output and completion time. In Stage 1, a new sample of size $n_1$ is collected from each system to obtain variance estimates $S_i^2=\sum_{\ell=1}^{n_1}(X_{i\ell}-{\bar{X}}_i(n_1) )^2/(n_1-1)$, where ${\bar{X}}_i(n)=\sum_{l=1}^n X_{il}/n$. Prior to Stage 2, obviously inferior systems are screened. In Stage 2, the workers iteratively visit the remaining systems and run additional replications, exchange statistics and independently perform screening over a subset of systems until either all but one are eliminated, or a pre-specified limit on sample size is reached. The screening rule and the limit on sample size are jointly chosen such that inferior systems can be eliminated efficiently, while the best system $k$ survives this stage with high probability regardless of the configuration of true means $\mu_1,\ldots,\mu_k$. Finally, in Stage 3, all systems surviving Stage 2 enter a [@Rinott1978] procedure where a maximum sample size is calculated, additional replications are simulated if necessary, and the system with the highest sample mean is selected as the best.
The sampling rules used in Stages 0, 1, and 3 are relatively straight forward, for they each require a fixed number of replications from each system. In Stage 2, where the procedure iteratively switches between simulation and screening, a sampling rule needs to be specified to fix the number of additional replications to take from each system before each round of screening. Prior to the start of the overall selection procedure we define increasing (in $r$) sequences $\{n_i(r): i=1,2,\ldots,k, r=0,1,\ldots \}$ giving the total number of replications to be collected for system $i$ by batch $r$, and let $n_i(0)=n_1$ since we include the Stage 1 sample in mean estimation. Following the discussion in §\[sect:BatchSizes\] where we recommend that batch size for system $i$ be proportional to $S_i/\sqrt{\bar{T_i}}$ in order to efficiently allocate simulation budget across systems, we use $$\begin{aligned}
n_i(r)=n_1+r\left\lceil\beta
\left(
\frac{S_i}{\sqrt{\bar{T_i}}}
\right)
/
\left(
\frac{1}{k}\sum_{j=1}^{k}
\frac{S_j}{\sqrt{\bar{T_j}}}
\right)
\right\rceil \label{eq:sampling_rule}\end{aligned}$$ where $\bar{T_i}$ is an estimator for simulation completion time of system $i$ obtained in Stage 0 if available, and $\beta$ is the average batch size and is specified by the user.
The parameters for the procedure are as follows. Before the procedure initiates, the user selects an overall confidence level $1-\alpha$, type-I error rates $\alpha_1$, $\alpha_2$ such that $\alpha=\alpha_1+\alpha_2$, an indifference-zone parameter $\delta$, Stage 0 and Stage 1 sample sizes $n_0,n_1\ge 2$, and average Stage 2 batch size $\beta$. The user also chooses $\bar{r}>0$ as the maximum number of iterations in Stage 2, which governs how much simulation budget to spend in iterative screening before moving to indifference-zone selection in Stage 3.
Typical choices for error rates are $\alpha_1 = \alpha_2 = 0.025$ for guaranteed PGS of 95%. The indifference-zone parameter $\delta$ is usually chosen within the context of the application, and is often referred to as the smallest difference worth detecting. The sample sizes $n_0$ and $n_1$ are typically chosen to be small multiples of 10, with the view that these give at least reasonable estimates of the runtime per replication and the variance per replication.
For non-normal simulation output, we recommend setting $\beta\ge 30$ to ensure normally distributed batch means. The parameter $\beta$ also helps to control communication frequency so as not to overwhelm the master with messages. Let $T_\text{sim}$ be a crude estimate of the average simulation time (in seconds) per replication, perhaps obtained in a debugging phase. Then ideally the master communicates with a worker every $\beta T_\text{sim}/c$ seconds. If every communication takes $T_\text{comm}$ seconds, the fraction of time the master is busy is $\rho=cT_\text{comm}/\beta T_\text{sim}$. We recommend setting $\beta$ such that $\rho\le 0.05$, in order to avoid significant waiting of workers.
We recommend choosing $\bar{r}$ such that a fair amount of simulation budget (no more than 20% of the sum of Rinott sample sizes) will be spent in the iterative screening stages. Note that a small $\bar{r}$ implies insufficient screening whereas a large $\bar{r}$ may be too conservative.
Under these general principles, our choices of $(\beta=100,\bar{r}=10)$ and $(\beta=200,\bar{r}=5)$ in the experiments in §\[sect:computation\] work reasonably well on our testing platform, but it is conceivable that other values could improve performance.
Finally, we define some quantities used in the iterative screening stages. Let $\eta$ be the solution to $$\begin{aligned}
E\left[2\bar{\Phi}\left(\eta\sqrt{R}\right)\right]=1-(1-\alpha_1)^{\frac{1}{k-1}}, \label{eq:Choice_eta}\end{aligned}$$ where $\bar{\Phi}(\cdot) $ denotes the complementary standard normal distribution function, and $R$ is the minimum of two i.i.d. $\chi^2$ random variables, each with $n_1-1$ degrees of freedom. Let the distribution function and density of such a $\chi^2$ random variable be denoted $F_{\chi^2_{n_1-1}}(x)$ and $f_{\chi^2_{n_1-1}}(x)$, respectively. Hence $R$ has density $
f_R(x)=2[1-F_{\chi^2_{n_1-1}}(x)]f_{\chi^2_{n_1-1}}(x).
$ Also, for any two systems $i\ne j$, define $$\begin{aligned}
t_{ij}(r)&=\left[\frac{\sigma^2_i}{n_i(r)}+\frac{\sigma^2_j}{n_{j}(r)}\right]^{-1},
&&Z_{ij}(r)=t_{ij}(r) [{\bar{X}}_i(n_i(r))-{\bar{X}}_{j}(n_{j}(r))],
\\
\tau_{ij}(r)&=\left[\frac{S^2_i}{n_i(r)}+\frac{S^2_{j}}{n_{j}(r)}\right]^{-1},
&&Y_{ij}(r)=\tau_{ij}(r) [{\bar{X}}_i(n_i(r))-{\bar{X}}_{j}(n_{j}(r))],
\\
a_{ij}(\bar{r}) &= \eta \sqrt{(n_1-1)\tau_{ij}(\bar{r})}.\end{aligned}$$
Good Selection Procedure under Unknown Variances {#sect:GSPsteps}
------------------------------------------------
1. **(Stage 0), optional** Master assigns systems to workers, so that each system $i$ is simulated for $n_0$ replications and the average simulation completion time $\bar{T_i}$ is reported to the master.
2. **(Stage 1)** Master assigns systems to load-balanced (using $\bar{T_i}$ if available) simulation groups $G_1^w$ for $w = 1,\ldots, c$. Let ${\mathcal{I}}\leftarrow{\mathcal{S}}$ be the set of surviving systems.
3. \[step:Stage1Main\] For $w=1,2,\ldots,c$ in parallel on workers:
1. Sample $X_{i\ell}$, $\ell=1,2,\ldots, n_1$ for all $i\in G_1^w$.
2. Compute Stage 1 sample means and variances ${\bar{X}}_i(n_1)$ and $S^2_i$ for $i\in G_1^w$.
3. \[step:NHHStage1Screening\] Screen within group $G_1^w$: system $i$ is eliminated (and removed from ${\mathcal{I}}$) if there exists a system $j\in G_1^w: j\neq i$ such that $
Y_{ij}(0)<
-a_{ij}(\bar{r})
$.
4. Report survivors, together with their Stage 1 sample means ${\bar{X}}_i(n_i(0))$ and variances $S_i^2$, to the master.
4. **(Stage 2)** Let $G_1\leftarrow{\mathcal{I}}$ be the set of systems surviving Stage 1. Master computes sampling rule using $S_i^2$ obtained in Stage 1, and divides $G_1$ to approximately load-balanced screening groups $G^w_2$ for $w = 1,\ldots, c$. Let $s_i\leftarrow 0, i\in G_1$ be the count of the number of batches simulated in Stage 2 for system $i$.
5. \[step:Stage2Main\] For $w=1,2,\ldots,c$ in parallel on workers, let $r_w \leftarrow 0$ be the count of the number of batches screened on worker $w$ (which is common to all systems in the screening), and iteratively switch between simulation and screening as follows (this step entails some communication with the master, the details of which are omitted):
1. \[step:CheckS1\] Check termination criteria with the master: if $|{\mathcal{I}}|=1$ (only one system remains) or $r_w\ge\bar{r}$ for all $w$ (each worker has screened up to $\bar{r}$, the maximum number of batches allowed), go to Stage 3; otherwise continue to Step \[step:Stage2Main\]\[step:checkSim\].
2. \[step:checkSim\] Decide to either simulate more replications or perform screening based on available results: check with the master if the $(r_w+1)$th iteration has completed for all systems $i\in G^w_2$ and $|G^w_2|>1$, if so, go to Step \[step:Stage2Main\]\[step:NHHscreening\], otherwise go to Step \[step:Stage2Main\]\[step:NHHsim\].
3. \[step:NHHsim\] Retrieve the next system $i$ in $G_1$ (not necessarily $G^w_2$) from the master and simulate it for an additional $n_i(s_i+1)-n_i(s_i)$ replications. Set $s_i\leftarrow s_i+1$. Report simulation results to the master. Go to Step \[step:Stage2Main\]\[step:CheckS1\].
4. \[step:NHHscreening\] Screen within $G^w_2$ as follows. Retrieve necessary statistics for systems in $G^w_2$ from the master (recall that a system in $G^w_2$ is not necessarily simulated by worker $w$). Let $r_w
\leftarrow r_w+1$. A system $i\in G^w_2$ is eliminated if $r_w\le
\bar{r}$ and there exists a system $j\in G^w_2:j\neq i$ such that $
Y_{ij}(r_w)
<
-a_{ij}(\bar{r})
$. Also use a subset of systems from other workers, e.g., those with the highest sample mean from each worker, to eliminate systems in $G^w_2$. Remove any eliminated system from $G^w_2$ and ${\mathcal{I}}$. Go to Step \[step:Stage2Main\]\[step:CheckS1\].
6. **(Stage 3)** Let $G_2\leftarrow{\mathcal{I}}$ be the set of systems surviving Stage 2. If $k':=|G_2|=1$, select the single system in $G_2$ as the best. Otherwise, set $h=h(1-\alpha_2,n_1,k')$, where the function $h(\cdot)$ gives Rinott’s constant (see e.g. [@Bechhofer1995 Chapter 2.8]). For each remaining system $i\in G_2$, compute $
N_i=\max\{ n_i(\bar{r}),\lceil (hS_i/\delta)^2 \rceil \},
$ and take an additional $\max\{N_i-n_i(\bar{r}),0\}$ sample observations in parallel. Once a total of $N_i $ replications have been collected in Stages 1 through 3 for each $i\in G_2$, select the system $K$ with the highest ${\bar{X}}(N_K)$ as the best.
Guaranteeing Good Selection
---------------------------
Our probabilistic guarantee on the final solution relies on the following assumption on the distribution of simulation output, which is common to the sequential R&S literature.
\[assum:mvn\] For each system $i = 1,2,\ldots,k$, the simulation output random variables $\{X_{i\ell}, \ell = 1,2,\ldots\}$ are i.i.d. replicates of a random variable $X_i$ having a normal distribution with finite mean $\mu_i$ and finite variance $\sigma^2_i$, and are mutually independent for different $i$.
In §\[sect:RandCompTime\] we gave conditions under which the simulation output generated by parallel cores satisfies the i.i.d. assumption.
We now formally state our good selection guarantee.
\[theorem:GoodSelection\] Under Assumption \[assum:mvn\], GSP selects a system $K$ that satisfies $\mu_k-\mu_K\le\delta$ with probability at least $1-\alpha$.
We discuss the key insights that yield the PGS guarantee here and defer the full proof of Theorem \[theorem:GoodSelection\] to §\[sect:GSPproof\].
First, we show that the best system, System $k$, survives the iterative screening in Stages 1 and 2 with probability at least $1-\alpha_1$, irrespective of whether the best solution is unique or not. Indeed, conditioning on the Stage 1 variance estimates $\{S_i^2:1=1,2,\ldots,k \}$, we can, for any system $i\ne k$, relate the batch statistics $Z_{ki}(r):r=0,1,\ldots, \bar{r}$ to a properly scaled Brownian motion with drift $\mu_k-\mu_i\ge 0$. Then, using the reflection principle of Brownian motion, we can upper-bound the probability that the scaled Brownian motion falls below some number $-a$ before some time $t$, or equivalently, the probability that $Y_{ki}(r)$ falls below $-a_{ki}(\bar{r})$ in some $r$th iteration where $r\le\bar{r}$, which is the criterion used to eliminate system $k$ in the iterative screening stages. The construction of continuation region parameter $\eta$ and the fact that $(n_1-1)S_i^2/\sigma_i^2\sim\chi^2_{n_1-1}$ for all $i$ jointly ensure that the unconditional probability of eliminating $k$ is no greater than $\alpha_1$.
Second, as Stage 3 is closely related to the [@Rinott1978] procedure with confidence level $1-\alpha_2$, it follows from Theorem 1 of [@NM1995] that $$P\left\lbrace\bar{X}_K(N_K) -
\bar{X}_i(N_i) - \left( \mu_K-\mu_i \right)>-\delta, \forall i\in
G_2: i\ne K \right\rbrace\ge 1-\alpha_2$$ where $K$ is the system with the highest sample mean after Stage 3. Therefore we conclude that if $k\in G_2$ (the best system $k$ survives Stages 1 and 2), then $P\left\lbrace\bar{X}_K(N_K) - \bar{X}_k(N_k) - \left( \mu_K-\mu_k
\right)>-\delta \right\rbrace\ge 1-\alpha_2$, that is, Stage 3 selects a good system $K$ such that $\mu_K\ge \mu_k-\delta$ with high probability.
Finally, we complete the proof by invoking a result from [@Nelson2001] that guarantees an overall PGS of $1-\alpha_1-\alpha_2$ for two-stage procedures.
The key difference between the screening methods used in GSP and the $\mathcal{KN}$ family procedures [@KimNels01; @HN2005; @Hong2006] is that the $\mathcal{KN}$ family relies on the PCS assumption ($\mu_k-\mu_i\ge \delta > 0$ for all $i\ne k$) to guarantee PCS, whereas our approach does not. Therefore, GSP works for any indifference-zone parameter $\delta>0$, and when there exist multiple systems $i$ such that $\mu_i\ge\mu_k-\delta$, GSP is guaranteed to select one such system with probability at least $1-\alpha$.
Choice of parameter $\eta$ {#sect:compute_eta}
--------------------------
One way to compute $\eta$, the solution to , is by integrating the LHS using Gauss-Laguerre quadrature and using bisection to find the root of . Alternatively, we may employ a bounding technique to approximate $\eta$ as follows. Indeed, the LHS of is $$\begin{aligned}
E\left[2\bar{\Phi}\left(\eta\sqrt{R}\right)\right]
&=\int_{y=0}^{\infty}2\bar{\Phi}(\eta\sqrt{y}) 2 [1-F_{\chi^2_{n_1-1}}(y)] f_{\chi^2_{n_1-1}}(y)\mathrm{d}y \nonumber
\\&\le \int_{y=0}^{\infty}4\bar{\Phi}(\eta\sqrt{y}) f_{\chi^2_{n_1-1}}(y)\mathrm{d}y \label{eq:CDFUpperBound}
\\&\le
\int_{y=0}^{\infty}4\frac{e^{-\eta^2y/2}}{\eta\sqrt{2\pi y}} \frac{y^{\frac{n_1-1}{2}-1} e^{-y/2}}{ 2^{\frac{n_1-1}{2}} \Gamma(\frac{n_1-1}{2})}\mathrm{d}y \label{eq:PhiBarApprox}
\\&= \frac{4\Gamma(\frac{n_1-2}{2})(\frac{2}{\eta^2+1})^{\frac{n_1-2}{2}}}{\sqrt{2\pi}\eta 2^{\frac{n_1-1}{2}}\Gamma(\frac{n_1-1}{2})}
\int_{0}^{\infty} \frac{(\frac{\eta^2+1}{2})^{\frac{n_1-2}{2}} y^{\frac{n_1-1}{2}-1-\frac{1}{2}} e^{-\frac{\eta^2+1}{2}y}}{\Gamma({\frac{n_1-2}{2}})} \mathrm{d}y\label{eq:GammaPDF}
\\&= \frac{2\Gamma(\frac{n_1-2}{2})}{\sqrt{\pi}\Gamma(\frac{n_1-1}{2})\eta ({\eta^2+1})^{\frac{n_1-2}{2}} }, \label{eq:bound_eta}\end{aligned}$$ where holds because distribution functions are non-negative and is inspired by a similar argument in [@Hong2006], follows from the fact that $\bar{\Phi}(x)\le {e^{-x^2/2}}/({x\sqrt{2\pi}})$ for all $x>0$, and the integrand in is the pdf of a Gamma distribution with shape $(n_1-1)/2$ and scale $2/(\eta^2+1)$, and hence integrates to 1.
Note that is an upper-bound on the left-hand side of . Setting to $1-(1-\alpha_1)^{\frac{1}{k-1}}$ and solving for $\eta$ yields an overestimate $\eta'$, which is more conservative and does not reduce the PGS. Furthermore, as is strictly decreasing in $\eta$, $\eta'$ can be easily determined using bisection.
The parameter $\eta$ determines the value of $a_{ij}(\bar{r})$, and hence how quickly an inferior system is eliminated in screening Steps \[step:Stage1Main\]\[step:NHHStage1Screening\] and \[step:Stage2Main\]\[step:NHHscreening\]. The value of $\eta$ therefore directly impacts the effectiveness of the iterative screening. Hence, it is desirable that $\eta$ does not grow dramatically as the problem gets bigger. Observe that can be further bounded by $$\begin{aligned}
\frac{2\Gamma(\frac{n_1-2}{2})}{\sqrt{\pi}\Gamma(\frac{n_1-1}{2})\eta ^{n_1-1} } :=C\eta^{1-n_1}. \label{eq:crude_eta}\end{aligned}$$ Setting to $1-(1-\alpha_1)^{\frac{1}{k-1}}$ implies that the right-hand side of must be small. After some further manipulations we have $$\begin{aligned}
\log(1-\alpha_1 )=(k-1)\log \left( 1-C\eta^{1-n_1} \right) \approx (k-1)(-C\eta^{1-n_1})
\label{eq:k_eta_relation}\end{aligned}$$ where the approximation holds because $\log(1-\epsilon)\approx -\epsilon$ for small $\epsilon>0$. It follows from that for fixed $\alpha_1$, the parameter $\eta$ grows very slowly with respect to $k$, at a rate of $k^{1/(n_1-1)}$. Therefore, the continuation region defined by $\eta$ and $\bar{r}$ as well as the power of our iterative screening are not substantially weakened as the number of systems increases, especially when $n_1$ or $k$ is large. In this regime, we should expect the total cost of this R&S procedure to grow approximately linearly with respect to the number of systems.
Computational Study {#sect:computation}
===================
In this section, we discuss our parallel computing environment and test problem, discuss two parallel implementations of GSP, and discuss the results of our numerical experiments.
Parallel Computing Environment and Test Problem
-----------------------------------------------
Our numerical experiments are conducted on Extreme Science and Engineering Discovery Environment (XSEDE)’s Stampede high-performance cluster. The Stampede cluster contains over 6,400 computer nodes, each equipped with two 8-core Intel Xeon E5 processors and 32 GB of memory and runs a Linux Centos 6.3 operating system [@TACCStampede]. Parallel programs are submitted through the Simple Linux Utility for Resource Management (SLURM) batch environment which enables users to specify the number of cores to use. The high-performance processors on Stampede are highly reliable and we have never seen a core failure.
We test R&S procedures on a throughput-maximization problem taken from `SimOpt.org`\
[@simoptlib]. In this problem, we solve $$\begin{aligned}
\max_x &\,\, E[g(x;\xi)] \label{eq:TpMaxObj}
\\\text{s.t. } r_1+r_2+r_3&=R \nonumber
\\ b_2+b_3&=B \nonumber
\\ x = (r_1,r_2,r_3,b_2,b_3) &\in \{1, 2, \ldots\}^5 \nonumber\end{aligned}$$ where the function $g(x;\xi)$ represents the random throughput of a three-station flow line with finite buffer storage in front of Stations 2 and 3, denoted by $b_2$ and $b_3$ respectively, and an infinite number of jobs in front of Station 1. The processing times of each job at stations 1, 2, and 3 are independently exponentially distributed with service rates $r_1$, $r_2$ and $r_3$, respectively. The overall objective is to maximize expected steady-state throughput by finding an optimal (integer-valued) allocation of buffer and service rate.
For each choice of the problem parameters $R,B\in \mathbb{Z}^+$, the number of feasible allocations is finite and can be easily computed. We consider three problem instances with very different sizes presented in Table \[tab:TpMax\]. Since the service times are all exponential, we can analytically compute the expected throughput of each feasible allocation by modeling the system as a continuous-time Markov chain. Furthermore, it can be shown that $E[g(r_1,r_2,r_3,b_2,b_3;\xi)] = E[g(r_3,r_2,r_1,b_3,b_2;\xi)] $ for any feasible allocation $(r_1,r_2,r_3,b_2,b_3)$, so the problem may have multiple optimal solutions. Therefore, this is a problem for which the PCS assumption $\mu_k-\mu_{k-1}\ge \delta>0$ can be violated and R&S procedures that only guarantee correct selection might be viewed as heuristics.
[rr r \*[3]{}[C[0.07]{}]{} \*[2]{}[C[0.085]{}]{} \*[1]{}[C[0.07]{}]{}]{} & Number of & Highest & &\
$(R,B)$ & systems $k$ & mean $\mu_k$ & $p=75$ & $p=50$ & $p=25$ & $\delta=0.01$ & $\delta=0.1$ & $\delta=1$\
$(20,20)$ & 3,249 & 5.78 & 3.52 & 2.00 & 1.00 & 6 & 21 & 256\
$(50,50)$ & 57,624 & 15.70 & 8.47 & 5.00 & 3.00 & 12 & 43 & 552\
$(128,128)$ & 1,016,127 & 41.66 & 21.9 & 13.2 & 6.15 & 28 & 97 & 866\
\[tab:TpMax\]
By default, in each simulation replication, we warm up the system for 2,000 released jobs starting from an empty system, before observing the simulated throughput to release the next 50 jobs. This may not be the most efficient way to estimate steady-state throughput compared to taking batch means from a single long run, but it suits our purpose which is to obtain i.i.d. random replicates from the $g(x;\xi)$ distribution in parallel. Due to the fixed number of jobs, the wall-clock time for each simulation replication exhibits low variability.
Parallel Implementations of GSP {#sect:implementation_parallel}
-------------------------------
In this section, we discuss two implementations of GSP proposed in §\[sect:GoodSelectionProcedure\]. Although we will primarily test them on Stampede, both procedures can be configured to run on a wide range of parallel platforms from multi-core personal computers to the Amazon EC2 cloud.
### MPI {#sect:implementation_mpi}
[ Message-Passing Interface (MPI) is a popular distributed-memory parallel programming framework with libraries available in C/C++ and Fortran and the de-facto standard for parallel programming on many high-performance parallel clusters including Stampede. Using MPI, programs operate in an environment where Assumptions 1 and 2 hold. The method by which parallel cores independently execute instructions and communicate through message-passing can be highly customized to serve different purposes. ]{}
The MPI implementation is a realization of GSP presented in §\[sect:GoodSelectionProcedure\], and is fully described in §\[sect:FullImplementation\_MPI\]. We designate one core as the master and let it control other worker cores. We observe that communication is fast on Stampede, taking only $10^{-6}$ to $10^{-4}$ seconds each time depending on the size of the message. Therefore, with an appropriate choice of the batch-size parameter $\beta$, the master remains idle most of the time so the workers are usually able to communicate with the master without much delay.
Our MPI implementation is designed primarily for high-performance clusters like Stampede and does not detect and manage core failures. As simulation output is distributed across parallel cores without backup, the MPI implementation is vulnerable to core failures which may cause loss of data and break the program. [In practice, such failures can occur for a number of reasons, including faulty hardware but also cores aborting due to being re-assigned to higher priority tasks by the system.]{} Therefore, for cheap and less reliable parallel platforms, the MPI implementation needs to be augmented with a “fault-tolerant” mechanism in order to allow the procedure to continue even if some cores fail. This motivates us to seek alternative programming tools such as MapReduce that handle core failures automatically. [Our MPI implementation uses the mvapich2 library and its source code and documentation is hosted in the open-access repository [@MPIRNS]. ]{}
### Hadoop MapReduce {#sect:implementation_hadoop}
MapReduce [@Dean2008] is a distributed computing model typically used to process large amounts of data. Conceptually, each MapReduce instance consists of a *Map* phase where *data entries* are processed by “Mapper” functions in parallel, and a *Reduce* phase where Mapper outputs are grouped by *keys* and summarized using parallel “Reducer” functions.
MapReduce has become a popular choice for data intensive applications such as PageRank and TeraSort, thanks to the following advantages.
- **Simplicity.** The MapReduce programming model allows its users to solely focus on designing meaningful Mappers and Reducers that define the parallel program, without explicitly handling the complex details of the message-passing and the distribution of workload to cores, a task which is completely delegated to the MapReduce package.
- **Portability.** Hadoop is a highly popular and portable MapReduce system that can be easily deployed with minimal changes on a wide range of parallel computer platforms such as the Amazon EC2 cloud.
- **Resilience to core failures.** On less reliable hardware where there is a non-negligible probability of core failure, the Apache Hadoop system can automatically reload any failed Mapper or Reducer job on another worker to guide the parallel job towards completion.
Despite these advantages, the use of MapReduce for computationally intensive and highly iterative applications, such as simulation and simulation optimization, is less documented. Moreover, most popular MapReduce implementations such as Apache Hadoop have limitations that may potentially reduce the efficiency of highly iterative algorithms such as parallel R&S procedures.
- **Synchronization.** By design, each Reduce phase cannot start before the previous Map phase finishes and each new MapReduce iteration cannot be initiated unless the previous one shuts down completely. Hence, a R&S procedure using MapReduce has to be made fully synchronous. If load-balancing is difficult, for instance as a result of random simulation completion times, then core hours could be wasted due to frequent synchronization.
- **Absence of Cache.** In Apache Hadoop, workers are not allowed to cache any information between Map and Reduce phases. As a result, the outputs generated by Mappers and Reducers are often written to a distributed file system (which are usually located on hard drives) before they are read in the next iteration. Compared to the MPI implementation where all intermediate variables are stored in memory, the MapReduce version could be slowed down by repeated data I/O. Moreover, the lack of cache requires the simulation program, including any static data and/or random number generators, to be initialized on workers before every MapReduce iteration.
- **Nonidentical Run Times.** By default, Apache Hadoop does not dedicate each worker to a single task. It may simultaneously launch several Mappers and Reducers on a single worker, run multiple MapReduce jobs that share workers on the same cluster, or even use workers that have different hardware configurations. In any of theses cases, simulation completion times may be highly variable and time-varying. Therefore, Stage 0 of GSP (estimation of simulation run time) is dropped from our MapReduce implementation.
[|>p[0.25]{}|>p[0.35]{}|>p[0.35]{}|]{} Task & MPI & Hadoop MapReduce\
Master & Explicitly coded & Automated\
Message-passing & Explicitly coded & Automated\
Synchronization & Once after each stage & More frequent: required in every\
&&iteration of Stage 2\
Simulation & Each worker simulates one system & Each worker simulates multiple\
&per iteration &systems per iteration\
Load-balancing & Via asynchronous & By assigning approximately equal\
&communications between the & number of systems (Mappers) to each\
&master and a single worker & worker in each synchronized iteration\
Batch statistics and & Always stored in workers’ memory & Written to hard disk after each\
random number seeds &&iteration\
\[tab:MPIvHadoop\_Implementations\]
Although there are specialized MapReduce variants such as “stateful MapReduce” that attempt to address these limitations [@Elgohary2012], we do not explore them as they are less available for general parallel platforms, at least at present. [However, some of these limitations (such as the lack of caching across multiple MapReduce rounds) are idiosyncratic to specific packages like Hadoop rather than the framework itself. Nevertheless, the a priori expectation is that, for a highly iterative procedure like ours, a highly optimized MPI approach will outperform a Hadoop one; thus our question is not which is fastest, but whether MapReduce can offer most of the speed of MPI along with its advantages discussed above.]{}
We propose a variant of GSP using iterative MapReduce as follows. In each Mapper function, we treat each surviving system as a single data entry, obtain an additional batched sample, and output updated summary statistics such as sample sizes, means, and variances. Each output entry is associated with a key which represents the screening group to which it belongs. Once output entries of Mappers are grouped by their keys, each Reducer receives a group of systems, screens amongst them, and writes each surviving system as a new data entry which in turn is used as the input to the next Mapper.
To fully implement GSP, MapReduce is run for several iterations. The first iteration implements Stage 1, where both ${\bar{X}}_i(n_1)$ and $S^2_i$ are collected. Then, a maximum number of $\bar{r}$ subsequent iterations are needed for Stage 2, with only $n_i(r)$ and ${\bar{X}}_i(n_i(r))$ being updated in each iteration. (Additional MapReduce iterations can be run where the best system from each group is shared for additional between-group screening.) The same Reducer can be applied in both Stages 1 and 2, as the screening logic is the same. Finally, a Stage 3 MapReduce features a Mapper that calculates the additional Rinott sample size, simulates the required replications, and a different Reducer that simply selects the system with the highest sample mean at the end. Full details of each stage are provided in §\[sect:FullImplementation\_Hadoop\].
Our MapReduce implementation is based on the native Java interface for MapReduce provided in Apache Hadoop 1.2.1. It is hosted in the open-access repository [@MAPREDRNS]. Table \[tab:MPIvHadoop\_Implementations\] summarizes some of the major differences between the MPI and Hadoop implementations.
Numerical Experiments
---------------------
We now demonstrate the practical performance of GSP by using it to solve the throughput maximization test problem.
### GSP vs Existing Parallel Procedures {#sect:GSPvNHHvNSGS}
[lllrr]{} Configuration & $\delta$ & Procedure & Wall-clock time (sec) & Total number of\
& & & & simulation replications ($\times 10^6$)\
3,249 systems & 0.01 & GSP & 14 & 2.3\
on 64 cores & & NHH & 14 & 2.5\
& & [NSGS$_p$]{} & 120 & 13\
& 0.1 & GSP & 3.4 & 0.57\
& & NHH & 2.6 & 0.44\
& & [NSGS$_p$]{} & 3.4 & 0.48\
57,624 systems & 0.01 & GSP & 720 & 130\
on 64 cores & & NHH & 520 & 89\
& & [NSGS$_p$]{} & 11,000 & 1600\
& 0.1 & GSP & 60 & 10\
& & NHH & 71 & 12\
& & [NSGS$_p$]{} & 150 & 23\
1,016,127 systems & 0.1 & GSP & 260 & 320\
on 1,024 cores & & NHH & 1,000 & 430\
& & [NSGS$_p$]{} & 1,400 & 1900\
\[tab:GSPvNHHvNSGS\]
GSP is motivated by an earlier computational study by [@Ni2014], which compares the performance of two parallel procedures, NHH [@Ni2013] and [NSGS$_p$]{} [@Ni2014]. NHH is a parallel procedure that adopts the fully-sequential serial procedure proposed in [@Hong2006], and only provides a PCS guarantee [@HN2014]. For problems where multiple optimal solutions exist, it is used as a heuristic because the PCS assumption does not hold. NHH can be viewed as a variant of GSP using a different screening method and without a Rinott-like Stage 3. [NSGS$_p$]{} is a parallel implementation of the NSGS procedure [@Nelson2001], and is a simplification of GSP without the iterative screening Stage 2.
We implement all three procedures using MPI and test them on different instances of the throughput maximization problem. We measure the performance of these procedures in terms of total wall-clock time and simulation replications required to find a solution, and report them in Table \[tab:GSPvNHHvNSGS\]. Preliminary runs on smaller test problems suggest that the variation in these two measures between multiple runs of the entire selection procedure are limited. Therefore we only present results from a single replication to save core hours.
[@Ni2014] argue that NHH tends to devote excessive simulation effort to systems with means that are very close to the best, whereas [NSGS$_p$]{} has a weaker screening mechanism but its Rinott stage can be effective when used with a large $\delta$, which is associated with higher tolerance of an optimality gap. GSP, by design, combines iterative screening with a Rinott stage. Like [NSGS$_p$]{}, we expect that GSP will cost less with a large $\delta$ as the Rinott sample size is $O(1/\delta^2)$, but its improved screening method should eliminate more systems than [NSGS$_p$]{}before entering the Rinott stage. Therefore, we expect GSP to work particularly well when a large number of systems exist both inside and outside the indifference zone. This intuition is supported by the outcomes of the medium and large test cases with $\delta=0.1$, when GSP outperforms both NHH and [NSGS$_p$]{}.
### A Comparison of MPI and Hadoop Versions of GSP
---------------- --------- --------- ----------------- ------------ --------------------- ------------- -----
Configuration $\beta$ Version Number of Wall-clock Utilization
replications time Simulation Screening
($\times 10^6$) (sec) ($\times 10^3$ sec) (sec) %
3,249 systems 100 HADOOP 0.46 460 0.34 0.14 1.2
on 64 cores MPI 0.50 3.0 0.18 0.01 94
200 HADOOP 0.63 280 0.41 0.10 2.3
MPI $0.69$ 4.1 0.25 0.01 95
57,624 systems 100 HADOOP 8.8 550 5.1 1.9 15
on 64 cores MPI $9.1$ 53 3.3 0.89 98
200 HADOOP 12 410 7.0 1.7 27
MPI $13$ 75 $4.7$ 0.83 98
1,016,127 100 HADOOP 280 1300 160 120 12
systems MPI 320 120 110 30 91
on 1,024 cores 200 HADOOP 340 810 190 89 23
MPI 380 140 140 29 97
\[-5ex\]
---------------- --------- --------- ----------------- ------------ --------------------- ------------- -----
: A comparison of two implementations of GSP using parameters $\delta=0.1$, $n_0=50$, $\alpha_1=\alpha_2=2.5\%$, $\bar{r} = 1000/\beta$. “Total time” is summed over all cores. (Results to 2 significant figures)
\[tab:MPIvHADOOP\]
We now focus on GSP and test its two implementations discussed in §\[sect:implementation\_parallel\]. Since Stage 0 is not included in the MapReduce implementation, we also remove it from the MPI version to have a fair comparison. Both procedures are tested on Stampede. While the cluster features highly optimized C++ compilers and MPI implementations, it provides relatively less support for MapReduce. Our MapReduce jobs are deployed using the myhadoop software [@MYHADOOP], which sets up an experimental Hadoop environment on Stampede.
[ Another difference is that we perform less screening in MPI than in Hadoop. In our initial experiments, we observed that the master could become overwhelmed by communication with the workers in the screening stages, and we fixed this problem by screening using only the $20$ best systems from other cores, versus the best systems from [*all*]{} other cores in Hadoop. While less screening is not a non-negligible effect, it will be apparent in our results that it is dominated by the time spent with simulation.]{}
Before we proceed to the results, we define core utilization, an important measure of interest, as $$\begin{aligned}
\text{Utilization}=\frac{\text{total time spent on simulation}}{\text{wall-clock time }\times\text{ number of cores}}.\end{aligned}$$ Utilization measures how efficiently the implementations use the available cores to generate simulation replications. The higher the utilization, the less overhead the procedure spends on communication and screening.
In Table \[tab:MPIvHADOOP\] we report the number of simulation replications, wall-clock time, and utilization for each of the GSP implementations. The MPI implementation takes substantially less wall-clock time than MapReduce to solve every problem instance, although it requires slightly more replications due to its asynchronous and distributed screening. The gap in wall clock times narrows as the batch size $\beta$ and/or the system-to-core ratio are increased. Similarly, the MPI implementation also yields much higher utilization, spending more than 90% of the total computation time on simulation runs in all problem instances. Compared to the MPI implementation, the MapReduce version utilizes core hours less efficiently but again its utilization significantly improves as we double batch size and increase the system-to-core ratio.
To further understand the low utilization, we give the number of active Mapper and Reducer jobs over an entire MapReduce run in Figure \[fig:hadoop\_profile\]. The plot reveals a number of reasons for low utilization. First, there are non-negligible gaps between Map and Reduce phases, which are due to an intermediary “Shuffle” step that collects and sorts the output of the Mappers and allocates it to the Reducers. Second, as the amount of data shuffled is likely to vary, the Reducers start and finish at different times. Third, owing to the varying amount of computing required for different systems, some Mappers take longer than others. In all, the strictly synchronized design of Hadoop causes some amount of core idleness that is perhaps inherent in the methodology, and therefore unavoidable. Nevertheless, the fact that utilization increases as average batch size $\beta$ or the system-to-core ratio increases suggests that the Hadoop overhead becomes less pronounced as the amount of computation work per Mapper increases. Therefore we expect utilization to also improve and become increasingly competitive with MPI’s for problems that feature a larger solution space or longer simulation runs.
### Robustness to Unequal and Random Run Times
The MapReduce implementation allocates approximately equal numbers of simulation replications to each Mapper and the simulation run times per replication are nearly constant for our test problem, so the computational workload in each MapReduce iteration should be fairly balanced. Indeed, in Figure \[fig:hadoop\_profile\] we observe that Mapper jobs terminate nearly simultaneously, which suggests that load-balancing works well. However, if the simulation run times exhibit enough variation that one Mapper takes much longer than the others, then we would expect synchronization delays that would greatly reduce utilization.
[lrrlrrr]{} Configuration & $\mu$ & $\sigma^2$ & Version & Wall-clock time (sec) & Utilization %\
3,249 systems & 7.4 & 0.5 & HADOOP & 280 & 2.3\
on 64 cores & & & MPI & 4.2 & 94\
& 6.6 & 2.0 & HADOOP & 280 & 2.0\
& & & MPI & 4.0 & 93\
57,624 systems & 7.4 & 0.5 & HADOOP & 400 & 27\
on 64 cores & & & MPI & 74 & 98\
& 6.6 & 2.0 & HADOOP & 400 & 26\
& & & MPI & 70 & 98\
1,016,127 systems & 7.4 & 0.5 & HADOOP & 850 & 25\
on 1,024 cores & & & MPI & 150 & 97\
& 6.6 & 2.0 & HADOOP & 850 & 22\
& & & MPI & 150 & 97\
\
\[-5ex\]
\[tab:MPIvHADOOP\_uneven\]
To verify this conjecture, we design additional computational experiments where variability in simulation run times is introduced by warming up each system for a random number $W$ of job releases (by default, we use a fixed 2,000 job releases in the warm-up stage). We take $W$ to be (rounded) log-normal, parameterized so that the average warm-up period is approximately 2,000, in the hope that the heavy tails of the log-normal distribution will lead to occasional large run times that might slow down the entire procedure. We also truncate the log-normal distributions from above at 20,000 job releases to avoid exceeding a built-in timeout limit in Hadoop. Parameters of the truncated log-normal distribution and the results of the experiment are given in Table \[tab:MPIvHADOOP\_uneven\].
We observe very similar wall-clock time and utilization in all instances compared to the base cases in Table \[tab:MPIvHADOOP\] where we used fixed warm-up periods. Both implementations seem quite robust against the additional randomness in simulation times, despite our intuition that the MapReduce version might be noticeably impacted due to additional synchronization waste. A potential explanation is that as each core is allocated at least 50 systems and each system is simulated for an average of 200 replications in each step, the variation in single-replication completion times is averaged out. Rather extreme variations would be required for MapReduce to suffer a sharp performance decrease. For problems with much longer simulation times and a lower systems-to-core ratio, the averaging effect might not completely cancel the variations across simulation run times.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was partially supported by NSF grant CMMI-1200315, and used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575.
**[Appendices]{}**
Proof of Theorem \[theorem:GoodSelection\]. {#sect:GSPproof}
===========================================
Proving Theorem \[theorem:GoodSelection\] requires the following lemmas, where we use $B_\Delta(\cdot)$ to denote a Brownian motion with drift $\Delta$ and volatility one.
(@Hong2006 [Theorem 1]) \[lemma:HongBMdistribution\] Let $m(r)$ and $n(r)$ be arbitrary nondecreasing integer-valued functions of $r=0,1,\ldots$ and $i$, $j$ be any two systems. Define $Z(m,n):=\left[\sigma_i^2/m + \sigma_{j}^2/n\right]^{-1} [ {\bar{X}}_i(m) - {\bar{X}}_{j}(n)]$ and $Z'(m,n):=B_{\mu_i-\mu_{j}}([\sigma_i^2/m + \sigma_{j}^2/n]^{-1})$. Then the random sequences $\{Z(m(r),n(r)):r=0,1,\ldots\}$ and $\{Z'(m(r),n(r)):r=0,1,\ldots\}$ have the same joint distribution.
\[lemma:lemma:bound\_a\_t\] Let $i\ne j$ be any two systems. Define $\tilde{a}_{ij}(\bar{r}):= \min\{S^2_i/\sigma^2_i,S^2_{j}/\sigma^2_{j} \}a_{ij}(\bar{r})$ and $\tilde{t}_{ij}(\bar{r}):= \min\{S^2_i/\sigma^2_i,S^2_{j}/\sigma^2_{j} \}\tau_{ij}(\bar{r})$. It can be shown [@Hong2006] that $\min\{S^2_i/\sigma^2_i,S^2_{j}/\sigma^2_{j} \} \le {t_{ij}(r)}/{\tau_{ij}(r)}$ for all $r\ge 0 $ regardless of the sampling rules $n_i(\cdot)$ and $n_{j}(\cdot)$. Therefore $
\tilde{a}_{ij}(\bar{r})
\le {t_{ij}(r)}a_{ij}(\bar{r})/{\tau_{ij}(r)}
\text{ and }
\tilde{t}_{ij}(\bar{r})
\le {t_{ij}(r)}\tau_{ij}(\bar{r})/{\tau_{ij}(r)}
\label{eq:tilde_a}
$ regardless of the sampling rules $n_i(\cdot)$ and $n_{j}(\cdot)$ for all $r\ge 0$.
(@Hong2006 [Lemma 4]) \[lemma:smallContRegion\] Let $g_1(\cdot)$, $g_2(\cdot)$ be two non-negative-valued functions such that $g_2(t')\ge g_1(t')$ for all $t'\ge 0$. Define symmetric continuations $C_m:=\{(t',x):-g_m(t')\le x \le g_m(t') \}$ and let $T_m:=\inf\{t':B_\Delta(t')\not\in C_m\}$ for $m=1,2$. If $\Delta\ge 0$, then $P[B_\Delta(T_1)<0]\ge P[B_\Delta(T_2)<0]$.
\[lemma:HallErrorProb\] By the reflection principle of Brownian motion, $P[\min_{0\le t'\le t}B_0(t')<-a]=2P[B_0(t)<-a]=2\bar{\Phi}(a/\sqrt{t})$ for all $a,t>0$.
(@Tamhane1977) \[lemma:TamhaneIneq\] Let $V_1, V_2,\ldots, V_k$ be independent random variables, and let $G^w (v_1 , v_2 ,\ldots , v_k )$, $j = 1, 2, \ldots, p$, be non-negative, real-valued functions, each one nondecreasing in each of its arguments. Then $$E\left[\prod_{j=1}^p G^w(V_1 , V_2 , \ldots, V_k )\right]\ge \prod_{j=1}^{p} E[G^w(V_1, V_2 , \ldots, V_k)].$$
(After @NM1995 and @Nelson2001 [Lemma 1]) \[lemma:secondstagePGS\] For any $G_2\subseteq \mathcal S$, Stage 3 guarantees to select a system $K\in G_2$ such that $\Pr \left[\max_{i\in G_2} \mu_i-\mu_K\le\delta\right]\ge 1-\alpha_2$. If, in addition, Stages 1 and 2 jointly guarantee that $\Pr[k\in G_2]\ge 1-\alpha_1$, then $$\Pr \left[\text{The procedure selects system }K: \mu_k-\mu_K\le \delta\right] \ge 1-\alpha_1-\alpha_2.$$
First, note that for any system $i$, it is well known [@Casella2002 page 218] that $\bar{X}_i(n_1)|S_i^2$ is normally distributed and $X_{i\ell}$ is independent of $S_i^2$ for all $\ell>n_1$. Furthermore, $\bar{T}_i$ is obtained in Stage 0 independently of all $X_{i\ell}$’s. Therefore, choosing the sampling rule based on $\bar{T}_i$ and $S_i^2$ does not affect the normality of the $\{\bar{X}_i(n_i(r)):r=0,1,\ldots,\bar{r}\}$ sequence.
For any two systems $i$ and $j$, let $KO_{ij}$ be the event that system $i$ eliminates system $j$ in Stages 1 or 2. It then follows that $$\begin{aligned}
\Pr&[KO_{ik} \text{ in Stages~1~or~2}]
&\\=&E[\Pr[KO_{ik} \text{ in Stages~1~or~2}|S^2_k,S^2_i]]
&\\\le&E[\Pr[Y_{ki}(\tau_{ki}(r))<-a_{ij}(\bar{r})
\text{ for some } r \le \bar{r}|S^2_k,S^2_i]]
\\&\text{ since system $i$ could be eliminated by some other system before it can eliminate system $k$}
&\\=&E[\Pr[Y_{ki}(\tau_{ki}(r))<-a_{ij}(\bar{r})
\text{ and } \tau_{ki}(r)\le \tau_{ki}(\bar{r})
\text{ for some } r |S^2_k,S^2_i]]
&\\=&E[\Pr[Z_{ki}(t_{ki}(r)) <-\frac{t_{ki}(r)}{\tau_{ki}(r)}a_{ij}(\bar{r})
\text{ and } t_{ki}(r)\le \frac{t_{ki}(r)}{\tau_{ki}(r)}\tau_{ki}(\bar{r}) \text{ for some } r|S^2_k,S^2_i]]
&\\=&E[\Pr[B_{\mu_k-\mu_i}(t_{ki}(r)) < -\frac{t_{ki}(r)}{\tau_{ki}(r)}a_{ij}(\bar{r}) \text{ and } t_{ki}(r)\le \frac{t_{ki}(r)}{\tau_{ki}(r)}\tau_{ki}(\bar{r}) \text{ for some } r|S^2_k,S^2_i]]
\text{ by Lemma~\ref{lemma:HongBMdistribution}}
\\\le&E[\Pr[B_{\mu_k-\mu_i}(t_{ki}(r)) < -\tilde{a}_{ij}(\bar{r}) \text{ and } t_{ki}(r)\le\tilde{t}_{ij}(\bar{r}) \text{ for some } r|S^2_k,S^2_i]]
\text{ by Lemmas~\ref{lemma:lemma:bound_a_t}~and~\ref{lemma:smallContRegion}
}
\\\le&E[\Pr[B_{\mu_k-\mu_i}(t) < -\tilde{a}_{ij}(\bar{r}) \text{ for some } t\le\tilde{t}_{ij}(\bar{r}) |S^2_k,S^2_i]]
\\\le&E[\Pr[B_{0}(t) < -\tilde{a}_{ij}(\bar{r}) \text{ for some } t\le\tilde{t}_{ij}(\bar{r})|S^2_k,S^2_i]]
\text{ since } \mu_k\ge \mu_i
\\=&
E\left[2\bar{\Phi}\left(\frac{\tilde{a}_{ij}(\bar{r})}{\sqrt{\tilde{t}_{ij}(\bar{r})}}\right)\right]
\text{ by Lemma~\ref{lemma:HallErrorProb}}
\\=& E\left[2\bar{\Phi}\left(\frac{a_{ij}(\bar{r})}{\sqrt{\tau_{ij}(\bar{r})(n_1-1)}}\sqrt{\min \left\lbrace \frac{(n_1-1)S_i^2}{\sigma^2_i},\frac{(n_1-1)S_k^2}{\sigma^2_k} \right\rbrace}\right)\right]
\\=& E\left[2\bar{\Phi}\left(\eta\sqrt{\min \left\lbrace \frac{(n_1-1)S_i^2}{\sigma^2_i},\frac{(n_1-1)S_k^2}{\sigma^2_k} \right\rbrace}\right)\right] \text{ by choice of $a_{ij}(\bar{r})$}
\\=&1-(1-\alpha_1)^{\frac{1}{k-1}}
\mbox{ by \eqref{eq:Choice_eta}, since } {(n_1-1)S_i^2}/{\sigma^2_i} \text{ and } {(n_1-1)S_k^2}/{\sigma^2_k} \text{ are i.i.d. } \chi^2_{n_1-1} \text{ random variables}.
\end{aligned}$$ Then, noting that simulation results from different systems are mutually independent, we have $$\begin{aligned}
\Pr [\text{system } k \in {\mathcal S}_1]
&= E\left[\Pr \left\lbrace\bigcap_{i=1}^{k-1} \overline{KO}_{ik}|X_{k1},X_{k2},\ldots\right\rbrace\right]
\\&=E\left[ \prod_{i=1}^{k-1} \Pr \left\lbrace \overline{KO}_{ik}|X_{k1},X_{k2},\ldots\right\rbrace \right]
\\&\ge \prod_{i=1}^{k-1} E\left[ \Pr \left\lbrace \overline{KO}_{ik}|X_{k1},X_{k2},\ldots\right\rbrace \right]
\text{ by Lemma~\ref{lemma:TamhaneIneq} }
\\&= \prod_{i=1}^{k-1} \Pr\left[ \overline{KO}_{ik} \right]
\ge \prod_{i=1}^{k-1} \left[1-\left(1-(1-\alpha_1)^{\frac{1}{k-1}}\right)\right] = 1-\alpha_1.
\end{aligned}$$
Finally, we invoke Lemma \[lemma:secondstagePGS\] to complete the proof.
Full Description of the MPI implementation {#sect:FullImplementation_MPI}
==========================================
The purpose of this section is to provide additional insight into our parallel codes. In Figures \[fig:master\_worker\_algo\_1\] through \[fig:master\_worker\_algo\_3\] we demonstrate in greater detail how the master core allocates and distributes systems, how random number streams are created and distributed together with the assigned systems to ensure independent sampling, and how simulation results are communicated between cores. We use the following notation for some subroutines in Figures \[fig:master\_worker\_algo\_1\] through \[fig:master\_worker\_algo\_3\]:
`Partition`(${\mathcal{S}}$, *Stage*) The master divides the set of systems ${\mathcal{S}}$ into disjoint partitions $\{G^w_{\textit{Stage}}:w=1,2,\ldots,c \}$:\
In *Stage 0*, all systems are simulated for $n_0$ replications to estimate simulation completion time. The master randomly permutes ${\mathcal{S}}$ (in case of long runtimes for some systems that are indexed closely) and assigns approximately equal numbers of systems to each $G^w_0$.\
In *Stage 1*, a fixed number $n_1$ of replications are required from each system. To balance the simulation work among workers, the master chooses $G_1^w$ such that the estimated completion time $\sum_{i\in G_1^w} n_1\bar{T}_i/n_0$ is approximately equal for all $w$.\
In *Stage 2*, both simulation and screening are performed iteratively. Simulation of a system is no longer dedicated to a particular worker, and $G^w_2$ is the set of systems that worker $w$ needs to screen. To load-balance the screening work, the master assigns approximately equal numbers of systems to each $G^w_2$.
`Collect`(*info*) The master collects *info* from all workers for all systems, in arbitrary order.
`Simulate`$(i,n,\text{\textit{info}})$ Worker $w$ simulates system $i$ for $n$ replications and records *info* using the next subsubstream in $U_w^i$.
$\texttt{Stat}_{i,r}$ The batch statistics for the $r$th batch of system $i$. This includes sample size $n_i(r)$ and sample mean ${\bar{X}}_{i}(n_i(r))$ as described in §\[sect:GoodSelectionProcedure\].
`BatchSize`($i,\beta$) The master calculates batch size $b_i$ system $i$ used in Stage 2. Following the recommendation from §\[sect:BatchSizes\], we let $$\begin{aligned}
b_i=\left \lceil \frac{S_{i} \sqrt{T_{i}}}{\frac{1}{|{\mathcal{S}}|}\sum_{j\in{\mathcal{S}}}S_{j} \sqrt{T_{j}}} \beta\right \rceil
\end{aligned}$$ where $\beta$ is a pre-determined average batch size.
`Screen`($G^w,r_0,r_1,$ *useothers*) Screen systems in $G^w$ from batches $r_0$ through $r_1$ inclusive. It can be checked that worker $w$ has received $\texttt{Stat}_{i,r}$ for all $i\in{\mathcal{S}}$, all $r\le r_1$ and stored the data in its memory.
A system $i\in G^w$ is eliminated if there exists system $j\in G^w_2:j\neq i$ and some $r':r_0\le r'\le r_1$ such that $r'\le \bar{r}$ and $
Y_{ij}(r')
<
-a_{ij}(\bar{r})
$ where $Y_{ij}$ and $a_{ij}$ are defined in §\[sect:GoodSelectionProcedure\].
In addition, if *useothers*$=true$ and ${\mathcal{W}}\ne\emptyset$, then for each $w'\in {\mathcal{W}}$ the worker also screens the systems in $G^w$ against system $i^*_{w'}$, the best system from worker $w'$, using batch statistics $\{\texttt{Stat}_{i^*_{w'},r'}:r'\le r_{w'} \}$ up to batch $\min\{ r_{w'}, r_1\}$.
`SendScreen(G^w,r_w)` and `RecvScreen(w)` Worker $w$ sends $r_w$ and screening results (updated $G^w$) to the master, which then updates $G^w$ and ${\mathcal{S}}$ on its own memory accordingly. The master also receives $r_w$ and lets $r_w^\text{screened}\leftarrow r_w$.
`Communicate`$()$ Worker sends a signal to master and waits for the master to receive the signal, before proceeding.
`SendSim(w,i,q_i,b_i)` and `RecvSim()` The master instructs worker $w$ to simulate the $q_i$th batch of system $i$, for $b_i$ replications. Worker $w$ receives $i$, $q_i$, $b_i$ from the master.
`SendOutput`$(i,q_i,\texttt{Stat}_{i,q_i})$ and `RecvOutput`$(w)$ Worker $w$ sends simulation output $\texttt{Stat}_{i,q_i}$ for the $q_i$th batch of system $i$ to the master. The master stores $\texttt{Stat}_{i,q_i}$ in memory.
`SendBest()` and `RecvBest(w)` Worker $w$ sends its estimated-best system $i^*_w$ (the one in $G^w$ with the highest batch mean) to the master, together with all batch statistics for system $i^*_w$, $\{\texttt{Stat}_{i^*_w,r}: r\le r_w \} $; the master receives $r_w$ and lets $r_w^\text{received}\leftarrow r_w$.
`CountBatch`$(w)$ The master finds the largest $r^\text{current}\ge r_w$ such that $\texttt{Stat}_{i,r}$ for all $i\in G^w$, $r_w<r\le r^\text{current}$ have been received by the master.
`SendAction(w,\text{flag}_w)` and `RecvAction()` The master sends an indicator $\text{flag}_w$ to worker $w$, where $\text{flag}_w=1$ indicates “simulate a batch” and $\text{flag}_w=2$ indicates “perform screening”.
`SendStats(w)` and `RecvStats()` The master sends $\texttt{Stat}_{i,r}$ for all $i\in G^w$, $r_w<r\le r^\text{current}$ to worker $w$; the worker receives $r^\text{current}$ and lets $r^\text{new} \leftarrow r^\text{current}$; the worker should have $\texttt{Stat}_{i,r}$ for all $i\in G^w$, $0<r\le r^\text{new}$ upon completion.
`SendBestStats(w)` and `RecvBestStats()` The master computes ${\mathcal{W}}=\{w'\ne w: |G_{w'}|>0) \}$ and sends ${\mathcal{W}}$ to worker $w$; the master then sends all available batch statistics for best systems, $\{\texttt{Stat}_{i^*_{w'},r}:{w'}\in {\mathcal{W}}, r\le r_{w'}^\text{received} \}$, to worker $w$.
Full Description of the Hadoop implementation {#sect:FullImplementation_Hadoop}
=============================================
We present in this section the full details of the MapReduce implementation of GSP.
Each Mapper reads a comma-separated string of varied length, denoted by $[\text{value 1}, \text{value 2},\ldots, \$\texttt{type}]$, where the last component $\$\texttt{type}$ is used to indicate the specific information captured in the string. A Mapper usually runs some simulation, updates batch statistics, and generates one or more `key`$: $ {`value`} pairs. All pairs under the same `key` are sent to the same Reducer, which is typically responsible for screening. A Reducer may generate one or more comma-separated strings which become the input to the Mapper in the next iteration.
Each system $i$ is coupled with `stream`$_i$ which is used by some random number generator and updated each time a random number is generated. The coupling of systems and `stream`s ensures that the random numbers generated for each system in each iteration are all mutually independent. We also assume that each system $i$ is preallocated to a particular screening group, as determined by the function `Group`$(i)$.
The procedure begins with Steps 1-3 which implements Stage 1, then enters Stage 2 where Steps 4 and 5 are run repeatedly for a maximum of $\bar{r}$ iterations. If multiple systems survive Stage 2, the procedure runs Steps 6 and 7 to finish Stage 3.
1. - **Map**: Estimate $S_i^2$
1. 1. 2. Initialize `stream`$_i$ with seed $i$; Simulate system $i$ for $n_1$ replications to obtain ${\bar{X}}_i(n_1)$ and $S_i^2$.
3. $i$: {${\bar{X}}_i(n_1)$, $S_i^2$, `stream`$_i$, \$S0}
- **Reduce**
1. 1. $i$: {${\bar{X}}_i(n_1)$, $S_i^2$, `stream`$_i$, \$S0}
2. Calculate $\sum_i S_i$.
3.
2. - **Map**: Calculate batch size
1. 1. 2. Calculate batch size $b_i$ using $b_i=\beta S_i/(\sum_i S_i/k)$.
3. `Group`($i$): {$i$, ${\bar{X}}_i(n_1)$, $n_1$, $b_i$, $S_i^2$, `stream`$_i$, \$Sim}
- **Reduce**: Screen within a group
1. 1. `Group`: {$i$, ${\bar{X}}_i(n_i)$, $n_i$, $b_i$, $S_i^2$, $_i$, \$Sim} for all $i$ in the `Group`
2. Screen all systems in the `Group` and find the one $i^*$ with the highest mean.
3. for each surviving system $i$, and \[$i^*$, ${\bar{X}}_{i^*}(n_{i^*})$, $n_{i^*}$, $b_{i^*}$, $S_{i^*}^2$, \$Best\] for the best system $i^*$
3. - **Map**: Share best systems between groups
1. 1. 2. Simply output to `Group`($i$).
3. `Group`($i$): {$i$, ${\bar{X}}_i(n_i)$, $n_i$, $b_i$, $S_i^2$, `stream`$_i$, \$Sim}
4. 5. Output to all groups.
6. `Group`: {$i^*$, ${\bar{X}}_{i^*}(n_{i^*})$, $n_{i^*}$, $b_{i^*}$, $S_{i^*}^2$, \$Best} for every `Group`
- **Reduce**: [Screen against the best systems from other groups]{}
1. 1. `Group`: {$i$, ${\bar{X}}_i(n_i)$, $n_i$, $b_i$, $S_i^2$, $_i$, \$Sim} for all $i$ in the `Group`, and `Group`: {$i^*$, ${\bar{X}}_{i^*}(n_{i^*})$, $n_{i^*}$, $b_{i^*}$, $S_{i^*}^2$, \$Best} from every other `Group`
2. Screen all systems in `Group` against the best systems from other groups.
3. for each surviving system $i$
4. - **Map**: Simulation
1. 1. 2. Simulate system $i$ for additional $b_i$ replications, update $n_i$, ${\bar{X}}_i(n_i)$, and `stream`$_i$.
3. `Group`($i$): {$i$, ${\bar{X}}_i(n_1)$, $n_1$, $b_i$, $S_i^2$, `stream`$_i$, \$Sim}
- **Reduce**: Screen within a group.
(Same as Step 2 Reduce.)
5. Screen against best systems from other groups.
(Same as Step 3.)
6. - **Map**: Determine Rinott sample sizes
1. 1. 2. Output to Reducer.
3. $i$: {${\bar{X}}_i(n_i)$, $n_i$, $S_i^2$, `stream`$_i$, \$Sim}
- **Reduce**
1. 1. $i$: {${\bar{X}}_i(n_i)$, $n_i$, $S_i^2$, `stream`$_i$, \$Sim}
2. Calculate Rinott sample size and divide the additional sample into batches. For each batch $j$, generate a substream `stream`$^j_i$ using `steam`$_i$.
3. , and for each batch $j$: \[$i$, `stream`$^j_i$, (size of batch $j$), \$S3\]
7. \[last-item\]
- **Map**: Simulate additional batches
1. 1. 2. Output to Reducer, since this is the batch statistics generated in Stage 2.
3. $1$: {$i$, ${\bar{X}}_i(n_i)$, $n_i$, \$S2}
4. 5. Simulate batch $j$ of system $i$ for the given batch size using `stream`$^j_i$, calculate batch sample mean ${\bar{X}}_i^j$.
6. $1$: {$i$, ${\bar{X}}_i^j$, (size of batch $j$), \$S3 }
- **Reduce**: Merge batches and find the best system
1. 1. (This step has only one Reducer) 1: {$i$, ${\bar{X}}_i(n_i)$, $n_i$, \$S2} and 1: {$i$, ${\bar{X}}_i^j$, (size of batch $j$), \$S3} for all system $i$ and all batch $j$
2. For each system $i$, merge all batches (including the one from Stage 2) to form a single sample mean.
3. Report the system $i^*$ that has the highest sample mean.
| {
"pile_set_name": "ArXiv"
} |
0.0in 0.0in 6.5in 9.2in -3 cm .75in 1.2ex
0.7cm
[**Gluon Radiation Off Scalar Stop Particles**]{}\
W. Beenakker, R. Höpker and P. M. Zerwas\
Deutsches Elektronen-Synchrotron DESY, D-22603 Hamburg, FRG\
ABSTRACT\
We present the distributions for gluon radiation off stop-antistop particles produced in $e^+e^-$ annihilation: $e^+e^- \to
\tilde t\, \bar{\tilde t} \,g$. For high energies the splitting functions of the fragmentation processes $\tilde t \to \tilde t \,g$ and $g \to \tilde t
\,\bar{\tilde t}$ are derived; they are universal and apply also to high-energy stop particles produced at hadron colliders.
#### Introduction.
Stop particles are exceptional among the supersymmetric partners of the standard-model fermions. Since the top quarks are heavy, the masses of the two stop particles $\tilde t_1$ and $\tilde t_2$, mixtures of the left (L) and right (R) squarks, may split into two levels separated by a large gap [@ellis]-[@olive]. The mass of the lightest eigenstate $\tilde t_1$ could be so low that the particle may eventually be accessible at the existing $p\bar p$ and even $e^+e^-$ storage rings. So far the result of search experiments at $e^+e^-$ colliders [@tristan; @opal] has been negative and a lower limit of $45.1$ GeV has been set at LEP [@opal] for the L/R mixing angle outside the band of $\cos^2\theta_t$ between 0.17 and 0.44 and for a mass difference between the $\tilde t_1$ and the lightest neutralino $\tilde\chi_1^0$ of more than $5$ GeV. The higher energy at LEP2 and dedicated efforts at the Tevatron will open the mass range beyond the current limits soon.
To begin, we briefly summarize the well-known theoretical predictions for the cross section of the production process \[Fig.\[feyn\](a)\] $$e^+\,e^- \to \tilde t_1\, \bar{\tilde t}_1$$ For a given value $\theta_t$ of the L/R mixing angle, the vertices of the $\tilde t_1$ pair with the photon and the $Z$ boson may be written as $i e_0 \tilde Q [p_{\tilde t_1} - p_{\bar{\tilde t}_1}]_\mu$, where $p_{\tilde t_1}$ and $p_{\bar{\tilde t}_1}$ are the 4-momenta of the stop and antistop squarks, and the charges read $$\begin{aligned}
\tilde Q_\gamma & = & - e_t \\
\tilde Q_Z & = & (\cos^2\theta_t -2 e_t \sin^2\theta_W)/\sin 2\theta_W\end{aligned}$$ respectively. $\theta_W$ is the standard electroweak mixing angle and $e_0=\sqrt{4\pi\alpha}$ is the electromagnetic coupling to be evaluated with $\alpha^{-1}(M_Z) = 129.1$ in the improved Born approximation [@swartz]. The $Z$ boson coupling vanishes for the L/R mixing angle $\cos^2 \theta_t \to 2 e_t \sin^2 \theta_W \approx 0.30$. Defining the $\gamma$ and $Z$ vector/axial-vector charges of the electron, as usual, by $e_e = -1$, $v_e = -1 +4 \sin^2 \theta_W$ and $a_e =
-1$, the cross section can be expressed in the compact form [@hikasa] $$\begin{aligned}
\sigma_B[e^+e^- \to \tilde t_1 \bar{\tilde t}_1] &=&
\frac{\pi\alpha^2}{s} \left[\tilde Q_\gamma^2 + \frac{(v_e^2
+a_e^2)\tilde Q_Z^2}{4\sin^2 2\theta_W}
\frac{s^2}{(s -M_Z^2)^2 + M_Z^2 \Gamma_Z^2}\right. \nonumber\\
& &\hphantom{\frac{\pi\alpha^2\beta^3}{s}a}\left.+\,\frac{v_e\tilde
Q_\gamma \tilde Q_Z}{\sin2\theta_W} \frac{s (s-M_Z^2)}{(s
-M_Z^2)^2 + M_Z^2 \Gamma_Z^2}\right] \beta^3
\label{born}\end{aligned}$$ where $\sqrt{s}$ is the center of mass energy and $M_Z,\,\Gamma_Z$ are the mass and the total width of the $Z$ boson, respectively. The $P$-wave excitation near the threshold gives rise to the familiar $\beta^3$ suppression, where $\beta = (1 -4 m^2_{\tilde t_1}/s)^{1/2}$ is the velocity of the stop particles. Angular momentum conservation enforces the $\sin^2\theta$ law, $\sigma_B^{-1} d\sigma_B/d\cos\theta
= \frac{3}{4}\sin^2\theta$, for the angular distribution of the stop particles with respect to the beam axis.
#### QCD corrections.
Gluonic corrections modify the cross section [@schwinger; @drees][^1]. The virtual corrections, Fig.\[feyn\](b), can be expressed by the form factor
$$\begin{aligned}
F(s) &=& \frac{4}{3}\frac{\alpha_s}{\pi}\left\{\frac{s -2 m_{\tilde
t_1}^2}{s\beta}\left[2\, \mbox{Li}_2(w)
+ 2\log(w)\log(1-w) - \frac{1}{2}\log^2(w)
+\frac{2}{3}\pi^2 - 2\log(w) \right.\right.\nonumber \\
& & \hphantom{\frac{4}{3}\frac{\alpha_s}{\pi}a}\left.\left.
-\,\log(w)\log\left(\frac{\lambda^2}{m_{\tilde t_1}^2}\right)\right]
-2 - \log\left(\frac{\lambda^2}{m_{\tilde t_1}^2}\right)\right\}\end{aligned}$$
where $\alpha_s$ is the strong coupling constant and the kinematical variable $w$ is defined as $w =(1-\beta)/(1+\beta)$. The form factor is infrared (IR) divergent. We have regularized this divergence by introducing a small parameter $\lambda$ for the gluon mass. The IR singularity is eliminated by adding the contribution of the soft gluon radiation \[Fig.\[feyn\](c)\], with the scaled gluon energy integrated up to a cut-off value $\epsilon_g = 2
E_g^{cut}/\sqrt{s} \ll 1$. The sum of the virtual correction ($V$) and the soft-gluon radiation ($S$) depends only on the physical energy cut-off $\epsilon_g$, $$\begin{aligned}
\sigma_{V+S} & = & \sigma_B\frac{4}{3}\frac{\alpha_s}{\pi}\left\{\frac{s -2
m_{\tilde t_1}^2}{s\beta}\left[\vphantom{\frac{1}{3}}
4\,\mbox{Li}_2(w) -2\log(w)\log(1+w)
+ 4\log(w)\log(1-w) \right. \right. \\
& &\left. \left. + \,\frac{1}{3}\pi^2
-2\log(w)\log(\epsilon_g) \right] +
\frac{4m_{\tilde t_1}^2 - 3 s}{s \beta}\log(w) +
\log\left(\frac{m_{\tilde t_1}^2}{s}\right) -2\log(\epsilon_g) -2 \right\}\end{aligned}$$
After including the hard gluon radiation, the dependence on the cut-off $\epsilon_g$ disappears from the total cross section. The total QCD corrections can finally be summarized in a universal factor [@drees] $$\sigma[ e^+e^- \to \tilde t_1 \,\bar{\tilde t}_1\, (g)]
= \sigma_B \left[1 +\frac{4}{3}\frac{\alpha_s}{\pi} f(\beta)\right]
\label{sigma}$$ with (Fig.\[fbeta\]) $$\begin{aligned}
f(\beta) &=& \frac{1+\beta^2}{\beta}\left[ \vphantom{\frac{1}{3}}
4\,\mbox{Li}_2(w) +
2\,\mbox{Li}_2(-w) +2\log(w)\log(1-w) +\log(w)\log(1+w)\right] \nonumber\\
& & -\,4\log(1-w) -2\log(1+w) +\left[3 + \frac{1}{\beta^3}\left(2
-\frac{5}{4}(1 + \beta^2)^2\right)\right]\log(w)
+\frac{3}{2}\frac{1+\beta^2}{\beta^2}\end{aligned}$$ Very close to the threshold the Coulombic gluon exchange between the slowly moving stop particles generates the universal Sommerfeld rescattering singularity [@sommer] $f \to \pi^2/2\beta$, which damps the threshold suppression, yet does not neutralize it entirely. Employing methods based on non-relativistic Green’s functions, an adequate description of stop pair production near threshold has been given in Ref.[@bigi1], which also takes into account screening effects due to the finite decay width of the stop particles. In the high-energy limit [@drees] the correction factor in eq.(\[sigma\]) approaches the value $( 1 +
4\alpha_s/\pi)$.
In this note we present a general analysis of hard gluon radiation. We also include stop fragmentation due to collinear gluon emission in the perturbative regime at high energies and we give an account of non-perturbative fragmentation effects.
For unpolarized lepton beams the cross section for gluon radiation off $\tilde t_1$ squarks $$e^+\,e^- \to \tilde t_1\, \bar{\tilde t}_1 \,g$$ depends on four variables: the polar angle $\theta$ between the momentum of the $\tilde t_1$ squark and the $e^-$ momentum, the azimuthal angle $\chi$ between the $\tilde t_1 \bar{\tilde t}_1 g $ plane and the plane spanned by the $e^\pm$ beam axis with the $\tilde t_1$ momentum \[see Ref.[@laer1]\], and two of the scaled energies $x(\tilde t_1)$, $\bar{x}(\bar{\tilde t_1})$, $z(g)$ in units of the beam energy. The energies are related through $x +\bar x +z
=2$ and vary over the intervals $\mu \leq x,\bar x \leq 1$ and $0\leq z \leq 1 -\mu^2$, where $\mu = 2m_{\tilde t_1}/\sqrt{s}$ denotes the squark mass in units of the beam energy. For the angles between the squark and gluon momenta we have $$\begin{aligned}
\cos\theta_{\tilde t_1 \bar{\tilde t}_1} & = & \frac{2 -2(x +\bar x)
+ x \bar
x +\mu^2}{\sqrt{(x^2-\mu^2)(\bar x^2 -\mu^2)}}\\
\cos\theta_{\tilde t_1 g} & = &\frac{2 -2 (x +z) +x
z}{z\sqrt{x^2-\mu^2}}\end{aligned}$$ The spin-1 helicity analysis of the cross section results in the following well-known angular decomposition [@laer2] $$\begin{aligned}
\frac{d\sigma}{dx\,d\bar x\,d\cos\theta\,d\chi/2\pi} &=&
\frac{3}{8}( 1+ \cos^2\theta) \frac{d\sigma^U}{dx\,d\bar x}
+\frac{3}{4}\sin^2\theta \frac{d\sigma^L}{dx\,d\bar x} \nonumber\\
& & -\frac{3}{2\sqrt{2}}\sin2\theta\cos\chi
\frac{d\sigma^I}{dx\,d\bar x}
+\frac{3}{4}\sin^2\theta \cos 2\chi \frac{d\sigma^T}{dx\,d\bar x}\end{aligned}$$ \[$U$ = transverse (no flip), $L$ = longitudinal, $I$ = trv$\ast$long, $T$ = trv$\ast$trv (flip)\]. If the polar and azimuthal angles are integrated out, the cross section is given by $\sigma = \sigma^U +\sigma^L$.
It is convenient to write the helicity cross sections as $$\frac{\beta^3}
{\sigma_B}\frac{d\sigma^j}{dx\,d\bar x} = \frac{\alpha_s}{4\pi}
\frac{S^j + \mu^2 N^j}{(1-x)(1-\bar x)}
\label{d2sigma}$$ The densities $S^j$ and $N^j$ are summarized in Table \[sandn\]; $p$ is the momentum of the $\tilde t_1$ squark, $\bar p$ and $k$ are the longitudinal momenta of $\bar{\tilde t}_1$ and $g$ in the $\tilde t_1$ direction, and $p_T$ is the modulus of the transverse $\bar{\tilde t}_1$, $g$ momentum with respect to this axis \[all momenta in units of the beam energy\]. Since $I,T$ correspond to $\gamma,Z$ helicity flips by 1 and 2 units, they are of order $p_T$ and $p_T^2$, respectively. Note that the threshold suppression is absent in the $U$, $I$, $T$ components and attenuated in the leading longitudinal $L$ term as expected from eq.(\[sigma\]).
-------------------------------- ------------------------------------------------------------------------------------ -----------------------------------------------------------------------------------------------------------------
\[-0.2cm\] $S^j$ $N^j$
\[-0.2cm\]
\[-0.2cm\] ${\displaystyle}U $ ${\displaystyle}\frac{{\textstyle}32}{{\textstyle}3}\,(1-x)\,(1-\bar x)$ ${\displaystyle}-\frac{{\textstyle}4}{{\textstyle}3}\,p_T^2\,\frac{{\textstyle}1-x}{{\textstyle}1-\bar x}$
\[0.4cm\] ${\displaystyle}L $ ${\displaystyle}\frac{{\textstyle}16 \beta^2}{{\textstyle}3}\,(1-z) $ ${\displaystyle}\frac{{\textstyle}4}{{\textstyle}3}\,\left[p_T^2\,
\frac{{\textstyle}1-x}{{\textstyle}1-\bar x} - \beta^2 \left(\frac{{\textstyle}1-x}{{\textstyle}1-\bar x}
+\frac{{\textstyle}1-\bar x}{{\textstyle}1-x} +2\right)\right] $
\[0.4cm\] ${\displaystyle}I $ ${\displaystyle}-\frac{{\textstyle}4\sqrt{{\textstyle}2}}{{\textstyle}3}\,p_T \,p$ ${\displaystyle}\frac{{\textstyle}2\sqrt{{\textstyle}2}}{{\textstyle}3}\,p_T \left(p - \bar p\,
\frac{{\textstyle}1-x}{{\textstyle}1-\bar x}\right) $
\[0.4cm\] ${\displaystyle}T $ ${\displaystyle}0$ ${\displaystyle}\frac{{\textstyle}2}{{\textstyle}3}\,p_T^2\,
\frac{{\textstyle}1-x}{{\textstyle}1-\bar x} $
\[-0.2cm\]
-------------------------------- ------------------------------------------------------------------------------------ -----------------------------------------------------------------------------------------------------------------
: Coefficients of the helicity cross sections in eq.(\[d2sigma\]). The energy and momentum variables are defined in the text.[]{data-label="sandn"}
#### Fragmentation.
In the limit where the gluons are emitted from fast moving squarks with small angles, the gluon radiation $$\label{brems}
\tilde t_1 \to \tilde t_1\, g$$ can be interpreted as a perturbative fragmentation process. From the form of the differential cross section $d\sigma/dz\,dp_T^2 $ we find in this limit for the splitting functions, in analogy to the Weizsäcker-Williams [@weiz] and Altarelli-Parisi splitting functions [@altarelli], $$\begin{aligned}
P[\tilde t_1 \to \tilde t_1; x] & = & \frac{\alpha_s}{2\pi}\,
\frac{8}{3}\, \frac{x}{1-x}\, \log\frac{Q^2}{m_{\tilde t_1}^2} \\
P[\tilde t_1 \to g; z] & = & \frac{\alpha_s}{2\pi}\, \frac{8}{3}\, \frac{
1-z}{z}\, \log\frac{Q^2}{m_{\tilde t_1}^2}
\nonumber\end{aligned}$$ As usual, $x$ and $z$ are the fractions of energy transferred from the $\tilde t_1$ beam to the squark $\tilde t_1$ and the gluon $g$ after fragmentation, respectively; $Q$ is the evolution scale of the elementary process, normalized by the squark mass rather than the QCD $\Lambda$ parameter \[in contrast to the light quark/gluon sector\]. As a consequence of angular-momentum conservation, the gluon cannot pick up the total momentum of the squark beam. \[Similar zeros have been found for helicity-flip fragmentation functions in QED/ QCD [@altarelli; @falk].\]
By using the crossing rules {$z \to 1, 1 \to x \}\mbox{ and }\{ 1-x \leftrightarrow 1-x$}, familiar from the analogous splitting functions in QED [@chen], we derive for the elementary gluon splitting process into a squark-antisquark pair $$g \to \tilde t_1 \bar{\tilde t_1}$$ the distribution $$P[g \to \tilde t_1; x] = \frac{\alpha_s}{2\pi}\, \frac{1}{2} \,x \,(1-x)
\,\log\frac{Q^2}{m_{\tilde t_1}^2}$$ after adjusting color and spin coefficients properly. This splitting function is symmetric under the $\tilde t_1 \leftrightarrow
\bar{\tilde t_1}$ exchange, i.e. $\{x \leftrightarrow
1-x\}$. The probability is maximal for the splitting into equal fractions $x = 1/2$ of the momenta, in contrast to spinor QED/QCD where the splitting into a quark-antiquark pair is proportional to $x^2 + (1-x)^2$ and hence asymmetric configurations are preferred.
The above splitting functions provide the kernels for the shower expansions in perturbative QCD Monte Carlos for $e^+e^-$ annihilation such as Pythia [@sjos] and Herwig [@webb]. They serve the same purpose in the hadron-hadron versions of these generators as well as Isajet [@baer]. Of course, the interpretation of the radiation processes as universal fragmentation processes becomes increasingly adequate with rising energy of the fragmenting squarks/gluons.
If the $\tilde t_1$ squark is lighter than the top quark, the lifetime will be long, $\tau \ge 10^{-20}$sec, since the dominant decay channel $\tilde t_1 \to t +\tilde\chi_1^0$ is shut off $[\tilde\chi_1^0 = LSP]$. The decay widths corresponding to the 2-body decay $\tilde t_1 \to c +\tilde\chi_1^0$ and 3-body slepton decays involve the electroweak coupling twice and hence will be very small [@hikasa]. As a result, the lifetime is much longer than the typical non-perturbative fragmentation time of order $1$ fm \[i.e. ${\cal O}\, (10^{-23}$sec)\] so that the squark has got enough time to form $(\tilde t_1\bar q)$ and $(\tilde t_1 q q)$ fermionic and bosonic hadrons. However, the energy transfer due to the non-perturbative fragmentation, evolving after the early perturbative fragmentation, is very small as a result of Galilei’s law of inertia. Describing this last step in the hadronization process of a $\tilde t_1$ jet by the non-perturbative fragmentation function [ *à la*]{} Peterson et al. [@pete] (which accounts very well for the heavy-quark analogue), we find $$D (x)^{NP} \approx \frac{4 \sqrt{\epsilon}}{\pi}\,\frac{1}{x\,[1 - 1/x
-\epsilon/(1-x)]^2}$$ with the parameter $\epsilon \sim 0.5\,\mbox{GeV}^2/m_{\tilde
t_1}^2$. Here, $x = E[(\tilde t_1\,\bar q)]/E[\tilde t_1]$ is the energy fraction transferred from the $\tilde t_1$ parton to the $(\tilde
t_1\,\bar q)$ hadron etc. The resulting average non-perturbative energy loss $$<1-x>^{NP}\sim\frac{2\sqrt{\epsilon}}{\pi}\left[
\log\left(\frac{1}{\epsilon}\right) -3\right]$$ is numerically at the level of a few percent.
Monte Carlo programs for the hadronization of $\tilde t_1$ squarks link the early perturbative fragmentation with the subsequent non-perturbative hadronization. The relative weight of perturbative and non-perturbative fragmentation can be characterized by the average energy loss in the two consecutive steps. The overall retained average energy of the $\tilde t_1$ squarks factorizes into the two components, $$<x> = <x>^{NP} <x>^{PT}$$ Summing up the energy loss due to multiple gluon radiation at high energies, we find in analogy to heavy-quark fragmentation [@bigi2] $$<x>^{PT} = \left[\frac{\alpha_s(m_{\tilde
t_1}^2)}{\alpha_s(E^2)}\right]^{-8/3b}$$ with $ b = (11 -2 n_f/3) +(-2 - n_f/3)$ being the LO QCD $\beta$ function including the colored supersymmetric particle spectrum. At high energies, the perturbative multi-gluon radiation has a bigger impact than the final non-perturbative hadronization mechanism, e.g. $<x>^{PT} \approx 0.93$ for a $\tilde
t_1$ beam energy $E=1\,\mbox{TeV}$ and $m_{\tilde t_1}=200\,\mbox{GeV}$ as compared to $<x>^{NP}\approx 0.98$. At low energies the two fragmentation effects are of comparable size.
After finalizing the manuscript, we received a copy of Ref.[@djou] in which the total cross sections for squark pair production in $e^+e^-$ annihilation have been discussed including squark-gluon and quark-gluino loops, yet not the gluon-jet distributions analysed in the present note.
We thank our colleagues at the LEP2 Workshop who demanded the analysis presented here to refine the experimental stop search techniques.
[xx]{} J. Ellis and S. Rudaz, Phys. Lett. [**B128**]{} (1983) 248. K. Hikasa and M. Kobayashi, Phys. Rev. [**D36**]{} (1987) 724. K.A. Olive and S. Rudaz, Phys. Lett. [**B340**]{} (1990) 74. J. Shirai [*et al.*]{} (Venus), Phys. Rev. Lett. [**72**]{} (1994) 3313. R. Akers [*et al.*]{} (Opal), Phys. Lett. [**B337**]{} (1994) 207. J.M.L. Swartz, SLAC-PUB-6710 (Nov. 1994). J. Schwinger, [*Particles, Sources and Fields*]{} vol. II (Addison-Wesley, New York 1973). M. Drees and K. Hikasa, Phys. Lett. [**B252**]{} (1990) 127. W. Beenakker, R. Höpker, M. Spira and P.M. Zerwas, Report DESY 94-212. A. Arhrib, M. Capdequi-Peyranere and A. Djouadi, Montreal Report UdeM-GPP-94-13. A. Sommerfeld, [*Atombau und Spektrallinien*]{} vol. 2 (Vieweg, Braunschweig 1939). I.I. Bigi, V.S. Fadin and V. Khoze, Nucl. Phys. [ **B377**]{} (1992) 461. E. Laermann, K.H. Streng and P.M. Zerwas, Z. Phys. [**C3**]{} (1980) 289. E. Laermann and P.M. Zerwas, Phys. Lett. [**B89**]{} (1980) 225. C.F.v. Weizsäcker, Z. Phys. [**88**]{} (1934) 612; E.J. Williams, Phys. Rev. [**45**]{} (1934) 729. G. Altarelli and G. Parisi, Nucl. Phys. [**B126**]{} (1977) 298. B. Falk and L.M. Sehgal, Phys. Lett. [**B325**]{} (1994) 509. M.-S. Chen and P.M. Zerwas, Phys. Rev. [**D12**]{} (1975) 187. T. Sjöstrand, Comp. Phys. Commun. [**82**]{} (1994) 74. G. Marchesini [*et al.*]{}, Comp. Phys. Commun. [**67**]{} (1992) 465; B.R. Webber, Cavendish-HEP-94/17. H. Baer, F. E. Paige, S. D. Protopopescu and X. Tata, Report on ISAJET 7.0/ ISASUSY 1.0, FSU-HEP 930329 and UH-511-764-93. C. Peterson, D. Schlatter, I. Schmitt and P.M. Zerwas, Phys. Rev. [**D27**]{} (1983) 105. I.I. Bigi, Yu.L. Dokshitser, V. Khoze, J.H. Kühn and P.M. Zerwas, Phys. Lett. [**B181**]{} (1986) 157.
[^1]: Since we focus on QCD gluon effects for light stop particles in the LEP range, we do not take into account quark-gluino loop effects, assuming the gluino to be heavy; these loop effects have been discussed for squark production at the Tevatron in Ref.[@been] and at $e^+e^-$ colliders in Ref.[@djou].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We apply a test for low-dimensional, deterministic dynamics to the Niño 3 time series for the El Niño Southern Oscillation (ENSO). The test is negative, indicating that the dynamics is high-dimensional/stochastic. However, application of stochastic forcing to a time-delay equation for equatorial-wave dynamics can reproduce this stochastic dynamics and other important aspects of ENSO. Without such stochastic forcing this model yields low-dimensional, deterministic dynamics, hence these results emphasize the importance of the stochastic nature of the atmosphere-ocean interaction in low-dimensional models of ENSO.'
title: 'ENSO dynamics: low-dimensional-chaotic or stochastic?'
---
Introduction
============
Prediction is the ultimate goal of meteorology and climatology and the issue of predictability is crucial. Prediction in these disciplines are mostly probabilistic, but there may be different rationales for a probabilistic description. These rationales are intimately linked to the various meanings of the concepts of determinism, deterministic chaos, and stochasticity. The fundamental laws of classical physics are deterministic, since the future and past of the state of such a system is uniquely given by the state at a given time. The majority of models of weather and climate are deterministic in this sense. After the discovery of deterministic chaos, it has been realized that the deterministic evolution may be sensitive to the initial conditions, in the sense that small perturbations in the initial conditions grow exponentially with time. Predictability in such systems is limited by this exponential growth rate, given by the largest positive Lyapunov exponent of the system. From a practical viewpoint the most important aspect of this insight is that predictability can be limited in this way, and may require a probabilistic description, not only in systems with a large number of degrees of freedom, but also in in simple, low-dimensional, nonlinear systems.
Before the advent of chaos theory, unpredictability was considered to be a practical consequence of our inability to specify and solve the evolution equations for the microscopic state of high-dimensional systems. However, these days high-speed computing allows us, not only to solve numerically low-dimensional, nonlinear problems, but also high-dimensional general circulation models of the climate system. Nevertheless, the problem of sensitivity to initial condition seems to persist, at least when it comes to weather prediction and climate variability on interannual to decadal time scales. The fluctuations of macroscopic variables around some time-averaged or ensemble-averaged state could be described in a probabilistic manner by representing them as stochastic processes or alternatively as low-dimensional chaotic processes. Thus, when dealing with data in climatology, either from observation or from large-scale simulations, the climate dynamicist will have to ask the question of whether prediction for the system/phenomenon of interest is better served by a low-dimensional chaotic model of climate variability, or by a high-dimensional (stochastic) model.
The answer to this question depends on whether the system/phenomenon represents a self-organization of the dynamics into an effective small number of degrees of freedom. One approach is of course to reduce the high-dimensional model to a low-dimensional one via a series of approximations and simplifications, i.e., via theoretical model reduction. In many cases, this is very demanding and many different reductions are possible, and it is hard to know whether a reduced model is what Nature abides to. Hence, there is a demand for methods by which it is possible to decide from the observation data whether the system dynamics can be uniquely projected onto an attractor in a low-dimensional phase space. Such methods exist, based on Taken’s time-delay embedding theorem [@T81], and from 1980 onwards there is a growing literature on reconstruction of chaotic attractors from time series and computation of attractor dimension and largest Lyapunov exponent [@DA]. The idea is to assume that the system state vector evolves according to the system of first-order ordinary differential equations describing the trajectory ${\bf z}(t)$ on the $d$-dimensional attractor in a $p>d$-dimensional phase space. If the system is autonomous and the attractor of the trajectory has dimension $d$, Takens’ time-delay method [@T81] can be used to construct an $m>2d$-dimensional embedding space on which the attractor can be mapped continuously and one-to-one. In practice this method works only if the attractor dimension $d$ is reasonably low. Dynamical systems with a large number of independent or weakly dependent degrees of freedom can only be described either by large-scale numerical simulation or by stochastic methods. For such systems the phase-space attractor is also high-dimensional and cannot be mapped one-to-one onto a low-dimensional time-delay embedding space. The computation of attractor dimension then typically fails to converge when embedding dimension $m$ is increased, but such convergence can be difficult to detect if the time series is short. [@KG92; @KG93] devised a direct test for the existence of low-dimensional deterministic dynamics which is useful for short time series. This is the kind of test that will be employed in this paper to the instrumental time series for the El-Niño Southern Oscillation (ENSO).
The remainder of the paper is structured as follows: In section (2) we briefly describe a time-delay equation for equatorial wave dynamics and ENSO, and in section (3) and (4) we review the time-delay phase space reconstruction technique and the test for determinism. In section (5) these techniques are employed to the Niño data and to numerical solutions to the time-delay equation with and without seasonality subtracted and with and without stochastic forcing. We also apply a superposed-epoch analysis to these data to highlight the characteristic waveforms of ENSO episodes as manifested in the Niño signal. The implications of our findings are also discussed in this section and summarized in section (6).
A time-delay equation for ENSO
==============================
An El-Niño episode is characterized by an increase in the sea-surface temperature (SST) in the eastern Pacific, with a strong impact at the coast of Peru. Generally, the SST in this area increases during the winter period, but occasionally (often every three to seven years) the temperature increase is more pronounced and this phase is called El-Niño. The opposite phase, a strong decrease in the SST, is coined La Niña. The intensity of El-Niño is higher than the intensity of La-Niña, and the SST distribution is positively skewed [@DK2005]. The increase in SST is followed by stronger precipitation, which is distributed over the entire Pacific basin, while during La Niña and normal conditions, the precipitation is heavier in the western Pacific. This is also due to the Southern Oscillation (SO) in the atmosphere, which is characterized by lower pressure over the western Pacific during La Niña and normal conditions. According to the Bjerknes hypothesis [@N2011], there is an initial warming in the SST in the eastern Pacific, which weakens the trade winds from the east and gives rise to westerly wind anomalies, which further increase SST. This positive feedback (which is the El-Niño phase of ENSO) is also accompanied by the formation of the eastward Kelvin wave. Concurrently, a westward Rossby wave is formed in the middle of the ocean basin, and after it is reflected at the western boundary of the ocean basin, an eastward Kelvin wave is formed, which has a cooling effect. This wave causes the La Niña phase of ENSO.
A large body of literature is concerned with the dynamical modeling of ENSO (a review can be found in [@Dijkstra]). An interesting class of models is described as delay-differential equations, which are linear and autonomous [@BH89], nonlinear, autonomous [@SS88], or nonlinear, periodically forced equations [@T94; @MCZ]. The latter will be our focus in this paper and can be formulated as an time-delay equation for the thermocline depth $h(t)$: $$\begin{aligned}
\frac{dh(t)}{dt}&=& a\tanh (kh\lbrace t-\frac{L}{2c_{K}}\rbrace) \\ \nonumber
&-&
d\tanh (kh\lbrace t-\lbrack\frac{L}{c_{K}}+
\frac{L}{2c_{R}}\rbrack \rbrace)
+c\cos(\Omega t), \label{eq1}\end{aligned}$$ where $k$ is an ocean-atmosphere coupling parameter, $c_{K}$ is the velocity of the wind-forced Kelvin mode, $L$ is the ocean basin width, and $c_{R}$ is the velocity of the Rossby wave. $\Omega$ is the frequency of the seasonal cycle, while $a$, $d$ and $c$ are constants. The cosine function in equation (\[eq1\]) accounts for the annual periodicity in the SST data. This delay differential equation has two time delays: $\tau_{1}=L/c_{K}+L/2c_{R}$, and $\tau_{2}=L/2c_{K}$. Here, $\tau_{1}$ is the summation over a time it takes a Rossby wave to travel from the middle of the ocean basin to the western boundary and then be reflected as a Kelvin wave, while $\tau_{2}$ is the transit time for the Kelvin wave which travels from the middle of the basin to reach the eastern Pacific.
Phase space reconstruction
==========================
Before we describe the test for determinism (or more precisely, for low-dimensional deterministic dynamics), we briefly explain how the phase space can be reconstructed from scalar time series of length $N$ by time-delay embedding. Suppose that the phenomenon under study can be described by a state vector ${\bf z}(t)$ in a phase space of dimension $p$, i.e., $\bf z$ evolves according to an autonomous system of 1st order ordinary differential equations: $$\frac{d{\bf z}}{dt}={\bf f}({\bf z}),\, \, \, {\bf f}: {\cal R}^p\rightarrow {\cal R}^p \label{(eq3)},$$ and that an observed time series $x(t)$ is generated by the measurement function $g: {\cal R}^p\rightarrow {\cal R}$: $$x(t)=g({\bf z}(t)). \label{(eq4)}$$ Further, assume that the dynamics takes place on an invariant set (an attractor) ${\cal A} \subseteq {\cal R}^p$ in phase space, and that this set has box-counting fractal dimension $d$. Since the dynamical system uniquely defines the entire phase-space trajectory once the state ${\bf z}(t)$ at a particular time $t$ is given, we can define uniquely an $m$-dimensional measurement function,
$${\bf g}: {\cal A}\rightarrow {\cal R}^m, \, \, {\bf g}({\bf z})=(x(t),x(t+\tau),\ldots,x(t+(m-1)\tau)), \label{(eq5)}$$
where the vector components are given by equation (\[(eq4)\]), and $\tau$ is a time delay of our choice. If the invariant set $\cal A$ is compact (closed and bounded), $g$ is a smooth function and $m>2d$, the map given by equation (\[(eq5)\]) is a topological embedding (a one-to-one continuous map) between $\cal A$ and ${\cal R}^m$. The condition $m>2d$ can be thought of as a condition for the image ${\bf g}({\cal A})$ not to intersect itself, i.e. to avoid that two different states on the attractor $\cal A$ are mapped to the same point in the $m$-dimensional embedding space ${\cal R}^m$. If such an embedding is achieved, the trajectory ${\bf x}(t)={\bf g}({\bf z})$ (where ${\bf g}({\bf z})$ is given by equation (\[(eq5)\])) in the embedding space is a complete mathematical representation of the dynamics on the attractor. Note that the dimension $p$ of the original phase space is irrelevant for the reconstruction of the embedding space. The important thing is the dimension $d$ of the invariant set $\cal A$ on which the dynamics unfolds.
Test for determinism
====================
![$\Lambda (\tau)$ averaged over results computed from an ensemble of ten realization of the O-U process (squares), and for the realizations with randomized phases (triangles). Error bars denote standard errors of the [*mean*]{} $\Lambda(\tau)$ (which is $(1/\sqrt{10})\times $the standard deviation of the distribution of the ten samples).](fig1.pdf){width="6cm"}
This method was recently successfully applied to the studies of magnetospheric organization during magnetospheric storms and substorms [@ZR1; @ZR2]. When a system is low-dimensional deterministic, the direction of the trajectory (its tangent) is a function of the position in the reconstructed phase space (from equation \[(eq3)\]). This means that trajectories emanating from points in a small neighborhood in phase space have almost parallel directions. On the other hand, corresponding trajectories in a stochastic or high-dimensional system have directions in a low-dimensional embedding space which are not uniquely dependent on the position in this space, and therefore the tangent can have a different direction the next time it recurs to the same neighborhood. Let $b$ denote a small time increment and envisage a portion of phase space spanning the entire attractor divided into an enumerable set of small “boxes" of size corresponding to the length of the trajectory increment:
$$\begin{aligned}
\Delta {\bf x}(t)&=&[ x(t+b)-x(t), x(t+\tau+b)-x(t+\tau),\ldots , \nonumber \\
& &x(t+(m-1)\tau+b)-x(t+(m-1)\tau)] , \label{(eq6)}\end{aligned}$$
The tangent for the $k$’th pass of the trajectory through box $j$ is the unit vector ${\bf u}_{k,j}=\Delta {\bf x}_{k,j}(t)/\vert \Delta {\bf x}_{k,j}(t)\vert$. The estimated averaged displacement vector in the box is
$${\bf V}_{j}=\frac{1}{n_{j}}\sum_{k=1}^{n_{j}}{\bf u}_{k,j}, \label{(eq7)}$$
where $n_j$ is the number of passes of the trajectory through box $j$. If the embedding dimension is sufficiently high and in the limit of vanishingly small box size, the trajectory directions should be aligned and the length $V_j\equiv \vert {\bf V}_{j}\vert=1$. In the case of deterministic dynamics and finite box size, ${\bf V}_{j}$ will not depend very much on the number of passes $n_j$, and $V_j$ will converge to $1$ as $n_j\rightarrow \infty$. In contrast, for the trajectory of a random process, where the direction of the next step is completely independent of the past, $V_j$ will decrease with $n_j$ as $V_j\sim n_j^{-1/2}$. The degree of determinism of the dynamics can be assessed by exploring the dependence of $ V_j$ on $n_j$. In practice, this can be done by computation of the averaged displacement vector length:
$$L_n\equiv\langle V_{j} \rangle_{n_j=n}, \label{(eq8)}$$
where the average is done over all boxes with same number $n$ of trajectory passes.
{width="6cm"}
Let us recall that we are describing a test to distinguish signals described by low-dimensional dynamical systems, i.e., signals that in the continuous-time limit are solutions to differential equations and therefore continuous, and signals described as a stochastic process. The archetype of a random and continuous stochastic process is the Wiener process (Brownian motion), which in discrete time is a random walk. The random walk has random increments, and hence the displacement vectors $\Delta {\bf x}(t)$ in an $m$-dimensional embedding space will have random directions. Hence, for a random walk $L_{n\rightarrow \infty}=0$. For finite $n$, however, there will be a finite statistical spread of $V_j$, and as shown in [@KG93], the average displacement of $n$ passes in [*m*]{}-dimensional phase space is $$L_n=R_n\equiv \frac{1}{\sqrt{n}}(\frac{2}{m})^{1/2}\frac{\Gamma\lbrack (m+1)/2\rbrack}{\Gamma(m/2)},\label{(eq9)}$$ where $\Gamma$ is the gamma function. The deviation in $\langle V_{j}\rangle$ between a given time series and a random-walk can be characterized by a single number given by the weighted average over all boxes of the quantity, $$\Lambda(\tau)\equiv \frac{1}{\sum_{j} n_{j}}\sum_{j} n_{j} \frac{\langle V_j\rangle ^{2}(\tau)-R_{n_j}^{2} }{1-R_{n_j}^{2}},\label{(eq10)}$$ where we have explicitly highlighted that the averaged displacement $\langle V_j\rangle(\tau)$ of the trajectory in the reconstructed phase space depends on the time delay $\tau$. For a completely deterministic signal we have $\Lambda(\tau)=1$, and for a completely random signal $\Lambda(\tau) =0$, hence this quantity can be considered as a measure of determinism.
In Figure 1, we show $\Lambda(\tau)$ averaged over ten numerical realizations of the Ornstein-Uhlenbeck (O-U) stochastic process, for embedding dimension is $m=8$, and $b=1$. The O-U process is described by the stochastic equation: $$dS_{t}=-\lambda S_{t}+\sigma dW_{t},$$ where $W(t)$ is the Wiener process. It is a more physically realistic random process for many phenomena than the Wiener process, since the damping term $-\lambda S_{t}$ makes it bounded. In Figure 1 we also show mean $\Lambda(\tau)$ computed from the same ten realizations, but after randomization of the phases of the Fourier coefficients. This randomization leaves the power spectral density, and hence the autocorrelation function, unchanged. Hence, $\Lambda(\tau)$ should also be unchanged for a random process, which Figure 1 demonstrates. On the other hand, for a signal from a low-dimensional chaotic system, which has to be nonlinear to be chaotic, the randomization of phases will destroy the nonlinear coupling between Fourier modes and make $\Lambda (\tau)$ more similar to a random signal, i.e., it will be reduced compared to $\Lambda (\tau)$ for the original signal. Examples of this were shown in [@KG92; @KG93].
Results
=======
We analyze Niño 3 data (see Figure 2) obtained from the sea ice and SST data set (HadISST1) [@R2003]. Niño 3 data is the area-averaged SST from 5S-5N and 150W-90W, with monthly resolution and the time span between 1871 and 2008. We compute $\Lambda$ as a function of time-delay, which is also used in the phase-space reconstruction procedure. The embedding dimension is $m=10$, $b=1$, i.e., the box size is equal to the average distance between two successive points on the reconstructed phase-space trajectory. It has been shown in [@KG93] that increasing embedding dimension can increase $\Lambda$ in deterministic systems, while it should not influence $\Lambda$ in stochastic systems. In principle, higher $m$ is better, but we also have to consider the number of data points available for the test, and this decreases with increasing $m$ (from equation \[(eq5)\]). In our analysis, $m=10$ seems a suitable choice of embedding dimension. In Figure 3 we observe that $\Lambda(\tau)$ exhibits spiky dips at those $\tau$ where the autocorrelation function $r(\tau)$ has an extremum. This is a spurious feature of the technique which is explained in [@KG93]. We also plot $\Lambda (\tau)$ averaged over an ensemble of ten surrogate time series with randomized phases.
From Figure 3, we can conclude that the dynamics underlying the Niño 3 data is dominated by a nonlinear, low-dimensional component, since $\Lambda (\tau)$ for the phase-randomized time series is considerably reduced compared to that computed for the original Niño 3 data. However, as we will demonstrate next, the nonlinear and low-dimensional behavior is rather trivial and derives from the seasonal cycle. This cycle can be represented by the climatology, which is the mean over all data for every month of the year. In Figure 4 we plot the Niño 3 climatology along with the sinusoidal climatology used in equation (\[eq1\]). Next, we subtract the climatology from the Niño 3 data, and compute $\Lambda(\tau)$ again (Figure 5 (a)). $\Lambda(\tau)$ is strongly reduced and randomization of phases makes no discernible difference, indicating the the positive test for determinism in the original Niño 3 time series is an effect of the seasonal cycle.
{width="6cm"}
![Climatology for Niño 3 time series (red) and the sinusoidal climatology used in equation (\[eq1\]) (blue).](fig4.pdf){width="6cm"}
![(a): $\Lambda(\tau)$ for original Niño 3 time-series with climatology subtracted (blue squares), and $\Lambda(\tau)$ averaged over ten realizations of phase-randomized versions of this time series (red triangles). The error bars of the latter denotes the standard deviation of $\Lambda(\tau)$ over the distribution of these ten samples. (b): $\Lambda(\tau)$ averaged over ten realizations of time series generated by equation (\[eq1\]) with a white-noise stochastic forcing and with climatology subtracted (blue), and the same for ten phase-randomized versions (red). In (b) the error bars denote the standard deviation of the [*mean*]{} $\Lambda(\tau)$.](fig5.pdf){width="6cm"}
{width="6cm"}
![$\Lambda(\tau)$ averaged over results computed from ten simulated time series like that shown in Figure 6 (a) with climatology included (blue squares). After phase-randomization (red triangles). The error bars denote the standard deviation of the [*mean*]{} $\Lambda(\tau)$.](fig7.pdf){width="6cm"}
{width="6cm"}
{width="6cm"}
{width="6cm"}
This implies that the Niño 3 time series with the seasonal cycle subtracted can be described a stochastic process similar to the O-U process for which results were presented in Figure 1. It does not imply, however, that the time-delay equation (\[eq1\]) has to be abandoned, since inclusion of stochastic forcing term may give rise to such stochastic dynamics and produce results compatible with the test for determinism in Niño 3 data. In the following we shall first show that such a time-delay equation with stochastic forcing is able to produce time series with these properties. Next, we shall demonstrate that in absence of stochastic forcing, the time-delay model tends to produce low-dimensional, deterministic dynamics, even when seasonality is removed.
In order to investigate the effect of such stochastic forcing we add a term $\sigma w(t)$, where $w(t)$ is a Gaussian white noise with unit variance, to equation (\[eq1\]) and solve the equation numerically with the parameters $a=0$, $d=1, k=100, c=1.3, \sigma=0.7$ and $\tau_{1}=0.2$. Figure 5(b) shows $\Lambda(\tau)$ ensemble-averaged over ten realizations of the simulation after the climatology has been subtracted. In the same plot we show $\Lambda(\tau)$ averaged over the ten corresponding phase-randomized realizations. The low values of $\Lambda(\tau)$, and their non-responsiveness to phase randomization, indicate that for these parameters the de-seasonalized solutions to the stochastically forced equation (\[eq1\]) have the character of a stochastic process similar to the de-seasonalized Niño 3 time series. A qualitative similarity is also apparent from Figure 6, where we plot a realization of this process along with the Niño 3 time series, both with the climatology subtracted. Both time series have been normalized to have unit variance. This similarity is also apparent when $\Lambda(\tau)$ is computed from the simulation without subtracting the seasonal cycle, as shown by comparing Figure 7 with Figure 3. The model described above, but without stochastic forcing ($\sigma=0$), was studied by [@G2008]. They used the same choice of parameters $a$, $c$ and $d$ as above, but varied the parameters $k$ and $\tau_1$. They computed and plotted the global maximum of the thermocline depth $h(t)$ as a function of $\tau_{1}$ and $k$, and observed discontinuities in this parameter space separating regions of regular solutions strictly governed by the seasonal forcing and more ENSO-like irregular solutions, indicating a structural instability of the model. In Figure 8 (a) we have plotted such a regular solution obtained for $\tau_{1}=0.4088$, and in Figure 8 (b) an irregular one on the other side of the parameter-space discontinuity; for $\tau_{1}=0. 4198$. In both of these cases, $c=1$, $k=100$ and $d=1$. This irregular solution exhibits a seasonal cycle of variable strength interrupted by stronger episodical oscillations with some reminiscence of observed ENSO episodes. However, by inspection of the time series the similarity to the Niño 3 signal is not too convincing. As can be seen from Figure 6, the solution with stochastic forcing added exhibits more qualitative similarity to the Niño 3 signal than the signal in Figure 8(b). This qualitative discrepancy is also reflected by the test of determinism. Figure 9(a) shows $\Lambda(\tau)$ for the signal in Figure 8(b) and for its phase-randomized version. The determinism is higher than the one for the Niño 3 signal shown in Figure 3(a), and it is less reduced by phase randomization. The former is obviously due to the higher degree of randomness apparent in the Niño 3 signal. The lack of reduction of determinism after phase-randomization of the Figure 8(b)-signal is due to the strong seasonal oscillation. Since a major fraction of the power resides in the annual Fourier component the phase randomization will change the phase of this component, but the signal will still have a deterministic appearance after this change. This is all quite trivial, so the crucial test is to compute determinism after de-seasonalization. For the signal in Figure 8(b) subtraction of the climatology does not offer an effective filtering of the seasonal component, since its amplitude varies a lot. For the same reason Fourier filtering is also ineffective. In Figure 8(c) we show the de-seasonalized signal after application of a mexican-hat wavelet filter. This signal represents the “true" ENSO episodes according to the time-delay model without stochastic forcing. The computed determinism $\Lambda(\tau)$ of this signal is shown in Figure 9(b). The determinism is very high, and slows clearly that the ENSO episodes in this model are the result of deterministic, low-dimensional dynamics. The strong reduction of determinism after phase-randomization demonstrates that this dynamics is nonlinear, and (since the timing of the episodes seems random) chaotic. This is in strong contrast to the results shown in Figure 5 for de-seasonalized Niño 3 and signals generated by the time-delay model with stochastic forcing, which show signals dominated by a stochastic component.
In the following, we go back to the time-delay simulation with stochastic forcing and demonstrate in two more examples its similarity to the Niño 3 data. In Figure 10(a) we plot a histogram for the Niño 3 data and for the simulation after the climatology has been subtracted (the signals in Figure 6). The ENSO events contribute to the tails of these distributions, which are somewhat heavier on the positive side due to the relative strength of El Niño compared to La Niña. The average time-evolution of ENSO events in Niño 3 and simulation can be investigated by means of a superposed-epoch analysis (conditional averaging). This analysis works as follows: Consider a signal $s(t)$. We define the onsets of ENSO events as the times $t_n$, $n=1,\ldots,N$, for which the signal ascends through two standard deviations from the mean. Then we produce an ensemble of conditional signals $s_n(\delta t)=s(t_n+\delta t)$ and produce the conditional average $\langle s_n(\delta t)\rangle \equiv (1/N)\sum_{n=1}^{N}s_n(\delta t)$. Figure 10(c) displays the average ENSO structure for Niño 3 (squares) and simulation (triangles) computed this way, and shows that the simulation predicts fairly well the average evolution of ENSO up to at least 15 months after the onset of the episode.
Conclusion
==========
From of a model for ENSO activity similar to the one studied here, [@ET95] came to the conclusion that depending on the strength of the coupling between the ocean and the atmosphere, the dynamics could undergo quasi-periodicity routes to chaos. The same authors suggested that the ENSO might be described as a low-dimensional chaotic dynamics driven by the seasonal cycle, where the appearance of the chaos is due to the nonlinear resonance between the natural oscillator of the atmosphere-ocean coupling and the seasonal cycle. Further, the irregular jumps of the state between different resonances should be the cause of the chaos. Unfortunately, SST time series in the eastern Pacific are too short for chaos to be proven by standard methods.
In this paper we have applied a simple test for determinism to show that Niño 3 data, which is the proxy of SST in the eastern Pacific, is most adequately described as a stochastic process after the seasonal cycle has been removed, implying that ENSO is a consequence of mainly stochastic and not low-dimensional, chaotic dynamics. A similar conclusion was made by [@BW2002], where a test for determinism was applied to the Southern Oscillation index (SOI) series. SOI measures a pressure difference between Tahiti and Darwin in the Pacific Ocean and can also be used as a proxy for the El-Niño dynamics.
Despite the stochastic nature of the Niño 3 signal, we have demonstrated here that low-dimensional, deterministic dynamics may also be involved. The equatorial-wave equation from [@T94] exhibits such dynamics without stochastic forcing, but by adding stochastic forcing to this equation and comparing determinism and average ENSO structure, with those of Niño 3 data, we conclude that this statistical-dynamical model can reproduce important aspects of ENSO dynamics.
Abarbanel, H. (1996), [**]{}Analysis of observed chaotic data, Institute for nonlinear science, Springer, New York. Battisti, D. S., and A. C. Hirst (1989), Interannual variability in the tropical atmosphere-ocean system: Influence of the basic state and ocean geometry, [*J. Atmos. Sci.*]{}, *46*, pp. 1687. Binder P. M, and C. A. Wilches (2002), Absence of determinism in El-Niño Southern Oscillation, [*Phys. Rev. E*]{}, *65*, doi: 10.1103/PhysRevE.65.055207. Dijkstra, H. A. (2010), [*Nonlinear Physical Oceanography. A dynamical System Approach to the Large Scale Ocean Circulation and El Niño*]{}, Kluwer. Ghil, M., I. Zaliapin, and S. Thompson (2008), A delay differential model of ENSO variability: parametric instability and the distribution of extremes, [*Nonlinear Processes in Geophysics*]{}, *15*, 417. Kaplan, D. T., and L. Glass (1992), Direct test for determinism in a time series, [*Phys. Rev. Lett.*]{}, *68*, 427.
Kaplan, D. T., and L. Glass (1993), Coarse-grained embedding of time series: random walks, Gaussian random processes, and deterministic chaos, [*Physica D*]{}, *64*, 431.
Kondrashov, D., S. Kravtsov, A. W. Robertson, and M. Ghil (2005), A hierarchy of data-based ENSO models, [*J. of Climate*]{}, *18*, 4425. Münnich, M., M. A. Cane, and S. E. Zebiak (1990), A study of self-excited oscillations of the tropical ocean-atmosphere system. Part 2: Nonlinear cases, [*J. Atmos. Sci.*]{}, *48*, 1238. Neelin, J. D. (2011), [*Climate change and climate modelling*]{}, Cambridge University Press. Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, A. Kaplan (2003), Global analysis of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century, [*J. Geophys. Res.*]{}, *108*, 4407, doi: 10. 1029/2002JD002670. Suarez, M. J., and P. S. Schopf (1988), A delayed action oscillator for ENSO, [*J. Atmos. Sci.*]{}, *45*, 3283. Sura, P. (2003), Stochastic analysis of Southern and Pacific ocean sea surface winds, [*J. Atmos. Sci.*]{}, *60*, 654. Takens, F. (1981), Detecting strange attractors in fluid turbulence, in: [*Dynamical Systems and Turbulence*]{} (1981) , edited by D. Rand, and L. S. Young, [*Lect. Notes Math.*]{}, *898*, 366. Tziperman, E., L. Stone, M. Cane, H. Jarosh (1994), El Niño chaos: Overlapping of resonances between the seasonal cycle and the Pacific ocean-atmosphere oscillator, [*Science*]{}, *264*, 72. Tziperman, E., M. A. Cane, and S. E. Zebiak (1995), Irregularity and locking to the seasonal cycle in an ENSO prediction model as explained by the quasi-periodicity route to chaos, [*J. Atmos. Sci.*]{}, *52*, 293. Živković, T., and K. Rypdal (2011), Low-dimensionality and predictability of solar wind and global magnetosphere during magnetic storms, [*J. Geophys. Res.*]{}, *116*, A10215, doi:10.1029/2011JA016547. Živković, T., and K. Rypdal (2012), Organization of the magnetosphere during substorms, [*J. Geophys. Res.*]{}, *117*, A05212, doi:10.1029/2011JA016878.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In a recent paper, Mazucheli et al. (2019) introduced the unit-Gompertz (UG) distribution and studied some of its properties. It is a continuous distribution with bounded support, and hence may be useful for modelling life-time phenomena. We present counter-examples to point out some subtle errors in their work, and subsequently correct them. We also look at some other interesting properties of this new distribution. Further, we also study some important reliability measures and consider some stochastic orderings associated with this new distribution.\
\
[**[Keywords]{}**]{} : Log concave; Reliability functions; Stochastic orders.\
\
[**[MSC Classification: Primary: ]{}**]{}62E15, [**[Secondary: ]{}**]{}62G30, 62N05.
author:
- |
M. Z. Anis[^1] & Debsurya De\
\
\
title:
- '**[A Revisit to the Unit-Gompertz Distribution]{}**'
- '**Some More Properties of the Unit-Gompertz Distribution** '
---
Introduction
============
During recent years there has been an increased interest in defining new generated classes of univariate continuous distributions. The works of Zografos and Balakrishnan (2009) and Ristić and Balakrishnan (2012) may be mentioned as examples. Earlier, Eugene et al. (2002) introduced a general class of distributions generated from the logit of the beta random variable. The so called $ T-X $ transformation introduced by Alzaatreh et al. (2013) is another such attempt. In a similar vein, Mazucheli et al. (2019) introduced the unit-Gompertz (UG) distribution and studied some of its properties. More specifically, they considered the random variable $ X=e^{-Y}, $ where $ Y $ has the Gompertz distribution. They erroneously claimed that its hazard rate function can admit all possible forms depending on the parameter. The support of this new distribution is $ (0, 1). $ It may be viewed as an alternative model for reliability studies where due to physical constraints such as design life of the system or limited power supply, distributions with a finite support might be required. As an application, Jha et al. (2020) consider the problem of estimating multicomponent stress-strength reliability under progressive Type II censoring when stress and strength variables follow unit Gompertz distributions with common scale parameter. Jha et al. (2019) consider reliability estimation in a multicomponent stress –strength based on unit-Gompertz distribution. Kumar et al. (2019) are concerned with inference for the unit-Gompertz model based on record values and inter-record times. However, some of the results presented in Mazucheli et al. (2019) are not entirely correct. We present counter-examples to point out these subtle errors in their work; and subsequently correct them in Section 2. We study conditional moments in Section 3. Other important properties of this distribution are investigated in Section 4. Reliability associated measures are studied in Section 5. Stochastic ordering are considered next in Section 6. Finally, Section 7 concludes the paper.
Counter-examples and Corrections
================================
For convenience, we shall stick to the notation of Mazucheli et al. (2019). Let $ Y \left( \alpha, \beta\right) $ be a non-negative random variable with Gompertz distribution having density function given by $$g\left( y\mid \alpha, \beta\right) = \alpha \beta\exp \left( \alpha+\beta y - \alpha e^{\beta y}\right),$$ where $ y>0; $ and $ \alpha > 0 $ and $ \beta >0 $ are shape and scale parameters, respectively. Using the transformation $$X= e^{-Y},$$ Mazucheli et al. (2019) obtained a new distribution with support on $ \left( 0, 1\right), $ which they refer to as the *unit-Gompertz distribution.* For completeness, we shall list down its pdf and cdf. The pdf of the unit-Gompertz distribution is given by $$f\left( x\mid\alpha, \beta\right) = \frac{\alpha\beta\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right] }{x^{1+\beta}}; \mbox {~~~}\alpha>0, \beta>0, x \in (0, 1) \label{pdf}$$ while its cdf is given by $$F\left( x\mid\alpha, \beta\right) =\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]; \label{cdf}$$ and hence, the survival function is given by $$\bar{F}\left( x\mid\alpha, \beta\right) =1-\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]; \label{sf}$$
Shape
-----
Mazucheli et al. (2019) erroneously stated (in their Proposition 1) that the pdf is log concave and unimodal over the *entire* support of $ X. $ As a counterexample, consider $ \alpha=0.25 ,$ and $ \beta=1. $ Let us consider the sign of the second derivative of $ \log f\left( x\mid \alpha, \beta\right) $ at $ x=0.50. $ Routine calculation shows that $ \frac{d^{2}}{dx^{2}}\log f\left( x\mid \alpha=0.25, \beta=1\right)\mid_{x=0.50} >0, $ contradicting Proposition 1 and Equation (6) of Mazucheli et al. (2019). We correct their Proposition 1 as follows:
The pdf of the UG distribution is log concave and unimodal for\
$ x \in \left( 0, \min {\left( \left( \alpha\beta\right) ^{\frac{1}{\beta}}, 1\right) }\right]. $
**Proof:** The second derivative of $ \log f\left( x\mid \alpha, \beta\right) $ is given by $$\frac{d^{2}}{dx^{2}}\log f\left( x\mid \alpha, \beta\right)=-\frac{\left( 1+\beta\right) }{x^{2}}\left( \frac{\alpha \beta}{x^{\beta}}-1\right) .$$ Now observe that $ \alpha>0, $ $ \beta>0 $ and $ 0<x<1; $ hence$ \frac{\left( 1+\beta\right) }{x^{2}} $ is always $ >0. $\
Hence, $\frac{d^{2}}{dx^{2}}\log f\left( x\mid \alpha, \beta\right)<0 $ if $ x\leq \left( \alpha\beta\right) ^{\frac{1}{\beta}}. $
This means that $ \log f\left( x\mid \alpha, \beta\right) $ is concave and unimodal for $ \alpha>0, $ $ \beta>0 $ and $ x \in \left( 0, \min {\left( \left( \alpha\beta\right) ^{\frac{1}{\beta}}, 1\right) }\right]. $ This completes the proof.
$\blacktriangleleft$
Clearly, we see that the graph of $ f(x;\alpha=2; \beta=1 ) $ is log-concave while the graph of $ f(x;\alpha=0.25; \beta=1 ) $ is not.
Mode
----
Consider the UG distribution with shape parameter $ \alpha = 3 $ and scale parameter $ \beta=1. $ According to equation (10) of Mazucheli et al. (2019), the modal point is $$x_{0}= \left( \frac{\alpha \beta}{1+\beta}\right) ^{\frac{1}{\beta}},$$ which in this particular case simplifies to 1.5. However, the support of the UG distribution is $ \left( 0, 1\right). $ The plot of the density function is given in Fig. 2 below.
\
It is easy to see from Fig. 2, that the mode of $ f(x;\alpha=3; \beta=1 ) $ is at $ x=1,$ while the mode of $ f(x;\alpha=1; \beta=1 )$ is at $ x=0.5. $ We formalize this and now correct their result.\
It is easy to see that the first derivative of $ \log f\left( x\mid \alpha, \beta\right) $ is given by $$\frac{d}{dx}\log f\left( x\mid \alpha, \beta\right)=-\frac{1+\beta}{x}+\frac{\alpha \beta}{x^{\beta+1}}.$$ Hence, the mode of $ f\left( x\mid \alpha, \beta\right) $ is $ x^{\star}, $ the root of the equation
$$\frac{d}{dx}\log f\left( x\mid \alpha, \beta\right) =0, \label{A10}$$
if $ x^{\star} \leq 1 $ where $ x^{\star}=\left( \frac{\alpha \beta}{1+\beta}\right) ^{\frac{1}{\beta}}. $ Hence the unique modal point is given by $$x_{mode}= \min \left( x^{\star}, 1\right) .$$
Hazard Rate
------------
Mazucheli et al. (2019) have erroneously mentioned (on page 27 of their paper) that $ \lim_{x \rightarrow 1} h(x)= \alpha \beta; $ and concluded that “monotonically increasing shapes are possible for all values of $ \alpha>1 $ and $ \beta\geq1"; $ and “possibly bathtub shapes of the hazard rate function will happen when $ \alpha\leq 0.5". $ Subsequently, they have used the result of Glaser (1980) to conclude (in their Theorem 3) that the hazard rate (HR) of the UG distribution is upside-down bathtub shaped. They have sketched the hazard rate plot for different values of $ \alpha $ and $ \beta $ (in their Figure 2); but, it is important to note that not a single one of these graphs seems to be upside-down bath-tub shaped. Fig. 3 shows the hazard plot of for a few selected values of $ \alpha$ and $\beta.$\
We see from Fig. 3 that the hazard rate is *not* upside-down bathtub shaped as enunciated in Theorem 3 of Mazucheli et al. (2019). This means that that their Theorem 3 is *not* correct. In fact, the shape of the hazard rate for the UG distribution *cannot* be obtained by appealing to Glaser’s results. This is because Glaser’s results are useful when the support of the distribution is $ \left( 0, \infty\right).$ However, in this case of the UG distribution, the support is the finite interval $ \left( 0, 1\right). $ Ghitany (2004) has obtained sufficient conditions to characterize the shape of the hazard rate when the support of the distribution is finite, say $ \left( 0, b\right). $ But even Ghitany’s theorem *cannot* be applied to the UG distribution because $ f\left( 1\right) = \alpha \beta; $ whereas Ghitany’s theorem demands that $ f(b)=0, $ where $ b $ is the left-end support of $ f. $\
For the UG distribution, we have the hazard rate $$h(x)= \frac{f(x)}{\bar{F}(x)}=\frac{\alpha\beta\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]}{x^{1+\beta}\left\lbrace 1-\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]\right\rbrace }.$$ Observe that $ \lim_{x \rightarrow 0+} h(x)=0 $ and $ \lim_{x \rightarrow 1-} h(x)= \infty. $ Hence, there exists $ 0<M<1 $ such that $ h(x) $ is increasing in $ \left( M, 1\right),$ suggesting that the hazard rate cannot be upside-down bath-tub shaped.
Conditional Moments
===================
The notions of conditional expectation (moment) and independence are routinely discussed in elementary probability and statistics courses at the undergraduate level. However, conditional moments are important in their own right, especially in probability theory and economics. Bryc (1996) considers conditional moment representations for dependent random variables. Domínguez and Lobato (2004) introduce simple and consistent estimation procedure for economic models directly based on the definition of conditional moments. For the UG distribution with parameters $ \alpha $ and $ \beta, $ we have the following theorem:
Let $ X $ follow the UG distribution with parameters $ \alpha $ and $ \beta. $ Then the conditional moment of $ X $ is given by $$E\left( X^{n} \mid X>x\right) =\frac{e ^{\alpha}\alpha^{n/\beta}\left[ \Gamma\left( 1-\frac{n}{\beta}; \alpha \right) -\Gamma\left( 1-\frac{n}{\beta}; \frac{\alpha}{x^{\beta}} \right) \right], }{1-\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]}$$ where $ \Gamma\left( s;x\right) $ is the upper incomplete gamma function defined by $$\Gamma\left( s; x\right) =\int_{x}^{\infty}t^{s-1}e^{-t} dt.\label{inc-up-gamma}$$
**Proof:** The conditional moment, $ E\left( X^{n} \mid X>t\right) $ can be written as $$E\left( X^{n} \mid X>t\right) =\frac{1}{S\left( t\right) }I_{n}^{*}\left( t\right),$$ where $$S\left( t\right)= 1- F\left( t\right)$$ and $$\begin{aligned}
I_{n}^{*}\left( t\right) & = & \int_{t}^{1}y^{n} f(y) dy \label{eqa1} \\
&=& \alpha \beta e^{\alpha}\int_{t}^{1}y^{n-1-\beta}\exp \left( -\frac{\alpha}{y^{\beta}}\right)dy \nonumber \\
&=& \alpha e^{\alpha}\int^{1/t^{\beta}}_{1}\frac{1}{z^{n/\beta}}e ^{-\alpha z} dz \nonumber \\
&=& e^{\alpha}\alpha^{n/\beta}\int^{\alpha/t^{\beta}}_{\alpha} u^{-n/\beta}e^{-u}du \nonumber \\
&=& e^{\alpha}\alpha^{n/\beta}\left[ \Gamma\left( 1-\frac{n}{\beta}; \alpha \right) - \Gamma\left( 1-\frac{n}{\beta}; \frac{\alpha}{t^{\beta}}\right)\right]; \label{eqn2} \end{aligned}$$ where $ \Gamma\left( s, x\right) $ is the upper incomplete gamma function defined in (\[inc-up-gamma\]) above. This completes the proof.
$\blacktriangleleft$
As applications of the concept of conditional moment, we may consider the evaluation of the mean residual life, the mean deviations about the mean and the median and the expected inactivity time. These are discussed in the subsequent sections.
Other Important Properties
==========================
We shall now look at some other other distributional properties not considered in Mazucheli et. al (2019). Specifically we shall consider the mean deviation, the entropy, the Lorenz, Bonferroni and Zenga curves and order statistics.
Mean Deviation
--------------
For empirical purposes, the shape of a distribution can be described by the so-called *first incomplete moment,* defined by $ m_{1}\left( z\right) =\int_{0}^{z}xf\left( x\right) dx=\mu - I_{1}^{*}\left( z\right), $ where $ I_{n}^{*}\left( z\right) $ is defined in (\[eqa1\]) above with $ n=1 $. This plays an important role in measuring inequality and is used to measure the dispersion and the spread in a population from the center. We shall first state a useful lemma.
The mean deviation about any arbitrary point $ x_{0} $ is given by $$\delta\left( x_{0}\right)=E\left( \lvert X- x_{0}\rvert\right)= \int_{0}^{1}\mid x - x_{0}\mid f\left( x\right) dx = 2x_{0}F\left( x_{0}\right) -x_{0}F\left(0\right) -\mu +2 I_{1}^{*}\left( x_{0}\right) - x_{0},$$ where $ F\left(0\right) $ is defined as 0.
The proof is simple and hence omitted.\
Then the mean deviation about the mean is given by
$$\delta\left( \mu\right) = E\left( \lvert X- \mu \rvert\right)=2\mu F\left( \mu\right) -2 m_{1}\left( \mu\right) =2\mu F\left( \mu\right) -2\mu +2 I_{1}^{*}\left( \mu\right);$$ and the mean deviation about the median $$\delta\left( M\right) =E\left( \lvert X- M \rvert\right)=\mu F\left( \mu\right) -2 m_{1}\left( M\right)=2M F\left( M\right) -\mu +2 I_{1}^{*}\left( M\right) -M$$ where $ \mu=E\left( X\right), $ $ M= Medain \left( X\right), $ $ m_{1}\left( z\right)= \int_{0}^{z}xf\left( x\right) dx $ is the first incomplete moment; and $ I_{1}^{*}(t) $ is as defined in (\[eqa1\]) above . The algebraic expressions for the mean and the median have already been obtained by Mazucheli et al. (2019); and hence $ \delta\left( \mu\right) $ and $ \delta\left( M\right) $ can be easily evaluated numerically.
Entropies
---------
An entropy is a measure of uncertainty of a random variable $ X. $ A large value of entropy implies greater uncertainty in the data. The concept of entropy is important in different subjects including communication theory, economics, physics, probability and statistics. Nanda and Chowdhury (2020) provide a useful review. Several measures of entropy have been studied and compared in the literature. Two popular entropy measures are the Shannon and R[é]{}nyi entropies (Shannon (1951)and R[é]{}nyi (1961)). The R[é]{}nyi entropy of a random variable $ X $ with pdf $ f\left( \cdot\right) $ is defined as $$I_{R}\left( \gamma\right) =\frac{1}{1-\gamma}\ln \int_{-\infty} ^{\infty}f^{\gamma}\left( x\right) dx,$$ for $ \gamma>0 $ and $ \gamma \neq 1; $ while the Shannon entropy is given by $ E\left[ -\ln f\left( X\right) \right] . $ It is a particular case of the R[é]{}nyi entropy for $ \gamma\uparrow 1. $\
First, we shall calculate the R[é]{}nyi entropy. Towards this end, we compute $$\int_{0}^{1}\left[ f\left( x\right)\right] ^{\gamma} dx= \left( \alpha \beta e^{\alpha}\right) ^{\gamma}\frac{1}{\beta}\left( \alpha \gamma\right) ^{\frac{1}{\beta}\left[ 1-\gamma\left( 1+\beta\right) \right] }\Gamma\left( \gamma+\frac{1}{\beta}\left(\gamma -1\right); \alpha\gamma\right)$$ where $ \Gamma\left( s; x\right) $ represents the upper incomplete gamma function defined in (\[inc-up-gamma\]) above.\
Then, the R[é]{}nyi entropy of $ X $ is given by $$\begin{aligned}
I_{R}\left( \gamma\right) & =& \frac{1}{1-\gamma}\left[\gamma \ln \left( \alpha \beta e^{\alpha}\right) +\ln \left( \Gamma\left( \gamma +\frac{1}{\beta}\left( \gamma-1\right) ;\alpha\gamma\right) \right)- \ln \beta +\frac{1}{\beta}\left\lbrace 1-\gamma\left( 1+\beta\right) \right\rbrace \ln (\alpha\gamma) \right]\nonumber \\
& = &\frac{1}{1-\gamma}\left[\alpha\gamma + \frac{\left( 1-\gamma\right)}{\beta} \ln \alpha - \left( 1-\gamma\right) \ln \beta +\frac{1}{\beta}\left\lbrace 1-\gamma\left( 1+\beta\right) \right\rbrace \ln (\gamma)\right. \nonumber\\
& & {}\mbox{~~~~~} + \left. \ln \left( \Gamma\left( \gamma +\frac{1}{\beta}\left( \gamma-1\right) ;\alpha\gamma\right) \right) \right] \label{renyi}
\end{aligned}$$
Similarly, the Shannon entropy is given by $$\begin{aligned}
E\left[ -\ln f\left( X\right) \right]&=& 1- \ln \left( \alpha \beta\right) - \left( 1+\beta\right) \frac{e^{\alpha}}{\beta}\Gamma\left( 0;\alpha\right).
\end{aligned}$$
The Shannon entropy can also be obtained by limiting $ \gamma \uparrow 1 $ in the R[é]{}nyi entropy obtained above.
Song (2001) has shown that the gradient of the R[é]{}nyi entropy $ I_{R}^{\prime}\left( \gamma\right)=\left( d/d\gamma\right)I_{R}\left( \gamma\right) $ is related to the log-likelihood by $ I_{R}^{\prime}\left( 1\right) =-\left( 1/2 \right) \mathrm{Var}\left[\left( \log f\left( X\right)\right) \right]. $ This equality and the fact that the quantity $-I_{R}^{\prime}\left( 1\right) $ remains invariant under location and scale transformations motivated Song to propose $-2I_{R}^{\prime}\left( 1\right) $ as a measure of the shape of a distribution. Taking the first derivative of (\[renyi\]) and then limiting $ \gamma\uparrow 1 $ using L’Hospital’s rule, one gets the expression $$\begin{aligned}
I_{R}^{\prime}\left( 1\right) &= & \frac{\beta +2}{2\beta}-\frac{1}{2}\left[ \alpha\left\lbrace \alpha -2\left( 1+\frac{1}{\beta}\right) \ln \alpha\right\rbrace - \left\lbrace \left[ \ln \alpha +e^{\alpha}\Gamma\left( 0; \alpha\right) \right] \left( 1+\frac{1}{\beta}\right) -\alpha \right\rbrace ^{2} \right.\\
& & \mbox{~~~~~~~~~~~~~~~~~~~~}+ \left. e^{\alpha}\left( 1+\frac{1}{\beta}\right)^{2}\int_{\alpha}^{\infty}e^{-t}\left( \ln t\right) ^{2}dt \right] \end{aligned}$$ for the measure proposed by Song (2001). This measure plays a similar role as the kurtosis measure in comparing the shapes of various densities and measuring heaviness of tails.
Lorenz, Bonferroni and Zenga curves
------------------------------------
The Lorenz curve, introduced by Lorenz (1905), was proposed to measure the concentration of wealth. However, since then it has been used in many other areas. See, for example, Aaberge (2000), Jacobson et al. (2005) and Groves-Kirkby et al. (2009) for its diverse use. In the field of reliability mention may be made of the works of Chandra and Singpurwalla (1984), Klefsjö (1984) and Pham and Turkkan (1994).
Similarly, Bonferroni (1930) proposed a curve to measure wealth and income inequality. This curve also has applications in demography, insurance and medicine. Giorgi and Crescenzi (2001) apply the Bonferroni curve to analyze life-testing and reliability.
More recently Zenga (2007) introduced a new index of income inequality, $ Z(x), $ based on the ratio between the lower mean and the upper mean. $ Z(x) $ can be also interpreted as the difference in average age of components which has survived beyond age $x$ from those which has failed before attaining age $x$, expressed in terms of average age of components exceeding age $ x. $ Hence, it can be viewed as a measure of proportional change in average age while switching over from survival before and after attaining age $ x. $ It is also related to the mean residual life function $ e_{F}\left( x\right) $ as follows:
$$Z(x)=\frac{1}{F\left( x\right) } \left[ 1-\frac{E\left( X\right) }{x+e_{F} \left( x\right)}\right].$$ Nair and Sreelakshmi (2012) discuss the Zenga curve in the context of reliability analysis.
These curves are defined as $$L(p)= \frac{1}{\mu}\int_{0}^{q} xf(x) dx;$$
$$B(p)= \frac{1}{p\mu}\int_{0}^{q} xf(x) dx;$$ and $$Z(x)= 1- \frac{\mu^{-}\left( x\right) }{\mu^{+}\left( x\right)}$$
respectively, where $ \mu^{+}\left( x\right)= E\left( X\mid X>x \right), $ $ \mu^{-}\left( x\right)= E\left( X\mid X\leq x\right), $ $q = F^{-1} ( p) $ and $ \mu=E(X). $ Then for the UG distribution, we have $$\begin{aligned}
I(q) & =& \int_{0}^{q} x\frac{\alpha \beta}{x^{1+\beta}}\exp \left[-\alpha\left( \frac{1}{x^{\beta}}-1\right) \right] dx\\
& =& \alpha^{1/\beta}e^{\alpha}\int_{\alpha/q^{\beta}}^{\infty}e^{-z}z^{1-1/\beta-1}dz\\
&=& \alpha^{1/\beta}e^{\alpha}\Gamma\left( 1-\frac{1}{\beta}; \frac{\alpha}{q^{\beta}}\right),\end{aligned}$$ where $ \Gamma (s; x) $ is as defined in (\[inc-up-gamma\]) above. Hence, for $ \beta>1 $ we have $$L(p)=\frac{\alpha^{1/\beta}e^{\alpha}}{\mu}\Gamma\left( 1-\frac{1}{\beta}; \frac{\alpha}{q^{\beta}}\right);$$
$$B(p)=\frac{\alpha^{1/\beta}e^{\alpha}}{p\mu}\Gamma\left( 1-\frac{1}{\beta}; \frac{\alpha}{q^{\beta}}\right);$$ and $$Z(x)=1- \frac{\Gamma\left( 1-\frac{1}{\beta};\frac{\alpha}{x^{\beta}}\right) }{\left[\Gamma\left( 1-\frac{1}{\beta};\alpha\right) - \Gamma\left( 1-\frac{1}{\beta};\frac{\alpha}{x^{\beta}}\right)\right] }\frac{\bar{F}(x)}{F(x)}.$$ $$Z(x)=1- \frac{\Gamma\left( 1-\frac{1}{\beta};\frac{\alpha}{x^{\beta}}\right) }{\left[\Gamma\left( 1-\frac{1}{\beta};\alpha\right) - \Gamma\left( 1-\frac{1}{\beta};\frac{\alpha}{x^{\beta}}\right)\right] }\frac{\left\lbrace 1-\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]\right\rbrace }{\exp \left[ -\alpha\left( 1/x^{\beta}-1\right) \right]}.$$
Order Statistics
----------------
It is well know that if $ X_{(1)} \leq X_{(2)}\leq \cdots \leq X_{(n)} $ denotes the order statistic of a random sample $ X_{1}, X_{2}, \cdots, X_{n} $ from a continuous population with cdf $ F_{X}(x) $ and pdf $ f_{X}(x), $ then the pdf of the $ j\mathrm{-th} $ order statistics is given by $$f_{X_{(j)}}(x)=\frac{n!}{(j-1)!(n-j)!}f_{X}(x)\left[F_{X}(x)\right] ^{j-1} \left[1-F_{X}(x)\right] ^{n-j},$$ for $ j=1, 2, \cdots , n. $ Hence, the pdf of the $ j\mathrm{-th} $ order statistic from the UG distribution will be given by $$f_{X_{(j)}}(x)=\frac{n!}{(j-1)!(n-j)!}\frac{\alpha \beta e^{\alpha j} e^{-\alpha j/x^{\beta}} } {x^{1+\beta}}\left[ 1- e^{\alpha}e^{-\alpha/x^{\beta}}\right] ^{n-j}.$$
The $ k\mathrm{-th} $ moment of $ X_{(j)} $ is obtained next. We have $$\begin{aligned}
E\left( X^{k}_{\left( j\right) }\right)&=& \int_{0}^{1}x^{k}f_{X_{(j)}}(x) dx\\
&=& \int_{0}^{1}x^{k}\frac{n!}{(j-1)!(n-j)!}\frac{\alpha \beta e^{\alpha j} e^{-\alpha j/x^{\beta}} } {x^{1+\beta}}\left[ 1- e^{\alpha}e^{-\alpha/x^{\beta}}\right] ^{n-j} dx\\
&=&\frac{n!}{(j-1)!(n-j)!}\alpha e^{\alpha j}\int_{1}^{\infty}t^{-k/\beta}e^{-\alpha jt}\left[ 1-e^{\alpha}e^{-\alpha t}\right] ^{n-j}dt\\
&=&\frac{n!}{(j-1)!(n-j)!}\alpha e^{\alpha j}\int_{1}^{\infty}t^{-k/\beta}e^{-\alpha jt}\sum_{r=0}^{n-j} {
\left(
\begin{array}{@{}c@{}}
n-j \\ r
\end{array}
\right)
}\left( -e^{\alpha}e^{-\alpha t}\right) ^{r}dt\\
&=&\frac{n!}{(j-1)!(n-j)!}\alpha e^{\alpha j}\sum_{r=0}^{n-j} {
\left(
\begin{array}{@{}c@{}}
n-j \\ r
\end{array}
\right)
}\left( -1\right) ^{r}e^{r \alpha}\int_{1}^{\infty}t^{-k/\beta}e^{-\alpha t\left( j+r\right) }dt\\
&=&\frac{\alpha e^{\alpha j}n!}{(j-1)!(n-j)!}\sum_{r=0}^{n-j} {
\left(
\begin{array}{@{}c@{}}
n-j \\ r
\end{array}
\right)
}\left( -1\right) ^{r}e^{r \alpha}\left\lbrace \alpha\left( j+r\right) \right\rbrace ^{\frac{k}{\beta}-1} \int_{\alpha\left( j+r\right) }^{\infty}y^{-k/\beta}e^{-y}dy\\
&=&\frac{\alpha e^{\alpha j}n!}{(j-1)!(n-j)!}\sum_{r=0}^{n-j} {
\left(
\begin{array}{@{}c@{}}
n-j \\ r
\end{array}
\right)
}\left( -1\right) ^{r}e^{r \alpha}\left\lbrace \alpha\left( j+r\right) \right\rbrace ^{\frac{k}{\beta}-1} \Gamma\left( 1-\frac{k}{\beta}; \alpha\left( j+r\right) \right), \end{aligned}$$ where $ \Gamma\left( s; x\right) $ is the upper incomplete gamma function defined in (\[inc-up-gamma\]) above. Note that, the moments exists only when $ k< \beta. $
Some Other Important Reliability Functions
==========================================
We shall now discuss the mean residual life (MRL), reversed hazard rate (RHR) and expected inactivity time (EIT) for the UG distribution. The RHR and EIT may be looked as the dual properties of the HR and the MRL functions. We shall also investigate the monotonicity of these duals. Finally, we shall discuss stress strength reliability for the UG distribution.
Mean Residual Life
------------------
An important ageing measure of interest is the mean residual life (MRL) function. It is defined as $$e_{F}\left( t\right) = E\left( X-t\mid X >t\right).$$
Physically, it measures the expected remaining life time for a unit having already survived up to time $ t. $ It is easy to see that $ e_{F}\left( 0\right) = E\left( X\right) =\mu,$ the mean of $ X. $ It can be calculated using the cdf or the pdf. Specifically, we have $$\begin{aligned}
e_{F}\left( t\right) & = & \frac{1}{\bar{F}\left( t\right)}\int_{t}^{\infty}\bar{F}\left( u\right)du \\
& = & \frac{\int_{t}^{\infty}uf\left( u\right) du}{\bar{F}\left( t\right) }-t \\
&=& \frac{1}{\bar{F}\left( t\right)}I_{1}^{*}\left( t\right) -t; \end{aligned}$$ where $ I_{1}^{*}\left( t\right) $ can be obtained from (\[eqa1\]) above with $ n=1.$ The expression of the MRL fucntion is quite complicated. However, since the hazard rate is increasing, it follows from Theorem 3 of Mi (1995) that the the MRL function $ e_{F}\left( t\right) $ is decreasing. Figure 4 shows the graph of the MRL function for some combinations of $ \alpha$ and $\beta. $\
Expected Inactivity Time
------------------------
The Expected Inactivity Time (EIT) (also known as the mean past lifetime function) of a non-negative continuous random variable $ X $ with cumulative distribution function $ F\left( x\right) $ is defined as $$I\left( x\right) =E\left( x-X\mid X\leq x\right) = \frac{1}{F\left( x\right)}\int_{0}^{x}F\left( y\right)dy.$$ Hence, $ I\left( x\right) $ defines the mean waiting time for a device that failed in the interval $ \left[ 0, x\right] . $ In other words, this conditional random variable shows the time elapsed from the failure of the component given that its lifetime is less than or equal to $x.$ The EIT is a dual property of the MRL; and is an important characteristic in many reliability applications. It has application in many disciplines such as survival analysis, actuarial studies and forensic science, to name but a few. It is also of interest while describing different maintenance strategies. Chandra and Roy (2001) have shown that the EIT function cannot decrease on $ \left( 0, \infty\right); $ while Kundu and Nanda (2010) have studied some of its reliability properties. The non-parametric smooth estimation of the EIT function has been studied by Jayasinghe and Zeephongsekul (2013).\
For the UG distribution, after detailed computation we
$$I\left( x\right) =\frac{e^{\alpha}\alpha^{1/\beta }}{\beta F\left( x\right)}\int_{\alpha /x^{\beta}} ^{\infty}e^{-u}\frac{du}{u^{1+1/\beta}}=\frac{e^{\alpha/x^{\beta}}\alpha^{1/\beta}}{\beta }\Gamma\left( \frac{-1}{\beta};\frac{\alpha}{x^{\beta}}\right),$$ which can be evaluated numerically.
Reversed Hazard Rate
--------------------
The reversed hazard rate (RHR) of a non-negative continuous random variable $ X $ with pdf $ f\left( x\right) $ and cdf $ F\left( x\right) $ at time $ x $ is defined as $$r\left( x\right) =\lim _{\Delta x\rightarrow 0}\frac{P\left( X>x-\Delta x\mid X\leq x\right) }{\Delta x}=\frac{f\left( x\right) }{F\left( x\right)}.$$ Thus, $r\left( x\right) $ defines the conditional probability of failure of a unit in $ \left( x-\Delta x,x\right) $ given that the failure had occurred in in $ \left[ 0, x\right]. $ The RHR is a dual property of the HR. However, it should be noted that the trend in RHR is *not* a direct indicator of the ageing pattern of a unit. The RHR has many interesting applications. Nanda and Shaked (2001) list the usefulness of the RHR while analyzing queuing systems. The RHR order arises naturally in economics and risk theory; see for example Eeckhoudt and Gollier (1995) and Veres-Ferrer and Pavía (2014). It is useful in estimating the survival function for left-censored lifetimes, see for example, Kalbfliesch and Lawless (1989). Irrespective of the shape of the hazard rate function, the RHR cannot increase on $ \left( 0, \infty\right) , $ as shown by Block et al. (1998). Testing the behaviour of the RHR is dealt with in Kayid et al. (2011).\
For the UG distribution, we have $$r\left( x\right) =\frac{\alpha \beta}{x^{1+\beta}}.$$
Relationship
------------
Chandra and Roy (2001) have proved the following:\
$ F\left( x\right) $ is log-concave $ \Leftrightarrow $ $ r\left( x\right) $ is decreasing $ \Rightarrow $ $I\left( x\right) $ is increasing.
We shall use the above result to prove the following theorem:
If $ X\backsim UG\left( \alpha, \beta\right) , $ then $ X $ has decreasing RHR (increasing EIT).
**Proof:** We have $$\frac{d^{2}}{dx^{2}}\log F\left( x\mid \alpha, \beta\right)=-\frac{\left( 1+\beta\right) \alpha \beta}{x^{2+\beta}}<0$$ since $ \alpha>0, $ $ \beta>0 $ and $ 0<x<1. $ Hence, $ F\left( x\mid \alpha, \beta\right) $ is log-concave. The theorem now follows from the result of Chandra and Roy (2001).
$ \blacktriangleleft $
Figures 5 and 6 show respectively, the expected inactivity time and the reversed hazard rate for the UG distribution with parameters $ \alpha =0.25; 0.50; 0.75; 1.0 $ and $ \beta =1. $
\
Stress Strength Reliability
---------------------------
Next, we derive the reliability $ R=Prob \left( Y<X\right) , $ where $ X \backsim UG \left( \alpha_{1}, \beta_{1}\right) $ and $ Y \backsim UG \left( \alpha_{2}, \beta_{2}\right) $ are independent random variables with distribution functions $ F_{X} $ and $ G_{Y} $ respectively. Notionally, we may think of $ X $ as the strength and $ Y $ as the stress. Probabilities of this form have many engineering applications. $$\begin{aligned}
R & = & Prob \left( Y<X\right)\\
& =& \int_{0}^{1} G_{Y}\left( x\right)f_{X} \left( x\right)dx\\
& =&\int_{0} ^{1}\frac{\alpha_{1}\beta_{1}}{x^{1+\beta_{1}}}\exp \left[-\alpha_{1} \left( \frac{1}{x^{\beta_{1}}}-1\right) \right] \exp \left[-\alpha_{2} \left( \frac{1}{x^{\beta_{2}}}-1\right) \right]dx\\
& =&\alpha_{1}\beta_{1}e^{\alpha_{1}+\alpha_{2}}\int_{0} ^{1}\frac{1}{x^{1+\beta_{1}}}\exp \left[- \left( \frac{\alpha_{1}}{x^{\beta_{1}}}+\frac{\alpha_{2}}{x^{\beta_{2}}}\right) \right]dx,\end{aligned}$$ which can be evaluated numerically. However, if the strength and stress distributions are independent random variables with with common scale parameter $ \beta, $ then we get a neat expression for $ R $ as $$\frac{\alpha_{1}}{\alpha_{1}+\alpha_{2}}$$
Stochastic Orderings
====================
Comparison of random variables based on their means, medians or variances is not very informative. The need to provide a more detailed comparison of two random quantities has been the origin of the theory of stochastic orders that has grown significantly during the last forty years. We shall begin by recalling some basic definitions.\
Let $X$ and $Y$ be random variables with distribution functions $F_{X}$ and $F_{Y},$ and survival functions $ \bar{F}_{X} $ and $ \bar{F}_{Y} $, respectively. Denote $ F^{-1}_{X}\left( u\right) = \sup \left\lbrace x: F_{X}\left( x\right) \leq u \right\rbrace $ and $ F^{-1}_{Y}\left( u\right) = \sup \left\lbrace x: F_{Y}\left( x\right) \leq u \right\rbrace $ for $ u \in \left[ 0, 1\right] $ the right continuous inverses of $F_{X}$ and $F_{Y}.$\
**Definition:** A random variable $ X $ is said to be smaller than a random variable $ Y $ in the
- usual stochastic order (denoted by $ X\leq_{st} Y $) if $ F_{X}\left( t\right)\geq F_{Y}\left( t\right) $ for all real $ t; $
- hazard rate order (denoted by $ X\leq_{hr} Y $) if $ \bar{F}_{X}\left( t\right)/ \bar{F}_{Y}\left( t\right) $ decreases in $ t; $
- reversed hazard rate order (denoted by $ X\leq_{rh} Y $) if $ F_{X}\left( t\right)/ F_{Y}\left( t\right) $ decreases in $ t; $
- mean residual life order (denoted by $ X\leq_{mrl} Y $) if $ \mu_{X}\left( t\right) \leq \mu_{Y}\left(t\right); $
- expected inactivity time order (denoted by $ X\leq_{eit} Y $) if $ I_{X}\left( t\right) \geq I_{Y}\left( t\right); $
- likelihood ratio order (denoted by $ X\leq_{lr} Y $) if $ f_{X}\left( t\right)/ f_{Y}\left( t\right) $ decreases in $t;$
- increasing convex order (denoted by $ X\leq_{icx} Y $) if $ \int_{x}^{+\infty} \bar{F}_{X}\left( t\right)dt \leq \int_{x}^{+\infty} \bar{F}_{Y}\left( t\right)dt, $ for all $ x, $ provided the two integrals exist;
- increasing concave order (denoted by $ X\leq_{icv} Y $) if $ \int_{-\infty}^{x} F_{X}\left( t\right)dt \geq \int_{-\infty}^{x} F_{Y}\left( t\right)dt, $ for all $ x, $ provided the two integrals exist;
- dispersive order (denoted by $ X\leq_{disp} Y $) if whenever $ 0 <\alpha \leq \beta <1; $
- stochastic-variability order (denoted as $ X\leq_{st:icx} Y $) if $ X\leq_{st} Y $ and for any increasing convex function $h,$ provided the two variances exist;
- harmonic mean residual life order (denoted by $ X\leq_{hmrl} Y $) if $$\left[ \frac{1}{x}\int_{0} ^{x}\frac{1}{m\left( u\right) }du\right] ^{-1}\leq \left[ \frac{1}{x}\int_{0} ^{x}\frac{1}{l\left( u\right) }du\right] ^{-1} \mbox{~~~~ for all $x>0$}$$ where $ m (u) $ and $ l (u) $ are the mrl functions of the random variables $ X $ and $ Y $ respectively;
- star-shaped order (denoted by $ X\leq_{ss} Y $) if $ E\left[ \phi\left( X\right) \right] \leq E\left[ \phi\left( Y\right) \right] $ for all star-shaped functions $ \phi:\left[ 0, \infty\right) \longrightarrow \phi:\left[ 0, \infty\right), $ provided the expectations exist.
- total time on test order (denoted by $ X\leq_{ttt} Y $) if and only if $$\int_{0}^{F^{-1}_{X}(p)}\bar{F}_{X}(x) dx\leq \int_{0}^{F^{-1}_{Y}(p)}\bar{F}_{Y}(x) dx, \mbox{~~~~} p \in (0, 1).$$
The following implications are well known: $$X\leq_{eit}Y \stackrel {(e)}\Longleftarrow X\leq_{rh}Y\stackrel {(c)}\Longleftarrow X\leq_{lr}Y \stackrel {(a)}\Rightarrow X\leq_{hr}Y \stackrel {(b)}\Rightarrow X\leq_{mrl}Y \stackrel {(d)}\Rightarrow X\leq_{hmrl}Y.$$ $$X\leq_{hr}Y \stackrel {(f)}\Rightarrow X\leq_{st}Y \stackrel {(g)}\Rightarrow X\leq_{ss}Y \stackrel {(h)}\Rightarrow X\leq_{icx}Y$$ $$X\leq_{st}Y \stackrel {(i)}\Rightarrow X\leq_{ttt}Y \stackrel {(j)}\Rightarrow X\leq_{icv}Y$$
The implications (a), (b) and (c) are given in Shaked and Shanthikumar (2007; page 43); implication (d) is in Shaked and Shanthikumar (2007; page 95) and implication (e) is given in Finkelstein (2002). The implications (f) is given in Shaked and Shanthikumar (2007; page 18) while the implications (g) and (h) can be found in Shaked and Shanthikumar (2007; page 205). The implications (i) and (j) can be found in Shaked and Shanthikumar (2007; page 224 and 225 respectively).\
The UG distributions are ordered with respect to the strongest likelihood ratio ordering as shown in the following theorem.\
Let $ X\backsim UG \left( \alpha_{1}, \beta\right) $ and $ Y\backsim UG \left( \alpha_{2}, \beta\right). $ If $ \alpha_{1} < \alpha_{2} ,$ then $$X\leqq_{lr} Y \left( X\leqq_{hr} Y; X\leqq_{rh} Y; X\leqq_{mrl} Y; X\leqq_{eit} Y \right) .$$
**Proof:** Let the corresponding pdfs be denoted by $ f_{X}\left( x\mid \alpha_{1}, \beta\right) $ and $ f_{Y}\left( x\mid \alpha_{2}, \beta\right) $ where $0<\alpha_{1} < \alpha_{2} $ are the respective shape parameters and $ \beta>0 $ is the common scale parameter.\
Now observe that $$\frac{f_{X}\left( x\mid \alpha_{1}, \beta\right) }{f_{Y}\left( x\mid \alpha_{2}, \beta\right)}=\frac{\alpha_{1}}{\alpha_{2}}\exp \left(\alpha_{1}-\alpha_{2}\right) \cdot \exp \left[ -x^{-\beta} \left(\alpha_{1}-\alpha_{2}\right)\right], \mbox{~~~} \alpha_{1}<\alpha_{2}.$$ Hence, for $ \alpha_{1}<\alpha_{2}, $ we have $$\frac{d}{dx}\left(\frac{f_{X}\left( x\mid \alpha_{1}, \beta\right) }{f_{Y}\left( x\mid \alpha_{2}, \beta\right)} \right)= \frac{\alpha_{1}}{\alpha_{2}}\exp \left(\alpha_{1}-\alpha_{2}\right) \cdot \beta \left(\alpha_{1}-\alpha_{2}\right) x^{-\beta - 1} \exp \left[ -x^{-\beta} \left(\alpha_{1}-\alpha_{2}\right)\right]<0.$$ This means $ {f_{X}\left( x\mid \alpha_{1}, \beta\right) }/{f_{Y}\left( x\mid \alpha_{2}, \beta\right)} $ is decreasing in $ x. $ Hence, $ X\leq_{lr}Y .$\
The remaining statements follow from the implications given above. This proves the theorem.
$\blacktriangleleft$
Conclusion
==========
This paper may be considered as an essentail follow-up paper of Mazucheli et al. (2019). It corrects some of the errors of the earlier paper. Some other important properties have been discussed. Since the proposed distribution can be used for modelling lifetime data, properties associated with lifetime distributions have been studied in the present work. It is hoped that this work will be a necessary complement to Mazucheli et al.(2019).
[99]{}
Mazucheli, J.; Menezes, AF; Dey, S. (2019) Unit-Gompertz Distribution with applications. Statistica. 79, No. 1. pp. 25-43.
Zografos, K.; Balakrishnan, N. (2009). On families of beta and generalized gamma-generated distributions and associated inference. Statistical Methopdology. 6. 344-362.
Ristić, MM; Balakrishnan, N. (2012). The gamma exponentiated exponential distribution. J. Stat. Comput. Simul. 82; 1191-1206.
Eugene, N; Lee, C; Famoye, F. (2002). Beta-Normal distributiona and its applications. Commun. Statist. - Theory Meth. 31. No. 4; pp 497-512.
Alzaatreh, A; Famoye, F; Lee, C. (2013). A new method for generating families of continuous distributions. Metron. 71. 63-79.
Chandra, NK; Roy, D. (2001). The reversed hazard rate function. Prob. in the Engineering and Informational Sciences. 15; pp. 95-102. Kundu, C. Nanda, AK. (2010). Some reliability properties of the inactivity time. Commun. Statist. - Theory Meth. 39. No. 5; pp 899-911.
Jayasinghe, C. Zeephongsekul, P. (2013). Non-parametric smooth estimation of the EIT function. Jor. Stat. Plan and Inf. 143, No. 5. pp 911-928. Nanda, AK; Shaked, M. (2001). The hazard rate and the reversed hazard rate orders, with application to order statistics. Ann. Inst. Stat. Math 53; No. 4. 853-864.
Veres-Ferrer, EJ; and Pavía, JM. (2014). On the relationship between the reversed hazard rate and elasticity. Statistical Papers volume 55, pp 275–284.
Kalbleisch, JD; Lawless, JF. (1989). Infeeence based on retrospective ascertainment: An analysis of the data on transfusion-related AIDS. Jor. American Stat. Assoc. 84; pp 360-372.
Block, H.; Savits, TH; Singh, H. (1989). The reversed hazard rate function. Prob. in the Engineering and Informational Sciences. 12; pp. 69-90.
Kayid, M; Al-Nahawati, H.; Ahmad, IA (2011) Testing behavior of the reversed hazard rate. Applied Mathematical Modelling. 35, No. 5; pp. 2508-2515.
Ghitany, ME. (2004). The monotonicity of the reliability measures of the beta distribution. Applied Mathematics Letters. 17; pp. 1277-1283.
Mi, J. (1995). Bathtub failure rate and upside-down bathtub mean residual life. IEEE Trans Rel. 44; No. 3. pp 388-391.
Gupta, RC; Akman, HO. (1995). Mean residual life function for certain types of non-monotonic ageing. Communications in Stat. - Stochastic Models. 11; No. 1. pp 219-225.
Tang, LC; Lu, Y; Chew, EP. (1999). Mean residual life of lifetime distributions. IEEE Trans Rel. 48; No. 1. pp 73-77.
Finkelstein, MS. (2002). On the reversed hazard rate. Rel Engg and System Safety. 78; pp 71-75.
Shaked, M.; Shanthikumar, JG. (2007). *Stochastic Orders*.Springer, New York.
Jha, MK; Dey, S; Alotaibi, RM; and Tripathi, YM. (2020). Reliability estimation of a multicomponent stress-strength model for unit Gompertz distribution under progressive Type II censoring. *Quality & Reliability Engineering International.* **36**; pp 965-987.
Bryc, W. (1996). Conditional moment representations for dependent random variables. *Electronic Journal of Probability.* **1**; Paper no. 7, 14 pp.
Domínguez, MA; and Lobato, IN. (2004). Consistent Estimation of Models Defined by Conditional Moment Restrictions. *Econometrica.* **72**; 1601-1615 .
Nanda, A. and Chowdhury, S. (2020) Shannon’s Entropy and Its Generalisations Towards Statistical Inference in Last Seven Decades *International Statistical Review*
Lorenz, M. (1905). Methods of measuring the concentration of wealth. American StatisticalAssociation, 9: 209–219 Aaberge, R. (2000). Characterizations of Lorenz curves and income distributions. Social Choice and Welfare, 17: 639–653 Bonferroni, E. (1930). Elementi di statistica generale. Libreria Seber,Firenze. Groves-Kirkby, C., Denman, A., and Phillips, P. (2009). Lorenz curve and Gini coefficient: Novel tools for analysing seasonal variation of environmental radon gas. Journal of Environmental Management, 90: 2480–2487 Jacobson, A., Milman, A., and Kammen, D. (2005). Letting the (energy) Gini out of the bottle:Lorenz curves of cumulative electricity consumption and Gini coefficients as metrics of energy distribution and equity. Energy Policy, 33: 1825–1832 Giorgi, GM and Crescenzi, M. (2001). A look at the Bonferroni inequality measure in a reliability framework. *Statistica.* **LXI**; 571-583.
Chandra, M; and Singpurwalla, ND. (1981). Relationships between some notions which are common to reliability theory and economics. *Mathematics of Operations Research.* **5**; 113-121. Klefsjö, B. (1984). Reliability interpretations of some concepts from economics. *Naval Research Logistics Quarterly.* **31**; 301-308.
Pham, TG; and Turkkan, N. (1994). The Lorenz and the scaled total-time-on-test transform curves: a unifies approach. *IEEE TRansactions on Reliability.* **43**; 76-84.
Kumar, D; Dey, S; Ormoz, E; MirMostafaee, S. M. T. K. (2019). Inference for the unit-Gompertz model based on record values and inter-record times with an application. *Rendiconti del Circolo Matematico di Palermo.*
Jha, MK; Dey, S; Tripathi, YM. (2019).Reliability estimation in a multicomponent stress–strength based on unit-Gompertz distribution. *International Journal of Quality & Reliability Management.*
Zenga, M. (200)7. Inequality curve and inequality index based on the ratios between lower and upper arithmetic means. *Statistica & Applicazioni.* **5**;3–27.
Nair, KRM; and Sreelakshmi, N. (2016). The new Zenga curve in the context of reliability analysis. *Communications in Statistics - Theory & Methods*. **45** (22); 6540-6552. K.-S. Song. (2001). R[é]{}nyi information, loglikelihood and an intrinsic distribution measure. *Journal of Statistical Planning and Inference*. **93**; 51-69.
[^1]: Corresponding Author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider a pairwise interacting quantum 3-body system in 3-dimensional space with finite masses and the interaction term $V_{12} + \lambda(V_{13} + V_{23})$, where all pair potentials are assumed to be nonpositive. The pair interaction of the particles $\{1,2\}$ is tuned to make them have a zero energy resonance and no negative energy bound states. The coupling constant $\lambda >0$ is allowed to take the values for which the particle pairs $\{1,3\}$ and $\{2,3\}$ have no bound states with negative energy. Let $\lambda_{cr}$ denote the critical value of the coupling constant such that $E(\lambda) \to -0$ for $\lambda \to \lambda_{cr}$, where $E(\lambda)$ is the ground state energy of the 3-body system. We prove the theorem, which states that near $\lambda_{cr}$ one has $E(\lambda)
= C (\lambda-\lambda_{cr})[\ln (\lambda-\lambda_{cr})]^{-1}+$h.t., where $C$ is a constant and h.t. stands for “higher terms”. This behavior of the ground state energy is universal (up to the value of the constant $C$), meaning that it is independent of the form of pair interactions.
author:
- 'Dmitry K. Gridnev'
title: 'Universal low-energy behavior in three-body systems'
---
\[1\][\#1]{} \[1\][\#1]{}
Introduction {#sec:1}
============
Universality plays an important role in physics. The interest to it is inspired by the striking similarity in behavior near the critical point among systems that are otherwise quite different in nature. For example, various substances, which exhibit liquid-gas phase transition, near the critical point obey the universal law $\rho_{gas} -\rho_c \to -A(T_c - T)^\beta$. Here $\rho_{gas}, \rho_c$ denote the density of gas and critical density respectively, $T,T_c$ are temperature and critical temperature, $A$ is a constant and $\beta$ is the so-called critical exponent [@hammer]. Amazingly, the value of $\beta \simeq 0.325$ is the same for many substances, which are completely different on the atomic level. Similar law with the same value of the critical exponent holds true for magnetization in ferromagnets as a function of temperature. Another example of universality is found in the ground state energy of the Bose gas as a function of density. In the low density limit it approaches an expression, which depends only on the scattering length but not on the overall form of pair interaction [@yngv].
Small quantum systems also exhibit universal features [@hammer]. One example of universality in the two-particle case concerns the behavior of the energy depending on the coupling constant near the threshold. Suppose that $E(\lambda)$ is the energy of an isolated non-degenerate state of the Hamiltonian $h(\lambda) = T + \lambda
V_{12}$ in 3-dimensional space and $E(\lambda) \to 0$ for $\lambda \to
\lambda_{cr}$. Then universally for $\lambda$ near $\lambda_{cr}$ one has $E(\lambda) =
c(\lambda - \lambda_{cr})^2 +$[h.t.]{} or $E(\lambda) \simeq c(\lambda - \lambda_{cr})+$[h.t.]{} depending on whether zero is an eigenvalue of $h(\lambda_{cr})$ or not, see [@klaus1]. **Universal** in this context means that this behavior up to a constant is true for all short range interactions independently of their form. “h.t.” is the shorthand notation for “higher terms” and $E(\lambda) = f(\lambda) +$[h.t.]{} for $f(\lambda) \to 0$ always implies that $E(\lambda) = f(\lambda) + \hbox{o}\bigl(f(\lambda)\bigr)$.
If two particles are set into an n-dimensional space the scenario depends on the space dimension [@klaus1]: for example, in 2-dimensional flatland the energy of the ground state energy can approach zero exponentially fast $E(\lambda) =\exp (-c(\lambda-\lambda_{cr})^{-1})+$[h.t.]{}, and in 4 dimensions the ground state energy approaches zero very slow, namely, $E(\lambda) = c(\lambda - \lambda_{cr})|\log (\lambda -
\lambda_{cr})|^{-1}+$[h.t.]{}. For a full account of possible scenarios see Table I in [@klaus1]. Another universality associated with 2-body system in 3-dimensional space relates to the wave function near the threshold. If the energy of a non-degenerate bound state near the threshold satisfies $E(\lambda) = c(\lambda - \lambda_{cr})^2+$[h.t.]{} then the wave function of this bound state $\psi(\lambda, x)$ approaches spherically symmetric expression [@schwinger] $$\label{24.1:1}
\left\| \psi(\lambda, x) - |E(\lambda)|^{1/4}\frac{e^{-|E(\lambda)|^{\frac 12}
|x|}}{\sqrt{2\pi}|x|}\right\| \to 0 ,$$ see Eq. (8) in [@3], where we have omitted the phase factor. For well-behaved short-range interactions Eq. (\[24.1:1\]) holds irrespectively of the form of the pair potential.
In 3-particle systems the notorious example of universality is the Efimov effect. Efimov’s striking and counterintuitive prediction [@vefimov] was that just by tuning coupling constants of the short-range interactions in the 3-body system one can bind an infinite number of levels, even though the two-body subsystems bind none. The infinitude of bound states was shown rigorously by Yafaev in [@yafaev]. Basing on the Yafaev’s method Sobolev [@sobolev] has proved that $$\lim_{\epsilon \to 0} |\ln \epsilon|^{-1}N(\epsilon) = \mathfrak{U}_0/2,$$ where $N(\epsilon)$ is the number of bound states with the energy less than $-\epsilon < 0$ and $\mathfrak{U}_0$ is the universal positive constant, which depends only on masses. Let us remark that in physics[@vefimov; @hammer; @helfrich] it is generally conjectured that $$\label{1.14;10}
\lim_{n \to \infty} \frac{|E_n|}{|E_{n+1}|} = e^{2\pi/s_0} ,$$ where $E_n$ is the energy of the $n$-th Efimov level and $s_0 = \pi \mathfrak{U}_0$. In physics this is termed as universal scaling of Efimov levels, see [@hoegkrohn; @lakaev] for the mathematical discussion of (\[1.14;10\]). Efmiov’s prediction was later confirmed experimentally in ultracold gases [@kraemer]. The so-called 4-body universality [@naturephysics] holds only approximately [@myfbs] and the question of finite range corrections is still being debated [@debate].
For further discussion it is useful to introduce the following mathematical notations. For an operator $A$ acting on a Hilbert space $D(A)$, $\sigma(A)$ and $\sigma_{ess} (A)$ denote the domain, the spectrum, and the essential spectrum of $A$ respectively [@reed]. $A > 0$ means that $(f, Af) > 0$ for all $f \in D(A)$, while $A \ngeq 0$ means that there exists $f_0 \in D(A)$ such that $(f_0, Af_0) < 0$. $\mathfrak{B}(\mathcal{H})$ denotes the set of bounded linear operators on the Hilbert space $\mathcal{H}$. For an interval $\Omega \subset \mathbb{R}$ the function $\chi_\Omega : \mathbb{R} \to \mathbb{R}$ is such that $\chi_\Omega (x) = 1$ if $x\in \Omega$ and $\chi_\Omega (x) = 0$ otherwise.
Recently a new type of universality in 3-body systems has been discovered in [@3]. Consider the Hamiltonian of the 3–particle system in $\mathbb{R}^3$ $$\label{hami}
H(\lambda) = H_0 + v_{12} + \lambda (v_{13} + v_{23}) ,$$ where $H_0$ is the kinetic energy operator with the center of mass removed, $\lambda >0$ is the coupling constant and none of the particle pairs has negative energy bound states. All particles are supposed to have a finite mass. The pair–interactions $v_{ik}$ are operators of multiplication by real $V_{ik} (r_i - r_k) \leq 0$ and $r_i \in \mathbb{R}^3$ are particle position vectors. For pair potentials we require like in [@3] that $$\gamma_0 := \max_{i=1,2}\max\left[ \int d^3 r \bigl| V_{i3} (r)\bigr|^2 , \int
d^3 r \bigl| V_{i3} (r)\bigr| (1+|r|)^{2\delta}\right] < \infty , \label{restr}$$ where $0 < \delta < 1/8$ is a fixed constant, and $$\label{restr3}
-b_1 e^{-b_2 |r|} \leq V_{12} (r) \leq 0 ,$$ where $b_{1,2} >0$ are constants.
We shall assume that the interaction between the particles $\{1,2\}$ is tuned to make them have a zero energy resonance and no negative energy bound states [@yafaev2]. This implies that in the absence of particle 3 the particles $\{1,2\}$ are “almost” bound, that is a bound state with negative energy appears if and only if the interaction $v_{12}$ is strengthened by a negligibly small amount. In mathematical terms this can be expressed as follows $$\begin{gathered}
H_0 + v_{12} >0 \label{14.1;2}\\
H_0 + (1+\varepsilon)v_{12} \ngeq 0 \quad \quad \textnormal{for all $\varepsilon >0$.} \label{14.1;3}\end{gathered}$$ Let $\lambda'_{1,2}$ be the values of the coupling constants such that $H_0 + \lambda'_1 v_{13}$ and $H_0 + \lambda'_2 v_{23}$ are at critical coupling in the sense of Def. 1 in [@2]. This means that $H_0 + \lambda'_i v_{i3} >0$ and $H_0 + (1+\varepsilon)\lambda'_i v_{i3} \ngeq 0$ for all $\varepsilon >0$ and $i=1,2$. Let us set $$\tilde \lambda := \min [\lambda'_1 , \lambda'_2]. \label{tildelambda}$$ We shall always assume that the coupling constant in (\[hami\]) satisfies the inequality $\lambda < \tilde \lambda$. In other words, the coupling constant takes the values for which the particle pairs $\{1,3\}$ and $\{2,3\}$ have neither zero energy resonances nor bound states with negative energy. Under these conditions $\sigma_{ess} (H(\lambda)) =[0, \infty)$ and $H(\lambda)$ has a finite number of bound states with negative energy as proved in [@yafaevnew].
Let $\lambda_{cr}$ be the value of the coupling constant such that $H(\lambda_{cr}) \geq 0$ but $H(\lambda_{cr} + \epsilon) \ngeq 0$ for all $\epsilon >0$. In Sec. 6 in [@1] it is proved that $\lambda_{cr} \in (0, \tilde \lambda)$. By the HVZ theorem (see [@reed], Vol. 4 and [@teschl]) for $\lambda \in (\lambda_{cr}, {\tilde \lambda})$ $$H(\lambda) \psi_\lambda = E(\lambda) \psi_\lambda ,$$ where $E(\lambda) <0$ is the ground state energy and $\psi_\lambda \in D(H_0)$. By Theorems 1, 3 in [@1] and Theorem 2 in [@3] $E(\lambda) \nearrow 0$ for $\lambda \searrow \lambda_{cr}$ but zero is not an eigenvalue of $H(\lambda_{cr})$. Moreover, $\psi_\lambda$ for $\lambda \searrow \lambda_{cr}$ totally spreads, that is $$\label{14.1;1}
\lim_{\lambda\to \lambda_{cr}} \int_{|x^2| + |y^2| < R} |\psi_\lambda (x,y)|^2 d^3 x d^3 y= 0 \quad \quad \textnormal{for all $R>0$} .$$ In (\[14.1;1\]) $x,y \in \mathbb{R}^3$ are Jacobi coordinates in the 3-body problem, which are shown in Fig. \[fig:1\] (left). They are defined as $$\begin{gathered}
x =\alpha^{-1} (r_2 - r_1 ) \nonumber, \\
y = \frac{\sqrt{M_{12}}}{\hbar} \left[ r_3 - \frac{m_1}{(m_1 + m_2)} r_1 - \frac{m_2}{(m_1 + m_2)} r_2\right] \nonumber , \end{gathered}$$ where $m_i$ denote particle masses, $\alpha = \hbar (m_1 + m_2)^{\frac 12} (2m_1 m_2)^{-\frac 12}$ and $M_{12} = (m_1 + m_2) m_3 /(m_1 + m_2 + m_3)$.
In [@3] it was proved that $\psi_\lambda$ for $\lambda\to \lambda_{cr}$ approaches in norm a universal expression, namely, $$\label{3bh3}
\psi_\lambda \to \frac{1}{\sqrt 2 \pi^{3/2} |\ln
|E(\lambda)||^{1/2}} \frac{ \bigl\{ |x|\sin(k_n |y|) + |y|\cos(k_n |y|)\bigr\}\exp(-|E(\lambda)|^{1/2}|x|)}{1+|x|^3|y|+|y|^3 |x|} .$$ In the limit the wave function $\psi_\lambda$ describes the state, in which average distances between all three particles go to infinity. (This is partly the reason why the short range details of pair interactions becomes unimportant).
By analogy with the 2-particle case it is natural to assume that $E(\lambda)$ would exhibit universal behavior near $\lambda_{cr}$. In this paper we shall prove Theorem \[th:main\], which states that it is indeed so and universally one has $E(\lambda)
= C (\lambda-\lambda_{cr})[\ln (\lambda-\lambda_{cr})]^{-1} +$[h.t.]{}, where $C >0$. It is important to stress that this result does not follow directly from (\[3bh3\]), if one merely tries to substitute (\[3bh3\]) into the equation $E(\lambda) = \bigl(\psi_\lambda , H(\lambda) \psi_\lambda \bigr)$. This is due to the error terms, which in spite of going to zero in norm may affect the resulting average, for a detailed explanation see Remark \[newremark\] in the next section. The obtained behavior of $E(\lambda)$ remarkably mimics that of the ground state energy of 2 particles in 4-dimensional space. The experimental observation of this type of universality can possibly be obtained in ultracold gas mixtures, see [@3] for discussion. When the pair interaction $v_{12}$ is not tuned there are 3 types of possible asymptotic behavior of $E(\lambda)$, which are listed in Theorem \[th:main1a\].
![Jacobi coordinates in the three-particle problem. $y$ points in the direction from the center of mass of the particles $\{1,2\}$ to particle $3$. $\zeta$ points in the direction from the center of mass of the particles $\{1,3\}$ to particle $2$. The picture shows only directions of the vectors, the scales are set in order to ensure that $H_0 =
-\Delta_{x}-\Delta_{y}$ and $H_0 =
-\Delta_{\eta}-\Delta_{\zeta}$ holds.[]{data-label="fig:1"}](Jacobi.eps){height="0.16\textheight"}
Main Result {#sec:2}
===========
The Hamiltonian $H(\lambda)$ in (\[hami\]) is self–adjoint on $D(H_0) = \mathcal{H}^2 (\mathbb{R}^{6}) \subset L^2 ( \mathbb{R}^{6})$, where $\mathcal{H}^2 (\mathbb{R}^{6})$ denotes the corresponding Sobolev space [@teschl; @liebloss]. The pair interaction between particles $\{1,2\}$ is tuned so that they have a zero energy resonance, that is Eqs. (\[14.1;2\])-(\[14.1;3\]) hold.
Our aim in this paper is to prove
\[th:main\] Suppose that $E(\lambda):= \inf \sigma(H(\lambda))$, then for $\lambda \searrow \lambda_{cr}$ one has $$\label{main}
E (\lambda) = C_0 \frac{(\lambda- \lambda_{cr})}{\ln (\lambda-\lambda_{cr})} + \textnormal{\hbox{o}} \left( \frac{(\lambda- \lambda_{cr})}{\left| \ln (\lambda-\lambda_{cr})\right|}\right) ,$$ where $C_0 > 0$ is a finite constant.
Before we proceed with the proof let us remark that 1) Eq. (\[main\]) is universal, meaning that up to a constant it does not depend on the details of pair interaction; 2) the function $E(\lambda)$ at $\lambda=\lambda_{cr}$ cannot be Taylor expanded in powers of $(\lambda-\lambda_{cr})^\alpha$ for any $\alpha >0$; 3) the three-body ground state energy in the 3-dimensional case has the same behavior near $\lambda_{cr}$ as the 2-body ground state energy in the 4-dimensional case (we do not have an explanation for this finding); 4) the constant $C_0$ depends on pair interactions and can be expressed through zero energy solutions of Birman-Schwinger operators; 5) the method of the proof is different from [@klaus1; @klaus2]: the method in [@klaus1; @klaus2], which uses the low energy expansions of the Birman-Schwinger operator, is not applicable here. The proof below hinges on Theorem \[th:main2\] in Section \[sec:3\], whose technical proof is heavily based on the results and methods in [@3].
For $\lambda \in (\lambda_{cr}, {\tilde \lambda} )$ there exists $\psi_\lambda \in
D(H_0)$, $\|\psi_\lambda\| = 1$ such that $H(\lambda) \psi_\lambda = E
(\lambda)\psi_\lambda$, besides we can assume that $\psi_\lambda >0$ because it is the ground state. $E (\lambda)$ is smooth and monotone increasing on $(\lambda_{cr}, {\tilde \lambda})$. Using perturbation theory [@reed; @kato] we obtain $$\label{24.1:9}
-\frac{dE (\lambda)}{d\lambda} = \||v_{13}|^{1/2} \psi_\lambda\|^2 +
\||v_{23}|^{1/2} \psi_\lambda\|^2 .$$ By Theorem \[th:main2\] for $\lambda$ close enough to $\lambda_{cr}$ there exists a constant $C_0 >0$ such that $$-\frac{C_0+\varepsilon }{\ln (-E (\lambda))} \geq -\frac{dE (\lambda)}{d\lambda} \geq -\frac{C_0-\varepsilon }{\ln (-E (\lambda))}$$ for any given $\varepsilon >0$. The last inequality can be equivalently rewritten as $$\label{24.1:3}
C_0 + \varepsilon \geq \frac{d}{d\lambda} \left( E(\lambda) \ln (-E(\lambda)) - E(\lambda)\right) \geq C_0 - \varepsilon .$$ Integrating (\[24.1:3\]) we obtain $$\label{24.1:4}
(C_0 + \varepsilon) (\lambda-
\lambda_{cr}) \geq E (\lambda) \ln (-E (\lambda)) - E(\lambda) \geq (C_0 - \varepsilon) (\lambda-
\lambda_{cr}) .$$ Let us set $$\label{24.1:6}
E (\lambda) =-f(\lambda)(\lambda - \lambda_{cr}) ,$$ where $f (\lambda) >0$. From (\[24.1:4\]) we get $$\label{24.1:5}
(C_0 + \varepsilon) \geq f (\lambda)[- \ln (-E (\lambda)) +1 ] \geq (C_0 - \varepsilon) .$$ Using that $E(\lambda) \to 0$ from (\[24.1:5\]) we conclude that $\lim_{\lambda \to \lambda_{cr}} f(\lambda ) = 0$. Again substituting (\[24.1:6\]) into (\[24.1:5\]) we obtain $$(C_0 + \varepsilon) \geq -f (\lambda)\ln (\lambda - \lambda_{cr}) - [ f(\lambda) \ln (f(\lambda)) -f(\lambda) ] \geq (C_0 - \varepsilon)$$ The term in square brackets in the last inequality goes to zero for $\lambda \to \lambda_{cr}$. Thus for $\lambda$ close to $\lambda_{cr}$ we have $$\label{24.1:8}
- \frac{C_0 + \varepsilon/2 }{\ln (\lambda - \lambda_{cr})} \geq f (\lambda) \geq - \frac{C_0 - \varepsilon/2 }{\ln (\lambda - \lambda_{cr})} .$$ Now (\[24.1:9\]) follows from (\[24.1:6\]), (\[24.1:8\]) since $\varepsilon >0 $ is arbitrarily small.
\[newremark\] We should explicitly warn against the direct “physicist’s approach”, when one substitutes in (\[24.1:9\]) instead of $\psi_\lambda$ the rhs of (\[3bh3\]). Eq. (\[3bh3\]) defines $\psi_\lambda$ up to error terms, which go to zero in norm when the energy goes to zero. There is otherwise no control of these error terms and one cannot exclude the situation, where, for example, $$\label{yasego}
\psi_\lambda = \Phi_\lambda + \bigl|\ln |E(\lambda)|\bigr|^{-\frac 18} \exp(-|x|^2 -|y|^2)$$ and $\Phi_\lambda$ equals the rhs of (\[3bh3\]). If one substitutes (\[yasego\]) into (\[24.1:9\]) one would find that the resulting $E(\lambda)$ would be very different from the form stated in Theorem \[th:main\].
Theorem \[th:main\] considers the case when the particles $\{1,2\}$ have a zero energy resonance. Now let us consider a more general situation and assume that the 3-particle system is described by the Hamiltonian (\[hami\]), where for simplicity we require that $V_{ik} \leq 0$ are bounded and have a compact support. We still require that $\lambda < {\tilde \lambda}$, [*i.e.*]{} $\lambda$ takes the values for which the subsystems $\{1,3\}$ and $\{2,3\}$ have no bound states with negative energy and no zero energy resonances. However, we do not impose restrictions on the spectrum of the particles $\{1,2\}$, which means that this pair determines the energy of the dissociation threshold $E_{thr}$, that is $$E_{thr} := \inf \sigma_{ess} (H(\lambda)) = \inf \sigma (H_0 + v_{12}) .$$ The critical coupling constant $\lambda_{cr}$ is the value of $\lambda$ for which the 3-body bound state, whose energy lies below $E_{thr}$, is about to be formed. Mathematically speaking $$\lambda_{cr} = \sup\{\lambda | \inf \sigma (H(\lambda)) = E_{thr}\} .$$ By the methods similar to [@oldstuff] one can prove that $\lambda_{cr} < {\tilde \lambda}$.
\[th:main1a\] One can distinguish 3 cases: (A) the pair $\{1,2\}$ has no negative energy bound states and no zero energy resonance; (B) the pair $\{1,2\}$ has no negative energy bound states but has a zero energy resonance; (C) the pair $\{1,2\}$ has at least one bound state with negative energy. Suppose that $E(\lambda):= \inf \sigma(H(\lambda))$, then for $\lambda \searrow \lambda_{cr}$ in each case one has $$\begin{aligned}
&(A) \quad \quad E(\lambda) - E_{thr}=E(\lambda)= - c (\lambda-\lambda_{cr}) + \mbox{\textnormal{h.t.}} \nonumber\\
&(B) \quad \quad E(\lambda) - E_{thr}=E(\lambda)= c (\lambda-\lambda_{cr})[\ln (\lambda-\lambda_{cr})]^{-1}+ \mbox{\textnormal{h.t.}} \nonumber\\
&(C) \quad \quad E(\lambda)-E_{thr} = - c (\lambda-\lambda_{cr})^2 + \mbox{\textnormal{h.t.}} \nonumber\end{aligned}$$ where $c > 0$ is a finite constant.
Case (B) follows from Theorem \[th:main\] and case (C) was proved in Theorem 3.2 in [@klaus2]. In case (A) let $\psi_\lambda$ denote the eigenfunction of $H(\lambda)$, which corresponds to the eigenvalue $E(\lambda)$. As follows from the proof of Theorem 2 in [@1] $\psi_\lambda \to \psi_0$ in norm, where $\psi_0 \in D(H_0)$ is the eigenfunction corresponding to the zero eigenvalue of $H(\lambda_{cr})$. Thus $\||v_{13}|^{1/2}\psi_\lambda\|^2 + \||v_{23}|^{1/2}\psi_\lambda\|^2 \to c$, where $c \in (0,\infty)$ and, hence, $dE/d\lambda \to -c$. The rest of the proof is trivial.
Asymptotics of Potential Energy Terms {#sec:3}
=====================================
Let $e(\mu)$ be the ground state energy of the Hamiltonian $h(\mu) = -\Delta_x + \mu v_{12} $ acting on $L^2 (\mathbb{R}^3)$, where $v_{12}$ is the operator of multiplication by $V_{12} (\alpha x)$ and $\mu \in \mathbb{R}_+$ is the coupling constant. Because the pair of particles $\{1,2\}$ has a zero energy resonance and no negative energy bound states the following expansion [@klaus1] $$e(\mu) = a^{-2} (\mu -1)^2 + \hbox{o} (\mu-1)$$ is true for $\mu \to + 1$. The positive constant $a$ in this expansion can be expressed through the zero energy solution of the Birman-Schwinger operator, namely [@klaus1], $$a = (4\pi )^{-1} \bigl\| |v_{12}|^{\frac 12} \phi_0\bigr\|_1^2 ,$$ where $\phi_0 \in L^2 (\mathbb{R}^3)$ is the unique nonnegative and normalized solution of the equation $$\label{1.16;1}
|v_{12}|^{\frac 12} \bigl[- \Delta_x + 0\bigr]^{-1} |v_{12}|^{\frac 12} \phi_0 = \phi_0,$$ see [@1] for details. As we shall see (Remark \[remark:3\] below), the constant $C_0$ in (\[main\]) can also be expressed through zero energy solutions of certain Birman-Schwinger operators.
The aim of this section is to prove
\[th:main2\] Suppose that the interaction between the particles $\{1,2\}$ is tuned to make them have a zero energy resonance and no negative energy bound states. Let $\psi_\lambda$ for $\lambda \in (\lambda_{cr}, \tilde \lambda)$ be the ground state wave function of $H(\lambda)$ defined in (\[hami\]), which corresponds to the ground state energy $E(\lambda)$. Then there exists $C_0 \in (0, \infty) $ such that $$\label{1.24:152}
\lim_{\lambda \to \lambda_{cr} + 0} \bigl|\ln |E(\lambda)|\bigr| \left\{ \bigl\||v_{13}|^{\frac12}\psi_\lambda \bigr\|^2 + \bigl\||v_{23}|^{\frac12}
\psi_\lambda \bigr\|^2 \right\} = C_0 .$$
(The same constant $C_0$ is used in Theorem \[th:main\]). We shall extensively use the results from [@3], therefore it is convenient to pass to sequences. Let $\lambda_n \in (\lambda_{cr}, {\tilde \lambda})$ be any sequence such that $\lambda_n \to \lambda_{cr}$ and $\psi_n\equiv \psi_{\lambda_n}$. Instead of (\[1.24:152\]) it suffices to prove that $$\label{1.24:15}
\lim_{n \to \infty} |\ln k_n| \left\{ \bigl\||v_{13}|^{\frac12}\psi_n\bigr\|^2 + \bigl\||v_{23}|^{\frac12}
\psi_n\bigr\|^2 \right\} = C_0/2 ,$$ where $k_n := |E(\lambda_n)|^\frac 12$. We need only to prove that the limit on the lhs of (\[1.24:15\]) exists and is positive. Note that all requirements R1-R3 in [@3] are satisfied and the sequence $\psi_n$ totally spreads (see Sec. Theorems 1,3 in [@1] for the proof).
Before we proceed with the proof let us introduce additional notations. Let $\mathcal{F}_{12}$ and $\mathbb{P}_0$ denote the partial Fourier transform and the projection operator, which act on $f(x,y)$ as follows $$\begin{gathered}
\hat f(x,p_y) = \mathcal{F}_{12} f(x,y) = \frac 1{(2 \pi )^{3/2}} \int d^3 y
\; e^{-ip_y \cdot \; y} f(x,y) ,\label{ay5} \\
[\mathbb{P}_0 f ] (x,y) = \phi_0 (x) \int f(x',y) \phi_0 (x') d^3 x' , \end{gathered}$$ and where $\phi_0$ is defined in (\[1.16;1\]). For a shorter notation let us denote $$M_n := \left\{ \bigl\||v_{13}|^{\frac12}\psi_n\bigr\|^2 + \bigl\||v_{23}|^{\frac12}
\psi_n\bigr\|^2 \right\}^{1/2} .$$ Similar to Eqs. (39)-(40) in [@3] we introduce the operator function $$\label{b12}
\tilde B_{12} (k_n) := \mathcal{F}^{-1}_{12} \xi_n (p_y)
\mathcal{F}_{12} ,$$ where $$\label{tail}
\xi_n (p_y) := \left\{ \begin{array}{ll}
|p_y|^{\delta/8} + (k_n)^{\delta/8} & \quad \mathrm{if} \;\; |p_y|
\leq 1 \\
1 + (k_n)^{\delta/8} & \quad \mathrm{if} \;\; |p_y| \geq 1 . \\
\end{array}
\right.
$$
\[lemma:1\] The sequences $\varphi^{(1)}_n := M_n^{-1}|v_{13}|^{\frac 12} \psi_n$ and $\varphi^{(2)}_n := M_n^{-1}|v_{23}|^{\frac 12} \psi_n$, where $\psi_n$ is defined in Theorem \[th:main2\], converge in norm. The sequence $\varphi^{(3)}_n := M_n^{-1}\tilde B_{12} (k_n)|v_{12}|^{\frac 12} \psi_n$ is uniformly norm-bounded.
From the Schrödinger equation for $\psi_n$ it follows that $$\lambda_n^{-1} \left( \begin{array}{c}
\varphi^{(1)}_n\\
\varphi^{(2)}_n
\end{array}\right)
=\mathcal{K} (k_n^2) \left( \begin{array}{c}
\varphi^{(1)}_n\\
\varphi^{(2)}_n
\end{array} \right) ,$$ where $\mathcal{K}(z)$ for $z>0$ is a bounded operator on $L^2 (\mathbb{R}^6) \oplus L^2 (\mathbb{R}^6)$ defined as $$\label{blacblac}
\mathcal{K}(z) := \left( \begin{array}{cc}
|v_{13}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{13}|^{\frac 12} &|v_{13}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{23}|^{\frac 12}\\
|v_{23}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{13}|^{\frac 12} &|v_{23}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{23}|^{\frac 12}
\end{array} \right) .$$ Like in Sec. II in [@4] one proves that $\mathcal{K}(z)$ for $z\to +0$ has a norm limit $\mathcal{K}(0)$, besides $\mathcal{K}(z)$ for $z\geq 0$ is a positivity preserving self-adjoint operator. Let us show that the off-diagonal terms in (\[blacblac\]) are compact operators on $L^2 (\mathbb{R}^6)$. By the resolvent identity $$\begin{gathered}
|v_{13}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{23}|^{\frac 12} = |v_{13}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{23}|^{\frac 12} \nonumber\\
+ \left[|v_{13}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{12}|^{\frac 12} \right] |v_{12}|^{\frac 12} \bigl( H_0 + v_{12} + z\bigr)^{-1} |v_{23}|^{\frac 12} = \left[ |v_{13}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{23}|^{\frac 12} \right] \nonumber\\
+ \left[|v_{13}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{12}|^{\frac 12} \right] \left[|v_{12}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{23}|^{\frac 12} \right] \nonumber\\
- \left[|v_{13}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{12}|^{\frac 12} \right] \left\{|v_{12}|^{\frac 12} \bigl( H_0 + v_{12}+z\bigr)^{-1} |v_{12}|^{\frac 12}\right\}
\left[|v_{12}|^{\frac 12} \bigl( H_0 + z\bigr)^{-1} |v_{23}|^{\frac 12} \right]. \label{gogol}\end{gathered}$$ On the rhs of (\[gogol\]) all operators in square brackets are Hilbert-Schmidt (the result of this sort goes back at least to [@ginibre], see also [@1] for the proof in present notations). Since the operator in curly brackets is bounded for $z>0$ the lhs of (\[gogol\]) is a compact operator. Thus by Weyl’s criterion [@teschl; @reed] $$\sigma_{ess} \bigl( \mathcal{K}(z) \bigr) = \sigma_{ess} \bigl( \mathcal{K}_{11}(z) \bigr) \cup \sigma_{ess} \bigl( \mathcal{K}_{22}(z) \bigr)
\subseteq [0, 1/{\tilde \lambda}] ,$$ where ${\tilde \lambda}$ was defined in (\[tildelambda\]). $(\varphi^{(1)}_n ,
\varphi^{(2)}_n )$ is a normalized eigenvector of $\mathcal{K}(k_n^2)$ corresponding to the eigenvalue $ \lambda_n^{-1}$. Due to the location of the essential spectrum $\|\mathcal{K}(k_n^2)\|$ equals the maximal eigenvalue of $\mathcal{K}(k_n^2)$. Because $\varphi^{(1)}_n , \varphi^{(2)}_n \geq 0$ we conclude due to the positivity preserving property that $\|\mathcal{K}(k_n^2)\| = \lambda_n^{-1}$ (see Theorem XIII.43 in Vol. 4 of [@reed]). Therefore, due to the norm convergence $\lambda_{cr}^{-1} = \|\mathcal{K}(0)\| $ is the maximal eigenvalue of $\mathcal{K}(0) $, which is isolated and non-degenerate. Let $(\varphi^{(1)}_\infty , \varphi^{(2)}_\infty )$ with $\varphi^{(1)}_\infty, \varphi^{(2)}_\infty \geq 0$ be the eigenvector of $\mathcal{K}(0) $, which corresponds to $\lambda_{cr}^{-1}$. Again by the norm convergence of $\mathcal{K}(k_n^2)$ we have that $\varphi^{(1)}_n \to \varphi^{(1)}_\infty$ and $\varphi^{(2)}_n \to \varphi^{(2)}_\infty$ in norm.
To prove that $\sup_n \|\varphi^{(3)}_n\| < \infty$ note that by Eq. (67) in [@3] $$\begin{gathered}
\varphi^{(3)}_n = \lambda_n \Bigl\{1- |v_{12}|^{\frac 12} \bigl(H_0 +
k_n^2\bigr)^{-1}|v_{12}|^{\frac 12} \Bigr\}^{-1}
|v_{12}|^{\frac 12} \tilde B_{12} (k_n) \bigl[H_0 + k_n^2\bigr]^{-1}|v_{13}|^{\frac 12}\varphi^{(1)}_n \nonumber\\
+ \lambda_n \Bigl\{1- |v_{12}|^{\frac 12} \bigl(H_0 +
k_n^2\bigr)^{-1}|v_{12}|^{\frac 12} \Bigr\}^{-1}
|v_{12}|^{\frac 12} \tilde B_{12} (k_n) \bigl[H_0 + k_n^2\bigr]^{-1} |v_{23}|^{\frac 12}\varphi^{(2)}_n . \end{gathered}$$ Without loosing generality it suffices to prove that $$\varphi^{(4)}_n := \Bigl\{1- |v_{12}|^{\frac 12} \bigl(H_0 +
k_n^2\bigr)^{-1}|v_{12}|^{\frac 12} \Bigr\}^{-1}
|v_{12}|^{\frac 12} \tilde B_{12} (k_n) \bigl[H_0 + k_n^2\bigr]^{-1}|v_{13}|^{\frac 12}\varphi^{(1)}_n \label{29.12:1}$$ is uniformly norm-bounded. Denoting $\hat \varphi^{(4)}_n \equiv \mathcal{F}_{12} \varphi^{(4)}_n $ we get $$\hat \varphi^{(4)}_n = \chi_{[0, \rho_0]} \Bigl( \sqrt{p_y^2 +k_n^2}\Bigr) \hat \varphi^{(4)}_n + \chi_{(\rho_0 , \infty)} \Bigl( \sqrt{p_y^2 +k_n^2}\Bigr) \hat \varphi^{(4)}_n =
\chi_{[0, \rho_0]} \Bigl( \sqrt{p_y^2 +k_n^2}\Bigr) \hat \varphi^{(4)}_n + \mathcal{O}(1),$$ where $\mathcal{O}(1)$ denotes the terms, which are uniformly norm-bounded. Here $\rho_0 >0$ is a fixed cutoff parameter. Following Lemma 11 in [@1] (the value of $\rho_0$ is also defined there) we can expand the operator in curly brackets in (\[29.12:1\]). This expansion gives (see Eqs. (73)-(74) in [@3]) $$\begin{gathered}
\chi_{[0, \rho_0]} \Bigl( \sqrt{p_y^2 +k_n^2}\Bigr) \hat \varphi^{(4)}_n =\chi_{[0, \rho_0]} \Bigl( \sqrt{p_y^2 +k_n^2}\Bigr)
a^{-1} \mathbb{P}_0 |v_{12}|^{\frac 12} \nonumber\\
\times \bigl(|p_y|^2 + k_n^2\bigr)^{-\frac 12}\xi_n (p_y) \bigl[-\Delta_x + p_y^2 + k_n^2\bigr]^{-1}\widehat{|v_{13}|^{\frac 12}} \hat \varphi^{(1)}_n + \mathcal{O}(1) , \end{gathered}$$ where $\widehat{|v_{13}|^{\frac 12}} := \mathcal{F}_{12} |v_{13}|^{\frac 12} \mathcal{F}^{-1}_{12}$. Thus $$\label{1.24:12}
\|\varphi^{(4)}_n\| \leq a^{-1} \left\| |v_{12}|^{\frac 12} \chi_{[0, \rho_0]} ( |p_y|) \bigl(|p_y|^2 + k_n^2\bigr)^{-\frac 12} \xi_n (p_y) \bigl[-\Delta_x + p_y^2 + k_n^2\bigr]^{-1}\widehat{|v_{13}|^{\frac 12}}\right\| + \mathcal{O}(1)$$ It remains to prove that the operator norm on the right hand side (rhs) of (\[1.24:12\]) is uniformly bounded. This can be trivially estimated through the Hilbert-Schmidt norm (c. f. proof of Lemma 9 in [@1]) $$\begin{gathered}
\left\| |v_{12}|^{\frac 12} \chi_{[0, \rho_0]} ( |p_y|) \bigl(|p_y|^2 + k_n^2\bigr)^{-\frac 12} \xi_n (|p_y|) \bigl[-\Delta_x + p_y^2 + k_n^2\bigr]^{-1}\widehat{|v_{13}|^{\frac 12}}\right\|^2 \nonumber \\
\leq c \int_{|p_y|\leq \rho_0} d^3 p_y \frac{\xi_n^2 (|p_y|)}{( p_y^2 + k_n^2)^{3/2}} , \label{1.24:14} \end{gathered}$$ where $c>0$ is a constant. The integral on the rhs of (\[1.24:14\]) is clearly convergent and uniformly bounded.
Following [@3] let us denote $$\label{Phi_n2}
\Phi_n^{(i)} := - \lambda_n \sqrt{|v_{12}|} \bigl[H_0 + k_n^2\bigr]^{-1} v_{i3} \psi_n \quad\quad (i=1,2)$$ and $$\label{gn}
g_n (y) = g_n^{(1)} (y) + g_n^{(2)} (y) ,$$ where $$\label{gni}
g_n^{(i)} (y) := \int d^3 x \phi_0 (x) \Phi_n^{(i)} (x, y) .$$ The functions $g(y), g_n^{(i)} (y) \in L^1 (\mathbb{R}^3) \cap L^2 (\mathbb{R}^3) $ coincide with the ones defined in Eqs. (93), (94) in [@3].
In [@3] it is proved that $$\label{1.16;3}
|\hat g_n (0)||\ln k_n|^{1/2} \to \frac{\sqrt 2 a}{R(0)} >0 ,$$ where $\hat g_n (p_y)$ is the Fourier image of $g_n$ defined in (\[gn\]) and $$R(0) = \int \frac{\phi_0 (x')}{|x'|} \bigl|V_{12} (\alpha x')\bigr|^{\frac 12} .$$ (\[1.16;3\]) follows from Eq. (90) in [@3] and the fact that the norm of the function on the rhs of that equation goes to one for $n \to \infty$ (see the text under Eq. (90) in [@3]).
Our aim is to show that the sequence $|\hat g_n (0)| /M_n$ converges and $\lim_{n \to \infty} |\hat g_n (0)| /M_n = C_1 >0$; then (\[1.24:15\]) follows with $C_0 = 4a^2 [C_1 R(0)]^{-2}$. Note that due to the positivity preserving property of $[H_0+k_n^2]^{-1}$ $$\label{1.24:17}
\frac{|\hat g_n (0)| }{M_n} = \frac{\|g^{(1)}_n\|_1 }{M_n} + \frac{\|g^{(2)}_n\|_1 }{M_n} ,$$ where $g^{(1,2)} (y) \in L^1 (\mathbb{R}^3) \cap L^2 (\mathbb{R}^3) $ are defined in Eqs. (93), (94) in [@3]. It suffices to prove the convergence of the first term on the rhs of (\[1.24:17\]) (the convergence of the second term is proved analogously). Similar to [@3] let us introduce Jacobi coordinates $\eta, \zeta$, which are pictured in Fig. \[fig:1\] (right) $$\begin{gathered}
\eta = \alpha'^{-1} (r_3 - r_1) \nonumber\\
\zeta = \left[\frac{2(m_1 +m_3) m_2}{\hbar^2 (m_1 + m_2 + m_3)} \right]^{\frac 12} \left( r_2 - \frac{m_1}{m_1+m_3} r_1 - \frac{m_3}{m_1+m_3} r_3\right) \nonumber , \end{gathered}$$ where $\alpha' = \hbar (m_1 + m_3)^{\frac 12} (2m_1 m_3)^{-\frac 12}$. The coordinates $(\eta , \zeta)$ and $(x,y)$ are connected through the orthogonal linear transformation $$\begin{gathered}
x = m_{x\eta} \eta + m_{x\zeta} \zeta , \nonumber\\
y = m_{y\eta} \eta + m_{y\zeta} \zeta,\nonumber\end{gathered}$$ where $m_{x\eta}, m_{x\zeta} \neq 0, m_{y\eta}, m_{y\zeta}$ are real and can be expressed through mass ratios in the system. For all $R >0$ we have $$\begin{gathered}
M_n^{-1} \left\| g^{(1)}_n\right\| _1 = \left\| \chi_{[0, R] }(|\eta|)\phi_0 \mathfrak{X}_n \varphi_n^{(1)}\right\| _1 \nonumber\\
+
M_n^{-1}\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} \left[H_0 + k_n^2 \right]^{-1} |v_{13}| \psi_n \right\| _1 , \label{1.24:21}\end{gathered}$$ where we have defined $$\mathfrak{X}_n := |v_{12}|^{\frac 12} \left[H_0 + k_n^2 \right]^{-1} |v_{13}|^{\frac 12} .$$ The operators $\mathfrak{X}_n : L^2 (\mathbb{R}^6) \to L^2 (\mathbb{R}^6) $ are norm-bounded and have a norm limit for $n \to \infty$, which we denote as $\mathfrak{X}_0$ (for the proof see f. e. Lemma 7 in [@1]). Thus by Lemma \[lemma:1\] $\mathfrak{X}_n \varphi_n^{(1)} \to \mathfrak{X}_0 \varphi_\infty^{(1)}$ in $L^2$ sense. Then $\chi_{[0, R] }(|\eta|)\phi_0 \mathfrak{X}_n \varphi_n^{(1)}$ converges to $\chi_{[0, R] }(|\eta|)\phi_0 \mathfrak{X}_0 \varphi_\infty^{(1)}$ in $L^1$ sense because by the Cauchy-Schwarz inequality $$\left\|\chi_{[0, R] }(|\eta|)\phi_0 \bigl( \mathfrak{X}_n \varphi_n^{(1)} - \mathfrak{X}_0 \varphi_\infty^{(1)}\bigr)\right\|_1 \leq \|\chi_{[0, R] }(|\eta|)\phi_0 \|
\left\|\mathfrak{X}_n \varphi_n^{(1)} - \mathfrak{X}_0 \varphi_\infty^{(1)}\right\| ,$$ where $\|\chi_{[0, R] }(|\eta|)\phi_0 \| $ is finite. Hence, the first term on the rhs of (\[1.24:21\]) converges for all $R >0$. Now the convergence of the sequence on the left hand side (lhs) of (\[1.24:21\]) follows from Lemmas \[lemma:2\], \[lemma:3\]. We have proved that the sequence on the lhs of (\[1.24:17\]) converges to $C_1 \in [0, \infty)$. It remains to show that $C_1 \neq 0$. This follows from the fact that $\varphi_\infty^{(1)}, \varphi_\infty^{(2)} \geq 0 $ and besides $\|\varphi_\infty^{(1)}\|^2 + \|\varphi_\infty^{(2)}\|^2 =1$, so at least one of the terms on the rhs of (\[1.24:17\]) converges to a positive value.
\[remark:3\] The proof of Theorem \[th:main2\] allows to express the constant $C_0$ in Theorem \[th:main\] in terms of zero energy solutions of Birman-Schwinger operators, namely, $C_0 = 4a^2 [R(0)C_1]^{-2}$, where $$C_1 = \bigl\|\phi_0 \mathfrak{X}_0 \varphi_\infty^{(1)}\bigr\|_1 + \bigl\|\phi_0 \mathfrak{Y}_0 \varphi_\infty^{(2)}\bigr\|_1$$ and by definition $\mathfrak{Y}_0$ is the operator norm limit $$\mathfrak{Y}_0 =\lim_{z \to +0} |v_{12}|^{\frac 12} \left[H_0 + z \right]^{-1} |v_{23}|^{\frac 12} .$$
\[lemma:2\] For $R \to \infty$ $$\sup_{n} M_n^{-1}\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} \left[H_0 + k_n^2 \right]^{-1} |v_{13}| \psi_n \right\| _1 \to 0 .$$
The proof is partly based on that of Lemma 4 in [@3], and we need to introduce additional notations from [@3]. $\mathcal{F}_{13}$ denotes the partial Fourier transform, which acts on $f(\eta,
\zeta) $ as $$\label{xw20}
\mathcal{F}_{13} f := \hat f(\eta, p_\zeta) = \frac 1{(2 \pi )^{3/2}} \int d^3
\zeta \; e^{-ip_\zeta \cdot \; \zeta} f(\eta, \zeta) .$$ Let us introduce the operator function $$\label{b13}
\tilde B_{13} (k_n) := \mathcal{F}^{-1}_{13} \tilde t_n (p_\zeta)
\mathcal{F}_{13} ,$$ where $$\label{tail2}
\tilde t_n (p_\zeta) = \left\{ \begin{array}{ll}
|p_\zeta|^{1-\delta} + (k_n)^{1-\delta} & \quad \mathrm{if} \;\; |p_\zeta|
\leq 1 \\
1 + (k_n)^{1-\delta} & \quad \mathrm{if} \;\; |p_\zeta| \geq 1 . \\
\end{array}
\right.
$$ Following Eq. (30) in [@3] we define $$F_n := \lambda_n \bigl[H_0 + k_n^2\bigr]^{-1} v_{13} \psi_n .$$ Then using Eq. (98) in [@3] we obtain the inequality $$\begin{gathered}
M_n^{-1}\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} \left[H_0 + k_n^2 \right]^{-1} |v_{13}| \psi_n \right\| _1 =
M_n^{-1} \lambda_n^{-1}\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} F_n \right\| _1 \nonumber\\
\leq M_n^{-1} \lambda_n^{-1} \sum_{i=1}^3\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} \tilde F_n^{(i)} \right\| _1 , \label{1.20;23}\end{gathered}$$ where $$\tilde F^{(i)}_n := \bigl[ H_0 + k_n^2\bigr]^{-1} |v_{13}|^{1/2}
\tilde B_{13}(k_n) \Psi^{(i)}_n \label{xw17}$$ and $$\begin{gathered}
\Psi^{(1)}_n := |v_{13}|^{1/2} \tilde B_{13}^{-1} (k_n) \bigl[ H_0 +
k_n^2\bigr]^{-1} |v_{12}| |\psi_n| , \label{psin1}\\
\Psi^{(2)}_n := \lambda_n |v_{13}|^{1/2} \tilde B_{13}^{-1} (k_n) \bigl[ H_0 +
k_n^2\bigr]^{-1} |v_{23}| |\psi_n| \label{psin2},\\
\Psi^{(3)}_n := \lambda_n |v_{13}|^{1/2} \bigl[ H_0 + k_n^2\bigr]^{-1} |v_{13}|^{1/2}
\Bigl\{1 - \lambda_n |v_{13}|^{1/2} \bigl[H_0
+ k_n^2\bigr]^{-1} |v_{13}|^{1/2} \Bigr\}^{-1} \nonumber \\\times\tilde B_{13}^{-1} (k_n) |v_{13}|^{1/2}
\bigl[ H_0 + k_n^2\bigr]^{-1} \bigl(|v_{12}| + \lambda_n
|v_{23}|\bigr)|\psi_n| . \label{psin3}\end{gathered}$$ Eqs. (\[xw17\]) and (\[psin1\])-((\[psin3\])) come from Eqs. (43), (44), (48) and (49) in [@3], where one has to take into account that all interaction potentials are nonpositive. By Eq. (103) in [@3] $$\label{xw23}
\bigl| \tilde F^{(i)}_n (\eta , \zeta) \bigr| \leq \frac 1{2^{7/2} \pi^{5/2}} \int d^3
\eta' \int d^3 p_\zeta \; \bigl| V_{13}(\alpha' \eta')\bigr|^{1/2}
\frac{e^{-\sqrt{p_\zeta^2 + k_n^2}|\eta-\eta'|}}{|\eta-\eta'|} \tilde t_n
(p_\zeta) \bigl| \hat \Psi^{(i)}_n (\eta', p_\zeta) \bigr| ,$$ where $\hat \Psi^{(i)}_n = \mathcal{F}_{13} \Psi^{(i)}_n $. Using the exponential falloff of $V_{12} (x)$ from (\[1.16;1\]) one can easily derive the inequality $$\label{20.1;15}
\phi_0 (x) \bigl|V_{12} (\alpha x) \bigr| \leq \tilde b_1 e^{-\tilde b_2 |x|},$$ where $\tilde b_{1,2}$ are constants. After substituting (\[20.1;15\]) and (\[xw23\]) into (\[1.20;23\]), interchanging the order of integration and applying the Cauchy-Shwarz inequality we get $$\begin{gathered}
M_n^{-1}\left\| \chi_{(R, \infty) }(|\eta|) \phi_0 |v_{12}|^{\frac 12} \left[H_0 + k_n^2 \right]^{-1} |v_{13}| \psi_n \right\| _1 \nonumber\\
\leq
\frac {\tilde b_1 \lambda_n^{-1}}{2^{7/2} \pi^{5/2}}
\sum_{i=1}^3 M_n^{-1} \bigl\| \Psi^{(i)}_n \bigr\|
\left\{ \int d^3 \eta' \int d^3 p_\zeta \; \bigl| V_{13}(\alpha' \eta')\bigr| \,
{\tilde t}^{\; 2}_n (p_\zeta) \, {\tilde J}^2 (\eta', p_\zeta) \right\}^{1/2} , \end{gathered}$$ where $${\tilde J} (\eta', p_\zeta):= \int_{|\eta| > R} d^3 \eta \int d^3 \zeta \;
\frac{e^{-\sqrt{p_\zeta^2 + k_n^2}|\eta-\eta'|}}{|\eta-\eta'|} e^{-\tilde b_2 |m_{x\eta} \eta +
m_{x\zeta} \zeta|} .$$ By Lemma \[lemma:4\] we only need to prove that $ \sup_{n} I_n \to 0 $ for $R \to \infty$, where we defined $$I_n := \int d^3 \eta' \int d^3 p_\zeta \; \bigl| V_{13}(\alpha' \eta')\bigr| \,
{\tilde t}^{\; 2}_n (p_\zeta) \, {\tilde J}^2 (\eta', p_\zeta) .$$ Let us split the last integral as $I_n = I_n^{(1)} + I_n^{(2)}$, where $$\begin{gathered}
I_n^{(1)} := \int_{|\eta'| \leq R/2} d^3 \eta' \int d^3 p_\zeta \; \bigl| V_{13}(\alpha' \eta')\bigr| \,
{\tilde t}^{\; 2}_n (p_\zeta) \, {\tilde J}^2 (\eta', p_\zeta) , \label{1.24:31}\\
I_n^{(2)} := \int_{|\eta'|>R/2} d^3 \eta' \int d^3 p_\zeta \; \bigl| V_{13}(\alpha' \eta')\bigr| \,
{\tilde t}^{\; 2}_n (p_\zeta) \, {\tilde J}^2 (\eta', p_\zeta) . \label{1.24:32}\end{gathered}$$ Clearly, we can write $$\label{1.24:37}
{\tilde J} (\eta', p_\zeta) \leq c_1 \int_{|\eta| >R} d^3 \eta \frac{e^{-\sqrt{p_\zeta^2 + k_n^2}|\eta-\eta'|}}{|\eta-\eta'|} ,$$ where $c_1 >0$ is a constant. For $|\eta'| \leq R/2$ and $|\eta| > R/2$ we have $|\eta - \eta'| \geq |\eta| - |\eta'| > |\eta|/2$, which gives the estimate $${\tilde J} (\eta', p_\zeta) \leq c_2 \frac{e^{-\sqrt{p_\zeta^2 + k_n^2}R/2}}{\sqrt{p_\zeta^2 + k_n^2}}
\left( R + \frac2{\sqrt{p_\zeta^2 + k_n^2}} \right) ,$$ where $c_2 >0$ is a constant. Substituting this estimate into (\[1.24:31\]) and using that $V_{13} (x) \in L^1 (\mathbb{R}^3)$ we obtain $$I_n^{(1)} \leq c_3 \int d^3 p_\zeta \, {\tilde t}^{\; 2}_n (p_\zeta) \frac{e^{-\sqrt{p_\zeta^2 + k_n^2}R}}{p_\zeta^2 + k_n^2} \left[R^2 + \frac4{p_\zeta^2 + k_n^2}\right],$$ where $c_3 >0$ is a constant. Substituting Eq. (40) from [@3] we get $$I_n^{(1)} \leq c_4 \int_0^1 s^{2-2\delta} e^{-sR} \left[R^2 + \frac4{s^2} \right] ds
+ c_4 \int_1^\infty e^{-sR} \left[R^2 + \frac4{s^2} \right]ds$$ where $c_4 >0$ is another constant. Let us set $A_\beta := \sup_{x\geq 0} x^{\beta} e^{-x}$, where $A_\beta$ is finite and depends only on $\beta$. Then $$\label{1.24:34}
I_n^{(1)} \leq c_4 R^{-\delta} (A_{2+\delta} + 4A_\delta) \int_0^1 s^{-3\delta} ds + c_4 R^2 e^{-R/2} \int_1^\infty e^{-sR/2}\bigl[1+ 4s^{-2}R^{-2}\bigr] ds = \hbox{o}(R)$$
It is easy to see that the terms on the rhs of (\[1.24:34\]) vanish for $R \to \infty$. Let us estimate $I_n^{(2)} $. From (\[1.24:37\]) we get $${\tilde J} (\eta', p_\zeta) \leq c_1 \int d^3 \eta \frac{e^{-\sqrt{p_\zeta^2 + k_n^2}|\eta-\eta'|}}{|\eta-\eta'|} \leq \frac{c'_1}{p_\zeta^2} ,$$ where $c'_1 >0$ is a constant. Substituting this into (\[1.24:32\]) we obtain $$I_n^{(2)} \leq c'_1 \left\{\int_{|\eta'|>R/2} d^3 \eta' \bigl| V_{13}(\alpha' \eta')\bigr| \right\} \int d^3 p_\zeta \; {\tilde t}^{\; 2}_n (p_\zeta) |p_\zeta|^{-4} = \hbox{o}(R),$$ where the integral in curly brackets goes to zero because $V_{13} (x) \in L^1 (\mathbb{R}^3)$ and the second integral is uniformly bounded for all $n$.
\[lemma:3\] Suppose that the sequence $a_n \in \mathbb{C}$ is such that $a_n = b_n (R) + c_n (R)$, where $ b_n (R), c_n (R) \in \mathbb{C}$ depend on a parameter $R >0$. Additionally assume that $ b_n (R)$ is convergent for all $R >0$ and $\limsup_{n \to \infty} |c_n ( R)| \to 0$ for $R \to \infty$. Then $a_n$ converges.
The proof is a trivial application of the Cauchy convergence criterion. For any $\varepsilon >0$ fix $N_1, R >0$ so that $|c_n ( R)| < \varepsilon/3$ for all $n > N_1$. Choose $N_2$ so that $|b_n (R) - b_m (R)| < \varepsilon/3$ for all $n, m > N_2$. Then $$|a_n - a_m| \leq |b_n (R) - b_m (R)| + |c_n (R)| + |c_m (R)| < \varepsilon$$ for all $n, m > \max (N_1, N_2)$.
\[lemma:4\] $\sup_n M_n^{-1} \|\Psi^{(i)}_n\| < \infty $, where $\Psi^{(i)}_n$ for $i=1,2,3$ are defined in Eqs.(\[psin1\])-(\[psin3\]).
We have $M_n^{-1} \Psi^{(2)}_n = \lambda_n \mathcal{T}^{(2)}_n \varphi_n^{(1)}$, where $$\mathcal{T}^{(2)}_n := \tilde B_{13}^{-1} (k_n) |v_{13}|^{1/2}_- \bigl[ H_0 + k_n^2\bigr]^{-1} |v_{12}|^{1/2}$$ are uniformly norm-bounded operators (they are defined in the same way as in Eq. (63) in [@3]). Thus by Lemma \[lemma:1\] $\sup_n M_n^{-1} \|\Psi^{(2)}_n \| < \infty$. From definition of $\Psi^{(1)}_n$ we have $$\label{1.24:41}
M_n^{-1} \Psi^{(1)}_n = \mathfrak{T}_n \mathfrak{D}^*_n \varphi_n^{(3)} ,$$ where $$\begin{gathered}
\mathfrak{T}_n := |v_{13}|^\frac12 \tilde B_{12}^{-1} \bigl[ H_0 + k_n^2\bigr]^{\frac{-3+\delta}4}\\
\mathfrak{D}_n:= |v_{12}|^\frac12 \tilde B_{13}^{-1} (k_n) \bigl[ H_0 + k_n^2\bigr]^{\frac{-1-\delta}4}. \end{gathered}$$ $\mathfrak{T}_n , \mathfrak{D}_n $ are bounded operators on $L^2 (\mathbb{R}^6)$. From (\[1.24:41\]) $\sup_n M_n^{-1} \| \Psi^{(1)}_n\| < \infty $ follows from $\mathfrak{T}_n , \mathfrak{D}_n $ being uniformly norm-bounded. Let us start with estimating the norm of $\mathfrak{T}_n$. Let us define operator functions $\mathfrak{T}^{(1)}_n, \mathfrak{T}^{(2)}_n : \mathbb{R}^3 \to \mathfrak{B}\bigl(L^2 (\mathbb{R}^3)\bigr)$, which act on $h(\eta) \in L^2 (\mathbb{R}^3)$ as follows $$\begin{gathered}
\mathfrak{T}^{(1)}_n (p_\zeta) h = |V_{13}(\alpha'\eta)|^{\frac 12}\left[\xi_n \bigl(m_{y\eta}(-i\nabla_\eta) + m_{y\zeta}p_\zeta\bigr) \right]^{-1}
\bigl[-\Delta_\eta +p_\zeta^2 + k_n^2\bigr]^{\frac{-3+\delta}4} \nonumber \\
\times \chi_{[0,1]}\bigl(|-i\nabla_\eta|\bigr) h , \\
\mathfrak{T}^{(2)}_n (p_\zeta) h = |V_{13}(\alpha'\eta)|^{\frac 12}\left[\xi_n \bigl(m_{y\eta}(-i\nabla_\eta) + m_{y\zeta}p_\zeta\bigr) \right]^{-1}
\bigl[-\Delta_\eta +p_\zeta^2 + k_n^2\bigr]^{\frac{-1-\delta}4} \nonumber \\
\times\chi_{(1,\infty)}\bigl(|-i\nabla_\eta|\bigr) h . \end{gathered}$$ The operators like $ \chi_{[0,1]}\bigl(|-i\nabla_\eta|\bigr)$ act in the sense described in Chapter 4 in [@traceideals]. It is easy to see that $$[\mathcal{F}_{13} \mathfrak{T}_n f] (\eta, p_\zeta) = \mathfrak{T}^{(1)}_n (p_\zeta) \hat f (\eta, p_\zeta) + \mathfrak{T}^{(2)}_n (p_\zeta) \hat f (\eta, p_\zeta) ,$$ where $\hat f (\eta, p_\zeta) \equiv \mathcal{F}_{13} f$. Thus $$\label{1.24:44}
\|\mathfrak{T}_n\| \leq \sup_{p_\zeta}\|\mathfrak{T}^{(1)}_n (p_\zeta)\| + \sup_{p_\zeta}\|\mathfrak{T}^{(2)}_n (p_\zeta)\| .$$ The operator norms on the rhs of (\[1.24:44\]) can be bounded by the trace ideals norms, which in turn can be bounded by Theorem 4.1 in [@traceideals]. $$\|\mathfrak{T}^{(1)}_n (p_\zeta)\| \leq \|\mathfrak{T}^{(1)}_n (p_\zeta)\|_2 \leq (2\pi\alpha')^{-\frac 32} \|V_{13}\|_1 [J_n^{(1)}(p_\zeta)]^{\frac 12},$$ where $$\begin{gathered}
J_n^{(1)}(p_\zeta) :=
\int_{|s|\leq 1 } \frac{d^3 s}{\xi_n^2 (m_{y\eta} s + m_{y\zeta}p_\zeta) \bigl[s^2 + p_\zeta^2 + k_n^2\bigr]^{\frac{3-\delta}2}}\leq
\int_{|s|\leq 1 } \frac{d^3 s}{\bigl| m_{y\eta} s + m_{y\zeta}p_\zeta\bigr|^{\frac \delta4} |s|^{3-\delta}} \nonumber\\
+ \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta}} \leq \left[\int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta/2}}\right]^{\frac{3-\delta}{3-\delta/2}}
\left[\int_{|s|\leq 1 } \frac{d^3 s}{\bigl| m_{y\eta} s + m_{y\zeta}p_\zeta\bigr|^{\frac 32 - \frac \delta4} }\right]^{\frac{\delta}{6-\delta}} + \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta}} \nonumber\\
\leq \left[\int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta/2}}\right]^{\frac{3-\delta}{3-\delta/2}}
\left[\int_{|s|\leq 1 } \frac{d^3 s}{m_{y\eta}| s|^{\frac 32 - \frac \delta4} }\right]^{\frac{\delta}{6-\delta}} + \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta}} . \label{1.24:46}\end{gathered}$$ In (\[1.24:46\]) we have used Hölders inequality. The integrals on the rhs of (\[1.24:46\]) are convergent and independent of $p_\zeta$ and $n$, hence, $\sup_{p_\zeta}\|\mathfrak{T}^{(1)}_n (p_\zeta)\| < \infty$. Similarly $$\|\mathfrak{T}^{(2)}_n (p_\zeta)\| \leq \|\mathfrak{T}^{(2)}_n (p_\zeta)\|_3 \leq (2\pi\alpha')^{-1} \|V_{13}\|_{3/2} [J_n^{(2)}(p_\zeta)]^{\frac 13},$$ where $$\begin{gathered}
J_n^{(2)}(p_\zeta) :=
\int_{|s|> 1 } \frac{d^3 s}{\xi_n^3 (m_{y\eta} s + m_{y\zeta}p_\zeta) |s|^{\frac{9-3\delta}{2}}}\leq
\int_{|s|> 1 } \frac{d^3 s}{\bigl| m_{y\eta} s + m_{y\zeta}p_\zeta\bigr|^{\frac{3\delta}{8}} |s|^{\frac{9-3\delta}{2}}} \nonumber\\
+ \int_{|s|> 1 } \frac{d^3 s}{|s|^{\frac{9-3\delta}{2}}} \leq \frac 1{m_{y\eta}^3}\int_{|s'|\leq 1} \frac{d^3 s'}{|s'|^{\frac{3\delta}{8}}}
+ 2 \int_{|s|> 1 } \frac{d^3 s}{|s|^{\frac{9-3\delta}{2}}} . \label{1.24:48}\end{gathered}$$ The integrals on the rhs of (\[1.24:48\]) obviously converge. Thus we find that $\sup_{p_\zeta}\|\mathfrak{T}^{(2)}_n (p_\zeta)\| < \infty$ and $\|\mathfrak{T}_n\|$ is uniformly bounded (note that $\|V_{13}\|_{3/2}$ in (\[1.24:48\]) is bounded because $V_{13} \in L^1 (\mathbb{R}^3) \cap L^2 (\mathbb{R}^3)$).
Now we pass to estimating $\|\mathfrak{D}_n\|$ and use the same method. Calculations similar to above ones give $$\begin{gathered}
\|\mathfrak{D}_n\| \leq (2\pi\alpha)^{-\frac 32} \|V_{12}\|_1 \left[\sup_{p_y} J_n^{(3)}(p_y)\right]^{\frac 12} \nonumber\\
+
\sup_{p_y} \left\| \bigl|V_{12}(\alpha x)\bigr|^{\frac 12} \left[\tilde t_n \bigl( m_{x\zeta} (-i\nabla_x) + m_{y\zeta} p_y \bigr)\right]^{-1}\right\| , \label{1.24:53}\end{gathered}$$ where $$\begin{gathered}
J_n^{(3)}(p_y) := \int_{|s|\leq 1 } \frac{d^3 s}{{\tilde t_n}^2 (m_{x\zeta} s + m_{y\zeta}p_y) |s|^{1+\delta}} \leq
\int_{|s|\leq 1 } \frac{d^3 s}{\bigl| m_{x\zeta} s + m_{y\zeta}p_y \bigr|^{2-2\delta} |s|^{1+\delta}} \nonumber\\
+ \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{1+\delta}} \leq \left[\int_{|s|\leq 1 } \frac{d^3 s}{\bigl| m_{x\zeta} s + m_{y\zeta}p_y \bigr|^{3-\delta}}\right]^{\frac{2-2\delta}{3-\delta}} \left[\int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta}}\right]^{\frac{1+\delta}{3-\delta}} + \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{1+\delta}} \nonumber\\
\leq \left[\int_{|s|\leq 1 } \frac{d^3 s}{m_{x\zeta} | s |^{3-\delta}}\right]^{\frac{2-2\delta}{3-\delta}} \left[\int_{|s|\leq 1 } \frac{d^3 s}{|s|^{3-\delta}}\right]^{\frac{1+\delta}{3-\delta}} + \int_{|s|\leq 1 } \frac{d^3 s}{|s|^{1+\delta}} . \label{1.24:51}\end{gathered}$$ The integrals on the rhs of (\[1.24:51\]) converge and it remains to estimate the operator norm in (\[1.24:53\]). $$\begin{gathered}
\left\| \bigl|V_{12}(\alpha x)\bigr|^{\frac 12} \left[\tilde t_n \bigl( m_{x\zeta} (-i\nabla_x) + m_{y\zeta} p_y \bigr)\right]^{-1}\right\| \nonumber\\
\leq
\left\| \bigl|V_{12}(\alpha x)\bigr|^{\frac 12} \left[\tilde t_n \bigl( m_{x\zeta} (-i\nabla_x) + m_{y\zeta} p_y \bigr)\right]^{-1} \chi_{[0,1]} \left( \bigl| m_{x\zeta} (-i\nabla_x) + m_{y\zeta} p_y \bigr| \right)\right\|_2 + \bigl\| V_{12}\bigr\|_\infty \nonumber\\
\leq \bigl\| V_{12}\bigr\|_\infty + (2\pi\alpha)^{-\frac 32} \|V_{12}\|_1 \left[\sup_{p_y} \int_{|m_{x\zeta}s + m_{y\zeta}p_y| \leq 1 } \frac{d^3 s}{\bigl|m_{x\zeta}s + m_{y\zeta}p_y\bigr|^{2-2\delta}}\right]^{\frac 12} \nonumber\\
\leq \bigl\| V_{12}\bigr\|_\infty + (2\pi\alpha m_{x\zeta})^{-\frac 32} \|V_{12}\|_1 \left[\int_{|s| \leq 1 } \frac{d^3 s}{|s|^{2-2\delta}} \right]^{\frac 12} , \end{gathered}$$ where we have again used Theorem 4.1 in [@traceideals]. Thus we find that $\sup_n \|\mathfrak{D}_n\| < \infty$ and, hence, $\sup_n M_n^{-1} \|\Psi^{(1)}_n \| < \infty$. Note that because all potentials are nonpositive we have $$M_n^{-1} \Psi^{(3)}_n = \lambda_n \mathcal{T}_n^{(1)} Q_n \left[M_n^{-1}\Psi^{(1)}_n + M_n^{-1}\Psi^{(2)}_n\right] ,$$ where $$\begin{gathered}
\mathcal{T}^{(1)}_n := |v_{13}|^{1/2} \bigl[ H_0 + k_n^2\bigr]^{-1} |v_{13}|^{1/2} , \\
Q_n := \Bigl\{1 - \lambda_n |v_{13}|^{1/2} \bigl[H_0
+ k_n^2\bigr]^{-1} |v_{13}|^{1/2} \Bigr\}^{-1} . \end{gathered}$$ Operators $ \mathcal{T}^{(1)}_n , Q_n$ are defined in the same way in Eqs. (62) and (46) in [@3]. In [@3] it was proved that $\sup_n \|\mathcal{T}_n^{(1)}\| < \infty$ and $\sup_n \|Q_n\| < \infty$. Thus $\sup_n M_n^{-1} \| \Psi^{(3)}_n\| < \infty$ as claimed.
[99]{}
E. Braaten, H.-W. Hammer, Phys. Rep. **428** 259 (2006)
E. H. Lieb, J. Yngvason, Phys. Rev. Lett. **80**, 2504 (1998)
M. Klaus and B. Simon, Ann. Phys. (N.Y.) [**130**]{}, 251 (1980)
J. Schwinger, Phys. Rev. **72**, 742 (1947)
D. K. Gridnev, J. Phys. A: Math. Theor. **46**, 115204 (2013)
V. Efimov, Phys. Lett. **B 33**, 563 (1970); Sov. J. Nucl. Phys. **12**, 589 (1971)
K. Helfrich and H.-W. Hammer, J. Phys. B: At. Mol. Opt. Phys. **44**, 215301 (2011)
D. R. Yafaev, Math. USSR-Sb. **23**, 535 (1974)
A. V. Sobolev, Commun. Math. Phys. **156**, 101 (1993)
S. Albeverio, R. H[ø]{}egh-Krohn and T. T. Wu, Phys. Lett. A [**83**]{}, pp. 105–109 (1981)
S. Albeverio, S. Lakaev and K. A. Makarov, Lett. Math. Phys. [**43**]{}, pp. 73–85 (1998)
Kraemer T *et al* 2006 *Nature* **440** 315
J. von Stecher, J. P. D‘Incao and C. H. Greene, Nature Physics **5**, 417 (2009)
D. K. Gridnev, [*The few-body universality is not exact for more than three particles*]{}, arXiv:1211.0433 (2012).
T. Frederico, A. Delfino, M. R. Hadizadeh, Lauro Tomio, M. T. Yamashita, Few-Body Syst. **54**, 559 (2013)
D. K. Gridnev, J. Phys. A: Math. Theor. **45** 395302 (2012); arXiv:1112.0112v2
D. K. Gridnev, J. Phys. A: Math. Theor. **45** 175203 (2012); arXiv:1111.6788v2
M. Klaus and B. Simon, Comm. Math. Phys. [**78**]{}, 153 (1980)
M. Reed and B. Simon, [*Methods of Modern Mathematical Physics*]{}, vol. 1 Academic Press/New York (1980), vol. 2 Academic Press/New York (1975) and vol. 4, Academic Press/New York (1978).
D. R. Yafaev, Notes of LOMI Seminars **51**, 203 (1975) (Russian)
D. R. Yafaev, Theor. Math. Phys., [**25**]{}, 1065-1072 (1975)
T. Kato, [*Perturbation Theory for Linear Operators*]{}, Springer–Verlag/Berlin Heidelberg (1995)
G. Teschl, [*Mathematical Methods in Quantum Mechanics; With Applications to Schrödinger Operators*]{}, Lecture Notes (2005), http://www.mat.univie.ac.at/ gerald/ftp/book-schroe/index.html; Graduate Studies in Mathematics, Vol. [**99**]{}, Amer. Math. Soc., Providence, 2009.
E. H. Lieb and M. Loss, [*Analysis*]{}, AMS (1997)
D. K. Gridnev, J. Phys. A: Math. Gen. **36**, 6725 (2003)
D. K. Gridnev, J. Math. Phys., **54**, 042105 (2013)
J. Ginibre and M. Moulin, Ann. Inst. Henri Poincar[é]{}, vol. XXI, No. 2, pp. 97-145 (1974)
B. Simon, *Trace Ideals and Their Applications*, Cambridge University Press, 1979
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the sine-square deformation (SSD) of free fermions in one-dimensional continuous space. On the basis of supersymmetric quantum mechanics, we prove the correspondence between the many-body ground state of the system with SSD and that of the uniform system with periodic boundary conditions. We also discuss the connection between the SSD in the continuous space and its lattice version, where the geometric correction due to the real-space deformation plays an important role in relating the eigenstates of the lattice SSD with those of the continuous SSD.'
address:
- 'Department of Physics, Faculty of Science, Niigata University, Japan'
- 'Department of Physics, Graduate School of Science, University of Tokyo, Japan '
author:
- Kouichi Okunishi
- Hosho Katsura
title: ' Sine-square deformation and supersymmetric quantum mechanics '
---
[*Keywords*]{}: sine-square deformation, SUSY quantum mechanics
Introduction
============
Recently, spatial deformations of interaction couplings in low-dimensional quantum many-body systems have attracted much attention [@GendiarKN2009; @HikiharaN2011; @GendiarDLN2011; @Katsura2011; @Katsura2011B; @ShibataH2011; @Hotta2; @Hotta; @Hikihara_Suzuki; @Tada:2014kza; @Ishibashi; @MaruyamaKH2011; @exponential; @rainbow; @cosh]. A smooth cutoff of the coupling encoded in the real space is generally expected to suppress the scattering at the boundaries [@VekicW1993]. Among the various schemes, the sine-square deformation (SSD) in one-dimensional (1D) quantum critical systems is of particular interest, because the many-body ground state of a system with SSD almost coincides with that of the uniform system with periodic boundary conditions (PBC). This is due to the nearly exact cancellation of the boundary scattering in the system with SSD [@GendiarKN2009; @HikiharaN2011]. In a class of 1D critical systems which are reducible to free fermions, moreover, the ground-state correspondence can be made exact. In fact, it was proved that the ground state of the spin-$1/2$ XY chain with SSD is identical to that of the uniform chain with PBC [@Katsura2011]. The same holds true for the tight-binding chains [@MaruyamaKH2011] and transverse field Ising chains at criticality [@Katsura2011B]. More generally, the ground-state correspondence between the SSD and uniform systems can be explained in terms of conformal field theory (CFT), where the Virasoro generator $L_{\pm 1} $ identified with the chiral part of the Hamiltonian of the SSD system [@chiral] annihilates the CFT vacuum [@Katsura2011B]. Another interesting aspect of the SSD is that it has a variety of applications; For instance, the nature of the quasi-localized edge excitations originating from the SSD is essential in a ‘grand-canonical’ approach for finite-size systems, where smooth magnetization curves of quantum spin systems can be obtained with very high accuracy [@Hotta2; @Hotta]. Besides, very recently, a further relation between the SSD and CFT, and its relevance to string theory is also pointed out in Refs. [@Tada:2014kza; @Ishibashi]
A key to understanding the ground-state correspondence between the SSD and uniform systems is that each chiral part of the SSD Hamiltonian annihilates the many-body ground state of the uniform system with PBC [@MaruyamaKH2011]. Here, we remark that the plane-wave representation of the single-particle basis, which is a suitable description for the uniform system with PBC, plays a very essential role in verifying the above statement. However, as shown numerically in Ref. [@ShibataH2011; @Hotta2; @Hotta], low-lying single-particle states in the SSD system seem to be localized around the center or edges of the chain, which would be rather puzzling. A natural question is: “What is the unified view of the physics behind the plane-wave description and the single-particle eigenstates of the SSD?" In this paper, we explore this question by examining the SSD of the free-fermion system defined in 1D continuous space. To this end, we formulate the SSD problem as the inverse problem of the quantum mechanical problem of a particle in the inverse sine-square ($1/\sin^2$) potential. An elegant way to solve the direct problem is to use supersymmetric (SUSY) quantum mechanics with the shape invariance [@Witten; @SUSYQM]. With this, we show that a set of the single-particle eigenstates of the SSD problem can be obtained in closed form for specific values of the parameter $\mu$, which will be introduced in the next section. Then we show that the Slater determinant obtained by filling these single-particle states up to the zero energy is identical to the Slater determinant corresponding to the ground state of the uniform system with PBC. Moreover, we discuss the relation of the SSD in a lattice with that in the continuous space, and clarify the role of the geometric correction involved in the single-particle wavefunctions of the SSD. Then, we demonstrate that a number of properties of the lattice SSD found in the previous study [@Hotta] can be well explained on the basis of the SSD wavefunctions obtained with the use of SUSY quantum mechanics.
The rest of the paper is organized as follows. In section 2, we define the SSD problem of free fermions in 1D continuous space. In section 3, we formulate the single-particle problem with SSD as the inverse problem of the $1/\sin^2$ potential problem. Then we obtain the single-particle eigenstates for specific values of the chemical potentials. In section 4, we prove the ground-state correspondence between the uniform and SSD systems by showing that the single-particle eigenstates for the SSD can be expressed as linear combinations of plane waves. In section 5, we briefly comment on the single-particle states with positive energy in the SSD system. In section 6, we compare the SSD in the continuous space with that in a lattice, and discuss the role of the geometric correction. In the last section, we summarize the results obtained and discuss the prospects for future researches.
Definition of the SSD problem in a one-dimensional ring {#sec:problem_setup}
=======================================================
We start with the definition of the SSD problem in 1D continuous space. Consider $N$ fermions on a ring of length $L$. We denote by ${\hat c}^\dagger (x)$ and ${\hat c}(x)$ the fermionic field operators at position $x$ ($0 \le x \le L$). They satisfy the canonical anticommutation relations: $\{ {\hat c}(x), {\hat c}^\dagger (y) \} = \delta (x-y)$, $\{ {\hat c}(x), {\hat c} (y) \} =\{ {\hat c}^\dagger (x), {\hat c}^\dagger (y) \}=0$. In analogy with lattice systems, the Hamiltonian for the system with a generic deformation is defined as $${\hat {\cal H}}_{\rm cont} = \int^L_0 \, dx\, {\hat c}^\dagger (x)\,
f(x) \left( -\frac{d^2}{dx^2} - \mu \right) {\hat c}(x)
\label{Hamiltonian}$$ where $\mu$ is the “chemical potential" introduced for later convenience, and $f(x)$ encodes the spatially varying couplings. Here we set $\hbar=2m=1$. If $f(x)=1$, the systems is nothing but the uniform ring with non-relativistic free fermions. For the case of SSD, $f(x)$ is defined as $$f(x) := \sin^2 \left(\frac{\pi }{ L} x\right).$$ Thus, the energy scale of the system smoothly approaches zero as $x \to 0$ and $x \to L$, implying that the ring is effectively disconnected at $x=0$ (mod $L$). In this sense, the boundary conditions of the SSD problem can be interpreted as open ones.
A general $N$-particle eigenstate of ${\hat {\cal H}}_{\rm cont}$ is of the form: $$|\Phi \rangle = \int^L_0 dx_1 \cdots \int^L_0 dx_N\,
{\hat c}^\dagger (x_1) \cdots {\hat c}^\dagger (x_N) |0\rangle
\Phi (x_1, ..., x_N)\, ,$$ where $|0\rangle$ is the state with no fermions. Since the Hamiltonian is quadratic in ${\hat c}$ and ${\hat c}^\dagger$, the $N$-body wavefunction in the coordinate representation $\Phi (x_1, ..., x_N)$ can be constructed as a determinant of single-particle wavefunctions $\varphi (x)$ which are determined by the Schrödinger-like equation: $$-f(x) \left( \frac{d^2}{dx^2} +\mu \right)\varphi(x) = \xi \varphi(x),
\label{ssdh}$$ where $\xi$ denotes the eigenvalue. Note that the Hamiltonian here is not self-adjoint in the standard sense and the appropriate inner product for the SSD problem will be introduced in section 5. For later convenience, we assume either periodic or antiperiodic boundary condition for $\varphi (x)$, depending on the total number of fermions: $\varphi (x+L) = (-1)^N \varphi (x)$. We denote by $\{\puni_n(x)\}$ and $\{\pssd_n(x)\}$ the sets of single-particle eigenfunctions of the uniform ($f(x)=1$) and SSD ($f(x)=\sin^2 \frac{\pi}{L}x$) systems, respectively. For the uniform system, $\{\puni_n(x)\}$ is of course described by the plane waves and is easy to obtain. For periodic boundary conditions, one finds that the linear space of $\{ \puni_n (x) \}$ is spanned by $$\exp \left( \pm i \frac{2 n \pi}{L}x \right),
\quad (n=0,1,2, ...),$$ and for antiperiodic boundary conditions, $$\exp \left( \pm i \frac{(2 n+1) \pi}{L}x \right),
\quad (n=0,1,2, ...).$$ In contrast, to obtain $\{ \pssd_n (x) \}$ for the SSD problem is an intriguing problem, because the translational symmetry is explicitly broken by the position-dependent factor $f(x)$. However, for specific values of the parameter $\mu$, one can obtain them in closed form by exploiting the hidden SUSY in the problem.
Single-particle eigenstates of the SSD problem {#sec:susy_solution}
==============================================
In this section, we present a systematic procedure to construct single-particle eigenstates of the SSD problem. To this end, we introduce the following parametrization of $\xi$: $$\xi = - \left(\frac{\pi}{L}\right)^2\beta (\beta-1),
\label{SSDspectrum}$$ and rewrite Eq. (\[ssdh\]) with $f(x)=\sin^2 \frac{\pi}{L}x$ as $$H_\beta\, \varphi (x) = \mu \varphi (x),
\label{SSDpotential}$$ where the “Hamiltonian" $H_\beta$ is defined as $$H_\beta = - \frac{d^2}{dx^2} + \left(\frac{\pi}{L}\right)^2
\frac{\beta (\beta-1)}{\sin^2 \frac{\pi}{L} x }.$$ In the standard setup, “energy eigenvalue" $\mu$ is determined for a given parameter $\xi$ (or, equivalently, $\beta$), where the corresponding eigenfunction has a wavepacket-like shape. On the other hand, however, what we need here is to determine the energy eigenvalue $\xi$ for a given chemical potential $\mu$. In this sense, the SSD problem can be formulated as the inverse problem of the $1/\sin^2$ potential problem.
It is known that the $1/\sin^2$ potential problem can be solved algebraically since the system has SUSY and shape invariance [@SUSYQM]. This property is also deeply related to the integrable many-body system with an inverse sine-square interaction solved by Sutherland [@Sutherland; @Kuramoto]. Here, we briefly summarize the SUSY quantum mechanics for the $1/\sin^2 $ potential problem. The eigenvalue problem of Eq. (\[SSDpotential\]) can be solved algebraically for non-negative integers $\beta=0,1,2,3 \cdots$. Let us start with defining intertwiners $$A_\beta = \frac{d}{dx} + W'_\beta(x),
\quad A^\dagger_\beta = -\frac{d}{dx} +W'_\beta(x)$$ with the superpotential $W_\beta(x) := - \beta \ln \left( \sin \frac{\pi}{L}x \right)$. Using these operators, the Hamiltonian $H_\beta$ in Eq. (\[SSDpotential\]) can be written as $$\begin{aligned}
H_\beta = A_\beta^\dagger A_\beta + \left(\frac{\pi}{L}\right)^2\beta^2.
\label{eq:quad}\end{aligned}$$ According to SUSY quantum mechanics, it is well-known that $A^{}_\beta A^\dagger_\beta$ and $A^\dagger_\beta A^{}_\beta$ form the SUSY-partner Hamiltonians, and the shape invariance ensures the recursive relation $$A^{}_\beta A_\beta^\dagger = A^\dagger_{\beta+1} A_{\beta+1}^{} + \left(\frac{\pi}{L}\right)^2 [(\beta+1)^2- \beta^2]
= H_{\beta+1} - \left(\frac{\pi}{L}\right)^2\beta^2.
\label{eq:recursion}$$ In addition, it is easily confirmed that $A_\beta \psi^0_\beta(x)=0$ with $\psi^0_\beta(x) := (\sin \frac{\pi}{L} x )^\beta$, which corresponds to the eigenfunction of $H_\beta$ with the lowest eigenvalue. Note that Eq. (\[eq:quad\]) implies that the eigenvalue of $H_\beta$ is bounded from below by $(\frac{\pi}{L})^2 \beta^2$. It follows from Eq. (\[eq:recursion\]) that, for a given non-negative integer $\beta$, the $\ell$th eigenstate of $H_\beta$ is constructed as $$\psi_\beta^\ell(x) = A^\dagger_\beta \cdots A^\dagger_{\beta+\ell-2}A^\dagger_{\beta+\ell-1}\psi^0_{\beta+\ell}(x).
\label{A_state}$$ The explicit form of the eigenfunction can be written as $$\psi_\beta^\ell(x) ={\cal N}^\ell_\beta\, C_\ell^\beta\left(\cos \frac{\pi}{L}x \right)\psi_\beta^0(x)
\label{gegenbauer}$$ with the coefficient $${\cal N}^\ell_\beta = \left( -\frac{\pi}{L} \right)^\ell
\frac{\ell!\, (2\beta+2\ell-1)!\, (\beta-1)! }
{2^\ell\, (2\beta+\ell-1)!\, (\beta+\ell-1)!},$$ where $C_\ell^\beta (z)$ denotes the Gegenbauer polynomial of the $\ell$th order [@tableofintegral; @Kuramoto]. The coefficient ${\cal N}^\ell_\beta$ can be obtained by demanding that $A_\beta \psi^\ell_\beta (x) = \left( \frac{\pi}{L} \right)^2 \ell (\ell+2\beta) \psi^{\ell-1}_{\beta+1} (x)$ and the following recursion relation for the Gegenbauer polynomials are consistent: $$\frac{d}{dz} C^\beta_\ell (z) = 2\beta\, C^{\beta+1}_{\ell-1} (z).$$ Then, the corresponding eigenvalue of $H_\beta$ is obtained as $$\mu= \left( \frac{\pi}{L} \right)^2(\beta+\ell)^2,
\label{chem_p}$$ for $\ell=0,1,2, \cdots$. Thus, if the sum of $\beta$ and $\ell$ is constant, say $\beta+\ell=q$, then $\mu$ takes the same value among $\psi^0_q (x), \psi^1_{q-1}(x), \cdots, \psi^{q-\beta}_{\beta}(x), \cdots, \psi^q_0 (x)$.
Here, we list in Table I the first few eigenfunctions of $H_\beta$ with $\beta=0, 1, ..., 5$, from which the relation between the $1/\sin^2$-potential problem and the SSD problem can be seen clearly. On the one hand, one finds the eigenfunctions of $H_\beta$ in the column labeled by $\beta$. On the other hand, the functions in the row labeled by $q$ give the solutions of the inverse problem (up to unimportant overall factors). For a particular chemical potential parameterized by $\mu= (\frac{\pi}{L})^2 q^2$, we obtain them as $\psi^{q-\beta}_{\beta} (x)$ ($\beta = 0, 1, \dots, q$) with the corresponding energy eigenvalues $\xi=-\left(\frac{\pi}{L}\right)^2\beta(\beta-1)$. Note that $\psi^0_q (x)$ can be thought of as the single-particle ground state of the SSD problem, while $\psi^{q}_{0}(x)$ and $\psi^{q-1}_{1} (x)$ are the plane-wave solutions with zero energy.
B.C. $q\backslash\beta$ 0 1 2 3 4 5
------ -------------------- ------------------------- ------------------------- ----------------------------------------------------------------------- ---------------------------------------------------- --------------------------------------------- ------------------------- --
AP 5 $\cos\frac{5 \pi x}{L}$ $\sin\frac{5 \pi x}{L}$ $\sin^2\frac{\pi x}{L}\cos\frac{\pi x}{L}(12\cos^2\frac{\pi x}{L}-5)$ $\sin^3\frac{\pi x}{L}(8\cos^2\frac{\pi x}{L}-1) $ $\sin^4\frac{\pi x}{L}\cos\frac{\pi x}{L} $ $\sin^5\frac{\pi x}{L}$
P 4 $\cos\frac{4 \pi x}{L}$ $\sin\frac{4 \pi x}{L}$ $\sin^2\frac{\pi x}{L}(6\cos^2\frac{\pi x}{L}-1) $ $\sin^3\frac{\pi x}{L}\cos\frac{\pi x}{L} $ $\sin^4\frac{\pi x}{L} $
AP 3 $\cos\frac{3 \pi x}{L}$ $\sin\frac{3 \pi x}{L}$ $\sin^2\frac{\pi x}{L}\cos\frac{\pi x}{L} $ $\sin^3\frac{\pi x}{L} $
P 2 $\cos\frac{2 \pi x}{L}$ $\sin\frac{2 \pi x}{L}$ $\sin^2\frac{\pi x}{L} $
AP 1 $\cos\frac{\pi x}{L}$ $\sin\frac{\pi x}{L}$
P 0 const.
In order to precisely see the correspondence between the solutions of the $1/\sin^2$-potential and the SSD problems, we further comment on the boundary conditions. In the literature, it is often assumed that $\beta \ge 1$ in Eq. (\[SSDpotential\]). Here, we have added the SUSY partners for $\beta=0$ and $1$. For both $\beta=0$ and 1, there is no potential term in $H_\beta$ and thus the system is uniform in $0 \le x \le L$. However, an important difference between them is that the boundary conditions for $\beta=0$ are Neumann, i.e., $\varphi'(0)=\varphi'(L)=0$, in contrast to the Dirichret ones for $\beta=1$. Thus, the eigenfunctions for $\beta=0$ and $1$ are described by cosine and sine functions, respectively. In terms of SUSY, the lowest eigenstate of $H_0$ is a singlet with zero energy, because it is annihilated by both $A_0$ and $A^\dagger_0$. Any other eigenstate of $H_0$ has the corresponding eigenstate of $H_1$, which give rise to a doublet of eigenstates with the same eigenvalue. Thus, we can express the plane-wave solutions for a ring of length $L$ by superposing the eigenfunctions of $H_0$ and $H_1$. We recall here that we have imposed periodic (antiperiodic) boundary conditions on $\varphi (x)$ depending on the total number of fermions $N$. Therefore, in Table I we should take the eigenfunctions in the row labeled by even (odd) $q$ when $N$ is even (odd). We also note that the boundary conditions for the eigenfunctions are indicated by P (periodic) and AP (antiperiodic) in Table I. In contrast to the cases of $\beta=0$ and $1$, the eigenfunctions (\[gegenbauer\]) with $\beta \ge 2$ enjoy $\varphi(0)=\varphi(L)=\varphi'(0)=\varphi'(L)=0$, which reflects the fact that they have wavepacket-like shapes and are disconnected at $x=0$ (mod $L$).
Equivalence of two Slater determinants {#sec: equivalence}
======================================
In this section, we give a proof that the Slater determinant of the plane waves, corresponding to the ground state of the uniform system with PBC, is identical to the Slater determinant of $\psi^{q-\beta}_\beta (x)$, which are the eigenfunctions of the SSD problem. This can be thought of as a continuous version of the ground-state correspondence proved in a class of 1D lattice models reducible to free fermions [@Katsura2011; @Katsura2011B; @MaruyamaKH2011]. In the following, we focus on the case where the total fermion number $N$ is even. It is, however, very straightforward to extend the proof to the case of odd $N$.
Let us first consider the many-body ground state of $N$ (even) fermions on the uniform periodic ring with $\mu = (\frac{\pi}{L})^2 N^2$, which is the Slater determinant obtained by filling the plane-wave states up to $\xi=0$. It is easy to see from Table I that the plane-wave states with even $q$ can be obtained as linear combinations of eigenfunctions in the columns labeled by $\beta=0$ and $1$, which satisfy the PBC. Since $\mu = (\frac{\pi}{L})^2 N^2$, one can easily confirm that the states in the $q=N$ row with $\beta=0$ and $1$ are just located at $\xi=0$. Thus, the many-body ground state is two-fold degenerate, because either one of the two $\xi=0$ states is occupied by a fermion. In the following, we assume that $\beta=1$ state is occupied. Then, the single-particle states involved in the Slater determinant is written as $$\begin{aligned}
\puni_n (x) = \left\{
\begin{array}{cc}
\cos \frac{2n \pi}{L}x \quad & (n=0, 1, ..., \frac{N}{2}-1), \\
& \\
\sin \frac{(2n-N+2)\pi}{L}x \quad & (n=\frac{N}{2}, \frac{N}{2}+1, ..., N-1),
\end{array}
\right.\end{aligned}$$ where $\cos \frac{N\pi}{L}x$ does not appear because we have assumed that $\beta=0$ state is empty. Note that, if we consider the case of odd $N$, the boundary conditions are antiperiodic, where we should take the wavefunctions only from $q=$odd sectors. In terms of $\puni_n (x)$, the $N$-fermion ground-state wavefunction is written as $$\Puni (x_1, ... , x_N) = \det_{1 \le i, j \le N} \, \left[ \puni_{i-1} (x_j) \right].
\label{SD1}$$
We next consider a many-body ground state of the SSD problem with the same setup, i.e., $N$ is even and $\mu= (\frac{\pi}{L})^2N^2$. As shown in the previous section, the single-particle eigenfunctions of the SSD problem is given by $$\pssd_{\beta} (x) = \psi^{N-\beta}_{\beta} (x) \quad (\beta=0,1,2, ..., N),$$ where $\psi^{N-n}_{n} (x)$ is defined in Eq. (\[A\_state\]). The many-body ground state is then described by the Slater determinant of $\psi_\beta^{N-\beta} (x)$ ($1\le \beta \le N$), which span the subspace of $\{ \pssd_n (x) \}^N_{n=0}$. Here, we again assume that not $\beta=0$, but $\beta=1$ state with $\xi=0$ is occupied by a fermion. Then, the $N$-fermion ground-state wavefunction is given by $$\Pssd (x_1, ..., x_N) = \det_{1 \le i,j \le N}\, \left[
\pssd_{i} (x_j) \right].
\label{SD2}$$
At first sight, one might think that the Slater determinants Eqs. (\[SD1\]) and (\[SD2\]) are not related to each other because $\{ \puni_n (x) \}$ and $\{ \pssd_n (x) \}$ look completely different. Nevertheless, we will show that the following relation $$\Pssd (x_1, ..., x_N) = C_N\, \Puni (x_1, ..., x_N)
\label{eq:theorem}$$ holds for any $N$ (even), where $C_N$ is a constant independent of $x_1, ..., x_N$. This means that the many-body ground states of the uniform and SSD systems are identical to each other. The above relation can be proved by noting that the two Slater determinants become equivalent if there exists a linear transformation between $\{ \puni_n (x) \}^{N-1}_{n=0}$ and $\{ \pssd_n (x) \}^{N}_{n=1}$ that is non-singular, i.e., the determinant of its transformation matrix is nonzero. For this purpose, however, the eigenfunctions in terms of the Gegenbauer polynomials (\[gegenbauer\]) are not so useful. In the following, we will consider a more direct treatment of $\psi_\beta^{N-\beta}(x)$. Let us rewrite the intertwiner $A^\dagger_\beta$ as $$A^\dagger_\beta = -\sin^{-\beta} (\tilde{x}) \frac{d}{dx} \sin^\beta (\tilde{x}),$$ where we have introduced the variable $\tilde{x} := \frac{\pi}{L} x $ for simplicity. This leads us to the Rodrigues formula for the eigenfunctions Eq. (\[A\_state\]), $$\psi_\beta^{N-\beta}(x)= \sin^{-\beta +1 } \tilde{x} \left( \frac{-1}{\sin \tilde{x}} \frac{d}{dx} \right)^{N-\beta} \sin^{2N-1} \tilde{x}.
\label{eq:Rod}$$ This implies that the single-particle states can be described in terms of only trigonometric functions of ${\tilde x}$. Note that the explicit form depends on whether $N$ and $\beta$ are even or odd. Here, recall that $N$ should be even for the PBC. We can thus write $N= 2\theta $ with a positive integer $\theta$. For $\beta=2\nu$ ($\nu=0, 1, ..., \theta$), we get from Eq. (\[eq:Rod\]) the following expansion: $$u_\beta(x) :=(-1)^{\theta-\nu}\left(\frac{\pi}{L}\right)^{-2(\theta-\nu)} \frac{(2\theta+2\nu-1)!}{(4\theta-1)!}\psi^{2(\theta-\nu)}_{2 \nu}(x)
= \sum_{n=\nu}^\theta a_{2\nu,2n} \sin^{2n} (\tilde{x}),
\label{eq:expansion1}$$ where $$a_{2\nu,2n} = \prod_{k=n}^{\theta-1}\frac{(2k+1)(k+1)-\nu(2\nu-1)}{2(k - \theta)(k+\theta)}$$ with $a_{2\nu,2\theta} := 1$. For $\beta=2\nu-1$ ($\nu=1,2, ..., \theta$), we also obtain another linearly-independent solution of odd parity as $$\begin{aligned}
v_\beta(x) :=&& (-1)^{\theta-\nu+1}\left(\frac{\pi}{L}\right)^{-2(\theta-\nu)-1} \frac{(2\theta+2\nu-2)!}{(4\theta-1)!}\psi^{2(\theta-\nu)+1}_{2\nu-1}(x) \nonumber \\
&&= \cos (\tilde{x}) \sum_{n=\nu}^{\theta} a_{2\nu-1,2n-1} \sin^{2n-1} (\tilde{x}),
\label{eq:expansion2}\end{aligned}$$ where $$a_{2\nu-1,2n-1} = \prod_{k=n}^{\theta-1}\frac{k(2k+1)-(2\nu-1)(\nu-1)}{2(k - \theta)(k+\theta)},$$ with $a_{2\nu-1,2\theta-1}:= 1$.
Since $\psi^{N-2\nu}_{2\nu}(x)$ ($\psi^{N-2\nu+1}_{2\nu-1}(x)$) and $u_{2\nu}(x)$ ($v_{2\nu-1}(x)$) are the same up to an overall constant, the only thing remaining to prove is to show that there is a linear transformation between the spaces of $\{ u_\beta(x), v_\beta(x)\}$ and $\{ \puni_n (x) \}^{N}_{n=0}$ such that it is invertible. From Eqs. (\[eq:expansion1\]) and (\[eq:expansion2\]), one finds the following relations: $$\begin{aligned}
\begin{pmatrix}
u_0(x) \\
u_2(x) \\
\vdots \\
u_N(x)
\end{pmatrix}
&=&
\begin{pmatrix}
a_{0, 0} & a_{0,2}& \cdots & a_{0,N} \\
& a_{2,2}& \cdots & a_{2,N} \\
& & \ddots& \vdots \\
& & & a_{N,N}
\end{pmatrix}
\begin{pmatrix}
1 \\
\sin^2(\tilde{x}) \\
\vdots \\
\sin^N(\tilde{x})
\end{pmatrix},
\label{vectorspaceu}\\
\begin{pmatrix}
v_1(x) \\
v_3(x) \\
\vdots \\
v_{N-1}(x)
\end{pmatrix}
&=&
\begin{pmatrix}
a_{1, 1} & a_{1,3}& \cdots & a_{1,N-1} \\
& a_{3,3}& \cdots & a_{3,N-1} \\
& & \ddots& \vdots \\
& & & a_{N-1,N-1}
\end{pmatrix}
\begin{pmatrix}
\cos(\tilde{x}) \sin(\tilde{x}) \\
\cos(\tilde{x}) \sin^3(\tilde{x}) \\
\vdots \\
\cos(\tilde{x}) \sin^{N-1}(\tilde{x})
\end{pmatrix},\end{aligned}$$ where matrix elements which are zero are left empty. These upper triangular matrices are non-singular and invertible since the diagonal element $a_{\beta, \beta}$ is nonzero for all $\beta$. Thus, the linear space of the SSD eigenfunctions is equivalent to that of $\{\sin^{2n}(\tilde{x}) \}$ and $\{\cos(\tilde{x})\sin^{2n-1}(\tilde{x}) \}$. According to the standard formula of the trigonometric functions [@sinexpansion], moreover, the linear spaces spanned by $\{\sin^{2n}(\tilde{x}) \}$ and $\{\cos(\tilde{x})\sin^{2n-1}(\tilde{x}) \}$ are respectively equivalent to those by $\{\cos(2n\tilde{x})\}$ and $\{\sin(2n\tilde{x})\}$. Hence, the linear space of $\{ \pssd_n (x) \}^{N}_{n=0}$ can be spanned by the basis elements of $\{\cos(2n\tilde{x})\}^{N/2}_{n=0}$ and those of $\{\sin(2n\tilde{x})\}^{N/2}_{n=1}$, indicating that the single-particle basis of the SSD problem is identical to that of the plane waves.
Here, we should recall that the $\xi=0$ state in the $\beta=0$ sector is assumed to be empty of fermion, which implies that $u_0(x)(\propto \psi_0^{N}(x))$ is not included in $\{ \pssd_n (x) \}^{N}_{n=1}$. However, we can easily check $u_0(x)=\sum_{n=0}^{N/2} a_{0,2n} \sin^n(\tilde{x})\propto \cos(N\tilde{x})$ [@redundant]. Thus, the removal of $u_0(x)$ from the SSD basis results in the linear space $\{ \puni_n (x) \}^{N-1}_{n=0}$, which is identical to that of the plane waves below the Fermi level in the uniform system. Thus, we have obtained the desired relation Eq. (\[eq:theorem\])
If the total fermion number $N$ is odd, we should take eigenfunctions in the row labeled by $q=N$, which are antiperiodic in $x$. The proof proceeds along the same lines as that of the $N=$even case, and will not be given here.
Single-particle states with positive energy
===========================================
The equivalence between the uniform ring and the SSD problem is established only for the ground states of many fermions. Here, we comment on single-particle excited states above the Fermi level, i.e., the positive energy ($\xi >0$) solutions of the SSD problem. From the correspondence between the SSD and inverse sine-square problems, one can see that they translate into the eigenfunctions of the attractive $1/\sin^2$ potential problem. As noted in Appendix A, the solution of the $1/\sin^2$ potential problem with the attractive coupling is classified into two cases: (i) $0<\xi <\frac{1}{4} (\frac{\pi}{L})^2 $ and (ii) $\frac{1}{4} (\frac{\pi}{L})^2 <\xi$. The case (i) corresponds to Eq. (\[SSDpotential\]) with $0<\beta<1$. For a non-integer value of $\beta$, however, the solution of Eq. (\[SSDpotential\]) is given by an analytic continuation of the Gegenbauer polynomial, which exhibits a singular behavior at the boundaries. Thus, the physical solution is permitted only in the limit of $\beta\to 0$ or $1$, which is reduced to the plane-wave solution of $\beta=0$ or 1. Recalling that the boundary condition for $\beta=0$ is Neumann and that for $\beta=1$ is the Dirichlet, one may think of the solution for $\beta=0$ as the zero-energy limit of a scattering state, while that for $\beta=1$ as the zero-energy limit of a bound state. For the case (ii), the attractive potential problem is ill-defined, where the energy of the single-particle ground state (in the sense of the potential problem) collapses to $-\infty$ [@Landau; @Sutherland]. Thus, the SSD problem does not have a proper single-particle excited state with positive energy, unlike the SSD for the lattice systems. We conclude that the SSD problem in continuous space is well defined only below the Fermi level.
Geometric factor and lattice SSD
================================
In this section, we discuss the relation between the above continuous-space SSD and its lattice version, in the former of which a geometric factor associated with the inner product of the SSD eigenfunctions is of particular importance. So far, we have not defined the inner product of the SSD eigenfunctions. This is because we should pay a particular attention to the geometric factor in the continuous SSD, unlike the SSD for lattice systems. For instance, the standard definition of the inner product for the plane waves fails to satisfy the orthogonality relation among $\psi^{N-\beta}_\beta(x)$ ($\beta=1,2, ..., N$), though the linear space of them is identical to that of the plane waves.
For the continuous SSD, the self-adjoint form of the differential equation tells us that an appropriate inner product for $\psi^{N-\beta}_\beta(x)$ is $$\langle\langle \psi^{N-\beta}_\beta,\, \psi^{N-\beta'}_{\beta'} \rangle\rangle := \frac{1}{L} \int_0^L \frac{dx}{\sin^2\frac{\pi}{L}x}\,
\psi^{N-\beta}_\beta(x)\, \psi^{N-\beta'}_{\beta'}(x) = I_{N,\beta}\, \delta_{\beta,\beta'},
\label{innerprod}$$ where the coefficient $I_{N,\beta}$ is given by $$I_{N,\beta} = 2^{2-2\beta}\, ({\cal N}^{N-\beta}_\beta)^2
\frac{(N+\beta-1)!}{(2\beta-1)\, (N-\beta)!\, [(\beta-1)!]^2},$$ the derivation of which is given in \[app:coeffI\]. The factor $\sin^2\frac{\pi}{L}x$ in the integrand in Eq. (\[innerprod\]) can be regarded as a geometric factor attributed to the spatial deformation. With this definition, the orthogonality of the single-particle wavefunctions for $\beta\ge 1$ is correctly recovered. Here, we should mention that $\psi_0^N(x) \propto \cos(\frac{N\pi}{L}x )$ for $\beta=0$ is not normalizable with Eq. (\[innerprod\]). We will discuss this problem later.
In order to see the role of the geometric factor, it is useful to consider the lattice version of the SSD problem, the Hamiltonian of which is defined as $$\begin{aligned}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
{\hat {\cal H}}_{\rm lat}= -\sum_{n=1}^{L} \sin^2 \left(\frac{\pi}{L}n \right) ({\hat c}_n^\dagger {\hat c}_{n+1} + {\hat c}^\dagger_{n+1} {\hat c}_n )
-\mu\sum_{n=1}^{L} \sin^2 \left[ \frac{\pi}{L} \left( n-\frac{1}{2} \right) \right]{\hat c}^\dagger_n {\hat c}_n ,
\label{latticeSSD}\end{aligned}$$ where ${\hat c}^\dagger_n$ and ${\hat c}_n$ are the creation and the annihilation operators, respectively, of a fermion at site $n$. The lattice spacing is set to unity so that the length of the chain $L$ is the same as the total number of sites. We numerically solve the single-particle problem of ${\hat {\cal H}}_{\rm lat}$ and discuss its relation to the continuous SSD. Let $\chi_m (n)$ be the $m$th single-particle eigenfunction of ${\hat {\cal H}}_{\rm lat}$. The level index $m$ runs from $0$ to $L-1$. The inner product between $\chi_m (n)$ and $\chi_{m'}(n)$ is well-defined and the following orthogonality holds: $\langle \chi_m,\, \chi_{m'} \rangle := \sum^{L}_{n=1} \chi_{m}(n) \chi_{m'}(n) =\delta_{m,m'}$. This suggests that, unlike Eq. (\[innerprod\]), the lattice SSD eigenfunctions do not involve any geometric factor. On the other hand, it is also expected that a naive continuum limit of the lattice SSD eigenfunction $\chi_m (n)$ may approach the eigenfunction of the continuous SSD problem $\psi_\beta^{N-\beta} (x)$. These seemingly contradictory facts can be reconciled by comparing the inner product for the continuous SSD with that for the lattice ones. This leads to the following correspondence: $$\psi_\beta^{N-\beta}(x) \sim \sin \left( \frac{\pi}{L} n \right) \chi_m (n) ,
\label{CLcomparison}$$ with $x\simeq n$, where $\sin ( \frac{\pi}{L} n)$ can be viewed as a geometric correction to $\chi_m$. For convenience, we fix the relation between $m$ and $\beta$ as $m = N-\beta$ so that the energy eigenvalues $\epsilon (m)$ of the lattice eigenfunctions are arranged in ascending order.
To verify Eq. (\[CLcomparison\]), we numerically diagonalize the lattice Hamiltonian (\[latticeSSD\]) for $L=101$ with $\mu=0$ and obtain the single-particle eigenfunctions $\chi_m (n)$. In this situation, we have $51$ energy levels below the zero energy: $m=0, \cdots, 50$, where the exact zero energy state corresponds to $m=50$. This can be seen in Fig. \[fig1\], where the eigenvalue spectrum is shown as a function of $m$. Note that, because of the finite-$L$ effect, the energy of the $m=49$ state slightly deviates from the zero energy to the low-energy side. Thus, the many-body ground state of the lattice SSD problem can be obtained by filling the levels $m=0, \cdots, 49$ with $N=50$ fermions.
![Eigenvalue spectrum of the lattice SSD model with $L=101$ and $\mu=0$. The energy eigenvalues $\epsilon (m)$ are arranged in ascending order. The Fermi level corresponds to $\epsilon (50)$, which is exactly at zero energy. The solid curve indicates $\epsilon (m) = -8.2 \times (m-49)(m-50)$ which is inferred from the spectrum of the continuous SSD analytically obtained in section 3.[]{data-label="fig1"}](fig1.eps){width="7cm"}
Before proceeding to the direct comparison of wavefunctions, we discuss the eigenvalue spectrum. Figure \[fig1\] shows the eigenvalue spectrum $\epsilon(m)$ for the lattice SSD problem with $L=101$ and $\mu=0$ in increasing order. In the figure, we also plot the spectrum of the continuous SSD in the negative energy region: $\epsilon (m)= -\kappa\beta(\beta-1)$ with $\beta=50-m$. The overall coefficient $\kappa= 8.2 \times 10^{-4}$ is determined so that the lattice and continuous results are on top of each other. Note that this value of $\kappa$ is clearly consistent with that of the continuous system $ \kappa=(\frac{\pi}{L})^2 \simeq 9.67 \times 10^{-4} $ for $L=101$. As depicted in the figure, we can see that the spectrum of the lattice SSD is well described by that of the continuous SSD in the entire energy range below the Fermi level. This should be contrasted to the fact that the linearization of the single-particle dispersion of the uniform lattice system, which is essential for taking the continuum limit, is valid only in the vicinity of the Fermi level. In this sense, the SSD for the lattice system efficiently smears out the lattice effect.
Using the above correspondence between the lattice and continuous SSD, we can further explain the various characteristics of the spectrum around the Fermi energy ($\epsilon=0$) reported in Ref. [@Hotta]. For instance, Eq. (\[SSDspectrum\]) for a fixed $\beta(=N-m)$ leads to $$\epsilon_L = -\frac{C}{L^2},
\label{FSS}$$ with $C := \pi^2\beta(\beta-1)$, which is in good agreement with the size dependence found in Ref. [@Hotta]. Note that this scaling relation contains no higher order correction of $O(L^{-3})$. Also from Eq. (\[SSDspectrum\]), we obtain the density of state (DOS) around the Fermi level as $$D(\epsilon) = \left|\frac{d \beta}{d\epsilon}\right| =\left(\frac{L}{\pi}\right)^2\frac{1}{2\beta-1}\simeq \frac{L}{2\pi} \frac{1}{\sqrt{|\epsilon|}}$$ for $|\epsilon| \gg 1/L^2$. For a sufficiently long chain, the DOS diverges with $|\epsilon|^{-1/2}$, which is also consistent with the numerical result in Ref. [@Hotta]. For $|\epsilon| \ll 1/L^2$, however, $D \simeq (\frac{L}{\pi})^{2}$(=const), which is a cutoff bound in the vicinity of the Fermi level.
![Comparison of single-particle wavefunctions between the lattice and continuous SSD systems. The chain length and the chemical potential are set to $L=101$ and $\mu=0$, respectively. In the corresponding continuous SSD, the total number of fermions is fixed as $N=50$. The lattice wavefunctions with the geometric correction ($\psi^{\rm ssd}_m (n)$) are plotted as open circles for $m=0, 1, 2, 25$, while the continuum wavefunctions ($\psi^{m}_{N-m}(x)$) are shown as solid lines. Note that the scaled position $\tilde{x} := \frac{\pi}{L}(n-\frac{1}{2})$ is used for $\psi^{\rm ssd}_m (n)$ for comparison. []{data-label="figwf"}](fig2_1.eps "fig:"){width="6cm"} ![Comparison of single-particle wavefunctions between the lattice and continuous SSD systems. The chain length and the chemical potential are set to $L=101$ and $\mu=0$, respectively. In the corresponding continuous SSD, the total number of fermions is fixed as $N=50$. The lattice wavefunctions with the geometric correction ($\psi^{\rm ssd}_m (n)$) are plotted as open circles for $m=0, 1, 2, 25$, while the continuum wavefunctions ($\psi^{m}_{N-m}(x)$) are shown as solid lines. Note that the scaled position $\tilde{x} := \frac{\pi}{L}(n-\frac{1}{2})$ is used for $\psi^{\rm ssd}_m (n)$ for comparison. []{data-label="figwf"}](fig2_2.eps "fig:"){width="6cm"} ![Comparison of single-particle wavefunctions between the lattice and continuous SSD systems. The chain length and the chemical potential are set to $L=101$ and $\mu=0$, respectively. In the corresponding continuous SSD, the total number of fermions is fixed as $N=50$. The lattice wavefunctions with the geometric correction ($\psi^{\rm ssd}_m (n)$) are plotted as open circles for $m=0, 1, 2, 25$, while the continuum wavefunctions ($\psi^{m}_{N-m}(x)$) are shown as solid lines. Note that the scaled position $\tilde{x} := \frac{\pi}{L}(n-\frac{1}{2})$ is used for $\psi^{\rm ssd}_m (n)$ for comparison. []{data-label="figwf"}](fig2_3.eps "fig:"){width="6cm"} ![Comparison of single-particle wavefunctions between the lattice and continuous SSD systems. The chain length and the chemical potential are set to $L=101$ and $\mu=0$, respectively. In the corresponding continuous SSD, the total number of fermions is fixed as $N=50$. The lattice wavefunctions with the geometric correction ($\psi^{\rm ssd}_m (n)$) are plotted as open circles for $m=0, 1, 2, 25$, while the continuum wavefunctions ($\psi^{m}_{N-m}(x)$) are shown as solid lines. Note that the scaled position $\tilde{x} := \frac{\pi}{L}(n-\frac{1}{2})$ is used for $\psi^{\rm ssd}_m (n)$ for comparison. []{data-label="figwf"}](fig2_4.eps "fig:"){width="6cm"}
Now, we turn to the comparison of the wavefunctions between the lattice and continuous SSD problems for $L=101$ and $N=50$. On the basis of Eq. (\[CLcomparison\]), we define the lattice wavefunction with the geometric correction as $$\psi^{\rm ssd}_m (n):=\sin(\tilde{x})\, \chi_m (n)$$ with $\tilde{x}= \frac{\pi}{L}(n-\frac{1}{2})$. The spatial profiles of $\psi^{\rm ssd}_m (n)$ for $m=0, 1, 2, 25$ are shown in Fig. \[figwf\]. The corresponding wavefunctions of the continuous SSD, $\psi_{N-m}^{m}(\tilde{x})$, for $m=0,1,2,25$ are also plotted as solid lines, where the overall normalization is adjusted so that $\psi^{\rm ssd}_m (n)$ and $\psi^{m}_{N-m}(x)$ are almost on top of each other. One can clearly see that the single-particle ground-state ($m=N-\beta=0$) is localized at the center of the chain and $\psi^{\rm ssd}_m (n)$ tends to delocalize as $m$ increases. It is remarkable that $\psi^{m}_{N-m}(x)$ well reproduces $\psi^{\rm ssd}(n)$ even when $m=25$, which exhibits a rapid oscillation as a function of ${\tilde x}$. Here, we note that the geometric correction becomes effective near the boundaries of the system. If we omit this correction, the lattice wavefunction significantly deviates from $\psi^{m}_{N-m}(x)$, as $\tilde{x}\to 0$ or $\pi$. We thus conclude that the correspondence (\[CLcomparison\]) between lattice and continuous SSD is well established not only at the level of energy eigenvalues, but also at the level of individual eigenfunctions. This implies that the continuum limit of the lattice is taken efficiently in the SSD system, which is consistent with the fact that the leading finite-size dependence of the excitation energy is proportional to $1/L^2$, instead of $1/L$ which is universal in CFT.
![The lattice SSD wavefunction at the Fermi level and geometric correction for the case of $L=101$ and $\mu=0$. The blue broken line with open circles indicates $\chi_{50} (n)$ as a function of the scaled position $\tilde{x}=\frac{\pi}{L}(n-\frac{1}{2})$. The red solid line with open circles represents $\psi^{\rm ssd}_{50} (n) = \sin(\tilde{x}) \chi_{50} (n)$.[]{data-label="zeroewf"}](fig3.eps){width="8cm"}
The geometric correction significantly affects the wavefunctions near the Fermi level, which is extended over the entire region of the chain. In Fig. \[zeroewf\], we show the lattice wavefunction $\chi_{50}(n)$ for $L=101$ with $\mu=0$, where $m=50$ corresponds to the zero energy state just on the Fermi level. In the figure, we also plot $\psi^{\rm ssd}_{50} (n) \propto \psi^{50}_0 (x)$ for comparison. Then, it is observed that the amplitude of the lattice wavefunction $\chi_{50} (n)$ increases toward the end sites $n=1$ and $L$, while $\psi^{\rm ssd}_{50} (n) \sim \psi_{0}^{50}(\tilde{x})$ retains a plane-wave behavior. As was mentioned before, $\psi_0^N(x)$ is not normalizable with respect to the inner product Eq. (\[innerprod\]) since $\psi_0^N(x)/\sin (N \tilde{x}) \sim \cot(N\tilde{x})$ diverges at $\tilde{x}=0$ and $\pi$. This divergence comes from the fact that $\psi_0^N(x)$ has a finite amplitude at the edges due to the Neumann boundary conditions. The result of Fig. \[zeroewf\] implies that the divergence of the inner product (\[innerprod\]) for $\psi_0^N$, which originates solely from the geometric factor, can be regularized by a lattice cutoff.
summary and discussions
=======================
In this paper, we have studied the sine-square deformation (SSD) of the free-fermion problem in one-dimensional continuous space. We formulated the single-particle problem of the SSD system as the inverse of the $1/\sin^2$ potential problem, where supersymmetric (SUSY) quantum mechanics is very effective. Exploiting both SUSY and the shape invariance, we explicitly obtained the complete set of the single-particle eigenfunctions of the SSD problem, and showed that its linear space is identical to that of the uniform periodic (antiperiodic) ring. Accordingly, we have shown that the many-body ground state of the system with SSD is the same as that of the uniform system with periodic (antiperiodic) boundary conditions. It would be interesting to generalize our theory to other class of superpotentials with the shape invariance. In addition, we have found that various properties of the lattice SSD problem are well explained by the analytic results for the SSD in the continuous space. In particular, the eigenfunctions of the lattice SSD problem can be reproduced from those of the continuous SSD by taking into account the geometric correction.
In the previous studies of the lattice systems that are reducible to free fermions [@Katsura2011; @Katsura2011B; @MaruyamaKH2011], the ground-state correspondence was proved using the plane-wave basis. Also, the correspondence was explained in the context of CFT [@Katsura2011B], where the SSD Hamiltonian is expressed as $L_0-(L_{+1}+L_{-1})/2$ and both the Virasoro generators $L_\pm$ annihilate the vacuum state $|0\rangle$. The present study clearly provides a complementary approach to the SSD problem, and various aspects of the SSD eigenfunctions can be clarified with the help of SUSY quantum mechanics. For example, we have demonstrated in the SSD system that the $1/L^2$ correction (\[FSS\]) of the excitation energy emerges with no higher order terms, instead of the universal $1/L$-size dependence in CFT. In this paper, nevertheless, SUSY quantum mechanics was just used to systematically construct the single-particle eigenfunctions of the SSD problem, and thus the role of the SUSY may not be clear at the level of field theory. Moreover, a very recent paper [@Ishibashi] pointed out that an infinite circumference limit of CFT can be obtained by adopting $L_0-(L_{+1}+L_{-1})/2$ as the Hamiltonian instead of $L_0$. Such singular behavior may be related to the normalizability of the zero energy state satisfying the Neumann boundary conditions that appear in the present continuous SSD. Therefore, it would be interesting to explore the relation between SUSY and CFT behind the SSD problems.
The authors would like to thank T. Hikihara, T. Nishino, K. Ohta, and T. Tada for useful discussions and comments. This work was supported in part by Grants-in-Aid Nos. 26400387, 25400407, and 15K17719 from the Ministry of Education, Culture, Sports, Science and Technology of Japan.
Formal solutions for positive energy {#app: positive_eng}
====================================
$0< \xi < \frac{1}{4}\left(\frac{\pi}{L}\right)^2 $
---------------------------------------------------
For Eq. (\[SSDpotential\]), $\xi$ becomes positive in $0<\beta<1$, which could be a candidate of a scattering state with positive energy. Taking account of a non-integer valued $\beta$, we introduce another parameterization $\beta=a+1/2$, which yields $$\xi = -\left(\frac{\pi}{L}\right)^2 \left( a^2-\frac{1}{4} \right).
\label{epsilon_a}$$ We then obtain the linearly-independent solutions of Eq. (\[SSDpotential\]) for a fixed $\mu = (\frac{\pi}{L})^2 q^2$ as $$\begin{aligned}
u(x) &=& (\sin \tilde{x})^{1/2 + a} \cos \tilde{x}\,
F \left(
\frac{3}{4} + \frac{a}{2} - \frac{q}{2},\,
\frac{3}{4} + \frac{a}{2} + \frac{q}{2};\,
\frac{3}{2};\, \cos^2 \tilde{x} \right)\label{hypergeo1}
\\
v(x) &=& (\sin \tilde{x})^{1/2 + a}\,
F \left(
\frac{1}{4} + \frac{a}{2} - \frac{q}{2},\,
\frac{1}{4} + \frac{a}{2} + \frac{q}{2};\,
\frac{1}{2};\, \cos^2 \tilde{x} \right),
\label{hypergeo2}\end{aligned}$$ where $F$ is Gauss’s hypergeometric function. Note that, if $a$ is a positive half integer, Eqs. (\[hypergeo1\]) and (\[hypergeo2\]) reduce to the Gegenbauer polynomials. Otherwise, these are not a polynomials in $\cos^2 {\tilde x}$.
Since Eqs. (\[hypergeo1\]) and (\[hypergeo2\]) with $|a|<1/2$ are not divergent in $0\le x \le L$, one might expect that the scattering state could be described by a linear combination of them. However, the derivative of $u(x)$ and $v(x)$ with respect to $x$ diverges at $\tilde{x}=0$ and $1$ when $|a|<1/2$, as far as $q$ is fixed at an integer [@Scarf]. Thus, there is no physical solution when $0 < \xi < \frac{1}{4}(\frac{\pi}{L})^2$. Of course, in the limit of $a\to \pm 1/2$, $u(x)$ and $v(x)$ approach to the trigonometric functions that are solutions for $\beta=0$ and $\beta=1$.
$\xi > \frac{1}{4}\left(\frac{\pi}{L}\right)^2$
-----------------------------------------------
In the context of the $1/\sin^2$ potential problem, the coupling corresponding to $\xi > \frac{1}{4}\left(\frac{\pi}{L}\right)^2$ is so strongly attractive that the single-particle ground state of $H_\beta$ is unstable, i.e., $\mu$ is not bounded from below. Formal solutions of Eq. (\[SSDpotential\]) for a certain fixed chemical potential $\mu$ are also given by Gauss’s hypergeometric function $$\begin{aligned}
u(x)&=&(\sin \tilde{x})^{1/2 + ia}\cos \tilde{x}\,
F \left(
\frac{3}{4} +i \frac{a}{2} - \frac{\mu}{2},\,
\frac{3}{4} +i \frac{a}{2} + \frac{\mu}{2};\,
\frac{3}{2};\, \cos^2 \tilde{x} \right)
\label{hypergeo3}\\
v(x)&=&(\sin \tilde{x})^{1/2 + ia}\,
F \left(
\frac{1}{4} + i\frac{a}{2} - \frac{\mu}{2},\,
\frac{1}{4} + i\frac{a}{2} + \frac{\mu}{2};\,
\frac{1}{2};\, \cos^2 \tilde{x} \right),
\label{hypergeo4}\end{aligned}$$ where we have adopted the parameterization $\xi = \left(\frac{\pi}{L}\right)^2 (a^2+\frac{1}{4})$. However, Eqs. (\[hypergeo3\]) and (\[hypergeo4\]) exhibit singular behavior in the vicinities of $\tilde{x}=0$ and $1$, reflecting the unstable ground state.
Derivation of $I_{N,\beta}$ in Eq. (\[innerprod\]) {#app:coeffI}
==================================================
In this Appendix, we present a detailed derivation of the coefficient $I_{N,\beta}$ appearing in Eq. (\[innerprod\]). Consider the squared norm of $\psi^{\ell}_{\beta} (x)$, where $\beta$ is assumed to be an integer and $\beta \ge 1$. Changing the variable from $x$ to $z:= \cos \frac{\pi}{L}x$, we have $$\langle\langle \psi^{\ell}_{\beta},\, \psi^{\ell}_{\beta} \rangle\rangle
=\frac{({\cal N}^{\ell}_{\beta})^2}{\pi}
\int^1_{-1} dz\, [C^{\beta}_{\ell} (z)]^2\, (1-z^2)^{\beta-\frac{3}{2}}.$$ To evaluate the above integral, we expand the Gegenbauer polynomials $C^{\beta}_\ell$ in terms of $C^{\beta-1}_{n}$ ($n=0,1,...,\ell$). The results depend on whether $\ell$ is even or odd: $$C^{\beta}_{2m} (z) = \sum^m_{k=0} \frac{\beta+2k-1}{\beta-1} C^{\beta-1}_{2k} (z),
\quad
C^{\beta}_{2m+1} (z) = \sum^m_{k=0} \frac{\beta+2k}{\beta-1} C^{\beta-1}_{2k+1} (z),$$ which can be proved using the formula obtained by Gegenbauer himself [@Keiner]. Then, using the orthogonality of $C^{\beta-1}_\ell (z)$: $$\int^1_{-1} dz\, C^{\beta-1}_{\ell} (z) C^{\beta-1}_{\ell'} (z) (1-z^2)^{\beta-3/2}
=\frac{\pi\, 2^{3-2\beta}\, (2\beta+\ell-3)!}{(\beta+\ell-1)\, \ell!\, [(\beta-2)!]^2}\,
\delta_{\ell,\ell'},$$ we arrive at $$\langle\langle \psi^{\ell}_{\beta} ,\, \psi^{\ell}_{\beta} \rangle\rangle
=2^{2-2\beta}\, ({\cal N}^{\ell}_{\beta})^2\,
\frac{(2\beta+\ell-1)!}{(2\beta-1)\, \ell!\, [(\beta-1)!]^2}.$$ The coefficient $I_{N,\beta}$ can be read off from the above formula by replacing $\ell$ with $N-\beta$.
References {#references .unnumbered}
==========
[99]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} A. Gendiar A, R. Krcmar R, and T. Nishino, Prog. Theor. Phys. [**122**]{}, 953 (2009); [**123**]{}, 393 (2010).
T. Hikihara and T. Nishino, Phys. Rev. B [**83**]{}, 060414 (2011).
A. Gendiar, M. Daniška, Y. Lee, and T. Nishino, Phys. Rev. A [**83**]{}, 052118 (2011).
H. Katsura, J. Phys. A: Math. Theor. [**44**]{}, 252001 (2011).
H. Katsura, J. Phys. A: Math. Theor. [**45**]{}, 115003 (2012).
N. Shibata and C. Hotta, Phys. Rev. B [**84**]{}, 115116 (2011).
C. Hotta and N. Shibata, Phys. Rev. B [**86**]{}, 041108(R) (2012).
C. Hotta, S. Nishimoto, and N. Shibata, Phys. Rev. B [**87**]{}, 115128 (2013).
T. Hikihara and T. Suzuki, Phys. Rev. A **87**, 042337 (2013).
T. Tada, Mod. Phys. Lett. A **30**, 1550092 (2015). N. Ishibashi and T. Tada, J. Phys. A: Math. Theor. **48**, 315402 (2015). I. Maruyama, H. Katsura, and T. Hikihara, Phys. Rev. B [**84**]{}, 165132 (2011).
K. Okunishi, J. Phys. Soc. Jpn. 76 063001 (2007); K. Okunishi and T. Nishino, Phys. Rev. B 82 144409 (2010).
G. Ram[í]{}rez, J. Rodr[í]{}guez-Laguna, and G. Sierra, J. Stat. Mech. P10004 (2014).
H. Ueda and T. Nishino, J. Phys. Soc. Jpn. [**78**]{}, 014001 (2009); H. Ueda, H. Nakano, K. Kusakabe, and T. Nishino, J. Phys. Soc. Jpn [**80**]{} 094001 (2011).
M. Vekić and S. R. White, Phys. Rev. Lett. [**71**]{}, 4283 (1993).
The meaning of “chiral” in this context is different from the original meaning in particle physics and field theory.
E. Witten, Nucl. Phys. B [**188**]{}, 513 (1981).
F. Cooper, A. Khare, and U. Sukhatme, Phys.Rept. [**251**]{}, 267 (1995).
B. Sutherland, J. Math. Phys. [**12**]{}, 246 (1971); [**12**]{}, 251 (1971).
Y. Kuramoto and Y. Kato,[*Dynamics of one-dimensional quantum systems*]{}, (Cambridge University Press, London, 2009).
See 8.934 in [*Table of integrals, series, and products, 7th edition*]{} by I. S. Gradshteyn and I. M. Ryzhik, (Academic Press, 2007).
See e.g. 1.332 in Ref. [@tableofintegral].
This relation can be checked by example in Table I.
L. D. Landau and I. E. Lifshitz, [*Quantum Mechanics*]{}, (Pergamon Press, 1965).
See J. Keiner, SIAM J. Sci. Comput. **31**, 2151 (2009), and references therein.
The Bloch state for a periodic array of the $1/\sin^2$ potential was constructed for $|a|<1/2$ in F. L. Scarf, Phys. Rev. [**112**]{}, 1137 (1958). The $k=0$ solution of the Bloch state corresponds to the present problem.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $P$ denote the class of languages accepted in polynomial time by Turing machine, let $NP$ denote the class of languages accepted in polynomial time by nondeterministic Turing machine, $coNP$ the class of complements of $NP$ languages, [*SAT*]{} the problem of Satisfiability, [*TAUT*]{} the problem of Tautologies. A central problem in computational complexity is to determine whether the class of $P$ and $NP$ are identical. In this paper, we first introduce the notion of reducibility from a language $L_1$ to another language $L_2$ which is basically the same as Karp’s, but to some extent is different, because our reducibility includes but not limited to be computable in polynomial-time. Our starting point is that we suspect there exists no polynomial-time reducibility from an arbitrary language $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to another arbitrary language $L_2\in P$ (resp. $L_2\in coP$). The whole discussions are divided into two parts, one is for $|\Sigma|\geq 2$ (general case), another is for $|\Sigma|=1$ (special case). The main contribution of the present paper is that a series of results are obtained. Specifically, we prove in general case that:
1. [Let $L_1\in NP-P$ and $L_2\in P$, then the complexity of problem on reducibility from $L_1$ to $L_2$ is $\Omega(m^{p(|\omega|)})$, $m\geq 2$ is a constant, where $p$ is a polynomial and $\omega\in\Sigma^*$ the input;]{}
2. [There exists no polynomial-time algorithm for [*SAT*]{};]{}
3. [An immediate corollary of (1) and (2) is that $P\neq NP$, which also can be deduced from (6), see Remark \[rk\_7.1\];]{}
4. [Let $L_1\in coNP-coP$ and $L_2\in coP$, then the complexity of problem on reducibility from $L_1$ to $L_2$ is $\Omega(m^{p(|\omega|)})$, $m\geq 2$ is a constant, where $p$ is a polynomial and $\omega\in\Sigma^*$ the input;]{}
5. [There exists no polynomial-time algorithm for [*TAUT*]{};]{}
6. [An immediate corollary of (4) and (5) is that $coP\neq coNP$;]{}
We next study the problem in special case. It is shown that:
1. [the complexity of problem of reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$) is $\mathcal {O}(p^{''}(|\omega|))$ where $p^{''}$ is a polynomial and $\omega\in\Sigma^*$ the input;]{}
2. [an immediate corollary is that $P=NP$ and $coP=coNP$ in the special case.]{}
However, the title of the paper will only reflect the general case.
author:
- 'Tianrong Lin[^1]'
title: ' $P\neq NP$'
---
\[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\]
Preliminaries {#sec_preliminaries}
=============
Let $\Sigma$ be a finite alphabet (with at least two elements), and let $\Sigma^*$ be the set of finite words over $\Sigma$, and $\Sigma^n$ the set of words over $\Sigma$ with length $n$. For any word $\omega\in\Sigma^*$, $|\omega|$ denotes the length of $\omega$, i.e., the number of elements in $\omega$. Further, for an alphabet $\Sigma$, $\Sigma^{\leq n}$ will denote the set $\{w\in\Sigma^* \,|\, |w|\leq n\}$.
We say a [*language over*]{} $\Sigma$ is a subset $L$ of $\Sigma^*$. For each word $\omega\in\Sigma^*$ there is a computation associated with a Turing machine $M$ (see [**Definition**]{} \[def\_turing\_machine\] below). We say that $M$ [*accepts*]{} $\omega$ if this computation terminates in the accepting state. Note that $M$ fails to accept $\omega$ either if this computation ends in the rejecting state, or if the computation fails to terminate. The [*language accepted by* ]{} $M$, denoted by $L(M)$, associating alphabet $\Sigma$, is defined by $$\begin{aligned}
\label{eq1}
L(M)&=&\{\omega\in\Sigma^*\,|\,M\,\,\,\mbox{accepts}\,\,\,\omega\}\end{aligned}$$
Further, denote by $t_{M}(\omega)$ the number of steps in the computation of $M$ on input $\omega$. If this computation never halts, then define $t_{M}(\omega)=\infty$. for $n\in\mathbb{N}$ we denote by $T_{M}(n)$ the [*worst case run time of*]{} $M$, i.e., $$\begin{aligned}
\label{eq2}
T_{M}(n)&=&\max\{t_{M}(\omega)\,|\,\omega\in\Sigma^n\}.\end{aligned}$$
We say that $M$ [*runs in polynomial time*]{} if there exists $k$ such that for all $n$, $T_{M}(n)\leq n^k+k$.[^2]
\[df\_p\_class\] $P$ is the class of languages accepted by Turing machines which operate in polynomial time, i.e., $$\begin{aligned}
\label{eq3}
P&=&\{L\,|\,L=L(M)\mbox{ for some Turing machine $M$ that runs in polynomial time}\}\end{aligned}$$
We supplement the definition of a Turing machine as follows. The Turing machine stated in the paper, unless otherwise indicated, is deterministic Turing machine. For the original definition, the reader is refer to Turing’s article [@Turing1936].
(see [@Cook2000])\[def\_turing\_machine\] Formally, a Turing machine $M$ is a tuple $(\Sigma,\Gamma,Q,\delta)$, where $\Sigma$, $\Gamma$, $Q$ are finite nonempty sets with $\Sigma\subseteq\Gamma$ and $b\in\Gamma-\Sigma$. The state set $Q$ contains three special states $q_0$, $q_{accept}$, $q_{reject}$ and $Q\cap\Gamma=\emptyset$. The [*transition function*]{} $\delta$ satisfies $$\begin{aligned}
\delta : (Q-\{q_{accept},q_{reject}\})\times\Gamma\longrightarrow Q\times\Gamma\times\{-1,1\}.
\end{aligned}$$ If $\delta(q,s)=(q',s',h)$, the interpretation is that, if $M$ is in state $q$ scanning the symbol $s$, then $q'$ is the new state, $s'$ is the symbol printed, and the tape head moves left or right one square depending on whether $h$ is $-1$ or $1$.
A [*configuration*]{} of $M$ is a string $xqy$ with $x,y\in\Gamma^*$, $y$ not the empty word, and $q\in Q$. Then interpretation of the configuration $xqy$ is that $M$ is in state $q$ with $xy$ on its tape, with its head scanning the left-most symbol of $y$.
If $C$ and $C'$ are configurations, then $C\stackrel{M}{\longrightarrow}C'$ if $C=xqsy$ and $\delta(q,s)=(q',s',h)$ and one of the following holds:
1. [$C'=xs'q'y$ and $h=1$ and $y$ is nonempty.]{}
2. [$C'=xs'q'b$ and $h=1$ and $y$ is empty.]{}
3. [$C'=x'q'as'y$ and $h=-1$ and $x=x'a$ for some $a\in\Gamma$.]{}
4. [$C'=q'bs'y$ and $h=-1$ and $x$ is empty.]{}
A configuration $xqy$ is [*halting*]{} if $q\in\{q_{accept},q_{reject}\}$. Note that for each non-halting configuration $C$ there is a unique configuration $C'$ such that $C\stackrel{M}{\longrightarrow}C'$.
Then [*computation*]{} of $M$ on input $\omega\in\Sigma^*$ is the unique sequence $C_0$, $C_1$, $\cdots$ of configurations such that $C_0=q_0\omega$ (or $C_0=q_0b$ if $\omega$ is empty word) and $C_i\stackrel{M}{\longrightarrow}C_{i+1}$ for each $i$ with $C_{i+1}$ in the computation, and either the sequence is infinite or it ends in a halting configuration. If the computation is finite, then the number of steps is one less than the number of configurations; otherwise the number of steps is infinite. We say the $M$ [*accepts*]{} $\omega$ if and only if the computation is finite and the final configuration contains the state $q_{accept}$.\
Let $R$ be a [*checking relation*]{}, which is simply a binary relation $R\subseteq\Sigma^*\times\Sigma^*$ for some finite alphabet $\Sigma$.[^3] We associate with each such relation $R$ a language $L_R$ over $\Sigma\cup\{\#\}$ defined by $$\begin{aligned}
\label{language_relation_r}
L_R&=&\{\omega\# y\,|\, R(\omega,y)\}\end{aligned}$$ where the symbol $\#\not\in\Sigma$. We say that $R$ is [*polynomial-time*]{} iff $L_R\in P$.
Now we define a language $L$ over $\Sigma$ by the condition that there is $k\in\mathbb{N}$ and a polynomial-time checking relation $R$ such that for all $\omega\in\Sigma^*$, $$\begin{aligned}
\label{df_np_languages}
\omega\in L&\Longleftrightarrow&\exists y\,(|y|\leq|\omega|^k\,\mbox{ and }\,R(\omega,y))\end{aligned}$$ where $|y|\leq|\omega|^k$, sometimes, it is also written as $|y|\leq p(|\omega|)$, $p$ stands for the polynomial $n^k$. We refer to $L$ as the language derived from $L_R$ by polynomial-bounded existential quantification. We also call $L_R$ the [*primal $P$-language*]{} of $L$.
\[df\_np\_class\] $NP$ is the class of languages $L$ derived from $L_R$ by polynomial-bounded existential quantification.
Originally, the class of $NP$ was defined in terms of nondeterministic Turing machines which have more than one possible move from a given configuration. The following Theorem due to Karp [@Karp1972] shows that two kinds of definition are equivalent. For the proof, the interested reader is refer to [@Karp1972] (see proof of Theorem 1 in [@Karp1972]).
$L\in NP$ if and only if $L$ is accepted by a nondeterministic Turing machine which operates in polynomial time.
Let $\prod$ be the class of functions from $\Sigma^*$ into $\Sigma^*$ computable in polynomial time by Turing machines, i.e., for any $\omega\in\Sigma^*$ and $\varphi\in\prod$, there exists a polynomial $p(\cdot)$ such that $T_{\varphi}(|\omega|)\leq p(|\omega|)$ (or $T_{\varphi}(|\omega|)\leq |\omega|^k+k$).[^4]
(Karp’s reducibility, cf. Definition 3 in [@Karp1972], or see Definition 3 in [@Cook2000])\[df\_polynomial\_reducibility\] Let $L_i$ be languages over $\Sigma$, $i=1,2$. We say that $L_1\propto_p L_2$ ($L_1$ is $p$-reducible to $L_2$) if there is a polynomial-time computable function $\varphi\in\prod$ such that $\varphi(\omega)\in L_2\Longleftrightarrow \omega\in L_1$ for all $\omega\in\Sigma^*$. [^5]
An obvious result is that, if $L_2\in P$ and $L_1\propto_p L_2$ then $L_1\in P$. We call $L_1$ and $L_2$ equivalent if $L_1\propto_p L_2$ and $L_2\propto_p L_1$. Call a language $L$ (polynomial) complete if $L\in NP$ and every language in $NP$ is $p$-reducible to $L$.
The following remarkable Theorem, attributed to Cook [@Cook1971], presents the first $NP$-complete language. For the proof we refer the interested reader to [@Cook1971] (see proof of Theorem 1 in [@Cook1971]).
\[cook\_theo\](Cook’s Theorem) If $L\in NP$, then $L\propto_p$ SAT.
We list the language [*SAT*]{} as follows (see [@Karp1972]).\
[*SAT*]{}\
INPUT: Clauses $C_1$, $C_2$, $\cdots$, $C_l$\
PROPERTY: The conjunction of the given clauses is satisfiable; i.e., there is a set\
$S\subseteq\{x_1,x_2,\cdots,x_n;\neg x_1, \neg x_2,\cdots,\neg x_n\}$ such that
1. [$S$ does not contain a complementary pair of literals,]{}
2. [$S\cap C_k\neq\emptyset$, $k=1$, $2$, $\cdots$, $l$.]{}
\[rk\_reducible\] In Karp’s reducibility (Definition \[df\_polynomial\_reducibility\]), it requires the functions $\in\prod$ computable in polynomial-time. This is a strong constraint, since in our subsequent study of reducibilities, we shall pay especial attention to all functions in $\bigvee$, consisting of $\varphi:\Sigma^*\rightarrow\Sigma^*$, which is computable by Turing machine, including but not limited to be computable in polynomial-time. Thus, we introduce the following
\[all\_reducibility\] Suppose that $L_i$, $i=1,2$, are two languages over $\Sigma$. Then $L_1\propto L_2$($L_1$ is reducible to $L_2$) if there is a computable function $\varphi\in\bigvee$, such that $\varphi(\omega)\in L_2\Longleftrightarrow \omega\in L_1 $ for all $\omega\in\Sigma^*$. We also call $\varphi$ a reducibility from $L_1$ to $L_2$.
The complement of a decision problem $L$, denoted by $\overline{L}$, will be defined as the decision problem whose answer is “yes" whenever the input is a “no" input of $L$, and vice versa, see [@Papad1994]. For instance, the complement of [*SAT*]{} (i.e., [*TAUT*]{}) is the problem: Given a Boolean expression $\phi$ in conjunctive normal form, is it unsatisfiable? Further, for any complexity class $\mathcal {C}$, $co\mathcal{C}$ denotes the class $\{\overline{L}\,|\,L\in\mathcal{C}\}$.
By the above explanations, now we also define a language $\overline{L}$ over $\Sigma$ by the condition that there is $k\in\mathbb{N}$ and a polynomial-time checking relation $R$ such that for all $\omega\in\Sigma^*$, $$\begin{aligned}
\label{df_conp_languages}
\omega\in\overline{L}&\Longleftrightarrow&\forall y\,(|y|\leq|\omega|^k\,\mbox{ and }\,\neg R(\omega,y))
\end{aligned}$$ where $|y|\leq|\omega|^k$, sometimes, it is also written as $|y|\leq p(|\omega|)$, $p$ stands for the polynomial $n^k$ and $\neg R(\omega,y)$ means $(\omega,y)\not\in R$. We refer to $\overline{L}$ as the language derived from $L_R$ by polynomial-bounded universal quantification, where $L_R=\{\omega\# y\,|\, R(\omega,y)\}$ is over $\Sigma\cup\{\#\}$.
\[df\_conp\_class\][^6] $coNP$ is the class of languages $\overline{L}$ derived from $L_R$ by polynomial-bounded universal quantification.
Dual to Theorem \[cook\_theo\], we have the following theorem with its proof omitted, i.e., [*TAUT*]{} is the standard example of a $coNP$-complete language
(Dual of Cook’s Theorem)\[dual\_cook\_theo\] If $\overline{L}\in coNP$, then $\overline{L}\propto_p$[*TAUT*]{}.
In the following, we will divide our discussions into two parts, one is for the general case, i.e., when the alphabet $\Sigma$ contains at least two element; and another is for the special case, i.e., when the alphabet $\Sigma$ contains only one element. To summarize, we will prove the following results:
In general case, i.e., when $|\Sigma|\geq 2$, we show the following
\[lower\_bouds\_reducibility\_1\] Let $L_1\in NP-P$ be an arbitrary language, let $L_2\in P$ be an arbitrary language. Then the complexity of problem of reducibility from $L_1$ to $L_2$ is $\Omega(m^{p(|\omega|)})$, where $m\geq 2$ is a constant, $\omega\in\Sigma^*$ the input.
\[theo\_sat\_not\_in\_p\] There exists no polynomial-time algorithm for [*SAT*]{}.
\[p\_not\_eq\_np\_cy\] $P\neq NP$.
This is a corollary of Theorem \[theo\_sat\_not\_in\_p\] and of the well known fact $P\subseteq NP$.
\[lower\_bouds\_reducibility\_2\] Let $L_1\in coNP-coP$ be an arbitrary language, let $L_2\in coP$ be an arbitrary language. Then the complexity of problem of reducibility from $L_1$ to $L_2$ is $\Omega(m^{p(|\omega|)})$, where $m\geq 2$ is a constant, $\omega\in\Sigma^*$ the input.
\[theo\_tau\_not\_in\_cop\] There exists no polynomial-time algorithm for [*TAUT*]{}.
\[p\_not\_eq\_conp\_cy\] $coP\neq coNP$.
This is a corollary of Theorem \[theo\_tau\_not\_in\_cop\] and of the fact $P=coP\subseteq coNP$.
In special case, i.e., when $|\Sigma|=1$, we prove that:
\[theo\_special\_case\] If $\Sigma$ contains only one element, then there exists polynomial-time reducibility from $L_1$ to $L_2$, where $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) and $L_2\in P$ (resp. $L_2\in coP$). An immediate corollary is that when $|\Sigma|=1$, $P=NP$ and $coNP=coP$.
The plan of the paper is the following. For convenience of the reader, almost all notation and notions are presented in this section. In Section \[sec\_related\_work\], we review some histories and importance of the problem, as well as possible ways and our approach to the problem. In order to prove our lower bounds on Theorem \[lower\_bouds\_reducibility\_1\] and Theorem \[lower\_bouds\_reducibility\_2\], we collect some lower bounds lemmas in Section \[sec\_lower\_bounds\], where an interesting property on reducibility from $L\in NP-P$ to its $L_R$ is also listed. Section \[reducibility\_1\] is devoted to the whole proof of Theorem \[lower\_bouds\_reducibility\_1\]. The non-existence of polynomial-time algorithm for [*SAT*]{} is discussed in Section \[non\_existence\_algorithm\_for\_sat\], where we get an important conclusion that $P\neq NP$ in general case. The Theorem \[lower\_bouds\_reducibility\_2\] is shown in Section \[reducibility\_2\] and non-existence of polynomial-time algorithm for [*TAUT*]{} is presented in Section \[non\_existence\_algorithm\_for\_taut\], where we claim that $coP\neq coNP$ in general case. We discuss the special case when the alphabet $\Sigma$ contains only one element and show different results in Section \[special\_case\]. Finally, in the last section of the paper we draw some conclusions.
Related work and importance {#sec_related_work}
===========================
A central problem in computational complexity is, as is well known, the $P$ versus $NP$ problem, which is to determine whether every language accepted by some nondeterministic algorithm (alternately, nondeterministic Turing machine) in polynomial time is also accepted by some (deterministic) algorithm (alternately, Turing machine) in polynomial time. In this Section, we will review its history, importance, and possible ways and attempts to resolve it, and we mention our proof ideas at the end of this section.
In 1971, Cook [@Cook1971] introduced a notion of $NP$-completeness as a polynomial-time analog of c.e.-completeness, except that the reduction used was a polynomial-time analog of Turing reducibility rather than of many-one reducibility (see [@HR1967] Chapter $7$). Besides the first well-known $NP$-complete problem of Satisfiability, Cook also showed in [@Cook1971] that several natural problems, including 3-SAT and subgraph isomorphism are $NP$-complete.
A year later stimulated by the work of Cook [@Cook1971], Karp used these completeness results to show in [@Karp1972] that $20$ other natural problems are $NP$-complete, forcefully demonstrating the importance of the subject. Thus far, there are many problems shown to be $NP$-complete, see excellent reference [@GJ1979] to this subject. Karp also introduced the now standard notation $P$ and $NP$ and redefined $NP$-completeness by using the polynomial-time analog of many-one reducibility, which has become standard. Meanwhile Levin [@Lev1973], independently of Cook [@Cook1971] and Karp [@Karp1972], defined the notion of “universal search problem", similar to the $NP$-complete problem, and gave six examples, which includes Satisfiability.
Although the $P$ versus $NP$ problem was formally defined in 1970s, there were previous inklings of the problems involved. A mention of the underlying problem occurred in a $1956$ letter written by K. Gödel to J. von Neumann. Gödel asked whether theorem-proving (now known to be $coNP$-complete) could be solved in quadratic or linear time (see [@Hart1989]). It is worth mentioning that, besides the classical version of the problem, there is one expressed in terms of the field of complex numbers, which catches interest in the mathematics community, see [@BCSS1998].
The importance of the $P$ versus $NP$ is well known, and one reason (see [@BGS1975] for other explanations) for this may be that $P$ and $NP$ are very natural classes of languages, invariant under reasonable changes of machine model. For example, $P$ or $NP$ is the same class whether defined by computations by one-tap Turing machines, multitape Turing machines, or random-access machines. Hence this question deals with the basic nature of computation and not merely with minor aspects of computer models. Furthermore, there are too many consequences if the question is resolved (see related survey in [@Cook2003]). We mention a few here, for example, if the first proof of $P=NP$ is present, it is possible that the proof is nonconstructive, in the sense that it does not yield an algorithm for any $NP$-complete problem. Or it might give an impractical algorithm, for instance running in time $n^{1000}$. These may disappoint complexity theorists. If it is proved by exhibiting a truly feasible algorithm for an $NP$-complete problem such as [*SAT*]{}, then it would be stunning. First, as mentioned in [@Cook2003], most of the hundreds of problems shown to be $NP$-complete can be efficiently reduced to [*SAT*]{}, so many of the optimization problems important to industry could be solved. Second, mathematics would be transformed, because computers could find a formal proof of any theorem which has a proof of reasonable length. Similar comments apply to the fundamental problems of artificial intelligence: planning, natural language understanding, vision, and so on. In each case success would depend on finding good algorithms for recognizing good results. Of course, there is a negative consequence of a feasible proof that $P=NP$, i.e., complexity based cryptography would become impossible, because all of these problems are efficiently reducible to [*SAT*]{}. If $P\neq NP$ is proved, this might just answer the most basic of a long list of important related questions that could keep complexity theorists busy far in the future.
Currently, there are three main ways have been tried to the problem, see survey of [@Cook2000]. A case is to assume that $P=NP$ and attempt to exhibit a polynomial-time algorithm for $3$-[*SAT*]{} or one of the other thousand or so known $NP$-complete problems. Because of their importance in industry, a vast number of engineers have attempted to find efficient algorithms for $NP$-complete problems, as well as a similar strong motivation for breaking the cryptographic schemes that assume $P\neq NP$ for their security. Another case is that, suppose $P\neq NP$, and there are two general methods that have been tried: diagonalization with reduction and Boolean circuit lower bounds.
Beginning with the Halting problem, the method of diagonalization with reduction has been used very successfully in computability theory to prove a host of problems undecidable, see [@Cook2000]. It has also been used successfully in complexity theory to prove super-exponential lower bounds for very hard decidable problems. For example, Presburger arithmetic, the first-order theory of integers under addition, is a decidable theory for which Fischer and Rabin [@FR1974] proved that any Turing machine deciding the theory must use at least $2^{2^{cn}}$ steps in the worst case, for some $c>0$. Returning back the $P$ versus $NP$ problem, it was, however, shown in [@BGS1975] that there is an oracle set $A$ relative to which $P=NP$, suggesting that diagonalization with reduction cannot be used to separate these two classes. Another to prove $P\neq NP$ is to prove a super-polynomial lower bound on the size of any family of Boolean circuits solving specific $NP$-complete problem, such as $3$-[*SAT*]{}. Combination of diagonalization and circuit lower bound methods is also used. For example, Yao [@Yao1985] presented exponential lower bounds on the size of depth-$k$ Boolean circuits for computing certain functions, which imply that there exists an oracle set $A$ such that, relative to $A$, all the levels in the polynomial-time hierarchy are distinct. However, all attempts to find even super-linear lower bounds for unrestricted Boolean circuits for “explicitly given" Boolean functions have met with total failure, see [@RR1997].
[**Our approach.**]{} Of central importance in computability theory is the notion of reducibility, this is also true in this paper. Specifically, our starting point is that we suspect there exists no polynomial-time reducibility from an arbitrary language $L_1\in NP-P$ to another arbitrary language $L_2\in P$. By Cook’s Theorem, it is well known that the reducibility from any language $L\in NP$ to [*SAT*]{} is polynomial-time. Furthermore, if the reducibility from any language $L_1\in NP-P$ to another arbitrary language $L_2\in P$ is also polynomial-time, then we assert that $P=NP$. This motivate us to study the complexity of problem on reducibility from $L_1\in NP-P$ to $L_2\in P$. It is shown that the complexity of the above problem is $\Omega(m^{p(|\omega|)})$, $m=|\Sigma|$ a constant and $\omega\in\Sigma^*$ the input. Such a result enable us to prove that [*SAT*]{} is not in $P$, or equivalently, there exists no polynomial-time algorithm for [*SAT*]{}.
Taking one further step, we also suspect there exists no polynomial-time reducibility from an arbitrary language $L_1\in coNP-coP$ to another arbitrary language $L_2\in coP$, which can be viewed as the second starting point. By Dual of Cook’s Theorem, the reducibility from any language in $coNP$ to [*TAUT*]{} is polynomial-time, if the reducibility from any language $L_1\in coNP-coP$ to another arbitrary language $L_2\in coP$ is also polynomial-time, then $coP=coNP$. Similarly, we prove that the complexity of problem of reducibility from any language $L_1\in coNP-coP$ to another arbitrary language $L_2\in coP$ is also $\Omega(m^{p(|\omega|)})$, $m=|\Sigma|$ a constant and $\omega\in\Sigma^*$ the input. As a consequence, we assert that there exists no polynomial-time algorithm for [*TAUT*]{}.
Finally we are left with the case of $|\Sigma|=1$. By carefully analyzing the cardinality $C$ of set $\Sigma^{\leq p(|\omega|)}$, we obtain a lower bound $\Omega(p(|\omega|))$ for the problem of reducibility from any language $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to any language $L_2\in P$(resp. $L_2\in coP$). Further, we show that there exists a reducibility from $L_1$ to $L_2$ with upper bound $\mathcal{O}(p^{''}(|\omega|))$, where $p^{''}(\cdot)$ is a polynomial greater than polynomial $p(\cdot)$. Therefore we show that, in the special case, there exists polynomial-time reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$), implying that $P=NP$ and $coP=coNP$ when $|\Sigma|=1$.
We would like to stress that the approach used in this paper had never been tried by anybody before in the literature.
Lower bounds {#sec_lower_bounds}
============
Recall that a lower bound on a [*problem*]{} is a $\Omega$ bound on the worst-case running time of [*any*]{} (i.e., [*all*]{}) algorithm that solves the problem. In this Section we list some lower bounds results about searching an unordered set that is crucial in the proof of some of our main results. We also study an interesting property on reducibility from any language $L\in NP-P$ to its primal $P$-language $L_R$.
\[lm\_lower\_boud\_searching\_1\] Given an unordered set $D$ of $N$ elements. Then any (deterministic) algorithm in the worst case to find whether there exists an element $d\in D$ with the property $\mathcal{P}$ needs at least $N$ times test, i.e., its complexity is $\Omega(N)$.
This following Lemma is so obvious, so the proof is omitted.
\[lm\_lower\_bound\_all\_searching\] Given an unordered set $D$ of $N$ elements. Then any (deterministic) algorithm in the worst case to find whether all elements in $D$ are with the property $\mathcal{P}$ is with the complexity $\Omega(N)$.
It is worth mentioning that, in computation theory, for any nondeterministic Turing machine $M_N$, there is a Turing machine $M_D$ such that the languages accepted by $M_N$ and $M_D$ are identical [@HMU2006](see Chapter 8, Theorem 8.11 of [@HMU2006]). That is to say, equivalently, for any $L\in NP$, there is an algorithm accepting it, but the algorithm may take exponentially more time than the nondeterministic algorithm, and it is also unknown whether or not this exponential slowdown is necessary. In this section, we present a property which is similar to that. Specifically, we assert that for any language $L_1\in NP-P$, there is a language $L_2\in P$, such that $L_1\propto L_2$, but the complexity of reducibility is exponential-time, i.e., the following
Let $L_1\in NP-P$ be a language over $\Sigma$. Then there exists a language $L_2\in P$ such that $L_1\propto L_2$. Moreover, the complexity of the reducibility is $\Omega(m^{p(|\omega|)})$ where $m=|\Sigma|\geq 2$ is a constant and $\omega\in\Sigma^*$ is the input, $p$ an polynomial.
The first half part of the above theorem is expounded as follows and the second half part can be deduced from Theorem \[lower\_bouds\_reducibility\_1\], hence we omit its proof here.
For any $L_1\in NP-P$, let $L_R$ be the primal $P$-language of $L_1$. And define the computable function $\varphi$ : $\Sigma^*\rightarrow(\Sigma\cup\{\#\})^*$ by $$\begin{aligned}
\varphi(\omega)&=&\omega\#y\end{aligned}$$ where $y\in\Sigma^*$ satisfies that $|y|\leq p(|\omega|)$ and $(\omega,y)\in R$ if $\omega\in L_1$ and otherwise $$\begin{aligned}
\varphi(\omega)&=&\omega\end{aligned}$$ It is clear that for any $\omega\in\Sigma^*$ $$\begin{aligned}
\omega\in L_1&\Longleftrightarrow&\omega\#y\in L_R\end{aligned}$$ i.e., $\varphi(\omega)=\omega\# y\in L_R\in P$. So $\varphi$ is a reducibility from $L_1$ to $L_R$. Then putting $L_2=L_R$ completes the assertion.
Reducibility from $L_1\in NP-P$ to $L_2\in P$ {#reducibility_1}
=============================================
The notion of [*reducibility*]{} is important in computability theory, and is an effective mean to establish certain [*unsolvability*]{} results, see [@HR1967] for more details. Roughly, one problem is [*reducible*]{} to another if an algorithm for solving the second problem yields an algorithm for solving the first. Further, Turing’s definition is : A language $L_1$ is [*Turing reducible*]{} to a language $L_2$ if and only if there is an oracle Turing machine $M$ that accepts $L_1$, where $M$ is allowed to make membership queries of the form $\omega\in L_2$, which are correctly answered by “oracle" for $L_2$, see [@Cook2000]. It is worth mentioning that reducibility plays an important role, for example, the notion of $NP$-complete, which is an analog of c.e.-complete (due to Post, see [@HR1967], p. 78), is based on more restricted notion of Turing reducibility [@Cook2000], i.e., a polynomial-time analog of Turing reducibility (see Definition \[df\_polynomial\_reducibility\]).
Comparing our reducibility with Karp’s, we can see from the Definition \[all\_reducibility\] and the Definition \[df\_polynomial\_reducibility\] that these two notions are basically the same, except that the former includes but not limited to polynomial-time one. The goal here is to present a proof to support our suspicion, i.e., we show that there exists no polynomial-time reducibility from any language $L_1\in NP-P$ to any language $L_2\in P$. In other words, we observe that, if there exists polynomial-time reducibility from [*SAT*]{} to a language in $P$, then [*SAT*]{} is in $P$, or there exists polynomial-time algorithm for [*SAT*]{}, which motivate us to further study the complexity of problem on reducibility from any language $L_1\in NP-P$ to any language $L_2\in P$.
The following lemma gives a characterization of the reducibility from a language in $NP$ but not in $P$ to a language in $P$. It provides a key step for the proof of the lower bound of the complexity on reducibility from any language in $NP-P$ to any language in $P$.
A auxiliary Lemma
-----------------
\[eq\_relation\] Let $L_1\in NP-P$, and $L_2\in P$ be two language over $\Sigma$ which are arbitrary. Then for any computable function $\varphi: \Sigma^*\rightarrow \Sigma^*$, $\varphi$ is a reducibility from $L_1$ to $L_2$ (i.e., $L_1\propto L_2$) iff $$\begin{aligned}
\label{eq1}
\exists y\,\, (|y|\leq p(|\omega|)\mbox{ and } R(\omega,y)) \Longleftrightarrow M \mbox{ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$ for any $\omega\in\Sigma^*$, where $p(\cdot)$ and $p'(\cdot)$ are polynomials and $R$ is a polynomial-time checking relation, $M$ is a Turing machine.[^7]
Since $L_1\in NP-P$, then by definition, there exists a polynomial $p(\cdot)$ and a polynomial-time checking relation $R$ such that $$\begin{aligned}
\label{eq2}
\omega\in L_1&\Longleftrightarrow& \exists y\,\,(|y|\leq p(|\omega|)\mbox{ and } R(\omega,y))
\end{aligned}$$ for all $\omega\in\Sigma^*$. Because $L_2\in P$, also by definition, there exists a Turing machine $M$ and a polynomial $p'(\cdot)$ such that $$\begin{aligned}
\label{eq3}
\varphi(\omega)\in L_2&\Longleftrightarrow& M \mbox{ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$
The “only if" part. If $\varphi$ is a reducibility from $L_1$ to $L_2$, then by Definition \[all\_reducibility\], we have $$\begin{aligned}
\label{eq4}
\omega\in L_1 &\Longleftrightarrow \varphi(\omega)\in L_2
\end{aligned}$$ for all $\omega\in\Sigma^*$.
Now (\[eq2\]), (\[eq3\]) and (\[eq4\]) lead to the following $$\begin{aligned}
\exists y\,\,(|y|\leq p(|\omega|)\mbox{ and } R(\omega,y) &\Longleftrightarrow& M \mbox{ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$ i.e., (\[eq1\]) is valid.
The “if" part. If (\[eq1\]) is valid, then we have $$\begin{aligned}
\omega\in L_1 \Longleftrightarrow \exists y\,\,((|y|\leq p(|\omega|)\mbox{ and } R(\omega,y))&\Longleftrightarrow& M \mbox{ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)\\
&\Longleftrightarrow&
\varphi(\omega)\in L_2
\end{aligned}$$ for all $\omega\in\Sigma^*$, where the first $\Longleftrightarrow$ is by $L_1\in NP-P$, the second $\Longleftrightarrow$ is by (\[eq1\]) and the third $\Longleftrightarrow$ is by $L_2\in P$.
This further implies that $\omega\in L_1\Longleftrightarrow \varphi(\omega)\in L_2$ for any $\omega\in\Sigma^*$, by Definition \[all\_reducibility\], it is not hard to see $\varphi$ is a reducibility from $L_1$ to $L_2$.
Now we are ready to present the proof of Theorem \[lower\_bouds\_reducibility\_1\] as follows.
Proof of Theorem \[lower\_bouds\_reducibility\_1\]
--------------------------------------------------
Let $\varphi:\Sigma^*\rightarrow\Sigma^*$ be any computable function. By Lemma \[eq\_relation\] just shown above, $\varphi$ is a reducibility from languages $L_1\in NP-P$ to $L_2\in P$ iff $$\begin{aligned}
\exists y\,\, (|y|\leq p(|\omega|)\mbox{ and } R(\omega,y)) &\Longleftrightarrow& M \mbox{ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$ for any $\omega\in\Sigma^*$ and for some Turing machine $M$.
For any input $\omega\in\Sigma^*$, we proceed to estimate $T_{\varphi}(|\omega|)$, which again is divided into two cases: $\varphi(\omega)\in L_2$ and $\varphi(\omega)\not\in L_2$.
- [The case of $\varphi(\omega)\in L_2$, i.e., there exists a Turing machine $M$ accepting $\varphi(\omega)$ with $T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)$. By Lemma \[eq\_relation\], it is equivalent to determining that there is a string $y\in\Sigma^{\leq p(|\omega|)}$ such that $(\omega,y)\in R$. Then by Lemma \[lm\_lower\_boud\_searching\_1\], this requires that $\varphi$ performs at least $|\Sigma^{\leq p(|\omega|)}|$ times test, and for each test it requires at least $1$ step to see whether $(\omega,y)\in R$ or not. Hence, [^8] $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq &|\Sigma^{\leq p(|\omega|)}|\times 1\mbox{ (by Lemma \ref{lm_lower_boud_searching_1})};
\end{aligned}$$]{}
- [The case of $\varphi(\omega)\not\in L_2$, i.e., there exists no Turing machine $M$ accepting $\varphi(\omega)$ with $T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)$. Again, by Lemma \[eq\_relation\], it is equivalent to determining that there exists no string $y\in\Sigma^{\leq p(|\omega|)}$ such that $(\omega,y)\in R$. By Lemma \[lm\_lower\_bound\_all\_searching\], this also requires $\varphi$ performing at least $|\Sigma^{\leq p(|\omega|)}|$ times test, and for each test it requires at least $1$ step to see whether $(\omega,y)\not\in R$ or not. Hence, $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq &|\Sigma^{\leq p(|\omega|)}|\times 1\mbox{ (by Lemma \ref{lm_lower_bound_all_searching})};
\end{aligned}$$]{}
In summary, we have $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq& |\Sigma^{\leq p(|\omega|)}|\times 1\\
&=&|\Sigma^{\leq p(|\omega|)}|
\end{aligned}$$ i.e., $\varphi$ is with the complexity $\Omega(C)$, where $C=|\Sigma^{\leq p(|\omega|)}|$ is the cardinality of set $\Sigma^{\leq p(|\omega|)}$. Now recall the definition of $\Sigma^{\leq n}$ and assume, without loss of generality, that $\Sigma=\{\sigma_1,\sigma_2,\cdots,\sigma_m\}$ where $m\geq 2$ is a constant, we compute the cardinality of $\Sigma^{\leq p(|\omega|)}$ as follows $$\begin{aligned}
C &=& |\Sigma^0|+|\Sigma^1|+\cdots +|\Sigma^{p(|\omega|)}|\\
&=&\sum\limits_{i=0}^{p(|\omega|)}|\Sigma^i|=\sum\limits_{i=0}^{p(|\omega|)}m^i\\
&=&\frac{m^{p(|\omega|)+1}-1}{m-1}=\frac{(m-1)m^{p(|\omega|)} + (m^{p(|\omega|)}-1)}{m-1}\\
&\geq&\frac{(m-1)m^{p(|\omega|)}}{m-1}=m^{p(|\omega|)}\qquad\mbox{( since $m^{p(|\omega|)}-1\geq 0$)}\end{aligned}$$ where $\Sigma^0=\{\epsilon\}$, and $\epsilon$ is empty word whose length is $0$.
Therefore, $\varphi$ is with complexity $\Omega(m^{p(|\omega|)})$. Since $\varphi$ is arbitrary, we get that, for any $\varphi\in\bigvee$, $\varphi$ is a reducibility from $L_1$ to $L_2$, then $\varphi$ is with complexity $\Omega(m^{p(|\omega|)})$, where $m=|\Sigma|\geq 2$ is a constant and $\omega\in\Sigma^*$ is the input. Consequently, this reaches the proof of Theorem \[lower\_bouds\_reducibility\_1\].
\[rm\_of\_lower\_bouds\_reducibility\_1\] Theorem \[lower\_bouds\_reducibility\_1\] implies that, for any languages $L_1\in NP-P$ and $L_2\in P$, there exists no polynomial-time computable function $\varphi:\Sigma^*\rightarrow\Sigma^*$ such that $\varphi(\omega)\in L_2\Longleftrightarrow \omega\in L_1$ for any $\omega\in\Sigma^*$. In other words, it is to say that any reducibility from $L_1$ to $L_2$ requires at least $\Omega(m^{p(|\omega|)})$ steps for any input $\omega\in\Sigma^*$.
\[rk\_4.2\] In the proof of Theorem \[lower\_bouds\_reducibility\_1\], we calculate the cardinality of $\Sigma^{\leq p(|\omega|)}$. As we can see, $C$ is strongly connect to the cardinality of $\Sigma$. In our opinion, $|\Sigma|\geq 2$ is a general case, i.e., $C$ is exponential in terms of $|\omega|$ (in fact, $|\Sigma|=2$ is enough for our discussion). The special case when $|\Sigma|=1$ will be performed careful analysis in the subsequent study.
No polynomial-time algorithm for [*SAT*]{} {#non_existence_algorithm_for_sat}
==========================================
As mentioned in Section \[sec\_related\_work\], one may assume that $P=NP$ and then to find a polynomial-time algorithm for $3$-[*SAT*]{} or just [*SAT*]{} or other so known $NP$-complete problems. Or, one may prove a polynomial upper bound on the size of any family of Boolean circuits solving specific $NP$-complete problem such as [*SAT*]{} or $3$-[*SAT*]{}. Indeed, under the assumption that $P=NP$, much research has already been done in the past $50$ years, but without success, see [@Cook2000] for the details. This causes people to doubt that $P$ may not be equal to $NP$. In this section, we will show that there exists no polynomial-time algorithm for [*SAT*]{}, implying that the success of the huge efforts to find polynomial-time algorithm for any $NP$-complete problem would be impossible.
Proof of Theorem \[theo\_sat\_not\_in\_p\]
------------------------------------------
We shall show the theorem by contradiction. Suppose there exists polynomial-time algorithm for [*SAT*]{}. Then by Definition \[df\_p\_class\], we have $$\begin{aligned}
SAT&\in& P.\end{aligned}$$ Clearly, for any language $L\in NP-P$, we have $$\begin{aligned}
L&\propto_p& SAT\mbox{ ( by Cook's Theorem, i.e., Theorem \ref{cook_theo} )}.\end{aligned}$$ However, by Theorem \[lower\_bouds\_reducibility\_1\] (or, by Remark \[rm\_of\_lower\_bouds\_reducibility\_1\])), we know that for any reducibility from $L_1\in NP-P$ to $L_2\in P$ (where $L_1$ and $L_2$ are arbitrary), its complexity is $\Omega(m^{p(|\omega|)})$ where $m\geq 2$ is a constant and $\omega\in\Sigma^*$ the input. This is a contradictory. Therefore, there exists no polynomial-time algorithm for [*SAT*]{}, i.e., [*SAT*]{} $\not\in P$. This completes the proof of Theorem \[theo\_sat\_not\_in\_p\].
Now, Corollary \[p\_not\_eq\_np\_cy\] follows immediately from Theorem \[theo\_sat\_not\_in\_p\] and $P\subseteq NP$.
Reducibility from $L_1\in coNP-coP$ to $L_2\in coP$ {#reducibility_2}
===================================================
As mentioned above, the complexity of problem of reducibility from any language in $coNP-coP$ to any language in $coP$ is our second starting point, which is similar to Section \[reducibility\_1\]. The goal here is to extend our discussion in Section \[reducibility\_1\] to this case. Again recall that for any language $L\in coNP$, $L\propto_p$[*TAUT*]{}, see Dual of Cook’s Theorem, if we can prove that there exists polynomial-time reducibility from [*TAUT*]{} to any language in $coP$ (i.e., [*TAUT*]{}$\,\,\propto_p L$ where $L\in coP$) then we can easily get that [*TAUT*]{} in $coP$, hence proving $coP=coNP$. This is not true, since any reducibility from $L_1\in coNP-coP$ to $L_2\in coP$ requires at least $\Omega(m^{p(|\omega|)})$ steps where $m=|\Sigma|\geq 2$ is a constant, which will be shown soon. The first step to do this is similar to Section \[reducibility\_1\], i.e., to give a characterization of when a computable function $\varphi$ in $\bigvee$ is eligible to be a reducibility from $L_1\in coNP-coP$ to $L_2 \in coP$.
The following lemma gives a characterization of the above problem. The reader may find that it is similar to the Lemma \[eq\_relation\]. In fact, we can call it Dual of Lemma \[eq\_relation\].
A auxiliary Lemma
-----------------
The proof of the following Lemma is similar to that of Lemma \[eq\_relation\], for the sake of completeness, we present the proof in detail.
\[eq\_relation\_co\] Let $L_1\in coNP-coP$, and $L_2\in coP$ be two language over $\Sigma$. Then for any computable function $\varphi: \Sigma^*\rightarrow \Sigma^*$, $\varphi$ is a reducibility from $L_1$ to $L_2$ (i.e., $L_1\propto L_2$) iff $$\begin{aligned}
\label{eq5}
\forall y\,\, (|y|\leq p(|\omega|)\mbox{ and } \neg R(\omega,y)) \Longleftrightarrow \mbox{ $M$ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$ for any $\omega\in\Sigma^*$, where $p(\cdot)$ and $p'(\cdot)$ are polynomials and $R$ is a polynomial-time checking relation, $M$ is a Turing machine.
Since $L_1\in coNP-coP$, then by definition, there exists a polynomial $p(\cdot)$ and a polynomial-time checking relation $R$ such that $$\begin{aligned}
\label{eq6}
\omega\in L_1&\Longleftrightarrow& \forall y\,\, (|y|\leq p(|\omega|)\mbox{ and } \neg R(\omega,y))
\end{aligned}$$ for all $\omega\in\Sigma^*$. Because $coP=P$ and $L_2\in coP$, then by definition, there exists Turing machine $M$ and polynomial $p'(\cdot)$ such that $$\begin{aligned}
\label{eq7}
\varphi(\omega)\in L_2&\Longleftrightarrow& \mbox{ $M$ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$
We show first the “only if" part. If $\varphi$ is a reducibility from $L_1$ to $L_2$, then by Definition \[all\_reducibility\], we have $$\begin{aligned}
\label{eq8}
\omega\in L_1 &\Longleftrightarrow \varphi(\omega)\in L_2
\end{aligned}$$ for all $\omega\in\Sigma^*$.
(\[eq6\]), (\[eq7\]) and (\[eq8\]) lead to the following $$\begin{aligned}
\forall y\,\, (|y|\leq p(|\omega|)\mbox{ and } \neg R(\omega,y)) &\Longleftrightarrow& \mbox{ $M$ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)
\end{aligned}$$ i.e., (\[eq5\]) is valid.
We prove next the “if" part. If (\[eq5\]) is valid, then we have $$\begin{aligned}
\omega\in L_1 \Longleftrightarrow \forall y\,\, (|y|\leq p(|\omega|)\mbox{ and } \neg R(\omega,y))&\Longleftrightarrow& \mbox{ $M$ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)\\
&\Longleftrightarrow& \varphi(\omega)\in L_2
\end{aligned}$$ for all $\omega\in\Sigma^*$, where the first $\Longleftrightarrow$ is by $L_1\in coNP-coP$, the second $\Longleftrightarrow$ is by (\[eq5\]) and the third $\Longleftrightarrow$ is by $L_2\in coP$.
This further implies that $\omega\in L_1\Longleftrightarrow \varphi(\omega)\in L_2$ for any $\omega\in\Sigma^*$, i.e., $\varphi$ is a reducibility from $L_1$ to $L_2$.
Proof of Theorem \[lower\_bouds\_reducibility\_2\]
--------------------------------------------------
Similar to proof of Theorem \[lower\_bouds\_reducibility\_1\], let $\varphi:\Sigma^*\rightarrow\Sigma^*$ be any computable function. By Lemma \[eq\_relation\_co\] above, for any input $\omega\in\Sigma^*$, $\varphi$ is a reducibility from $L_1\in coNP-coP$ to $L_2\in coP$ iff $$\begin{aligned}
\forall y\,\, (|y|\leq p(|\omega|)\mbox{ and } \neg R(\omega,y)) &\Longleftrightarrow& \mbox{ $M$ accepts $\varphi(\omega)$ and } T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)\end{aligned}$$ for any $\omega\in\Sigma^*$ and for some Turing machine $M$.
For any input $\omega\in\Sigma^*$, we also proceed to estimate $T_{\varphi}(|\omega|)$, which again is divided into two cases: $\varphi(\omega)\in L_2$ and $\varphi(\omega)\not\in L_2$.
- [The case of $\varphi(\omega)\in L_2$, i.e., there exists a Turing machine $M$ accepting $\varphi(\omega)$ with $T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)$. By Lemma \[eq\_relation\_co\], it is equivalent to determining whether for all string $y\in\Sigma^{\leq p(|\omega|)}$ such that $(\omega,y)\not\in R$. Then by Lemma \[lm\_lower\_bound\_all\_searching\], this requires that $\varphi$ performs at least $|\Sigma^{\leq p(|\omega|)}|$ times test, and for each test it requires at least $1$ step to see whether $(\omega,y)\in R$ or not. Hence, $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq &|\Sigma^{\leq p(|\omega|)}|\times 1\mbox{ (by Lemma \ref{lm_lower_bound_all_searching})};
\end{aligned}$$]{}
- [The case of $\varphi(\omega)\not\in L_2$, i.e., there exists no Turing machine $M$ accepting $\varphi(\omega)$ with $T_M(|\varphi(\omega)|)\leq p'(|\varphi(\omega)|)$. Again, by Lemma \[eq\_relation\_co\], it is equivalent to determining that there exists string $y\in\Sigma^{\leq p(|\omega|)}$ such that $(\omega,y)\in R$. By Lemma \[lm\_lower\_boud\_searching\_1\], this also requires $\varphi$ performing at least $|\Sigma^{\leq p(|\omega|)}|$ times test, and for each test it requires at least $1$ step to see whether $(\omega,y)\in R$ or not. Hence, $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq &|\Sigma^{\leq p(|\omega|)}|\times 1\mbox{ (by Lemma \ref{lm_lower_boud_searching_1})};
\end{aligned}$$]{}
In summary, we have $$\begin{aligned}
T_{\varphi}(|\omega|)&\geq& |\Sigma^{\leq p(|\omega|)}|\times 1\\
&=&|\Sigma^{\leq p(|\omega|)}|
\end{aligned}$$ i.e., $\varphi$ is with the complexity $\Omega(C)$, where $C=|\Sigma^{\leq p(|\omega|)}|$ is the cardinality of set $\Sigma^{\leq p(|\omega|)}$. Similarly, we assume, without loss of generality, that $\Sigma=\{\sigma_1,\sigma_2,\cdots,\sigma_m\}$, and compute the cardinality of $\Sigma^{\leq p(|\omega|)}$ in the following $$\begin{aligned}
C &=& |\Sigma^0|+|\Sigma^1|+\cdots +|\Sigma^{p(|\omega|)}|\\
&=&\sum\limits_{i=0}^{p(|\omega|)}|\Sigma^i|=\sum\limits_{i=0}^{p(|\omega|)}m^i\\
&=&\frac{m^{p(|\omega|)+1}-1}{m-1}\\
&\geq&m^{p(|\omega|)}\end{aligned}$$ where $\Sigma^0=\{\epsilon\}$, and $\epsilon$ is empty word whose length is $0$.
Hence, $\varphi$ is with complexity $\Omega(m^{p(|\omega|)})$. Since $\varphi$ is arbitrary, we have that for any reducibility $\varphi$ from $L_1$ to $L_2$, $\varphi$ is with complexity $\Omega(m^{p(|\omega|)})$, where $m=|\Sigma|\geq 2$ is a constant and $\omega\in\Sigma^*$ is the input. Therefore, Theorem \[lower\_bouds\_reducibility\_2\] follows.
\[rk\_6.1\] As mentioned just now, Theorem \[lower\_bouds\_reducibility\_2\] implies that, for any languages $L_1\in coNP-coP$ and $L_2\in coP$, there exists no polynomial-time computable function $\varphi:\Sigma^*\rightarrow\Sigma^*$ such that $\varphi(\omega)\in L_2\Longleftrightarrow \omega\in L_1$ for any $\omega\in\Sigma^*$. That is, any reducibility from $L_1$ to $L_2$ requires at least $\Omega(m^{p(|\omega|)})$ steps for any input $\omega\in\Sigma^*$.
In proof of Theorem \[lower\_bouds\_reducibility\_2\], we calculate the cardinality of $\Sigma^{\leq p(|\omega|)}$. Similarly, $C$ is strongly connect to the cardinality of $\Sigma$. In our opinion, $|\Sigma|\geq 2$ is a general case, i.e., $C$ is exponential with respect to $|\omega|$(in fact, $|\Sigma|=2$ is enough for our discussion). The special case when $|\Sigma|=1$ will be performed careful analysist in the subsequent study.
No polynomial-time algorithm for [*TAUT*]{} {#non_existence_algorithm_for_taut}
===========================================
In similar to the case discussed in Section \[non\_existence\_algorithm\_for\_sat\], one may also suppose that $coP=coNP$ and thereby to design a polynomial-time algorithm for [*TAUT*]{}, a standard example of a $coNP$-complete language. Although we do not know whether much research has already been tried previously. Now, we suggest not to do that, because we will show that no polynomial-time algorithm for [*TAUT*]{} exists. We have seen in Section \[non\_existence\_algorithm\_for\_sat\] the proof of non-existence of polynomial-time algorithm for [*SAT*]{} is via Cook’s Theorem and Theorem \[lower\_bouds\_reducibility\_1\]. Similarly, to prove that there exists no polynomial-time algorithm for [*TAUT*]{} is also by contradiction via the Dual of Cook’s Theorem and Theorem \[lower\_bouds\_reducibility\_2\]. We now proceed to prove Theorem \[theo\_tau\_not\_in\_cop\] as follows
Proof of Theorem \[theo\_tau\_not\_in\_cop\]
--------------------------------------------
Similar to proof of Theorem \[theo\_sat\_not\_in\_p\], we shall show the theorem by contradiction. Suppose there exists polynomial-time algorithm for [*TAUT*]{}. Then by Definition \[df\_p\_class\], we have $$\begin{aligned}
TAUT&\in& coP=P.\end{aligned}$$ Then for any language $L\in coNP-coP$, we have $$\begin{aligned}
L&\propto_p& TAUT\mbox{ ( by Dual of Cook's Theorem, i.e., Theorem \ref{dual_cook_theo} )},\end{aligned}$$ But Theorem \[lower\_bouds\_reducibility\_2\] says that there exists no polynomial-time reducibility from $L_1\in coNP-coP$ to $L_2 \in coP$, see Remark \[rk\_6.1\], which is a contradictory. Therefore, there exists no polynomial-time for [*TAUT*]{}. This finishes the proof of Theorem \[theo\_tau\_not\_in\_cop\].
\[rk\_7.1\] Now, Corollary \[p\_not\_eq\_conp\_cy\] follows immediately from Theorem \[theo\_tau\_not\_in\_cop\] and $P=coP\subseteq coNP$. Further note that, Corollary \[p\_not\_eq\_np\_cy\] can also be deduced from Theorem \[theo\_tau\_not\_in\_cop\]. To see this, assume that $P=NP$, then we obtain $coNP=NP=P=coP$, which contradicts to $coNP\neq coP$, hence Corollary \[p\_not\_eq\_np\_cy\] follows immediately.
The special case of $|\Sigma|=1$ {#special_case}
================================
Now we are left with the case that the alphabet $\Sigma$ contains only one element. Although we have shown that in general case, the complexity of problem on reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$) is $\Omega(m^{p(|\omega|)})$ where $m\geq 2$ is a constant, $\omega\in\Sigma^*$ the input, we need carefully analyze the cardinality $C$ of set $\Sigma^{\leq p(|\omega|)}$ when $|\Sigma|=1$ to see whether the similar result holds or not. If not, then moving one step further, we need to find the upper bound of a reducibility to see whether it is a polynomial upper bound. We proceed to prove in this section that, when $\Sigma$ contains only one element, there exists polynomial-time reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$), which implies that in this case $P=NP$ and $coP=coNP$. However, as the reader see, the title of the paper will only reflect the general case.
Proof of Theorem \[theo\_special\_case\]
----------------------------------------
First, compute the cardinality $C$ of $\Sigma^{\leq p(|\omega|)}$ by setting $|\Sigma|=1$, similar to that in Section \[reducibility\_1\] and Section \[reducibility\_2\] $$\begin{aligned}
C &=& |\{0\}^0|+|\{0\}^1|+\cdots +|\{0\}^{p(|\omega|)}|\\
&=&\sum\limits_{i=0}^{p(|\omega|)}1^i=1+\underbrace{1+\cdots +1}_{p(|\omega|)}\\
&=&p(|\omega|)+1\end{aligned}$$ That is, for any reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$), it is with complexity $\Omega(p(|\omega|))$, where $p$ an polynomial and $\omega\in\Sigma^*$ the input.
We prove next that there exists a reducibility $\varphi\in\bigvee$ running in $\mathcal{O}(|\omega|^k+k)$ for $k > 0$ from $L_1$ to $L_2$ as follows. To do this, first assume $p(n)=n^{k_1}+k_1$. Note that there is only one element in $\Sigma$, assume $\Sigma=\{0\}$. Since $R$ is a polynomial-time checking relation, we may assume that there is a polynomial $p'(n)=n^{k_2}+k_2$ and a Turing machine for $L_R$ such that $T_M(|\omega|)\leq p'(|\omega|)$ for all $\omega\in\Sigma^*$.
1. for the case of reducibility from $L_1\in NP-P$ to $L_2\in P$ : given an input $\omega$, then we verify all string whose length is less than $p(|\omega|)$, i.e., to see whether there is a string $\pi$ among $\epsilon$, $0$, $00$, $\cdots$, $\underbrace{00\cdots 0}_{p(|\omega|)}$ such that $\omega\#\pi\in L_R$. This requires at most $p'(|\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}|)$ steps for each string listed above. Hence, we get $$\begin{aligned}
T_{\varphi}(|\omega|)&\leq& (p(|\omega|)+1)\times p'(|\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}|)
\end{aligned}$$
Clearly, $$\begin{aligned}
p'(|\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}|)&=&|\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}|^{k_2}+k_2\\
&\leq&\left(2|\omega|^{k_1}+(k_1+1)\right)^{k_2}+k_2\\
&=&\sum\limits_{i=0}^{k_2}{k_2\choose i}2^i|\omega|^{ik_1}(k_1+1)^{k_2-i}+k_2
\end{aligned}$$ Hence, $$\begin{aligned}
T_{\varphi}(|\omega|)&\leq&(|\omega|^{k_1}+(k_1+1))\times \left(\sum\limits_{i=0}^{k_2}{k_2\choose i}2^i|\omega|^{ik_1}(k_1+1)^{k_2-i}+k_2\right)\\
&=&p^{''}(|\omega|)
\end{aligned}$$ where ${k_2\choose i}$ denotes binomial coefficient and $p^{''}(\cdot)$ is a polynomial.
2. [for the case of reducibility from $L_1\in coNP-coP$ to $L_2\in coP$ : it is similar to the above case. Given an input $\omega$, then we verify that for all string whose length is less than $p(|\omega|)$, i.e., $\epsilon$, $0$,$00$, $\cdots$, $\underbrace{00\cdots 0}_{p(|\omega|)}$, to see whether $(\omega,\epsilon)\not\in R$, $\cdots$, $(\omega,\underbrace{00\cdots 0}_{p(|\omega|)})\not\in R$. This is done by verifying that $\omega\#\not\in L_R$, $\omega\#0\not\in L_R$, $\omega\#00\not\in L_R$, $\cdots$, $\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}\not\in L_R$. For each string, it also requires at most $p'(|\omega\#\underbrace{00\cdots 0}_{p(|\omega|)}|)$ steps. Therefore $$\begin{aligned}
T_{\varphi}(|\omega|)&\leq&(|\omega|^{k_1}+(k_1+1))\times \left(\sum\limits_{i=0}^{k_2}{k_2\choose i}2^i|\omega|^{ik_1}(k_1+1)^{k_2-i}+k_2\right)\\
&=&p^{''}(|\omega|)
\end{aligned}$$ where ${k_2\choose i}$ denotes binomial coefficient and $p^{''}(\cdot)$ is a polynomial. ]{}
In summary, $$\begin{aligned}
p(|\omega|)\leq T_{\varphi}(|\omega|)\leq p^{''}(|\omega|)\end{aligned}$$ for all $\omega\in\{0\}^*$. Hence, there exists polynomial-time reducibility from $L_1$ to $L_2$, which implies that when $\Sigma=\{0\}$, $$\begin{aligned}
{\it SAT}&\propto_p& L\qquad\mbox{( $L\in P$)}\\
{\it TAUT} &\propto_p& L'\qquad\mbox{ ( $L'\in coP$)}
\end{aligned}$$ Therefore, in this case, it is true that $P=NP$ and $coNP=coP$.
Conclusions {#sec_conclusions}
===========
The main goals of this paper is to investigate the complexity of problem on the reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$), and to obtain a lower bound of $\Omega(m^{p(|\omega|)})$ ($m\geq 2$ is a constant) for these problems, thus proving that there exists no polynomial-time algorithm for [*SAT*]{} (resp. [*TAUT*]{}). Our reducibility is basically the same as Karp’s, but Karp’s mainly focus on polynomial-time reducibility and our definition includes but not limited to polynomial-time one. The whole discussion is divided into two parts, one is for general case, i.e., when the alphabet $\Sigma$ contains at least $2$ elements. In fact, for our purpose, the alphabet $\Sigma$ contains $2$ elements is enough for our discussion. Another is for special case, i.e., when the alphabet contains only $1$ elements. In the context, we begin with collecting some lower bound results, and also present an assertion with respect to reducibility from any language $L\in NP$ to its primal $P$-language.
The main contributions of this paper are that showing a series of results : first we obtain a lower bound for the problem on reducibility from any language $L_1\in NP-P$ to any language $L_2\in P$. Afterwards, by the lower bound result we show that there exists no polynomial-time algorithm for [*SAT*]{}, hence showing that [*SAT*]{} is not in $P$, implying further that $P\neq NP$. Moving one step further, we extend the techniques used above to investigate the reducibility from any language $L_1\in coNP-coP$ to any language $L_2\in coP$. Similar lower bound is obtained and by this, we can present a proof of Theorem saying that there exists no polynomial-time algorithm for [*TAUT*]{}, a standard example of $coNP$-complete language. These results are dual of the results obtained above. We point out that $P\neq NP$ and $coP\neq coNP$ are immediate corollaries of the above results. Furthermore, the conclusion of $P\neq NP$ also can be deduced from $coP\neq coNP$. By carefully analyzing, we show that, when the alphabet $\Sigma$ contains only $1$ element, the complexity of problem on reducibility from $L_1\in NP-P$ (resp. $L_1\in coNP-coP$) to $L_2\in P$ (resp. $L_2\in coP$) is greater than $p(|\omega|)$ and less than $p^{''}(|\omega|)$, where $p$ and $p^{''}$ are polynomial, implying that there exists polynomial-time reducibility from $L_1$ to $L_2$, which shows that $P=NP$ and $coP=coNP$ in the special case. However, the title of the paper attempts not to reflect this.
Finally, note that the complexity of problem on reducibility from $L_1\in NP-P$ to $L_2\in P$ is identical to that of problem on reducibility from $L_1\in coNP-coP$ to $L_2\in coP$, i.e., both of them are $\Omega(m^{p(|\omega|)})$, $m\geq 2$ a constant and $\omega\in\Sigma^*$ the input. We strongly suspect that $NP=coNP$ in the case of $|\Sigma|\geq 2$, but currently without proofs.
[1]{}
S. Cook. . In Proceedings of the Third Annual ACM Symposium on Theory of Computing, pages 151–158, ACM, New York, 1971.
S. Cook. . April, 2000. Available at <http://www.claymath.org/sites/default/files/pvsnp.pdf>.
S. Cook. . Journal of the ACM, 50 (1), 27–29.
R. M. Karp. . In Complexity of Computer Computations, R. E. Miller and J. W. Thatcher, eds., Plenum Press, New York, 1972, 85–103.
L. Levin. (in Russian), Problemy Peredachi Informatsii 9 (1973), 265–266. English translation in B. A. Trakhtenbrot, [*A survey of Russian approaches to Perebor (brute-force search) algorithms*]{}, Annals of the History of Computing 6 (1984), 384–400.
A. Turing. . Proc. London Math. Soc. 42 : 230–265, 1936.
C. Papadimitriou. . Addison–Wesley, Reading, MA, 1994.
M. R. Garey and D. S. Johnson. . W. H. Freeman and Co., San Francisco, 1979.
J. Hartmanis. . Bulletin of the European Association for Theoretical Computer Science. 38 : 101–107, 1989. Available at <https://ecommons.cornell.edu/bitstream/handle/1813/6910/89-994.pdf>.
H. Rogers, Jr. . McGraw-Hill, New York, 1967.
A. Yao. . Proc. IEEE Symp. Found. Comput. Sci., 1985, pp. 1–10.
M. J. Fischer and M. O. Rabin. . Complexity of Computation 7, AMS, Providence, RI, 1974, 27–41.
T. Baker, J. Gill, and R. Solovay. . SIAM J. Comput., 4 (4) : 431–442, 1975.
A. A. Razborov and S. Rudich. . J. Comput. System Sci. 55 (1997) 24–35.
L. Blum, F. Cucker, M. Shub, and S. Smale. . Springer–Verlag, New York, 1998.
J. E. Hopcroft, R. Motwani, and J. D. Ullman. . 3rd edition, Addision-Wesley, New York, 2006.
[^1]: [*E-mail address*]{}: [email protected]
[^2]: Occasionally, we write $T_M(|\omega|)\leq |\omega|^k+k$ for any $|\omega|\in\mathbb{N}$, $\omega\in\Sigma^*$, and $p(\cdot)$ stands for $n^k +k$.
[^3]: Or, the relation is defined as $R\subseteq\Sigma^*\times\Sigma_1^*$ for two different finite alphabets $\Sigma$ and $\Sigma_1$, but it has no influence to our results.
[^4]: Note that $\varphi$ is computable in polynomial-time by Turing machines, we can view $\varphi$ as a Turing machine.
[^5]: More generally, $L_i$ can be over different alphabet $\Sigma_i$. For example, $L_i\subseteq\Sigma_i^*$, $i=1,2$ and $\varphi:\Sigma_1^*\rightarrow\Sigma_2^*$ is computable in polynomial-time.
[^6]: Now it is obvious that $coNP=\{\overline{L}\,|\,L\in NP\}$ and ${\it TAUT}=\overline{{\it SAT}}\in coNP$.
[^7]: The reducibility from $L_1\in NP-P$ to $L_2\in NP-P$ is different. Let $\varphi$ : $\Sigma^*\rightarrow \Sigma^*$ be a computable function, then $\varphi$ is a reducibility from $L_1\in NP-P$ to $L_2\in NP-P$ if and only if $$\begin{aligned}
\exists y\,\,(|y|\leq p_1(|\omega|)\mbox{ and } R_1(\omega,y))\Longleftrightarrow \exists z\,\,(|z|\leq p_2(|\varphi(\omega)|)\mbox{ and } R_2(\varphi(\omega),z))
\end{aligned}$$ See proof of Cook’s Theorem (i.e., Theorem \[cook\_theo\]) in [@Cook1971].
[^8]: Without Lemma \[lm\_lower\_boud\_searching\_1\], we can prove that for any input $\omega\in\Sigma^*$, in the case of $\varphi(\omega)\in L_2$, $T_{\varphi}(|\omega|)\geq m^{|\omega|}$, where $m=|\Sigma|$ a constant. To show this, let us assume that $|\omega|=n$ and $\varphi$ can test less than $m^{|\omega|}$ times producing the correct result for any input with length $n$. Then there is necessarily an element between $1$ and $m^n$ that $\varphi$ did not look at. Then assume, without loss of generality, that the $i$th element. Since $\varphi$ is deterministic, running $\varphi$ with input $\omega'\in\Sigma$ with $|\omega'|=n$ and the $i$th element is $y$ such that $(\omega',y)\in R$ and for any $y'$ in other places with $(\omega',y')\not\in R$ will produce incorrect result, which is a contradictory.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A space curve in a Euclidean 3-space $\mathbb E^3$ is called a rectifying curve if its position vector field always lies in its rectifying plane. This notion of rectifying curves was introduced by the author in [@c6]. In this present article, we introduce and study the notion of rectifying submanifolds in Euclidean spaces. In particular, we prove that a Euclidean submanifold is rectifying if and only if the tangential component of its position vector field is a concurrent vector field. Moreover, rectifying submanifolds with arbitrary codimension are completely determined.'
address: '*Michigan State University Department of Mathematics 619 Red Cedar Road, East Lansing, Michigan 48824–1027, U.S.A.*'
author:
- 'Bang-Yen Chen'
title: Differential geometry of rectifying submanifolds
---
Introduction
============
Let $\mathbb E^3$ denote Euclidean 3-space with its inner product $\left<\;\,,\;\right>$. Consider a unit-speed space curve $x : I\to \mathbb E^3$, where $I=(\alpha,\beta)$ is a real interval. Let ${\bf x}$ denote the position vector field of $x$ and ${\bf x}'$ be denoted by [**t**]{}.
It is possible, in general, that ${\bf t} '(s)=0$ for some $s$; however, we assume that this never happens. Then we can introduce a unique vector field ${\bf n}$ and positive function $\kappa$ so that ${\bf t}'=\kappa {\bf n}$. We call ${\bf t}'$ the [*curvature vector field*]{}, ${\bf n}$ the [*principal normal vector field*]{}, and $\kappa$ the [*curvature*]{} of the curve. Since ${\bf t}$ is of constant length, ${\bf n}$ is orthogonal to ${\bf t}$. The [*binormal vector field*]{} is defined by ${\bf b}={\bf t}\times{\bf n}$, which is a unit vector field orthogonal to both ${\bf t}$ and ${\bf n}$. One defines the [*torsion*]{} $\tau$ by the equation ${\bf b}'=-\tau{\bf n}$.
The famous Frenet-Serret equations are given by $$\begin{aligned}
\label{E:S-F} &\begin{cases}{\bf t}'=\hskip.38in\kappa {\bf n}\\ {\bf n}'=-\kappa {\bf t} \hskip.3in+\tau{\bf b}
\\{\bf b}'=\hskip.3in-\tau{\bf n}.\end{cases}\end{aligned}$$ At each point of the curve, the planes spanned by $\{{\bf t},{\bf n}\}$, $\{{\bf t},{\bf b}\}$, and $\{{\bf n},{\bf b}\}$ are known as the [*osculating plane*]{}, the [*rectifying plane*]{}, and the [*normal plane*]{}, respectively.
From elementary differential geometry it is well known that a curve in $\mathbb E^3$ lies in a plane if its position vector lies in its osculating plane at each point, and lies on a sphere if its position vector lies in its normal plane at each point. In view of these basic facts, the author asked the following simple geometric question in [@c6]:
When does the position vector of a space curve $\hbox{\bf x}: I\to \mathbb E^3$ always lie in its rectifying plane?
The author called such a curve a [*rectifying curve*]{} in [@c6]. The author derived many fundamental properties of rectifying curves. In particular, he completely classifies all rectifying curves in [@c6]. It is known that rectifying curves related with the notions constant-ratio curves and convolution (cf. [@c2001; @c2002a; @c2002b; @c2003a; @c2003b]). Furthermore, the author and F. Dillen established in [@cd05] a simple link between rectifying curves and the notion of centrodes in mechanics. Moreover, they showed in [@cd05] that rectifying curves are indeed the extremal curves which satisfy the equality case of a general inequality. Since then rectifying curves have been studied by many authors, see [@Cam16; @II2003; @II2007; @II2008; @II2014; @Lucas2015; @O2009; @Yi16; @Yu2007] among many others. For the most recent survey on rectifying curves, see [@c16].
In this article, we extend the notion of rectifying curves to the notion of rectifying submanifolds in a very natural way. Many fundamental properties of rectifying submanifolds are obtained. In particular, we prove that a Euclidean submanifold is rectifying if and only if the tangential component of its position vector field is a concurrent vector field. Moreover, rectifying submanifolds with arbitrary codimension are completely determined.
Preliminaries
=============
Let $x: M\to \mathbb E^m$ be an isometric immersion of a Riemannian manifold $M$ into the Euclidean $m$-space $\mathbb E^m$. For each point $p\in M$, we denote by $T_pM$ and $T^\perp_p M$ the tangent and the normal spaces at $p$.
There is a natural orthogonal decomposition: $$\label{2.1} T_p{\mathbb E}^{m}=T_pM\oplus T^\perp_p
M.$$
Denote by $\nabla$ and $\tilde\nabla$ the Levi-Civita connections of $M$ and ${\mathbb E}^{m}$, respectively. The formulas of Gauss and Weingarten are given respectively by (cf. [@cbook; @book11]) $$\begin{aligned}
\label{2.2} &\tilde \nabla_XY=\nabla_X Y+h(X,Y),
\\& \label{2.3} \tilde\nabla_X\xi =-A_\xi X+D_X\xi \end{aligned}$$ for vector fields $X,\,Y$ tangent to $M$ and $\xi$ normal to $M$, where $h$ is the second fundamental form, $D$ the normal connection, and $A$ the shape operator of $M$.
For a given point $p\in M$, the [*first normal space*]{}, of $M$ in $\mathbb E^m$, denoted by ${\rm Im}\,h_p$, is the subspace defined by $$\begin{aligned}
\label{2.4} {\rm Im}\,h_p ={\rm Span}\{h(X,Y):X,Y\in T_p M\}.\end{aligned}$$
For each normal vector $\xi$ at $p$, the shape operator $A_\xi$ is a self-adjoint endomorphism of $T_pM$. The second fundamental form $h$ and the shape operator $A$ are related by $$\label{2.5} \<A_\xi X,Y\>=\<h(X,Y), \xi\>,$$ where $\<\;\, ,\;\>$ is the inner product on $M$ as well as on the ambient Euclidean space.
The [*equation of Gauss*]{} of $M$ in $\mathbb E^m$ is given by $$\begin{aligned}
\label{2.6} R(X,Y;Z,W)=\<\sigma(X,W),\sigma(Y,Z)\>-\<\sigma(X,Z),\sigma(Y,W)\> \end{aligned}$$ for $X,Y,Z,W$ tangent to $M$, where $R$ denotes the curvature tensors of $M$.
The covariant derivative ${\bar \nabla}h$ of $h$ with respect to the connection on $TM \oplus T^{\perp}M$ is defined by $$\begin{aligned}
\label{2.7}({\bar\nabla}_{X}h)(Y,Z)=D_{X}(h (Y,Z))-h(\nabla_{X}Y,Z)-h(Y,\nabla_{X}Z).\end{aligned}$$
The [*equation of Codazzi*]{} is $$\begin{aligned}
\label{2.8}({\bar\nabla}_{X}h)(Y,Z)= ({\bar\nabla}_{Y}h)(X,Z).\end{aligned}$$
It follows from the definition of a rectifying curve $x:I\to \mathbb E^3$ that the position vector field $\x$ of $x$ satisfies $$\label{2.9} {\bf x}(s)=\lambda(s){\bf t}(s)+\mu(s){\bf b}(s)$$ for some functions $\lambda$ and $\mu$.
For a curve $x:I\to \mathbb E^3$ with $\kappa(s_0)\ne 0$ at $s_0\in I$, the first normal space at $s_0$ is the line spanned by the principal normal vector ${\bf n}(s_0)$. Hence, the rectifying plane at $s_0$ is nothing but the plane orthogonal to the first normal space at $s_0$. Therefore, for a submanifold $M$ of $\mathbb E^m$ and a point $p\in M$, we call the subspace of $T_p\mathbb E^m$, orthogonal complement to the first normal space ${\rm Im}\,\sigma_p$, the [*rectifying space of*]{} $M$ at $p$.
\[D:2.1\] [A submanifold $M$ of a Euclidean $m$-space $\mathbb E^m$ is called a]{} [*rectifying submanifold*]{} [if the position vector field $\x$ of $M$ always lies in its rectifying space. In other words, $M$ is called a rectifying submanifold if and only if $$\label{2.10} \<\x(p),{\rm Im}\, h_p\>=0$$ holds at every $p\in M$.]{}
[A non-trivial vector field $Z$ on a Riemannian manifold $M$ is called a [*concurrent vector field*]{} if it satisfies $$\begin{aligned}
\label{2.11}\nabla_X Z=X\end{aligned}$$ for any vector $X$ tangent to $M$, where $\nabla$ is the Levi-Civita connection of $M$.]{}
Lemmas
======
By a [*cone*]{} in $\mathbb E^m$ with vertex at the origin we mean a ruled submanifold generated by a family of lines passing through the origin. A submanifold of $\mathbb E^m$ is called a [*conic submanifold*]{} with vertex at the origin if it is an open portion of a cone with vertex at the origin.
There exists a natural orthogonal decomposition of the position vector field $\x$ at each point for a Euclidean submanifold $M$; namely, $$\x=\x^T+\x^N,$$ where $\x^T$ and $\x^N$ denote the tangential and normal components of $\x$, respectively. Let $|\x^T|$ and $|\x^N|$ be the length of $\x^T$ and $\x^N$, respectively.
\[L:3.1\] Let $\,x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold into the Euclidean $m$-space ${\mathbb E}^{m}$. Then $\x=\x^T$ holds identically if and only if $M$ is a conic submanifold with the vertex at the origin.
Let $\,x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold into the Euclidean $m$-space ${\mathbb E}^{m}$. If $\x=\x^T$ holds identically, then $e_1=\x/|\x|$ is a unit vector field tangent to $M$.
Put $\x=\rho e_1$. Since $\tilde\nabla_{e_1}e_1$ is perpendicular to $e_1$, we find from $$\tilde\nabla_{e_1}\x=e_1,\;\; \tilde\nabla_{e_1}\x=(e_1\rho)e_1+\rho
\tilde\nabla_{e_1}e_1,$$ that $\tilde\nabla_{e_1}e_1=0$. Therefore, the integral curves of $e_1$ are some open portions of generating lines in $\mathbb E^m$. Moreover, because $\x=\x^T$, the generating lines given by the integral curves of $e_1$ pass through the origin. Consequently, $M$ is a conic submanifold with the vertex at the origin.
The converse is clear.
\[L:3.2\] Let $\,x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold into the Euclidean $m$-space ${\mathbb E}^{m}$. Then $\x=\x^N$ holds identically if and only if $M$ lies in a hypersphere centered at the origin.
Let $\,x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold into the Euclidean $m$-space ${\mathbb E}^{m}$. If $\x=\x^N$ holds identically, then we get $$Z\! \<\x,\x\>=2\<\right.\! \tilde\nabla_Z\x,\x\left.\!\>=2\<Z,\x^N\>=0$$ for any $Z\in TM$. Thus $M$ lies in a hypersphere centered at the origin.
The converse is obvious.
In views of Lemma \[L:3.1\] and Lemma \[L:3.2\] we make the following.
[A rectifying submanifold $M$ of $\mathbb E^m$ is called ]{} [*proper*]{} [if its position vector field $\x$ satisfies $\x\ne \x^T$ and $\x\ne \x^N$ at every point on $M$.]{}
\[L:3.3\] Let $M$ be a proper rectifying submanifold of $\mathbb E^m$ with $\dim M=n$. Then we have $$\begin{aligned}
\label{3.5} m> n+\dim\, ({\rm Im}\,h_p)\end{aligned}$$ for each $p\in M$.
Let $M$ be a proper rectifying submanifold of $\mathbb E^m$. If $m=n+\dim\, ({\rm Im}\,h_p)$, then we get $\x=\x^T$ which is a contradiction.
In views of Lemma \[L:3.1\] and Lemma \[L:3.2\], we are only interested on proper rectifying submanifolds.
Characterization and classification of rectifying submanifolds
==============================================================
First, we give the following simple characterization of rectifying submanifolds.
\[T:4.1\] If the position vector field $\x$ of a submanifold $M$ in ${\mathbb E}^{m}$ satisfies $\x^N\ne 0$, then $M$ is a proper rectifying submanifold if and only if $\x^T$ is concurrent vector field on $M$.
Let $x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold into the Euclidean $m$-space ${\mathbb E}^{m}$. Consider the orthogonal decomposition $$\begin{aligned}
\label{4.1} \x=\x^T +\x^N\end{aligned}$$ of the position vector field $\x$ of $M$ in $\mathbb E^m$.
From and formulas of Gauss and Weingarten, we find $$\begin{aligned}
\label{4.2} Z=\tilde\nabla_Z \x=\nabla_Z \x^T +h(Z,\x^T)-A_{\x^N}Z +D_Z\x^N\end{aligned}$$ for any $Z\in TM$. After comparing the tangential components in , we obtain $$\begin{aligned}
\label{4.3} A_{\x^N}Z=\nabla_Z \x^T - Z.\end{aligned}$$
Assume that $M$ is a proper rectifying submanifold. Then we have $\x^T\ne 0$ and $\x^N\ne 0$ . Moreover, it follows from the Definition \[D:2.1\] that $$\begin{aligned}
\label{4.4} \<\x, h(X,Y)\>=0\end{aligned}$$ for $X,Y\in TM$. So we get $A_{\x^N}=0$. Hence, we obtain from that $$\begin{aligned}
\label{4.5} \nabla_Z \x^T = Z,\end{aligned}$$ which shows that $\x^T$ is a concurrent vector field on $M$.
Conversely, if $\x^T$ is a concurrent vector field on $M$, then we find from and that $A_{\x^N}=0$. Therefore we obtain . Consequently, $M$ is a proper rectifying submanifold due to $\x^N\ne 0$ by assumption.
Next, we give the following classification of rectifying submanifolds.
\[T:4.2\] If $M$ is a proper rectifying submanifold of ${\mathbb E}^{m}$, then with respect to some suitable local coordinate systems $\{s,u_2,\ldots,u_n\}$ on $M$ the immersion $\,x$ of $M$ in $\mathbb E^m$ is of the form $$\label{4.6} x(s,u_2,\ldots,u_n)=\sqrt{s^2+c^2}\,
Y(s,u_2,\ldots,u_n),\;\; \<Y,Y\>=1,\; c>0,$$ such that the metric tensor $g_Y\! $ of the spherical submanifold defined by $Y$ satisfies $$\begin{aligned}
\label{4.7} g_Y=\frac{c^2}{(s^2+c^2)^2}ds^2+\frac{s^2}{s^2+c^2}\sum_{i,j=2}^n g_{ij}(u_2,\ldots,u_n) du_i du_j.\end{aligned}$$
Conversely, the immersion given by - defines a proper rectifying submanifold.
Let $x:M\to {\mathbb E}^{m}$ be an isometric immersion of a Riemannian $n$-manifold $M$ into the Euclidean $m$-space ${\mathbb E}^{m}$. Assume that $M$ is a proper rectifying submanifold. Then holds.
After comparing the normal components of , we obtain $$\begin{aligned}
\label{4.8} &D_Z\x^N=-h(Z,\x^T),\end{aligned}$$ for $Z\in TM$.
It follows from and that $\<\x,D_Z \x^N\>=0$. Hence we get $$Z\! \<\x^N,\x^N\>=0,$$ which implies that $\x^N$ is of positive constant length, say $c$. From we obtain $$\label{4.11} \<A_{\x^N} X,Y\>=\<\x^N,h(X,Y)\>=\<\x,h(X,Y)\>=0.$$ Hence we have $A_{\x^N}=0$. Let us put $\rho=|\x^T|$ and $e_1=\x^T/\rho$. We may extend $e_1$ to a local orthonormal frame $e_1,\ldots,e_n$.
We put $$\label{4.9} \nabla_X e_i=\sum_{j=1}^n \omega_i^j(X)e_j,\;\; i=1,\ldots,n.$$ For $j,k=2,\ldots,n$, we find $$\label{4.10} 0=e_k\! \<\x,e_j\>=\delta_{jk}+ \<\x,\nabla_{e_k}e_j\>+\<\x,h(e_j,e_k)\>,$$ Since $h(e_j,e_k)=h(e_k,e_j)$, equation gives $$\omega^1_j(e_k)=\omega^1_k(e_j),\;\; j,k=2,\ldots,n.$$ Hence, it follows from the Frobenius theorem that the distribution $\mathcal D$ spanned by $e_2,\ldots,e_n$ is an integrable distribution.
On the other hand, the distribution $\mathcal D^\perp={\rm Span}\,\{e_1\}$ is also integrable since it is of rank one. Therefore, there exist local coordinate systems $\{s,u_2,\ldots,u_n\}$ on $M$ such that $e_1=\partial/\partial s$ and $\partial/\partial u_2,\ldots,\partial/\partial u_n$ span the distribution $\mathcal D$.
Let us put $$\label{4.12}\x^T=\varphi e_1$$ with $\varphi=|\x^T|$. By taking the derivative of $\varphi=\<\x,e_1\>$ with respect to $e_j$ for $j=1,\ldots,n$, we also have $$\label{4.13}e_j\varphi=\delta_{1j}+\<\x,h(e_1,e_j)\>.$$ Combining and gives $$\label{4.14}e_j\varphi=\delta_{1j},\quad j=1,\ldots,n.$$ Therefore, we obtain $\varphi=\varphi(s)$ and $\varphi'(s)=1$ which imply $\varphi(s)=s+b$ for some constant $b$. Thus, after applying a suitable translation on $s$ if necessary, we have $\varphi=s$. Consequently, the position vector field satisfies $$\label{4.15}\x=se_1+\x^N.$$ By combining and $|\x^N|=c$, we find $$\label{4.16}\<\x,\x\>= s^2+c^2,$$ where $c$ is a positive number. Hence we may put $$\label{4.17} x(s,u_2,\ldots,u_n)=\sqrt{s^2+c^2}\,
Y(s,u_2,\ldots,u_n),$$ for some $\mathbb E^m$-valued function $Y=Y(s,u_2,\ldots,u_n)$ satisfying $\<Y,Y\>=1$.
Using and the fact that $e_1=\partial/\partial s$ is orthogonal to the distribution $\mathcal D$, we obtain that $$\label{4.18} \<Y_s,Y_s\>={c^2\over {(s^2+c^2)^2}},\;\; \<Y_s,Y_{u_j}\>=0, \;\; j=2,\ldots,n.$$ Therefore, the metric tensor $g_Y$ of the spherical submanifold defined by $Y$ takes the following form: $$\begin{aligned}
\label{4.19} g_Y=\frac{c^2}{(s^2+c^2)^2}ds^2+\sum_{i,j=2}^n g_{ij}(s,u_2,\ldots,u_n) du_i du_j.\end{aligned}$$
On the other hand, it follows from Theorem \[T:4.1\] that $\x^T=se_1$ is a concurrent vector field. Thus, we find from that $$\begin{aligned}\label{4.20} e_1= \nabla_{e_1}\x^T=\nabla_{e_1} se_1=e_1+s\nabla_{e_1}e_1.\end{aligned}$$ Hence we get $\nabla_{e_1}e_1=0$, which implies that the integral curves of $e_1$ are geodesic in $M$. Therefore, the distribution $\mathcal D^\perp$ spanned by $e_1$ is a totally geodesic foliation.
From we have $$e_i= \nabla_{e_i} \x^T=s\nabla_{e_i}e_1, \;\; i=2,\ldots,n,$$ which implies that $$\label{4.22}\omega^j_1(e_i)=\frac{\delta_{ij}}{s}, \;\; i,j=2,\ldots,n,$$ where $\delta_{ij}=1$ or 0 depending on $i=j$ or $i\ne j$.
From we conclude that $\mathcal D$ is an integrable distribution whose leaves are totally umbilical in $M$. Moreover, the mean curvature of leaves of $\mathcal D$ are given by $s^{-1}$. Since the leaves of $\mathcal D$ are hypersurfaces in $M$, it follows that the mean curvature vector fields of leaves of $\mathcal D_2$ are parallel in the normal bundle of $M$ in $\mathbb E^m$. Therefore, $\mathcal D$ is a spherical foliation. Consequently, by a result of [@H] (or Theorem 4.4 of [@book11 page 90]) we conclude that $M$ is locally a warped product $I\times_{s} F$, where $F$ is a Riemannian $(n-1)$-manifold. Thus, the metric tensor $g$ of $M$ takes the form $$\label{4.23}g=ds^2+s^2 g_F,$$ where $g_F$ is the metric tensor of $F$. Now, by applying , and , we may conclude that the metric tensor $g_Y$ can be expressed as .
Conversely, let us consider a submanifold $M$ of $\mathbb E^m$ defined by $$\label{4.24} x(s,u_2,\ldots,u_n)=\sqrt{s^2+c^2}\,
Y(s,u_2,\ldots,u_n),\;\; \<Y,Y\>=1,\; c>0,$$ such that the metric tensor $g_Y\! $ satisfies $$\begin{aligned}
\label{4.25} g_Y=\frac{c^2}{(s^2+c^2)^2}ds^2+\frac{s^2}{s^2+c^2}\sum_{i,j=2}^n g_{ij}(u_2,\ldots,u_n) du_i du_j.\end{aligned}$$ Then it follows from that $$\begin{aligned}\label{4.26} &\frac{\partial \x}{\partial s}=\frac{sY}{\sqrt{s^2+c^2}}+\sqrt{s^2+c^2}\, Y_s,\;\;
\\&\frac{\partial \x}{\partial u_j}=\sqrt{s^2+c^2}\,
Y_{u_j},\;\; j=2,\ldots,n, \end{aligned}$$ where $Y_s=\partial Y/\partial s$ and $Y_{u_j}=\partial Y/\partial u_j$. It follows from , and that the metric tensor $g_M$ of $M$ is given by $$\label{4.27}g_M=ds^2+s^2 \sum_{i,j=2}^n g_{ij}(u_2,\ldots,u_n) du_i du_j.$$
Now, by an easy computation, we find from that $$\label{4.28}\nabla_{\frac{\partial}{\partial s}}\frac{\partial}{\partial s}=0, \;\; \nabla_{\frac{\partial}{\partial u_j}} \frac{\partial}{\partial s}=\frac{1}{s}\frac{\partial}{\partial u_j},\;\; j=2,\ldots,n.$$
Since $\<Y,Y\>=1$, and imply that $$\label{4.29} \<\right.\! \x, \x_{u_j}\! \left.\>=0, \;\; j=2,\ldots,n.$$ Therefore, we obtain $\x^T=s \frac{\partial}{\partial s}$. Now, it is easy to verify that $\x^T$ is a concurrent vector field on $M$. Moreover, it is direct to show that the normal component of $\x$ is given by $$\x^N=\frac{c^2}{\sqrt{s^2+c^2}}\,Y-s\sqrt{s^2+c^2}\, Y_s,$$ which is alway non-zero everywhere on $M$. Consequently, $M$ is a proper rectifying submanifold, according to Theorem \[T:4.1\].
Theorem \[T:4.2\] extend Theorem 3 of [@c6].
If we put $s=\tan^{-1}\(\frac{t}{c}\)$, then becomes $$\begin{aligned}
\label{4.30} g_Y=dt^2+\sin^2t \sum_{j,k=2}^n g_{jk}(u_2,\ldots,u_n)du_j du_k.\end{aligned}$$ For $n=2$, we get $g_Y=dt^2+(\sin^2t) du^2$ from , which is the metric tensor of a spherical coordinate system $(t,u)$ on $S^2(1)$. Hence, for $n=2$, $Y=Y(t,u)$ is nothing but an isometric immersion from an open portion of $S^1(1)$ into $S^{m-1}(1)\subset \mathbb E^m$. Therefore, there exist many spherical submanifolds in $\mathbb E^m$ whose metric tensor is given by . Consequently, there exist many rectifying submanifolds in $\mathbb E^m$ according to Theorem \[T:4.2\].
Some properties of rectifying submanifolds
==========================================
Finally, we provide some basic properties of proper rectifying submanifolds.
\[T:5.1\] Let $M$ be a proper rectifying submanifold of ${\mathbb E}^{m}$. Then
- $|\x^T|=s+b$ for some constant $b$.
- $|\x|^2=s^2+c_1 s+c_2$ for some constants $c_1$ and $c_2$.
- $\x^N$ is of constant length.
- $A_{\x^N}=0$.
- The curvature tensor $R$ satisfies $R(\x^T,Y)=0$ for any $Y \in TM$.
- The sectional curvature $K$ of $M$ satisfies $K(\x^T,Z)=0$ for any unit vector $Z$ perpendicular to $\x^T$.
Statements (a), (b), (c) and (d) are already done in the proof of Theorem \[T:4.2\]. Clearly, statement (f) follows immediately from statement (e).
Now, we prove statement (e). This can be done as follows. By applying and we have $$\begin{aligned}\label{5.1} R(\x^T,Y,Z;W)\, &=\<h(\x^T,W),h(Y,Z)\>-\<h(\x^T,Z),h(Y,W)\>
\\&=\<D_Z\x^N,h(Y,W)\>-\<D_W \x^N,h(Y,Z)\>
\\&=-\<\x^N,D_Z h(Y,W)\>+\<\x^N,D_W h(Y,Z)\>.\end{aligned}$$ Therefore, after applying and equation of Codazzi, we derive from that $$\begin{aligned} R(\x^T,Y,Z;W)\, &=\<\x^N,(\bar\nabla_W h)(Y,Z)\>-\<\x^N,(\bar\nabla_Z h)(Y,W)\>=0,\end{aligned}$$ which gives statement (e).
Statement (a), (b) and (c) of Theorem \[T:5.1\] extend the corresponding results obtained in Theorem 1 of [@c6].
One may define rectifying submanifolds in a pseudo-Euclidean space in the same as Definition \[D:2.1\]. We will treat rectifying submanifolds in pseudo-Euclidean spaces in a separate article.
[20]{}
Cambie, S., Goemans, W. and Van den Bussche, I., Rectifying curves in the $n$-dimensional Euclidean space. Turkish J. Math. [**40**]{} (2016), no. 1, 210–223.
Chen, B.-Y., Geometry of Submanifolds. Marcel Dekker, New York, 1973.
Chen, B.-Y., Constant-ratio hypersurfaces. Soochow J. Math. [**21**]{} (2001), 353–361.
Chen, B.-Y., Geometry of position functions of Riemannian submanifolds in pseudo-Euclidean space. J. Geom. [**74**]{} (2002), 61–77.
Chen, B.-Y., Convolution of Riemannian manifolds and its applications. Bull. Austral. Math. Soc. [**66**]{} (2002), no. 2, 177–191.
Chen, B.-Y., When does the position vector of a space curve always lie in its rectifying plane?. Amer. Math. Monthly [**110**]{} (2003), no. 2, 147–152.
Chen, B.-Y., More on convolution of Riemannian manifolds. Beiträge Algebra Geom. [**44**]{} (2003), 9–24.
Chen, B.-Y., Constant-ratio space-like submanifolds in pseudo-Euclidean space. Houston J. Math. [**29**]{} (2003), no. 2, 281–294
Chen, B.-Y., [Pseudo-Riemannian geometry, $\delta$-invariants and applications]{}. World Scientific, 2011.
Chen, B.-Y., Topics in differential geometry associated with position vector fields on Euclidean submanifolds, (submitted for publication).
Chen, B.-Y. and Dillen, F., Rectifying curves as centrodes and extremal curves. Bull. Inst. Math. Acad. Sinica [**33**]{} (2005), no. 2, 77–90.
Gungor, M. A. and Tosun, M., Some characterizations of quaternionic rectifying curves. Differ. Geom. Dyn. Syst. [**13**]{} (2011), 89–100.
S. Hiepko, [*Eine innere Kennzeichnung der verzerrten Produkte*]{}, [Math. Ann.]{} [**241**]{} (1979), no. 3, 209–215.
Ilarslan, K., Nesovic, E. and Petrovic-Torga[s]{}ev, M., Miroslava Some characterizations of rectifying curves in the Minkowski 3-space. Novi Sad J. Math. [**33**]{} (2003), no. 2, 23–32.
Ilarslan, K. and Nesovic, E., On rectifying curves as centrodes and extremal curves in the Minkowski 3-space. Novi Sad J. Math. [**37**]{} (2007), no. 1, 53–64.
Ilarslan, K. and Nesovic, E., Some characterizations of rectifying curves in the Euclidean space $\mathbb E^4$. Turkish J. Math. [**32**]{} (2008), no. 1, 21–30.
Ilarslan, K. and Nesovic, E., Some relations between normal and rectifying curves in Minkowski space-time. Int. Electron. J. Geom. [**7**]{} (2014), no. 1, 26–35.
Lucas, P. and Ortega-Yagues, J. A., Rectifying curves in the three-dimensional sphere. J. Math. Anal. Appl. [**421**]{} (2015), no. 2, 1855–1868.
Ozbey, and Oral, M., A study on rectifying curves in the dual Lorentzian space. Bull. Korean Math. Soc. [**46**]{} (2009), no. 5, 967–978.
Yilmaz, B., Gok, I. and Yayli, Y., Extended rectifying curves in Minkowski 3-space. Adv. Appl. Clifford Algebr. [**26**]{} (2016), no. 2, 861–872.
Yücesan, A., Ayyildiz, N. and Coken, A. C., On rectifying dual space curves. Rev. Mat. Complut. [**20**]{} (2007), no. 2, 497–506.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we study the selectivity of the potassium channel KcsA by a recently developed image-charge solvation method(ICSM) combined with molecular dynamics simulations. The hybrid solvation model in the ICSM is able to demonstrate atomistically the function of the selectivity filter of the KcsA channel when potassium and sodium ions are considered and their distributions inside the filter are simulated. Our study also shows that the reaction field effect, explicitly accounted for through image charge approximation in the ICSM model, is necessary in reproducing the correct selectivity property of the potassium channels.'
address: ' Department of Mathematics and Statistics, University of North Carolina at Charlotte, Charlotte, NC 28223, USA'
author:
- 'Katherine Baker, Duan Chen, and Wei Cai'
bibliography:
- 'kcsaICSMkab.bib'
title: 'Investigating the Selectivity of KcsA Channel by an Image Charge Solvation Method (ICSM) in Molecular Dynamics Simulations'
---
Introduction
============
Ion channels are membrane-spanning proteins that form a pathway for the movement of ions through the cell membrane and they play significant roles in a wide variety of biological processes. Some examples of their many functions include the control of secretion of hormones into the bloodstream, generating electrical impulses that establish information transfer in the nervous system, and controlling the pace of the heart and other muscles [@UofIwebsite]. The idea for assuming the existence of a means for transporting ions from the exterior of a cell to the interior was proposed 63 years ago with Hodgkin and Huxleyś study of the electrical activity in squid giant axon [@Jordan-2005; @Hodgkin-Huxley-1952]. They showed that both sodium and potassium ions contributed to the ionic current and that their fluxes were in opposite directions. Twenty years later Hladky and Haydon used small antibiotic gramicidins to actually prove the existence of an ionic pathway [@Jordan-2005; @Hladky-Haydon-1972]. Properties of ion channels and their functions in manipulating electric currents by conducting different ionic species heavily depend on their molecular structure in the presence of a complicated surrounding solvent environment. In past decades, great technical strides in many diverse areas of science culminated in the completion of x-ray crystal structures of ion channels. Meanwhile, for theoretical studies, a hierarchy of multi-scale mathematical models, from molecular dynamics [@AMBER; @CHARMM22], Brownian dynamics [@Chung:2002], and Poisson-Nernst-Planck (PNP) theories [@Chen:1997; @Coalson:2005], were developed to study functions of ion channels.
Potassium channels are specialized proteins able to facilitate and regulate the conduction of ions, $\mathrm{K^{+}}$ ions in particular, through cell membranes [@UofIwebsite; @Egwolf-Roux-2010]. In 1998, MacKinnon et al.[@Jordan-2005; @Hodgkin-Huxley-1952] successfully obtained the crystal structure of KcsA (potassium crystallographically sited activation) channel at a resolution of 2.0 Å, allowing a direct laboratory observation of the selectivity filter and binding sites of $\mathrm{K^{+}}$ ions. KcsA is comprised of around 560 residues (Fig. \[figallatom\]) which form four identical subunits (Fig. \[figproteins\]), each containing two alpha-helices connected by a loop of approximately 30 amino acids. These proteins combine to form three primary sections of the channel: the opening pore on the cytoplasmic side of the cell interior, a small cavity ([of 5Å in radius]{}) filled with water and a mix of sodium ($\mathrm{Na^{+}}$) and potassium ($\mathrm{K^{+}}$) ions, and the selectivity filter. The selectivity filter, [could be as narrow as 2Å in radius]{}, comprised of four specific cation binding sites and formed by the backbone carbonyl groups of conserved residues threonine (T), valine (V), glycine (G), and tyrosine (Y), allows fast conduction of $\mathrm{K^{+}}$ while being highly selective for potassium ions over sodium ions (Fig. \[figaminos\]) [@Egwolf-Roux-2010].
![An all-atom model of KcsA from a) Side view and b) extracellular end view.[]{data-label="figallatom"}](allatomboth.pdf)
![View of the four identical subunits of the channel[]{data-label="figproteins"}](proteinboth.pdf)
![The amino acids that make up the selection filter: blue – threonine(T), red – valine(V), yellow – glycine(G), green – tyrosine(Y).[]{data-label="figaminos"}](aminoboth.pdf)
The relatively complete functioning components (gating, selectivity, conductance) and available high resolution structure of KcsA channel makes it the most interesting case attracting investigation in biological studies and mathematical modeling. However, modeling the selectivity of KcsA channel is extremely challenging due to the complicated ion-water-protein interactions. Counterintuitively, the $\mathrm{Na^{+}}$ ion has the same ionic valence as the $\mathrm{K^{+}}$ ion does but with a smaller ionic radius, nevertheless it is the one that is rejected by the narrow selectivity filter. In [@Doyle-1998], it was suggested that the small diameter of the selectivity filter required dehydration of the cations entering the filter. To compensate for the cost of the dehydration, the carbonyl oxygen atoms from the amino acids in the filter take the place of the water oxygen atoms. The relative rigidity of the filter precludes this action in the case of the $\mathrm{Na^{+}}$ ions with a smaller radius but stronger binding of water shell. However, a later study suggests that the selectivity can be explained by the fact that the smaller $\mathrm{Na^{+}}$ does not bind to the $\mathrm{K^{+}}$ sites in a thermodynamically favorable way [@Thompson2009].
To understand the selectivity mechanism of the filter, molecular dynamics (MD), being an explicit atomistic model, is a naturally suitable method to model these characteristics but the computation remains very expensive even with current computer powers due to the large number of degree of freedom and the necessarily small time step ($10^{-15}$ seconds) versus the ion permeation time scale ($10^{-6}$ seconds) [@Chung:2002]. The mean-field theory, such as the PNP model, has successfully enhanced the model efficiency and reproduced the channel conductance with carefully calibrated parameters. Further, the PNP theory was strengthened recently to adopt the capacity of modeling selectivity, by taking into account the ion-size effect and nonlocal dielectric property of the solvent [@Burger:2007; @Liu:2011; @Zhou:2011; @xied:2013; @lub:2014; @Gillespie:2003]. Therefore, a highly efficient model with the ability to retain the ion-water-protein interactions is indispensable to investigate the properties, especially the selectivity, of the KcsA channel.
In this paper, we will conduct a study of the selectivity filter in the KcsA channel with our recently developed Image Charge Solvation Method (ICSM), implemented in the open-source Tinker MD package [@tinker]. The ICSM is a hybrid explicit/implicit solvation model developed to accurately account for the reaction field of the solute-solvent environment. In contrast to the traditional full MD simulations [@Biggin-2001; @Shirivastava02; @Berneche00], most of the solvent in the system outside a designated spherical region is modeled as a dielectric continuum, while only a limited number of particles (protein, water molecules, and ions) inside the sphere are given an atomistic description. The reaction field effect on the permanent charges of the protein and mobile ions, due to the solvent/membrane surroundings, are accounted for by a multiple image charge approximation [@cai-dieSphImg-2007; @deng-ionic-CCP2007; @deng-ionic-JCP2007]. Therefore, in the ICSM method, the treatment of electrostatic interaction does not use the periodic condition as in the EWALD approach where infinitely many artificial periodic copies of a simulated system, the ion-channel filter in this case, will occur. The effect of the dielectric exterior to the simulation sphere is accounted for easily by the image-charge method. Moreover, there is no requirement of system neutrality in the hybrid ICSM model, not like EWALD-sum based MD simulations. The efficiency, robustness, and capability of the ICSM have been tested for homogeneous water system and solvation of ions [@Lin-2009].
Here in this paper, we will use this hybrid solvation model to investigate the positioning of sodium ions and potassium ions inside the selectivity filter of the KcsA channel, to evaluate its likelihood of conducting the ions, and thus the selective functions of this potassium channel. In order to balance the accuracy and efficiency, cell membrane and intra/extra cellular solvents are modeled as a dielectric continuum, most of the permanent charges on the channel proteins are assumed as rigid, while the molecular dynamics is applied to the selectivity filter, as well as ions and waters in the cavity chamber. First of all, the fundamental physical properties of the KcsA channel, the electrostatic landscape is obtained by the ICSM and compared to existing results in literatures; four binding sites for $\mathrm{K^{+}}$ are identified. Then, different combinations of $\mathrm{Na^{+}}$ and $\mathrm{K^{+}}$ in the selectivity filter and cavity chamber are tested to show that the movement of the ion in the cavity in fact depends on which ions are in the selectivity filter, which will in turn determine what ions will eventually be transported through the whole channel. Along with these investigations, the inclusion of the reaction field in the hybrid ICSM will also be examined and will be shown to be indispensable for accurate portrayal of the ion conduction process.
The rest of paper is organized as follows: Section \[sec:icsm\] is a review of the ICSM. Simulation results of the selectivity filter of the KcsA are given in Section \[sec:results\], and the paper ends with concluding remarks in Section \[sec:conclusion\].
ICSM - Image Charge Solvation Method {#sec:icsm}
====================================
The original setup of the ICSM is shown in a schematic of Figure \[figicsmsystem\]. A regular truncated octahedron (TO) with a size $R_{c}$ (the distance between its center and farthest corner) is employed as the main simulation box. Inside a spherical domain (dashed circle with radius $a$) labeled as region I, the solute molecule is placed in the central area of the TO. Region II is the remainder of the TO box with Region I excluded, and this region contains the solvent. Both the solute and solvent in regions I and II are described explicitly. In order to reduce the possible surface effects produced by the boundary of regions of atomistic and continuum descriptions, a buffer layer of thickness $\tau$ is constructed by adding a larger spherical domain containing the TO box. The area outside the TO box but inside the larger sphere is denoted as region III and it is filled with periodic images of the solvent molecules from Region II. Finally, the remaining solvent outside the larger spherical cavity of radius $R_{c}+\tau$ is modeled implicitly as a dielectric continuum, whose reaction field effect on any physical charge inside the larger sphere is approximated by multiple image charges (located outside the larger sphere) and the later will be included in the calculation of electrostatic interactions with the physical charges inside the spherical cavity of radius $R_{c}+\tau$.
The multiple image charge approximation of the reaction field inside an dielectric sphere extends the Kelvin image concept [@Jackson-1999] for a conducting sphere. To simulate the solvation of a protein inside a dielectric sphere containing water molecules, both the source charges and the field points are inside the sphere. For each source charge, in addition to an image point charge at the Kelvin image inversion point, there will be other image charges distributed along a ray starting from the inversion point [@cai-dieSphImg-2007] and a short summary of the multiple image charge approximation of the reaction field is given below.
![A schematic of the ICSM system. Region I, indicated by a broken circular line, contains the solute which can have a maximum diameter of $d=(\sqrt{3}-\sqrt{5}/2)L-2\tau$. Particles in this area are not imaged. Region II contains particles which have periodic images in Region III. Region III contains the nearest periodic images of the particles in Region II. The polarization of the solvent by the source charge $q$ at position $\mathbf{r}_{s}$ results in the reaction field $\Phi_{RF}(r,\theta)$ that is approximated by the potential created by the image charges, $q_{k}^{\prime}$ and $q_{i}^{\prime},i\geq2$, located at positions $r_{k},r_{i},i\geq2$.[]{data-label="figicsmsystem"}](icsmsystem.pdf)
Given a local volume $\Gamma$ of a spherical shape of radius $R,R=R_{c}+\tau$ in our case, and dielectric permittivity $\varepsilon_{\mathrm{i}},$ embedded in an infinite solvent of dielectric permittivity $\varepsilon_{\mathrm{o}}$, the total electrostatic potential $\Phi(\mathbf{r})$ satisfies the Poisson and Poisson-Boltzmann equation:
$$\begin{aligned}
\varepsilon_{\mathrm{i}}\Delta\Phi(\mathbf{r})=-\rho_{in}(\mathbf{r}),\quad &
\mathrm{if\ \ }\mathbf{r}\in V_{\mathrm{in}},\label{poissona}\\
(\Delta-\lambda^{2})\Phi(\mathbf{r})=0,\quad & \mathrm{if\ \ }\mathbf{r}\in
V_{\mathrm{out}},\label{poissonb}$$
where the charge distribution inside the solute domain $V_{\mathrm{in}}$, $\rho_{\mathrm{in}}(\mathbf{r})=\sum_{i}q_{i}\delta(\mathbf{r}-\mathbf{r}_{i})$ contains all explicit charges of the solute and solvent molecules. And for the implicit region $V_{\mathrm{out}}$, $\lambda$ is the inverse Debye-Hückel screening length [@Lin-2009] and $\lambda=0$ if the medium is considered as simple dielectric with no mobile ion density as in this work. Here $\delta(\cdot)$ denotes the Dirac delta function. The two equations are accompanied with the following interface and boundary conditions: $$\begin{aligned} \Phi(\mathbf{r}^+)& =\Phi(\mathbf{r}^-)\\ \epsilon_i\nabla \Phi\cdot \vec{n}|_{\mathbf{r}^+}& =\epsilon_o\nabla \Phi\cdot \vec{n}|_{\mathbf{r}^-}\\ \Phi(\infty)& =0, \end{aligned}$$ where $\mathbf{r^{+}}$ and $\mathbf{r^{+}}$ represent the limiting value of $\mathbf{r}$ when approaching to the interface $\Gamma$ from two sides, and $\vec{n}$ is the unit normal direction vector.
Due to the principle of superposition, we only need to consider the case of one single source charge in $V_{\mathrm{in}}$, i.e. $\rho_{\mathrm{in}}(\mathbf{r})=q\delta(\mathbf{r}-\mathbf{r}_{s})$, and the total potential can be written as $\Phi=\Phi_{\mathrm{S}}+\Phi_{\mathrm{RF}}$ where $\Phi
_{\mathrm{RF}}$ is the primary field that results from the source charge $q$ at $r_{s}$ and $\Phi_{\mathrm{RF}}$ is the reaction field from the exterior dielectric medium with dielectric constant $\varepsilon_{\mathrm{o}}$ and the inverse Debye-Hückel screening length $\lambda=0$, respectively [@Baker-2013; @Lin-2009]. The reaction field can be approximated by a set of discrete image charges based on Gauss-Radau quadratures as [@cai-dieSphImg-2007]:
$$\Phi_{\mathrm{RF}}(\mathbf{r})\approx\frac{{q}_{\mathrm{k}}^{\prime}}{4\pi\varepsilon_{\mathrm{i}}|\mathbf{r}-\mathbf{r}_{\mathrm{k}}|}+\sum\limits_{m=2}^{N_{i}}\frac{q_{m}^{\prime}}{4\pi\varepsilon_{\mathrm{i}}|\mathbf{r}-\mathbf{r}_{m}^{\prime}|}, \label{phi}$$
where the subscript $\mathrm{k}$ is for the Kelvin image and $m\geq2$ represents the remainder of the discrete image charges, ${q}_{\mathrm{k}}^{\prime}={q}_{\mathrm{k}}+q_{1}^{\prime}=(1+\omega_{1}\varepsilon
_{\mathrm{i}}/2\varepsilon_{\mathrm{o}}){q}_{\mathrm{k}}$, ${q}_{\mathrm{k}}=\gamma\frac{R}{r_{s}}q,\gamma=\frac{\varepsilon_{\mathrm{i}}-\varepsilon
_{\mathrm{o}}}{\varepsilon_{\mathrm{i}}+\varepsilon_{\mathrm{o}}}$ and for all $m\geq1$ $$\ q_{m}^{\prime}=\frac{\varepsilon_{\mathrm{i}}(\varepsilon_{\mathrm{i}}-\varepsilon_{\mathrm{o}})}{2\varepsilon_{\mathrm{o}}(\varepsilon
_{\mathrm{i}}+\varepsilon_{\mathrm{o}})}\frac{\omega_{m}r_{m}}{R}q,\quad
r_{m}=r_{\mathrm{k}}\left( \frac{2}{1-s_{m}}\right) ^{1+\varepsilon
_{\mathrm{i}}/\varepsilon_{\mathrm{o}}}.$$ Here $\{s_{m},\omega_{m}\}_{m=1}^{N_{i}}$ represent the points and the weights of the Gauss-Radau quadrature [@gautschi-ORTHPOL-1994]. Image approximation can also be derived for ionic solvent media [@deng-ionic-JCP2007] [@Xucai:2009][@Xuz:2015].
Simulation Results {#sec:results}
==================
In the present study, we use the KcsA channel with the PDB ID 2A9H [@Yu-2005]. The data from the PDB was converted to a Tinker input file using the built-in program pdbxyz.x, with the associated toxin removed and five water molecules added to the central cavity.
Simulation setup for the selectivity filter
-------------------------------------------
The ICSM is modified to simulate the selectivity filter inside the KcsA channel. The channel pore of the KcsA, including the cavity, selectivity filter, and residential ions are of greatest interest, so they are placed inside Region I in the ICSM schematic of Fig. 4, along with a small portion of the protein structure (see Fig.\[figsyswithprot\]). Meanwhile, most of the channel protein, some of the cell membrane and surrounding solvent are located in Region II. The dielectric permittivity inside the TO box is denoted as $\varepsilon_{\mathrm{i}}$ while $\varepsilon_{\mathrm{o}}$ is the dielectric constant outside the larger sphere of radius $R_{c}+\tau$. As the difference between the two dielectric constants (and therefore the surface effect) is small, the buffer layer thickness $\tau$ is set to be zero.
Figure \[figschemchannel\] gives a zoomed-in view of the channel pore of the KcsA. The channel is set to run along the x-axis and its total length is taken to be 60Å. It contains three regions: the channel gate, a cavity chamber, and a selectivity filter. In our coordinates, the channel starts from -30Å and ends at 30Å . The positions of four standard binding sites for the ions in the selectivity filter (labeled S1 - S4), which are located between the carbonyl oxygen of the amino acids that make up the filter (shown in blue).
![A schematic showing the setup for the KcsA system. The red dots are ions. The gray area is the dielectric continuum with $\varepsilon_{o}=2$ to 80 , the area in the circle of a diameter 30 Å is the productive region. The dielectric inside the TO box is $\varepsilon_{i}=1$. The area between the dashed black lines represents the size of the cellular membrane.[]{data-label="figsyswithprot"}](syswithprot_3.pdf)
![Example of channel with three ions. One ion typically remains in the water-filled cavity and two in the filter area. The blue dots represent the carbonyl oxygen from the amino acids that make up the selectivity filter.[]{data-label="figschemchannel"}](schemchannel.pdf)
The system is first minimized with all particles allowed to move freely using the minimize program in Tinker and then run for 200ps for an initial equilibrium. For subsequent simulations, we use the ACTIVE keyword in the key file to lock down all atoms of the channel except the (sixteen) amino acids making up the selection filter, the waters, and the ions. The parameters contained in the Amber99 force field [@Amber-1981; @Amber-2006], also included in Tinker, will be used for the MD simulations. The Velocity Verlet algorithm was chosen for the time integration with a Nose-Hoover bath set at 300K. The time step is taken to be 2 fs, and the trajectory is recorded every $0.1$ ps for analysis.
Fundamental characteristics {#sec:fundamental}
---------------------------
Before the selectivity of the KcsA channel is investigated, we use the ICSM to simulate some fundamental characteristics of the channel, comparing with the results in the literature.
*Case 1. Profile of electrostatic landscape of the channel:* We first compute the channel permanent potential using the program analyze.x of the Tinker package. This routine was modified to output the potential energy acting on a particular ion as the latter moves through the channel beginning at $x=-30$Åand ending at $x=30$Å. The ion was moved $2$Åat a time and the potential energy was calculated on the ion at that position. The results for the channel prior to minimization and equilibration are shown in Figure \[figpermpot\], which closely resembles the channel permanent potential used by Jung, et al. for their studies on ERINP [@Yung-2009]. In our calculation we found energy minima at -10.4 Å, 11.5Å, and 15.6Å, respectively. Next, we allowed all atoms to move freely in the system and minimized the system using the minimization program in Tinker. The final configuration will be then used as the initial input for the subsequent numerical simulations in this paper. The interval $[-5\mathring{A},5\mathring{A}]$ in the channel is considered as the filter, while $[6\mathring{A},10\mathring{A}]$, $[10\mathring{A},13\mathring{A}]$, $[13\mathring{A},16\mathring{A}]$, and $[16\mathring{A},19\mathring{A}]$ are regarded as site S1, S2, S3 and S4, respectively.
![Potential energy profile of KcsA prior to minimization.[]{data-label="figpermpot"}](permpot.pdf){width="75.00000%"}
*Case 2.* *Preferred positioning for two $\mathrm{K^{+}}$ inside the channel:* It was reported in [@Zhouyf:2001] that two $\mathrm{K^{+}}$ ions would stay at site 1 (S1) and site 3 (S3) when a third ion is far away. Otherwise they will occupy sites S2 and S4 when the third ion is relatively close to the filter entrance. To verify this conclusion by the ICSM, we run 100 ps simulations for several permutations of the positions of only two potassium ions in the channel. One ion is located in the water cavity (around $x=0.418$Å ) and the other is placed in different positions (10-19Å ) in the selectivity filter. The detailed positions of the ions we tested are summarized in Table \[table: table1\]. The starting positions for the ions in the filter were set at the midpoint between the carbonyl oxygen of the filter amino acids. Regardless of their starting positions in the filter, the two potassium ions come to rest at S1 (interval \[6Å, 11Å\]) and S3 (interval \[14Å, 16Å\]) , and the relaxation period for the ions is very short. By approximately $3$ ps, the ions and waters in the filter reach their approximate final positions. An example of the trajectories for four of these runs is plotted in Figure \[2-3ions\] (top). In this figure and all remaining figures, the red horizontal lines indicate the position of the oxygens for the Threonine (T), Valine (V), Glycine (G), Tyrosine (Y), and Glycine (G). The intervals between TV, VG, GY, and YG are considered as S1, S2, S3, and S4, respectively. In this instance, *both of the ions move into the filter and come to rest at positions S1 and S3*. And this configuration will be a transitional phase between an ion exiting the channel and another one entering when three ions are present in the channel for transport.
*Case 3.* *Preferred positioning for three $\mathrm{K^{+}}$ inside the channel:* In this case, we add a third potassium ion into the filter (interval \[10Å, 16Å\]) in addition to the two $\mathrm{K^{+}}$ at the various places as in the previous study. The detailed positions of the three ions are listed in Table \[table: table2\]. After a similar relaxation period, two of the ions rest at site S2 (interval \[11Å, 14Å\]) and site S4 (interval\[16Å, 19Å\]), and another ion stays in the water cavity but close to the entrance of the selectivity filter (around $x=5$Å). Figure \[2-3ions\] (bottom) shows the trajectories of the four permutations of the positions of the three potassium ions listed in Table \[table: table2\]. In the final states, two ions stay between V-G and between Y-G, respectively. The three ion configuration can be considered to be a “steady” or normal state for the potassium channel and we note that *the final positions for the two ions in the selectivity filter are in fact S2 and S4*, which are very close to the energy minima found in the empty channel in Case 1 above.
![Initial and final positions for tests with (A) - two potassium ions in the channel and (B) - three potassium ions in the channel (the one in the cavity is not plotted). The various colors represent different starting positions. The sites S1-S4 are situated in between the five red lines, which are corresponding to the position of filter residues. The unit for the horizontal time axis is ps.[]{data-label="2-3ions"}](2v3iontest.pdf){width="\textwidth"}
\[c\][|c|c|c|c|c|c|c|c|c|]{}& & & &\
& Initial & Final & Initial & Final & Initial & Final & Initial & Final\
K1 & 0.418 & 8.635 & 0.418 & 9.049 & 0.418 & 8.957 & 0.418 & 8.923\
K2 & 18.544 & 14.739 & 15.800 & 14.717 & 13.324 & 15.387 & 10.347 & 13.451\
\[c\][|c|c|c|c|c|c|c|c|c|]{}& & & &\
& Initial & Final & Initial & Final & Initial & Final & Initial & Final\
K1 & 0.418 & 5.310 & 0.418 & 5.404 & 0.418 & 8.562 & 0.418 & 5.092\
K2 & 13.324 & 11.808 & 10.347 & 11.813 & 10.347 & 12.047 & 15.800 & 11.695\
K3 & 18.544 & 16.280 & 15.800 & 15.727 & 18.544 & 16.656 & 18.544 & 16.385\
Selectivity and positioning of potassium versus sodium ions {#sec:selectivity}
-----------------------------------------------------------
Our simulations in Section \[sec:fundamental\] indicate that in the “normal state” of three potassium ions in the channel, two of them are located at site S2 and site S4, while the third one stays in the area around $x=5$Å, which is at the entrance to the selectivity filter. This is consistent with the results in [@Biggin-2001; @Shirivastava02; @Berneche00], and this location corresponds to the favorable location for a $\mathrm{K^{+}}$ ion formed by the P helix dipoles [@Biggin-2001; @Roux-1999].
The KcsA channel dominantly conducts potassium over sodium ions, although the two type of ions have the same charge. To investigate the ability of the ICSM system to distinguish between $\mathrm{K^{+}}$ ions and $\mathrm{Na^{+}}$ ions, we repeat the three ion tests with sodium ions in the selectivity filter instead of potassium ions, in order to simulate the preferred positionings of $\mathrm{Na^{+}}$ ions and $\mathrm{Na^{+}}$/$\mathrm{K^{+}}$ mixtures.
*Case 1*: *Preferred positioning of three $\mathrm{Na^{+}}$ inside the channel*: Figure \[fignatestspure\]A shows the simulation results if three $\mathrm{Na^{+}}$ ions are placed in the channel. In contrast to the case of three-$\mathrm{K^{+}}$ (black curves), the two $\mathrm{Na^{+}}$ ions (dashed grey) in the selectivity filter rest at the sites S1 and S3 (instead of S2 and S4 as for the case of $\mathrm{K^{+}}$), while the third $\mathrm{Na^{+}}$ ion stays around $x=2.5$Å in the cavity, which is toward the center of the cavity and relatively away from the filter entrance. This simulation suggests that when two $\mathrm{Na^{+}}$ are in the filter, they are relatively away from the channel exit, meanwhile the third $\mathrm{Na^{+}}$ has difficulty approaching the filter entrance.
![Positioning of same ions in the selectivity filter. (A) two $\mathrm{Na^{+}}$ in filter and a $\mathrm{Na^{+}}$ in the cavity; (B) two $\mathrm{Na^{+}}$ in filter and a $\mathrm{K^{+}}$ in the cavity; (C) two $\mathrm{K^{+}}$ in filter and a $\mathrm{Na^{+}}$ in the cavity. The black solid lines represent case of 3 potassium ions for comparison, and the dotted color lines are for a mix of potassium and sodium ions where green (circles) is potassium and blue (+) is sodium. A schematic of the ion positions is given under each figure. The unit for the horizontal time axis is ps.[]{data-label="fignatestspure"}](natestspure.pdf){width="80.00000%"}
*Case 2*: *One $\mathrm{K^{+}}$ in cavity and two $\mathrm{Na^{+}}$ ions in filter*: With Case 1 in mind, we replace the $\mathrm{Na^{+}}$ in the cavity by a $\mathrm{K^{+}}$ and keep the two $\mathrm{Na^{+}}$ in the filter. From Fig. \[fignatestspure\]B we can clearly see that the $\mathrm{K^{+}}$ in the cavity pushes the $\mathrm{Na^{+}}$ at site S1 in Case 1 into site S2, hence the two $\mathrm{Na^{+}}$ ions in the filter take sites S2 and S3. On the other hand, the $\mathrm{K^{+}}$ itself stays in the cavity around $x=2.5\mathring{A}$, which is relatively away from the filter entrance.
*Case 3. One $\mathrm{Na^{+}}$ in cavity and two $\mathrm{K^{+}}$ ions in filter*: As shown in Figure \[fignatestspure\] C, in this case, two potassium ions are already situated in the selectivity filter and a sodium ion is located on the cellular side of the water-filled cavity. The two $\mathrm{K^{+}}$ ions in the filter take the sites S2 and S4, as for the three-$\mathrm{K^{+}}$ configuration. Comparing to the $\mathrm{K^{+}}$ in the cavity in the three-$\mathrm{K^{+}}$ system, the position of the $\mathrm{Na^{+}}$ ion is closer to the entrance of the filter and on the other side of the carbonyl oxygen for the tyrosine.
*Case 4. $\mathrm{Na^{+}}$ and $\mathrm{K^{+}}$ mixtures in filter*: Figure \[fignatestsmixed\] shows the permutation of $\mathrm{Na^{+}}$ and $\mathrm{K^{+}}$ mixtures in the filter region, while in the cavity we have either $\mathrm{Na^{+}}$ or $\mathrm{K^{+}}$. Figs. \[fignatestsmixed\]A-D list the situations of $\mathrm{K^{+}Na^{+}K^{+}}$, $\mathrm{Na^{+}Na^{+}K^{+}}$, $\mathrm{K^{+}K^{+}Na^{+}}$, and $\mathrm{Na^{+}K^{+}Na^{+}}$, respectively. The positions of the ions are ordered from the cavity to the exit of the filter. It can be seen from the simulations that no ion can occupy the site S4 in any of the ion combinations.
![Positioning of ion mixtures in the selectivity filter (from the cavity to the filter exit direction). (A) $\mathrm{K^{+}Na^{+}K^{+}}$; (B) $\mathrm{Na^{+}Na^{+}K^{+}}$; (C) $\mathrm{K^{+}K^{+}Na^{+}}$; (D) $\mathrm{Na^{+}K^{+}Na^{+}}$. The black solid lines represent case of 3 potassium ions for comparison, and the dotted color lines are for a mix of potassium and sodium ions where green (circles) is potassium and blue (+) is sodium. A schematic of the ion positions is given under each figure. The unit for the horizontal time axis is ps.[]{data-label="fignatestsmixed"}](natestsmixed.pdf){width="80.00000%"}
From the above simulations it can be summarized that in a three-$\mathrm{Na^{+}}$ system, two sodium ions are occupying in the filter region and stay at sites S1 and S3 positions, while the third ion in the cavity is not able to move toward the proximity of the filter. The $\mathrm{Na^{+}}$ at site S1 is able to proceed to site S2 if the sodium ion in the cavity is replaced by a potassium ion but the other $\mathrm{Na^{+}}$ in the filter stays at site S3. Alternatively, when a $\mathrm{Na^{+}}$ in the water cavity faces two $\mathrm{K^{+}}$ in the selectivity filter, it is closer to the filter entrance than the position of the $\mathrm{K^{+}}$ in the cavity in a three $\mathrm{K^{+}}$ system. Therefore, we conjecture that $\mathrm{Na^{+}}$ ions are able to approach the filter entrance and translocate from site S1 to S2, but no evidence shows that they are able to move to site S4. This may be one reason for the reduced possibility of $\mathrm{Na^{+}}$ ion conducting through the channel, consistent with the fact that the KcsA channel is designed preferentially for three potassium ions in the setting for potassium transporting through the channel as shown in Section 3.2.
Effect of reaction field in ICSM
--------------------------------
We also investigated the importance of the long range reaction field by running same configurations with the image charges and without image charges.
*Case 1. Simulation of the three-$\mathrm{K^{+}}$system without the reaction field:* We first consider the stable configuration of three potassium ions in the channel without taking into account the reaction field. For comparisons, the corresponding simulation results with reaction field are also provided. Fig. \[figrftests\] A show that without the reaction field, the two $\mathrm{K^{+}}$ ions in the selectivity filter are still occupying sites S2 and S4. However, the final resting position for the third ion, which is initially in the water cavity, is approximately at $x=8.3$Å when the reaction field is neglected. In other words, this potassium ion is able to move closer to entrance of the filter and eventually takes site S1(between the residues T and V). Therefore, comparing to the results with reaction field for the three-$\mathrm{K^{+}}$ system where two ions take sites S2 and S4 in the filter while one stays in the water cavity at $x=5.5$Å. Without the reaction field the three potassium ions will take sites S1, S2, and S4; this is against the current biological conclusions.
*Case 2. Simulation of the three-$\mathrm{Na^{+}}$ system without the reaction field:* The positions of three $\mathrm{Na^{+}}$ ions in the channel are plotted in Fig.\[figrftests\]B when the simulation is implemented without the reaction field. For a better comparison, the approximate final positions of these ions are shown in Table \[table3\]. Recall the simulations with the reaction field, two sodium ions in the selectivity filter take the sites S1 and S3, while the third ion stays in the cavity around $x\approx2.6$Å. But without the reaction field, all three sodium ions are able to enter the filter region and they occupy the sites S1, S2, and S3, respectively. Specifically, the ion in the cavity (sodium 1) moves toward the filter and rests at $x\approx8.19$Å if the reaction field is absent.
![Positioning of ion combinations in the channel (from the cavity to the filter exit direction) without reaction field. (A) $\mathrm{K^{+}K^{+}K^{+}}$; (B) $\mathrm{Na^{+}Na^{+}Na^{+}}$; (C) $\mathrm{K^{+}Na^{+}Na^{+}}$; (D) $\mathrm{Na^{+}K^{+}K^{+}}$. The black solid lines represent each case with the reaction field for comparison, and the dotted color lines are for a mix of potassium and sodium ions where green (circles) is potassium and blue (+) is sodium. A schematic of the ion positions is given under each figure. The unit for the horizontal time axis is ps.[]{data-label="figrftests"}](rftests2.pdf){width="80.00000%"}
\[c\][|c|c|c|c|]{}& Sodium 1 & Sodium 2 & Sodium 3\
With RF & 2.59Å& 8.00Å& 15.66Å\
No RF & 8.18Å& 11.99Å& 15.74Å\
*Case 3. Simulation of $\mathrm{K^{+}\slash Na^{+}}$ mixtures in the channel without the reaction field:* For a further investigation of the effect of the reaction field, the case of two $\mathrm{Na^{+}}$ in the selectivity filter and a $\mathrm{K^{+}}$ in the cavity is considered. In section \[sec:selectivity\], it is observed that the two $\mathrm{Na^{+}}$ in the filter take sites S2 and S3, and the $\mathrm{K^{+}}$ ion stays in the cavity around the position $x\approx2.6$Å. While ignoring the reaction field, the $\mathrm{K^{+}}$ ion enters the filter and takes the position at site S1. Thus, in this case, the selectivity filter will contain three residential ions as $\mathrm{K^{+}}$ occupies S1 and the two $\mathrm{Na^{+}}$ ions occupy sites S2 and S3, as shown in Fig.\[figrftests\]C. On the other hand, for the case of two $\mathrm{K^{+}}$ in the filter and one $\mathrm{Na^{+}}$ in the cavity, the results without the reaction field are very close to the ones with the reaction field, however, the position of the $\mathrm{Na^{+}}$ in the cavity fluctuates considerably more (see Fig.\[figrftests\]D).
From the above studies it can be summarized that the cavity ion, no matter $\mathrm{K^{+}}$ or $\mathrm{Na^{+}}$, is able to move into the filter on the other side of the threonine to position $x\approx8.2$Å if the reaction field is neglected. This fact implies that there will be three ions in the selectivity filter, and this does not agree with the biological observations. Therefore, we can see clearly that the reaction field is critical for the fidelity of the ICSM to reproduce the physical mechanism of the KcsA channel in its selectivity and positioning of the ions.
Conclusions {#sec:conclusion}
===========
In this paper, we have shown that the Image-Charge Solvation Method (ICSM) is applicable to study the selectivity filter of a KcsA channel where only a local spherical region around the filter section of the channel is modeled in atomistic details and the effect of the rest of the channel and membrane and solvents can be modeled as a continuum dielectric, whose reaction field is approximated by a simple multiple image charge representation.
Our numerical simulations have produced the following results:
- The preferred positioning for two potassium ions inside the channel are in sites *S1 and S3.*
- The preferred positioning for three potassium ions inside the channel are one in the cavity and two at sites *S2* and *S4*, which is the characteristic positioning of three ions for conducting the potassium ions by the KcsA channel.
- The ICSM algorithm is able to distinguish sodium and potassium ions when investigating the selectivity of the KcsA channel. In contrast to a three-$\mathrm{K^{+}}$ system, the two $\mathrm{Na^{+}}$ ions in the selectivity filter occupy sites *S1* and *S3* instead of *S2* and *S4*, and the third $\mathrm{Na^{+}}$ in the water cavity stays relatively away from the filter entrance. Among all the permutations of 3 ions, except for the case of 3 potassium ions, no ion was found to be able to move to site *S4,* this is consistent with the fact that the KcsA channel is indeed a potassium channel designed to transport potassium ions.
- The reaction field from the dielectric environment outside the atomistic region containing the filter is critical in the accurate representation of the selectivity function of the filter and the correct predictions of the ion positioning inside the channel.
In future research, the following issues will be addressed: the layered dielectric media outside the simulation region comprised of membrane and ionic solvents as for the time being the latter is simply ignored; the I-V curve of the KcsA channel calculated by ICSM molecular dynamics simulations.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors acknowledge the support of the US Army Office of Research (Grant No. W911NF-14-1-0297), the National Science Foundation (DMS-1315128), and the NSFC (grant number 91230105) for the work in this paper. Authors also thank the insightful comments from Drs. Xiaolin Cheng and Chun Liu on various aspects of the simulations for KcsA channels.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Composite fermion wavefuctions have been used to describe electrons in a strong magnetic field. We show that the polynomial part of these wavefunctions can be obtained by applying a normal ordered product of suitably defined annihilation and creation operators to an even power of the Vandermonde determinant, which can been considered as a kind of a non-trivial Fermi sea. In the case of the harmonic interaction we solve the system exactly in the lowest Landau level. The solution makes explicit the boson-fermion correspondence proposed recently.'
author:
- 'V. Ruuska'
- 'M. Manninen'
title: Composite fermions from the algebraic point of view
---
Introduction {#intro}
============
Composite fermions provide a unified approach to fractional quantum Hall effect (FQHE) [@Tsui] and other liquid states of the two-dimensional electron gas [@Halperin] in a strong magnetic field. They have also been used to describe rotating states of quantum dots [@Jain1]. The composite fermion wavefunctions proposed by Jain [@Jain2; @Jain3] are obtained in the following manner. First multiply a wavefunction $\phi$ of [*non-interacting*]{} electrons by the Jastrow factor $$D^k= \prod_{i<j} (z_i-z_j)^{2k}$$ and then project the product to the lowest Landau level. The ansatz was motivated by a heuristic derivation in which an even number of flux quanta were attached to the electrons and a mean-field approximation was applied to the resulting state. Numerical studies for finite systems have supported the ansatz. Recently similar numerical studies have provided additional information about quantum dots. In particular, rotating systems of a number of identical particles with repulsive interactions confined in a trap have been studied. It has been found that the systems form vortices and that the general features of the vortex formation are common to both boson and fermion systems [@Manninen1; @Manninen2].
In this article we study particles in a harmonic trap using analytical and algebraic means.The physical realization could be a two-dimensional electron gas in a strong magnetic field or a quantum dot. The mathematical model is basically the same. In particular, we shall utilize the method of spectrum generating algebras, which is reviewed in section \[ladder\]. Composite fermions are considered in section \[composite\]. Our result is a simple expression for the composite fermion wavefunctions. The polynomial parts of these wavefunctions are obtained by applying a normal ordered product of suitably defined annihilation and creation operators to an even power of the Vandermonde determinant, which can been considered as a kind of a non-trivial Fermi sea. Finally, in section \[harmonic\] we give the exact solution for the harmonic interaction in the lowest Landau level. The LLL spectrum turns out to be qualitatively different from that of the complete solution, which is also known for the harmonic interaction [@Poles; @Lawson]. Nevertheless, using the exact eigenstates we show that there is a simple correspondence between the boson and fermion systems as anticipated by Toreblad et al. [@Manninen1].
Spectrum generating algebras {#ladder}
============================
It often happens in quantum mechanics that energy eigenstates can be obtained from each other by applying suitable ladder operators. We say that $K$ is a [*ladder operator*]{} if the commutator of the Hamiltonian $H$ and $K$ is a multiple of $K$, that is, $[H,K]=\lambda K$. Then it follows for an energy eigenstate $\phi$ with energy $\epsilon$ that $K\phi$ is also an energy eigenstate with energy $\epsilon + \lambda$ since $$H(K\phi) = K(H\phi)+\lambda K \phi = (\epsilon + \lambda) K\phi.$$ If $K_1$ and $K_2$ are ladder operators such that $[H,K_1]=\lambda_1 K$ and $[H,K_2]=\lambda K_2$, then $$\begin{aligned}
[H,[K_1,K_2]]&=[[H,K_1],K_2]+[K_1,[H,K_2]] \\
&=[\lambda_1 K_1, K_2],+ [K_1,\lambda_2 K_2] \\
&=(\lambda_1+\lambda_2)[K_1,K_2].\end{aligned}$$ Hence the ladder operators form a Lie algebra $\mathfrak {L} (H)$. Clearly all the symmetries of the Hamiltonian, that is, the operators commuting with the Hamiltonian, belong to $\mathfrak {L} (H)$. Loosely speaking, a subalgebra $\mathfrak {g}$ of $\mathfrak {L} (H)$ is called a [*spectrum generating algebra*]{} if the physical state space is an irreducible representation of $\mathfrak {g}$. Instead of making the definition precise, we try to illustrate the concept by the following familiar examples.
The best-known example is the harmonic oscillator, the Hamiltonian of which is $$H= -\frac {1}{2} \nabla^2 + \frac{1}{2} {\mathbf {x}}^2$$ where we have chosen the length and energy units to be $\sqrt {\hbar/m\omega }$ and $\hbar \omega$, respectively. Here $m$ is the mass of the particles and $\omega $ is the angular frequence of the oscillator. In the one-dimensional case all the energy eigenstates can be generated from the unique ground state $$\phi_0(x)=\pi^{-1/4}e^{-\frac{x^2}{2}}$$ using the creation operator $$a^{\dagger} =\frac{1}{\sqrt{2}}(\hat x - \frac {i}{\hbar} \hat p)
=\frac{1}{\sqrt{2}} (x - \frac {d}{dx}).$$ The normalized eigenfunctions are $$\label{onedim}
\phi_n =\frac{1}{\sqrt{n!}}(a^{\dagger})^n\phi_0
=\frac{1}{(2^n \sqrt{\pi} n!)^{1/2}}(x - \frac {d}{dx})^n e^{-\frac{x^2}{2}}.$$ We have the commutation relations $[H,a^{\dagger}]=a^{\dagger}$ and $[H,a]=-a$ where the annihilation operator $$a =\frac{1}{\sqrt{2}} (x + \frac {d}{dx}).$$ In this case the appropriate spectrum generating algebra $\mathfrak {g}$ is the Heisenberg algebra generated by $a$ and $a^{\dagger}$ with the canonical commutation relation $[a,a^{\dagger}] = 1$.
This approach readily generalizes to higher dimensional oscillators; we only need one pair of creation and annihilation operators for each degree of freedom. Interestingly, we can do something slightly different in the plane. Let us use the standard notation $$\frac{d}{dz}=\frac{1}{2}(\frac{\partial}{\partial x}-i\frac{\partial}{\partial y}), \qquad
\frac{d}{d\bar{z}}=\frac{1}{2}(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y})$$ Because of the rotational symmetry the Hamiltonian and the angular momentum operator $$\label{rotation}
L=\hbar(z\frac{d}{dz}-\bar {z} \frac {d}{d \bar{z}})=-i\hbar(x\frac{\partial}{\partial y}-y\frac{\partial}{\partial x})$$ commute and we can find a basis of common eigenvectors, which we may express in the following forms: $$\label{twodim}
\begin{split}
\phi_{mn}&=\frac{1}{\sqrt{2\pi m!n!}} \, e^{\frac{1}{2}\vert z \vert^2} \frac{d^m}{dz^m} \frac{d^n}{d\bar{z}^n} e^{-\vert z \vert^2} \\
&=\frac{1}{\sqrt{2\pi m!n!}} \, (e^{\frac{1}{2}\vert z \vert^2 } \frac{d}{dz} e^{-\frac{1}{2}\vert z \vert^2})^m \,
(e^{\frac{1}{2}\vert z \vert^2} \frac{d}{d\bar{z}} e^{-\frac{1}{2}\vert z \vert^2})^n\,
e^{-\frac{1}{2}\vert z \vert^2} \\
&=\frac{(-1)^{m+n}}{\sqrt{2\pi m!n!}} \, (\frac{1}{2} \bar {z}-\frac{d}{dz})^m \,(\frac{1}{2} z- \frac{d}{d\bar{z}})^n \, e^{-\frac{1}{2}\vert z \vert^2}
\end{split}$$ Hence the generators of the spectrum generating algebra can be chosen to be $$\begin{aligned}
b^{\dagger} &= \frac{1}{2} z - \frac {d}{d\bar {z}}, &
{\bar {b}}^\dagger &= \frac{1}{2} \bar {z} - \frac {d}{dz }, \\
b &= \frac{1}{2}\bar {z} + \frac {d}{dz}, &
\bar {b} &= \frac{1}{2} z + \frac {d}{d \bar {z} }\end{aligned}$$ with the commutation relations $[b,\bar {b}]= [b^{\dagger},{\bar {b} }^{\dagger}] = [b,{\bar {b} }^{\dagger}] =
[b^{\dagger}, \bar {b}] =0 $ and $[b,b^{\dagger} ]= [\bar {b},{ \bar {b} }^{\dagger} ]=1 $. This is again the Heisenberg algebra but the advantage is that the algebra is spectrum generating not only for $H$ but also for $L$. Using the expression $$H= \frac{1}{2}(b b^{\dagger} + b^{\dagger} b + \bar {b} {\bar {b}}^{\dagger} +{\bar {b}}^{\dagger} \bar {b} )$$ and equation it is easy to show the necessary commutation relations: $$\begin{aligned}
&[H,b^{\dagger}] = b^{\dagger}, &[H,{\bar {b}}^{\dagger}] &= {\bar {b}}^{\dagger},
&[H,b] &= - b, &[H,\bar {b}] &= - \bar {b} \\
&[L,b^{\dagger}] = \hbar b^{\dagger}, &[L,{\bar {b}}^{\dagger}] &= -\hbar {\bar {b}}^{\dagger},
&[L,b] &= - \hbar b, &[L,\bar {b}] &= \hbar \bar {b}.
\end{aligned}$$ The eigenstates as well as the action of the above generators on them are shown in Fig. \[spectrum\].
![The eigenstates of the two-dimensional harmonic oscillator in the (L,H) plane. The action of the annihilation and creation operators on the eigenstates is also shown graphically. \[spectrum\]](ruuskafig1.eps)
Composite fermions {#composite}
==================
Composite fermions [@Jain2; @Jain3] provide a unified approach to fractional quantum Hall effect (FQHE) [@Tsui] and other liquid states of the two-dimensional electron gas [@Halperin] in a strong magnetic field. They have also been used to describe rotating states of quantum dots [@Jain1]. Let us consider electrons confined to the plane and subjected to a magnetic field $H$ perpendicular to it. Assume that the strong magnetic field makes the electrons effectively spinless. In the symmetric gauge $$\mathbf{A} = \frac{1}{2} H(y\mathbf{e}_x-x\mathbf{e}_y)$$ the single-body Hamiltonian $$H_s=\frac{1}{2m}(\frac{\hbar}{i}\mathbf{\nabla}-\frac{e}{c}\mathbf{A})^2$$ takes the form $ H_s = H -\omega L$ where the angular frequence of the harmonic oscillator is $$\omega = \frac{eH}{2mc}.$$ Note that $\omega$ is half the usual cyclotron frequence but this choice complies with the conventions adopted for the harmonic oscillator. Now the common eigenstates $\phi_{m,n}$ of $H$ and $L$ are also the eigenstates of $H_s$. However, the eigenenergies change and the ground state becomes infinitely degenerate. All the states $$\phi_{0,n} = \frac{1}{\sqrt{2\pi n!}} \, z^n \, e^{-\frac{1}{2}\vert z \vert^2}$$ forming the lowest Landau level (LLL) have zero energy. The Hilbert space of LLL states may be identified with the Bargmann space $$\mathcal {H}_0= L^2_{hol}(\mathbb {C} \, ; \, e^{-\vert z \vert^2} dm(z))$$ of square-integrable holomorphic functions with respect to the weight function $e^{-\vert z \vert^2}$ while the entire state space is $$\mathcal {H}= L^2(\mathbb {C} \, ; \, e^{-\vert z \vert^2} dm(z)).$$ Here $dm(z)$ means the Lebesgue measure. Including the exponential factor into the measure reduces the calculations to polynomial calculus. This technique was already inherent in Laughlin’s work [@Laughlin1], and was explained by Girvin and Jach [@GirvinJach1].
Let $\operatorname{P}$ stand for the projection $\mathcal{H} \to \mathcal{H}_0$. Since the functions $$\chi_n(z) =\frac{1}{\sqrt{2\pi n!}} \, z^n$$ constitute an orthonormal basis of $\mathcal {H}_0 $, the kernel of the projection $\operatorname{P}$ with respect to the weighted measure is $$K(z,w)=\sum_n \chi_n(z) \overline {\chi_n(w)} =\frac{1}{2\pi} \sum_n \frac {1}{n!} z^n \bar{w}^n = \frac{1}{2\pi} e^{z \bar{w} }.$$
The composite fermion trial wavefunctions are obtained in the following manner. First multiply a wavefunction $\phi$ of [*non-interacting*]{} electrons by the Jastrow factor $$D^k= \prod_{i<j} (z_i-z_j)^{2k}$$ and then project the product to the LLL. Here $k$ is a positive integer. All indices will run from $1$ to the number $N$ of the particles unless otherwise stated. Our goal is to make the calculation of the composite fermion wavefunctions more explicit. The wavefunctions of the non-interacting electrons are anti-symmetrized products of the eigenfunctions $\phi_{m,n}$. The projection can be done one complex variable at a time. Consider the polynomial parts of the eigenfunctions without the normalizing factor $$\varphi_{mn} = \sqrt{2\pi m!n!} \, \phi_{mn} \, e^{\frac{1}{2} \vert z \vert^2}=
e^{\vert z \vert^2}\frac{d^m}{dz^m} \frac{d^n}{d\bar{z}^n} e^{-\vert z \vert^2}.$$ Then $\varphi_{mn}$ are holomorphic polynomials and we have to calculate the projection $\operatorname{P}(\psi \varphi_{m,n})$. Here $\psi$ is the Jastrow factor or more generally an arbitrary holomorphic polynomial considered as a function of one variable only. We have $$\label{first}
\begin{split}
\operatorname{P}(\psi \varphi_{mn}) &= \int K(z,w) \, \psi(w) \, [e^{\vert w \vert^2}\frac{d^m}{dw^m} \frac{d^n}{d{\bar{w}}^n}
e^{-\vert w \vert^2}] \, e^{-\vert w \vert^2} \, dm(w) \\
&= \frac{1}{2\pi} \int e^{z\bar{w}} \psi(w) \, \frac{d^m}{dw^m} \frac{d^n}{d{\bar{w}}^n} e^{-\vert w \vert^2} \, dm(w) \\
&= -\frac{1}{2\pi} \int \frac{d}{d\bar{w}} [e^{z\bar{w}} \psi(w)] \, \frac{d^m}{dw^m} \frac{d^{n-1}}{d{\bar{w}}^{n-1}} e^{-\vert w \vert^2} \, dm(w) \\
&= -\frac{1}{2\pi} \int z e^{z\bar{w}} \psi(w) \, \frac{d^m}{dw^m} \frac{d^{n-1}}{d{\bar{w}}^{n-1}} e^{-\vert w \vert^2} \, dm(w) \\
&=-z \operatorname{P}(\psi \varphi_{m,n-1}) = \dots = (-z)^n \operatorname{P}(\psi \varphi_{m,0}).
\end{split}$$ by induction. On the other hand, $$\begin{split}
\label{second}
\operatorname{P}(\psi \varphi_{m,0}) &= \int K(z,w) \, \psi(w) \, [e^{\vert w \vert^2}\frac{d^m}{dw^m}
e^{-\vert w \vert^2}] \, e^{-\vert w \vert^2} \, dm(w) \\
&= \frac{1}{2\pi} \int e^{z\bar{w}} \psi(w) \, \frac{d^m}{dw^m} e^{-\vert w \vert^2} \, dm(w) \\
&= \frac{1}{2\pi} \int e^{z\bar{w}} \psi(w) \, (-\bar{w})^m \, e^{-\vert w \vert^2} \, dm(w) \\
&= (-\frac{d}{dz})^m \, [\frac{1}{2\pi} \int e^{z\bar{w}} \psi(w) \, e^{-\vert w \vert^2} \, dm(w)] \\
&= (-\frac{d}{dz})^m \, [\operatorname{P}(\psi \varphi_{0,0})]= (-\frac{d}{dz})^m \, \psi.
\end{split}$$ Combining the equations and we obtain $$\operatorname{P}(\psi \varphi_{mn}) = (-1)^{n+m} z^n \frac{d^m}{dz^m} \, \psi.$$ In general, an N-particle state of non-interacting electrons is an antisymmetrized product $\phi= \operatorname{A}[\phi_{\alpha_1\beta_1} \cdots \phi_{\alpha_N\beta_N}]$ of the single-particle wave-functions. Here the antisymmetrization operator is defined as $$\operatorname{A}[\Phi](z_1,\dots,z_N)=\sum_{\sigma\in S_n} (-1)^{\operatorname{sign}(\sigma)} \Phi(z_{\sigma(1)},\dots,z_{\sigma(N)})$$ where the summation is over all the permutations of the indices $1,\dots,N$. The above calculation easily generalizes to several variables. We can express the result using the multi-index notation $$z^{\alpha} = z_1^{\alpha_1} \cdots z_N^{\alpha_N} \quad \text{ and } \quad
\frac{\partial^{\beta}}{\partial z^{\beta}} = \frac {\partial^{\vert \beta \vert}}
{\partial z_1^{\beta_1} \cdots \partial z_N^{\beta_N} }$$ where $\vert \beta \vert = \beta_1 + \cdots + \beta_N$. The composite fermion wavefunction corresponding to $\phi$ and Jastrow factor $D^k$ is given up to normalization by $$\label{cpfunction}
\phi^{CF}_{\alpha, \beta,k} = \operatorname{A}[ z^{\alpha} (\frac {d^{\beta}}{dz^{\beta}} \, D^k) ] e^{-\frac {1}{2} \vert z \vert^2}.$$ The formula is analogous to the equations and in the sense that all the states can be created by applying the multiplication and partial derivative operators to the given ground states, Jastrow factors. Consequently, it seems natural to search for spectrum generating algebras even for interacting fermions. However, to proceed further we have to choose the actual form of the interaction.
Exact results for the harmonic interaction {#harmonic}
==========================================
It was already known for the nuclear physicists of the fifties that the harmonic interaction potential $$\label{harmint}
V = \lambda \sum_{i<j} \vert z_i - z_j \vert^2$$ in a harmonic trap is exactly diagonalizable [@Lawson]. The model has also been applied to quantum dots [@johnson]. A complete set of eigenfunctions is difficult to write down due to the degeneracy problems but the eigenenergies and their degeneracies are known [@Poles]. The general solution has no restrictions with respect to the angular momentum, whereas the composite fermion wavefunctions are by definition restricted to the lowest Landau level. The LLL approximation is usually justified by physical considerations related either to the strong magnetic field or to the rotational state. We shall find the exact solutions of the projected Hamiltonian in the lowest Landau level. Our result generalizes the observation by Girvin and Jach [@GirvinJach1; @GirvinJach2], who noticed that Laughlin’s functions $$\Psi_m = \prod_{i < j} (z_i -z_j)^m e^{-\frac{1}{2}\vert z \vert^2}$$ for an odd integer $m$ are exact energy eigenstates for $V$. Comparison with the complete solution will hopefully shed light on the subtleties of the LLL approximation. At least the LLL solution makes explicit the correspondence between the boson and fermion systems suggested in [@Manninen1].
To start with, a technical remark is in order. Note that $\lambda$ must be chosen negative to make the interaction repulsive. Moreover, the absolute value of $\lambda$ must be small enough to make the energy spectrum bounded from below. For the LLL Hamiltonian even this is not sufficient. However, if we fix the total angular momentum, a (degenerate) ground state can be found. The projection $\hat V=\operatorname{P}V \operatorname{P}$ of the interaction to the Bargmann space $\mathcal {H}_0$ was calculated by Girvin and Jach [@GirvinJach2]. It is given by the formula $$\begin{split}
\hat V &= \lambda \sum_{i,j} (\frac {\partial}{\partial z_i}- \frac {\partial}{\partial z_j})(z_i - z_j) \\
&= N(N-1)\lambda + \lambda \sum_{i,j} (z_i - z_j)(\frac {\partial}{\partial z_i}- \frac {\partial}{\partial z_j})
\end{split}$$ Multiplication operator $z_k$ is not quite a ladder operator for $\hat V$ since $$\begin{split}
[\hat V, z_k] &= \lambda \, [\sum_{i<j} (z_i - z_j)(\frac {\partial}{\partial z_i}- \frac {\partial}{\partial z_j}), z_k] \\
&= \lambda \, \sum_{i<j} (z_i - z_j)(\delta_{ik} - \delta_{jk}) \\
&= N\lambda \, z_k - \sum_i z_i.
\end{split}$$ However, it follows from the above calculation that $$[\hat V, z_k-z_l] = N \lambda (z_k-z_l)$$ for all $k$ and $l$. Hence multiplication of an eigestate by the linear factors $z_k - z_l$ creates new eigestates with eigenvalues increased by $N \lambda$. We also have $$[\hat V, \sum_k z_k] = 0$$ so that the operator $\sum_k z_k$ creates states which are denegenerate with the original one. We can start with the constant function $\varphi_B = 1$, which is an eigenstate for $\hat V$ with eigenvalue $N(N-1)\lambda $, and multiply it by the above linear factors to obtain new eigenstates. However, in order to qualify as proper physical states, they also have to obey the usual permutation symmetries. Let us consider the bosonic case first. The question is which symmetric polynomials can be represented as linear combinations of the products of the factors $e_{kl} = z_k-z_l$. Since the factors $e_{kl}$ span the space $C= \{ z_1 + \dots + z_N=0 \}$, which is isomorphic to the Cartan subalgebra of the Lie algebra $\mathfrak {su} (N)$ as a representation of the symmetric group $S_n$, the problem is solved by the classical invariant theory [@Bourbaki]. As long as we are restricted to $C$, the subalgebra of symmetric polynomials obtained from $e_{kl}$ is freely generated by the power sums $$q_k = z_1^k + \cdots + z_N^k$$ for $1< k \leq N$. Now $q_k$ is obtained by restricting a linear combination of the products of $e_{kl}$ from ${\mathbb {C}}^N$ to $C$. The polynomials must be invariant in the translations parallel to $(1,1,\dots,1)$ since the factors $e_{kl}$ are. Hence the only possible extension of $q_k$ to ${\mathbb {C}}^N$ is $$p_k = (z_1-\bar z)^k + \cdots + (z_N - \bar z)^k$$ where $\bar z = (z_1+\dots+z_N)/N$ is the center-of-mass. If we allow $p_1=z_1+\dots+z_N$ among the generators, then by dimensional counting we obtain a complete basis of the symmetric polynomials. The difference is that the factor $p_1$ does not have effect on the eigenvalue while the other power sums $p_k$ increase it by $Nk\lambda$. Including the Gaussian factors we obtain the boson eigenstates $$\psi_\alpha^B = \prod_{k=1}^N p_k^{\alpha_k} e^{-\frac{1}{2} \vert z \vert^2}$$ where $\alpha_k$ are non-negative integers. The corresponding eigenenergies are $$E_\alpha^B= N(N-1)\lambda + N\lambda\sum_{k=2}^N k\alpha_k.$$ Disbelievers can check the result by a straight-forward calculation. Once the answer is known, it is not that hard since $\hat
V$ is a derivation up to an additive constant. The fermion states can be obtained by exactly the same reasoning starting from the ground state $$\varphi_F = \prod_{i<j} (z_i - z_j)$$ with the eigenvalue $N(N-1)(1+N/2) \lambda$. The fermion eigenstates are $$\psi_\alpha^F = \prod_{k=1}^N p_k^{\alpha_k} \prod_{i<j} (z_i - z_j) e^{-\frac{1}{2}\vert z \vert^2}$$ with eigenenergies $$E_\alpha^B= N(N-1)(1+N/2)\lambda + N\lambda\sum_{k=2}^N k\alpha_k.$$ Observe that the fermion states differ from the boson states only by the factor $\varphi_F$, while the corresponding eigenenergies are shifted by a constant depending only on the number of particles and the strength $\lambda $ of the interaction. Practically all states are degenerate due to the huge symmetry of the interaction potential.
Conclusions
===========
We have given a short review on the spectrum generating algebras and applied them to derive the composite fermion wavefunctions. As an example we studied the exactly solvable model of harmonic interparticle interactions and derived the wavefunctions and energy spectra for spinless fermions and bosons. In this case there is a one-to-one correspondense between the fermionic and bosonic wavefunctions: Any fermion wavefunction is obtained from the corresponding boson wavefunction by multiplying it with a simple Slater determinant.
[99]{} D.C. Tsui, H.L. Stormer, and A.C. Gossard, Phys. Rev. Lett. [**48**]{}, 1559 (1992).
B.I. Halperin, P.A. Lee, and N. Read, Phys. Rev. [**472**]{}, 7312 (1993).
J.K. Jain and T. Kawamura, Europhys. Lett. [**29**]{}, 321 (1995).
J.K. Jain, Phys. Rev. Lett.", [**63**]{}, 199 (1989).
J.K. Jain, Phys. Rev. B [**41**]{}, 7653 (1990).
M. Toreblad, M. Borgh, M. Koskinen, M. Manninen, and S.M. Reimann, Phys. Rev. Lett. [**93**]{}, 090407 (2004).
S.M. Reimann and M. Manninen, M., Rev. Mod. Phys. [**74**]{}, 1283 (2002).
M.A. Zaluska-Kotur, M. Gajda, A. Orlowski, and J. Mostowski, Phys. Rev. A [**61**]{}, 033613 (2000).
R.D. Lawson, [*Theory of Nuclear Shell Model*]{}, (Oxford Press, Oxford 1980).
N.F. Johnson, Phys. Rev. B [**46**]{}, 2636 (1992).
R.B. Laughlin, Surf. Sci. [**141**]{}, 11 (1984).
S.M. Girvin and T. Jach, Phys. Rev. B [**28**]{}, 4506 (1983).
S.M. Girvin and T. Jach, Phys. Rev. B [**29**]{}, 5617 (1984).
N. Bourbaki, [*Groupes et algèbres de Lie (Ch. 5)*]{}", (Hermann, Paris 1968).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
[**See [*http://www.xassist.org/ixocat\_hri*]{} for figures and future updates of this catalog. NOTE – some IXO numbers have changed since the original astro-ph submission. This paper is now accepted into the ApJ Supplements, and Tables 1 and 2 comprise version 2.0 of our ROSAT HRI IXO catalog.**]{}\
\
ROSAT, and now [*Chandra*]{}, X-ray images allow studies of [*extranuclear*]{} X-ray point sources in galaxies other than our own. X-ray observations of normal galaxies with ROSAT and Chandra have revealed that off-nuclear, compact, Intermediate-luminosity (L$_X[2$-$10~{\rm keV}] \ge$10$^{39.0}$ erg s$^{-1}$) X-ray Objects (IXOs, a.k.a. ULXs \[Ultraluminous X-ray sources\]) are quite common. Here we present a catalog and finding charts for 87 IXOs in 54 galaxies, derived from all of the ROSAT HRI imaging data for galaxies with $cz \le$ 5000 km s$^{-1}$ from the Third Reference Catalog of Bright Galaxies (RC3). We have defined the cutoff L$_X$ for IXOs so that it is well above the Eddington luminosity of a 1.4 M$_\odot$ black hole (10$^{38.3}$ erg s$^{-1}$), so as not to confuse IXOs with “normal” black hole X-ray binaries. This catalog is intended to provide a baseline for follow-up work with [*Chandra*]{} and XMM, and with space- and ground-based survey work at wavelengths other than X-ray. We demonstrate that elliptical galaxies with IXOs have a larger number of IXOs per galaxy than non-elliptical galaxies with IXOs, and note that they are not likely to be merely high-mass X-ray binaries with beamed X-ray emission, as may be the case for IXOs in starburst galaxies. Approximately half of the IXOs with multiple observations show X-ray variability, and many (19) of the IXOs have faint optical counterparts in DSS optical B-band images. Follow-up observations of these objects should be helpful in identifying their nature.
author:
- 'E. J. M. Colbert and A. F. Ptak'
title: 'A Catalog of Candidate Intermediate-luminosity X-ray Objects'
---
‘=11 versim\#1\#2[2.45pt]{} ‘=12
Introduction
============
In the early 1980s, surveys of [*normal*]{} galaxies with the Einstein X-ray satellite revealed intermediate-luminosity (L$_X$ $\sim$ 10$^{39}$$-$10$^{40}$ erg s$^{-1}$) X-ray sources which were seemingly located in the centers of spiral galaxies (cf. Fabbiano 1989). This was very interesting, since Seyfert nuclei (active galactic nuclei \[AGNs\] in nearby spirals) are typically much more luminous (L$_X$ $\sim$ 10$^{42}$$-$10$^{44}$ erg s$^{-1}$) and black hole X-ray binaries (BH XRBs) are much less luminous (typically L$_X$ $\lapprox$ 10$^{38}$ erg s$^{-1}$). It was not clear whether these intriguing sources were underluminous accreting supermassive BHs, over-luminous XRBs located near the galactic nucleus, or an entirely new type of astrophysical object altogether (e.g., Colbert et al. 1995). In the 1990s, ROSAT observations with the High Resolution Imager (HRI) showed that these Intermediate-luminosity X-ray Objects (IXOs, a.k.a. “Ultraluminous X-ray sources,” or ULXs) are compact X-ray sources, and are quite common in the local Universe – present in the nuclear regions of $\sim$50% of nearby normal galaxies observed with the ROSAT HRI (Colbert & Mushotzky 1999; Roberts & Warwick 2000, hereafter RW2000). Many IXOs were found to be offset from the optical nucleus (e.g., Colbert & Mushotzky 1999).
As an example, the “near-nuclear” IXO in NGC 1313 is actually displaced $\sim$1$'$ ($\sim$ 1.5 kpc) from the optical nucleus. Such a displacement is not expected if the object is a supermassive BH (M $\gapprox$ 10$^6$ M$\odot$), since dynamical friction would likely cause the BH to sink to the center of the galaxy in a Hubble time (cf. Tremaine, Ostriker & Spitzer 1975). For this reason, IXO BH masses are estimated to be $\lapprox$10$^5$ M$\odot$. If the X-rays are emitted isotropically and the luminosities are sub-Eddington, the mass of the central object is required to be $\gapprox$ 10 M$_\odot$. Therefore IXOs, if powered by accretion and not beamed (cf. King et al. 2001 and Koerding, Falcke & Markoff 2002 for discussions on beaming), have [*intermediate-mass black holes*]{} (IMBHs). If this is true, then IXOs are extremely interesting objects since they are the only tracers of this peculiar class of black holes. Formation of IMBHs is not well understood – for example, supernova explosions produce BHs $\lapprox$ 20 M$\odot$. The merging of smaller mass black holes has been proposed as a formation mechanism for IMBHs (e.g. Taniguchi et al. 2000, Miller & Hamilton 2002). Continued observational and theoretical work on IXOs will hopefully help unravel their mysterious nature.
There are currently many important, unanswered questions about IXOs. For example: Do they really contain IMBHs? How did IXOs form? Are they binary systems? Are IXOs merely “normal” stellar-mass BH XRBs with beamed X-ray emission? How are the BHs fueled? Are IXOs more numerous in galaxies with high star formation rates? If so, is that because there are more seed black holes to merge into IMBHs, more fuel, or for some other reason? Are IXOs in elliptical galaxies physically different types of objects than IXOs in spiral galaxies?
Since IXOs are usually found displaced from the center of the galaxy, much of the current work on IXOs stems from comparisons of their properties with those of “normal” BH XRBs and “micro-quasars” (e.g., Colbert et al. 1999, Makishima et al. 2000, Roberts et al. 2002, and Koerding, Falcke & Markoff 2002). Reviews of the properties of BH XRBs and micro-quasars can be found in Tanaka et al. (1995) and Mirabel (1998), respectively.
Thus far, most of the effort to understand IXOs has been focussed on a few well-known objects (e.g., the M82 IXO, N5204 X-1, NGC 3628 X-1). Comparatively fewer studies have been done at wavelengths other than X-ray, and this is partly due to a lack of a complete catalog of IXOs. Optical surveys will allow studies of the properties of the accreting gas (star or dense gas cloud) immediately around the BH, and of the local galactic environment (e.g. globular clusters, gaseous nebulae, young star clusters). This could then provide clues to how IXOs might have formed, or at least how they might have become “active” X-ray emitters. Here we present a catalog of 87 IXO candidates that were carefully selected from all of the public ROSAT HRI images of galaxies from the Third Reference Catalog of Bright Galaxies (hereafter RC3, de Vaucouleurs et al. 1991).
Data Reduction
==============
Our IXO catalog is based on a cross-correlation between all of the X-ray point sources in ROSAT HRI images and galaxies listed in v3.9b of RC3.
Distances to RC3 galaxies were extracted from galaxy distance catalogs (see references listed in Table 1), or were computed from the recessional velocity $cz$ using H$_0 =$ 75 km s$^{-1}$ Mpc$^{-1}$, if $cz >$ 1000 km s$^{-1}$. Galaxies with unlisted values of $cz$ and no known distances were rejected. We also rejected very distant galaxies, by rejecting galaxies with $cz >$ 5000 km s$^{-1}$ (66.7 Mpc for H$_0 =$ 75 km s$^{-1}$ Mpc$^{-1}$). After selection, the number of RC3 galaxies in our sample was 9999. We have revised the distance to NGC 3031 (M81) from 1.4 Mpc (Tully 1998) to the more accurate and much larger Cepheid distance of 3.55 Mpc (Paturel et al. 2002). We did not find any other of our galaxy distances from Tully (1998) that were highly discrepent with recent Cepheid distances.
A list of 5465 ROSAT HRI datasets was compiled, based on all of the public HRI data available on the FTP archive at NASA/GSFC on February 18, 2001. The positions of the pointing centers for each HRI image were cross-correlated with the positions of each of the 9999 galaxies listed in our RC3 galaxy sample. Since the HRI FOV is 36$'$ in diameter, we selected HRI images for reduction if they had pointing centers offset $\le$ 40$'$ from any of the 9999 accepted RC3 galaxies. After selection, our data sample consisted of 1883 unique ROSAT HRI datasets.
Reduction of the ROSAT HRI data was done using the semi-automated X-ray data reduction package XASSIST (http://www.xassist.org). The first step of the reduction process consists of finding point sources in the images. Each of the raw 1883 HRI images was analyzed using a wavelet point source detection algorithm (the “detect” feature in XASSIST) to construct a list of potential point sources. After the detection algorithm, a more rigorous spatial fitting test was performed for each point source, which consisted of fitting a 2D Gaussian model and a sloping background model to the image at each of the source positions. The resulting sources were then flagged by XASSIST as being point-like, extended, or asymmetric. Only point-like sources with “significance” (S/N) greater than 3.0 were kept. The final list consisted of 4281 point sources. These sources were then cross-correlated with RC3, and sources within 10$"$ of each other were identified as the same source. Count rates were then converted to 2$-$10 keV fluxes using PIMMS and an assumed power-law spectrum with $\Gamma=$1.7 (see also Table 1 notes). As noted by Strickland et al. (2001) and Roberts et al. (2002), many Chandra spectra of IXOs are consistent with this assumed spectral model. Isotropic luminosities were then calculated from the fluxes, assuming that the IXOs were located at the distance of the corresponding galaxy. A filtered list was then made of all sources with L$_X(2-10~{\rm keV}) \ge 10^{39}$ erg s$^{-1}$ and positions within two R$_{25}$ radii of an RC3 galaxy. The optical fields around the potential IXOs were then individually inspected and sources that were obvious mis-associations (e.g., nuclei of other nearby galaxies, and sources located far from edge-on disk galaxies, perpendicular to the disk) were rejected. This list was then cross-correlated with the NED and Simbad databases to further filter out QSOs and X-ray supernova. Our final catalog of 87 IXOs is listed in Table 1.
Finding Charts
==============
In Figures 1$-$15, we show contour plots of the ROSAT HRI X-ray emission from the 87 IXOs, overlayed on B-band Digital Sky Survey (DSS) images from STScI. The datasets used for the contour plot are listed in Table 2.
Comments on Individual Objects
==============================
IXO 1 and IXO 2
---------------
Both of these objects are located near the elliptical galaxy NGC 720 (Figure 1). They have faint optical counterparts in the DSS images, but the counterparts are not listed in either the NED or the Simbad databases. See also Fabbiano, Trinchieri & Kim (1992, hereafter FKT92).
IXO 3
-----
This object is located at the outer edge of the disk in the edge-on spiral galaxy NGC 891 (Figure 1). It is also known as NGC 891 X-5 (RW2000). Count rates from two separate HRI observations of IXO 3 show $\sim$50% variability (Table 2). IXO 3 was detected with ASCA (cf. Ueda et al. 2001), but was not detected with the ROSAT PSPC (e.g., Read, Ponman & Strickland 1997, hereafter RPS97).
IXO 4
-----
IXO 4 is located in the disk of the face-on spiral galaxy NGC 1042 (Figure 1). It is quite luminous, with an estimated 2$-$10 keV luminosity of 1 $\times$ 10$^{40}$ erg s$^{-1}$. See also FKT92 and Ueda et al. (2001).
IXO 5
-----
IXO 5 is also located in the disk of a face-on spiral galaxy (NGC 1073, see Figure 1). There are two other bright X-ray point sources positioned within this galaxy. These other two sources are coincident with known QSOs (see Figure 2).
IXO 6
-----
This object is positioned in the outer disk (or halo) of the early-type galaxy NGC 1291 (Figure 2). See also RPS97.
IXOs 7 and 8
------------
IXO 7 is a well-studied X-ray source (Fabbiano & Trinchieri 1987, Petre et al. 1994, Colbert et al. 1995, Miller et al. 1998, Schlegel et al. 2000b). Its ASCA spectra show spectral variability (e.g., Colbert & Mushotzky 1999), similar to that of a black-hole XRB. It is located at the northern end of the bar, $\sim$1 kpc from the center of the face-on spiral galaxy NGC 1313 (Figure 2). IXO 7 is commonly referred to as NGC 1313 X-1.
IXO 8 (NGC 1313 X-2) is positioned in the outer disk of NGC 1313, $\sim$7’ south from X-1 (Figure 2). See Makishima et al. (2000) and Mizuno et al. (2001) for ASCA spectral analyses of this source.
The count rates from numerous HRI observations of these two IXOs show that they are variable by a factor of $\sim$2 (Table 2).
IXOs 9 $-$ 14
-------------
IXOs 9$-$14 are positioned in the outskirts of the elliptical galaxy NGC 1316 (Fornax A; see Figure 2). IXOs 9, 10 and 14 have faint optical counterparts in the DSS image, but there are no catalogued sources in NED or Simbad at those positions. Kim, Fabbiano, & Mackie (1998) have published an extensive study of the ROSAT X-ray sources in Fornax A. See also FKT92.
IXOs 15 and 16
--------------
These two objects are positioned very near each other, in the outer disk of the spiral galaxy NGC 1365; see Figure 3). Possible optical counterparts are visible in the DSS images. See Turner, Urry & Mushotzky (1993) and Singh (1999) for more detailed discussions of these X-ray objects. See also FKT92.
IXOs 17, 18, 19 and 20
----------------------
These four objects are positioned near the elliptical galaxy NGC 1399 (see Figure 3). IXOs 17 and 18 are near the (extended) nuclear X-ray source and are not well resolved by the HRI. IXO 19 is located much further out in the halo, and is coincident with a faint optical counterpart seen in the DSS image. However, there is no catalogued entry for the counterpart in NED or Simbad. IXO 20 is also coincident with an optical DSS counterpart. IXO 19 shows variability by a factor of $\sim$3, while IXO 20 shows $\sim$50% variability between two HRI observations (Table 2). See also Jones et al. (1997).
IXO 21
------
IXO 21 is positioned in the outer regions of the irregular galaxy NGC 1427A (Figure 3). See also Jones et al. (1997).
IXO 22
------
IXO 22 is positioned in the spiral arms of the nearby, face-on spiral galaxy IC 342 (Figure 3). It is commonly known as IC 342 X-1. It is coincident with an optical counterpart seen in the DSS image, but NED and Simbad have no catalog entry for the counterpart. IC 342 has a low Galactic latitude ($b \approx$10$^\circ$), so the counterpart could be a Galactic star. ASCA studies of IXO 22 (e.g., Makishima et al. 2000, Kubota et al. 2001, Mizuno et al. 2001) show that this source is quite variable. See also Fabbiano & Trinchieri (1987), Bregman, Cox & Tomisaka (1993), Colbert & Muzhotzky (1999), RW2000, and Lira et al. (2000).
IXO 23
------
This object is positioned in the outskirts of the elliptical galaxy NGC 1553 (Figure 4). See Blanton, Sarazin & Irwin (2001) for a discussion of the Chandra X-ray point sources in NGC 1553.
IXO 24
------
IXO 24 is located near the center of the face-on spiral galaxy NGC 1566 (Figure 4). This source is surrounded by strong extended (diffuse?) X-ray emission from the disk of NGC 1566.
IXOs 25, 26 and 27
------------------
These three objects are located in the spiral arms of the galaxy NGC 1672 (see Figure 4). See Brandt, Halpern & Iwasawa (1996) and De Naray et al. (2000) for more detailed discussions of these X-ray sources. Count rates from two HRI observations of IXO 26 show variation by a factor $\sim$2.5.
IXO 28
------
This IXO is located in the disk of the spiral galaxy NGC 1792 (Figure 4). See Dahlem et al. (1994) for a ROSAT analysis of the point sources in NGC 1792.
IXO 29
------
IXO 29 is located in the outskirts of the spiral galaxy NGC 1961 (Figure 5). See also FKT92, Pence & Rots (1997), and RW2000.
IXO 30
------
IXO 30, located $\sim$2$'$ west of the center of the Seyfert galaxy Mrk 3 (Figure 5), is discussed by Turner, Urry & Mushotzky (1993) and Morse et al. (1995).
IXO 31
------
This IXO in the dwarf galaxy Holmberg II (Figure 5) is well studied in X-rays (e.g., Colbert & Mushotzky 1999, Zezas, Georgantopoulos, & Ward 1999, and Miyaji et al. 2001). It is located in the middle of an H II region complex, and is also coincident with a luminous radio source (cf. Miyaji et al. 2001).
IXOs 32 and 33
--------------
These two objects are positioned in the outskirts of the early-type spiral galaxy NGC 2775 (Figure 5). They are both coincident with faint DSS sources, although the optical counterparts are not listed in NED or Simbad. See also RW2000.
IXO 34
------
IXO 34 is a well-known X-ray source (M81 X-9) near the dwarf galaxy Holmberg IX, a companion to NGC 3031 (M81; see Figure 6). It has been hypothesized to be a superbubble (Miller 1995), a background quasar (Ishisaki et al. 1996; Ezoe et al. 2002), and an accreting intermediate-mass black hole (Wang 2002). See La Perola et al. (2001) for a discussion of this X-ray source. IXO 34 is highly variable. As can be seen from the HRI count rates listed in Table 2, it shows variation in count rate by a factor of $\sim$3. See also Fabbiano (1988), FKT92, Makishima et al. (2000), RW2000, Lira et al. (2000), Immler & Wang (2001), and Mizuno et al. (2001).
IXO 35
------
IXO 35 is located $\sim$2$'$ north of the elliptical galaxy NGC 3226 (see Figure 6). There is a possible optical counterpart to IXO 35 in the DSS image, but it is not cataloged in NED or Simbad. See also RW2000.
IXO 36
------
This object is located $\sim$2.5$'$ to the west of the center of the galaxy merger system NGC 3256 (Figure 6).
IXO 37
------
This IXO is located in the outskirts of the elliptical galaxy IC 2597, which is in the Hydra cluster (see Figure 6). There is a faint optical sources near the IXO in the DSS image. The IXO is also positioned $\sim$10$"$ from a radio source (Andernach et al. 1988).
IXO 38
------
IXO 38 is located $\sim$1.5$'$ to the northwest of the center of the spiral galaxy NGC 3310 (Figure 7). See Zezas, Georgantopoulos, & Ward (1998) for more details of this X-ray source. See also RW2000.
IXO 39
------
IXO 39 is located near the nucleus of the edge-on starburst galaxy NGC 3628 (Figure 7). It is also known as NGC 3628 X-1. See Dahlem, Heckman & Fabbiano (1995) and Strickland et al. (2001) for discussions of this well-known IXO. See also Bregman & Glassgold (1982), FKT92, Yaqoob et al. (1995), RPS97, Dahlem, Weaver & Heckman (1998), and RW2000.
IXO 40
------
This X-ray object is positioned $\sim$5$'$ to the north of the elliptical galaxy NGC 3923 (Figure 7). It is coincident with an optical counterpart in the DSS image, but the counterpart is not listed in NED or Simbad.
IXO 41
------
IXO 41 is positioned $\sim$0.5$'$ to the west from the Irregular galaxy UGC 7009 (Figure 7).
IXO 42
------
IXO 42 is located in the disk of the edge-on spiral galaxy NGC 4088 (Figure 8). It is also known as NGC 4088 X-1 (RW2000).
IXOs 43 and 44
--------------
These two IXOs are positioned $\sim$4$'$ and $\sim$2$'$ west, respectively, of the Seyfert nucleus in NGC 4151 (see Figure 8). See Warwick, Done, & Smith (1995) for a more detailed discussions of these two objects. See also RW2000.
IXO 45
------
IXO 45 is positioned $\sim$3$'$ to the southeast of the elliptical galaxy NGC 4203 (Figure 8). It is also positioned near the QSO Ton 1480. See also RW2000.
IXO 46
------
This X-ray object is located in the disk of the spiral galaxy NGC 4254 (Figure 8). There is a bright optical counterpart visible in the DSS image, but it is not catalogued in NED or Simbad.
IXOs 47, 48 and 49
------------------
These three IXOs are located in the disk of the nearby spiral galaxy NGC 4321 (M100; see Figure 9). IXOs 47 and 48 are sources H13 and H14 (respectively) in the ROSAT HRI study by Immler, Pietsch, & Aschenbach (1998). See also RW2000.
IXOs 50 and 51
--------------
IXO 50 is a slightly resolved X-ray source located just east of the nucleus of the elliptical galaxy NGC 4374 (see Figure 9). IXO 51 is positioned $\sim$3.5$'$ to the northeast of NGC 4374. There is also an X-ray bright QSO near NGC 4374. See also FKT92 and Finoguenov & Jones (2001).
IXO 52
------
IXO 52 is positioned $\sim$2$'$ north of the elliptical galaxy NGC 4393 (Figure 9). There is a faint optical counterpart visible in the DSS image.
IXO 53
------
This IXO is located east of the dwarf galaxy NGC 4395 (see Figure 10). It is also known as NGC 4395 X-2 (Lira, Lawrence, & Johnson 2000). See also Iwasawa et al. (2000) and RW2000.
IXO 54
------
IXO 54 is positioned $\sim$3$'$ northeast of the edge-on spiral galaxy NGC 4438 (Figure 10). See also RW2000.
IXOs 55 $-$ 61
--------------
These seven IXOs are positioned in the halo of the elliptical galaxy NGC 4472 (Figures 10 and 11). See also Irwin & Sarazin (1996).
IXO 62
------
IXO 62 is located in the galaxy NGC 4485, which is a small companion to the spiral galaxy NGC 4490 (see Figure 11). Several HRI observations of IXO 64 show that it is variable by a factor of $\sim$2 (Table 2). There are several extended X-ray sources in the disk of NGC 4490 that have been resolved into individual point sources with Chandra (Roberts et al. 2002). See also Colbert & Mushotzky (1999) and RW2000.
IXOs 63 and 64
--------------
IXOs 63 and 64 are positioned in the halo of the elliptical galaxy NGC 4552 (see Figure 11).
IXO 65 and 66
-------------
These two IXOs are positioned in the spiral galaxy NGC 4559 (see Figure 11). IXO 65 is located $\sim$3$^\prime$ southwest of the galaxy center, while IXO 66 is located near the galaxy center. IXO 65 shows X-ray variability (Table 2) and has an optical counterpart in the DSS image. See also Vogler, Pietsch & Bertoldi (1997) and RW2000.
IXO 67
------
IXO 67 is positioned in the halo of the edge-on spiral galaxy NGC 4565, $\sim$1$'$ from the nucleus (see Figure 12). See Vogler, Pietsch, & Kahabka (1996) and Wu et al. (2002) for more discussions of IXO 66. See also Mizuno et al. (1999) and Makishima et al. (2000).
IXO 68
------
IXO 68 is positioned in the disk of the edge-on spiral galaxy NGC 4631 (see Figure 12). See Vogler & Pietsch (1996) for a more detailed discussion of the X-ray properties of these IXOs. Count rate measurements from several HRI observations of IXO 68 show that it is variable by a factor of $\sim$2. See also RPS97, Dahlem, Weaver & Heckman (1998), and RW2000.
IXOs 69 $-$ 71
--------------
These three IXOs are positioned in the halo of the elliptical galaxy NGC 4649 (M60; see Figure 12). Count rate measurements from two HRI observations of IXO 73 show variability of $\sim$50% (Table 2). See also FKT92.
IXOs 72 and 73
--------------
IXOs 72 and 73 are located in the irregular galaxy NGC 4861 (see Figure 12). IXO 73 varied by a factor of $\sim$5 between two separate HRI observations (Table 2). See also Stevens & Strickland (1998).
IXO 74
------
IXO 74 is positioned in the outer disk of the spiral galaxy NGC 5055 (Figure 13). It is also known as NGC 5055 X-2 (RW2000). This X-ray source is quite variable. It shows a factor of $\sim$2 variability between two HRI observations (Table 2). See also RPS97.
IXOs 75 and 76
--------------
These two IXOs are positioned in the halo/bulge of NGC 5128 (Cen A; see Figure 13). IXO 75 is coincident with a bright optical counterpart in the DSS image, but the counterpart is not catalogued in NED or Simbad. See Turner et al. (1997) for a more detailed discussion of these two X-ray sources. Count rates from multiple HRI observations of these two sources show variability by a factor of $\sim$2 (Table 2). See Steinle, Dennerl & Englhauser (200) for a detailed ROSAT analysis of these point sources.
IXO 77
------
IXO 77 is also known as NGC 5204 X-1 (e.g., Lira et al. 2000, Roberts et al. 2001). It is positioned near an H II region complex in the irregular galaxy NGC 5204 (Figure 13). See also Colbert & Mushotzky (1999) and RW2000.
IXOs 78 $-$ 81
--------------
These four IXOs are located in the spiral arms of NGC 5194 (M51; see Figure 13). See Ehle, Pietsch, & Beck (1995) and RPS97 for more detailed discussions of these IXOs. From the measured count rates from multiple HRI observations of these IXOs (Table 2), we find that IXO 79 varies by a factor of $\sim$3, while IXO 81 shows $\sim$50% variability. See also Palumbo et al. (1985), Fabbiano & Trinchieri (1987), FKT92, Marston et al. (1995), and RW2000.
IXO 82
------
This IXO is located in the outer disk of NGC 5236 (M83; see Figure 14). See Ehle et al. (1998) and Immler et al. (1999) for detailed X-ray studies of M83. The measured count rates for two HRI observations of this sources differ by a factor of $\sim$2 (Table 2). See also Lira et al. (2000).
IXO 83
------
IXO 83 is located in the outer disk of NGC 5457 (M101; see Figure 14). See RPS97 and Wang, Immler & Pietsch (1999) for ROSAT analyses of M101. IXO 83 shows $\sim$50% variability between four difference HRI observations (Table 2). See also RW2000 and Lira et al. (2000).
IXO 84
------
This IXO is positioned near NGC 5774 and NGC 5775 in the NGC 5775 group (see Figure 14). See also RW2000.
IXO 85
------
IXO 85 is located near the center of the face-on spiral galaxy NGC 6946 (see Figure 14). There are many very luminous X-ray SNRs in NGC 6946, although IXO 85 is not classified as such. See Schlegel, Blair & Fesen (2000a) for details. It shows variability by a factor of $\sim$2.5 between two different HRI observations (Table 2). See also Colbert & Mushotkzy (1999), RW2000, and Lira et al. (2000).
IXO 86
------
IXO 86 is located in the outer disk of the spiral galaxy NGC 7314 (see Figure 15). See also Radecke (1997).
IXO 87
------
This X-ray object is located in the disk of the Seyfert galaxy NGC 7590 (see Figure 15). See also Ueda et al. (2001).
Discussion and Summary
======================
We have found 87 IXOs in 54 different galaxies. We find the following statistical results from perusal of the optical images and from the information listed in Table 1.
IXOs are found both in spiral and elliptical galaxies, although elliptical galaxies have a higher number of IXOs per galaxy (mean of 2.3 per galaxy for T-type $<$ 0, [*vs.*]{} 1.6 for T-type $\ge$ 0). For example, the elliptical galaxy NGC 1316 (Fornax A) has an astounding 7 IXOs in its halo.
Nineteen of the 87 IXOs are positioned near faint optical sources in the DSS images, and many more are located in the disks of spiral galaxies, or are known from more recent optical observations to have counterparts (e.g., NGC 5204 X-1, Roberts et al. 2001; see also Pakull et al. 2002). Follow-up studies of these potential optical counterparts using [*Chandra*]{} X-ray positions would be very useful for understanding the nature of IXOs. IXOs are, in general, quite variable: 51% (20 of 39) of the objects with two or more HRI observations show $\gapprox$50% variability between the observations.
We note here that the IXOs in this catalog are selected strictly by their X-ray luminosity and proximity to nearby RC3 galaxies. It is possible that some of the objects listed in the catalog are background quasars, as can be deduced from the number of QSOs seen in the finding charts (Figures 1 $-$ 15). Thus, this catalog should be viewed as a list of IXO [*candidates*]{}, especially for objects positioned at large projected radii from the galaxy nucleus. On the other hand, many of the X-ray sources, such as those in NGC 1313 and NGC 5194, for example, are very likely associated with their host galaxies.
Although an overabundance of IXOs are often found in starburst galaxies (e.g. Roberts et al. 2002), we have shown here that there is a significant population (35 of the 87) of IXOs in the halos of elliptical galaxies, where there is no recent star formation. This suggests that IXOs in elliptical galaxies may not be merely young, high-mass XRBs with mildly beamed X-ray emission (e.g., King et al. 2001). A more extensive analysis of the properties of these ROSAT IXOs, together with a study of their host galaxies, will be presented in a later work.
The current version and future versions of this catalog are available on the World Wide Web (http://www.xassist.org/ixocat\_hri).
The Digitized Sky Survey (DSS) was produced at the Space Telescope Science Institute under U.S. Government grant NAG W-2166. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This research has also made use of the SIMBAD database, operated at CDS, Strasbourg, France. EJMC acknowledges support from NASA LTSA grant NAG 59878. We thank the anonymous referee and Manfred Pakull for useful comments.
Andernach, H., Tie, H., Sievers, A., Teuter, H.-P., Junkes, N., & Wielebinski, R. 1988, A&AS, 73, 265 Blanton, E. L., Sarazin, C. L., & Irwin, J. A. 2001, ApJ, 552, 106 Brandt, W. N., Halpern, J. P., & Iwasawa, K. 1996, MNRAS, 281, 687 Bregman, J., & Glassgold, A. 1982, ApJ, 263, 564 Colbert, E. J. M., & Mushotzky, R. F. 1999, ApJ, 519, 89 Colbert, E. J. M., Petre, R., Schlegel, E. M., & Ryder, S. D. 1995, , 446, 177 Dahlem, M., Hartner, G. D., & Junkes, N. 1994, ApJ, 432, 598 Dahlem, M., Heckman, T. M., & Fabbiano, G. 1995, ApJ. 442, L49 Dahlem, M., Weaver, K. A., & Heckman, T. M. 1998, ApJS, 118, 401 \[De Naray et al. 2000[den00]{} De Naray, P., Brandt, W., Halpern, J., & Iwasawa, K. 2000, AJ, 119, 612 de Vaucouleurs, G., de Vaucouleurs, A., Corwin, H. G., Buta, R. J., Paturel, G., & Fouque, P. 1991, Third Reference Catalog of Bright Galaxies (RC3; New York: Springer-Verlag) Dickey, J. M., & Lockman, F. J. 1990, ARA&A, 28, 215 Ehle, M., Pietsch, W., & Beck, R. 1995, A&A, 295, 289 Ehle, M., Pietsch, W., Beck, R., & Klein, U. 1998, A&A, 329, 39 Ezoe, Y., Iyomoto, N., & Makishima, K. 2000, PASJ, 53, 69 Fabbiano, G. 1988, ApJ, 325, 544 Fabbiano, G. 1989, ARA&A, 27, 87 Fabbiano, G., & Trinchieri, G. 1987, ApJ, 315, 46 Finoguenov, A., & Jones, C. 2001, ApJ, 547, L107 Fullmer, L., & Lonsdale, C. 1989, Cataloged Galaxies and Quasars in the IRAS Survey (JPL Pub. D-1932, Version 2, Appendix B) Immler, S., Pietsch, W., & Aschenbach, B. 1998, A&A, 331, 601 Immler, S., Vogler, A., Ehle, M., & Pietsch, W., 1999, A&A, 352, 415 Immler, S., & Wang, Q. 2001, ApJ, 554, 202 Irwin, J. A., & Sarazin, C. L. 1996, ApJ, 471, 683 Ishisaki, Y. et al. 1996, PASJ, 48, 237 Iwasawa, K., Fabian, A., Almaini, O., Lira, P., Lawrence, A., Hayashida, K., & Inoue, H. 2000, MNRAS, 318, 879 Jones, C., Stern, C., Forman, W., Breen, J., David, L., Tucker, W., & Franx, M. 1997, ApJ, 482, 143 Kim, D.-W., Fabbiano, G., & Mackie, G. 1998, ApJ, 497, 699 King, A. R., Davies, M. B., Ward, M. J., Fabbiano, G., & Elvis, M. 2001, ApJ, 552, L109 Koerding, E., Falcke, H., & Markoff, S. 2002, A&A, 382, L13 La Perola, V., Peres, G., Fabbiano, G., Kim, D. W., & Bocchino, F. 2001, ApJ, 556, 47 Lira, P., Lawrence, A., & Johnson, R. 2000, MNRAS, 319, 17 Marston, A. P., Elmegreen, D., Elmegreen, B., Forman, W., Jones, C., & Flanagan, K. 1995, ApJ, 438, 663 Miller, B. W. 1995, ApJ, 446, L75 Miller, M. C., & Hamilton, D. P. 2002, MNRAS, 330, 232 Miller, S., Schlegel, E., Petre, R., & Colbert, E. 1998, AJ, 116, 1657 Mirabel, F. 1998, in 19th Texas Symposium on Relativistic Astrophysics and Cosmology (Paris: CEA Saclay), ed. Paul, Montmerle, & Aubourg, p. 265 Miyaji, T., Lehmann, I., & Hasinger, G. 2001, AJ, 121, 3041 Morse, J. A., Wilson, A. S., Elvis, M., & Weaver, K. A. 1995, ApJ, 439, 121 Pakull, M. W., & Mirioni, L. 2002, in “New Visions of the X-ray Universe in the XMM-Newton and Chandra Era,” (ESTEC: The Netherlands) (astro-ph/0202488) Palumbo, G. G. C., Fabbiano, G., Fransson, C., & Trinchieri, G. 1985, ApJ, 298, 259 Paturel, G., Theureau, G., Fouque, P, Terry, J. N., Musella, I, & Ekholm, T. 2002, A&A, 383, 398 Pence, W. D., & Rots, A. H. 1997, ApJ, 478, 107 Petre, R., Okada, K., Mihara, T., Makishima, K., & Colbert, E. J. M. 1994, PASJ, 46, L115 Radecke, H.-D. 1997, A&A, 319, 18 Read, A. M., Ponman, T. J., & Strickland, D. K. 1997, MNRAS, 286, 626 Roberts, T. P., Goad, M. R., Ward, M. J., Warwick, R. S., & Lira, P. 2002, in “New Visions of the X-ray Universe in the XMM-Newton and Chandra Era,” (ESTEC: The Netherlands) (astro-ph/0202017) Roberts, T. P., Goad, M. R., Ward, M. J., Warwick, R. S., O’Brien, P. T., Lira, P., & Hands, A. D. P. 2001, MNRAS, 325, L7 Roberts, T., & Warwick, R. 2000, MNRAS, 315, 98 Schlegel, E., Blair, W., & Fesen, R., 2000, AJ, 120, 791 Schlegel, E., Petre, R., Colbert, E., & Miller, S. 2000, AJ, 120, 2373 Singh, K. P. 1999, MNRAS, 309, 991 Steinle, H., Dennerl, K., & Englhauser, J. 2000, A&A, 357, L57 Stevens, I. R., & Strickland, D. K. 1998, MNRAS, 294, 523 Strickland, D. K., Colbert, E. J. M., Heckman, T. M., Weaver, K. A., Dahlem, M., & Stevens, I. R. 2001, ApJ, 560, 707 Tanaka, Y., & Lewin, W. H. G. 1995, in X-ray Binaries (ed. Lewin, Paradijs, & van den Heuvel), Ch. 3 Taniguchi, Y., Shioya, Y., Tsuru, T., & Ikeuchi, S. 2000, PASJ, 52, 533 Tremaine, S. D., Ostriker, J. P., & Spitzer, L., Jr. 1975, ApJ, 196, 407 Tully, R. B. 1988, Nearby Galaxies Catalog (Cambridge: Cambridge Univ. Press) Turner, T. J., George, I. M., Mushotzky, R. F., & Nandra, K. 1997, ApJ, 475, 118 Turner, T. J., Urry, C. M., & Mushotzky, R. F. 1993, ApJ, 418, 653 Vogler, A., & Pietsch, W. 1996, A&A, 311, 35 Vogler, A., Pietsch, W., & Bertoldi, F. 1997, A&A, 318, 768 Vogler, A., Pietsch, W., & Kahabka, P. 1996, A&A, 305, 74 Wang, Q. D. 2002, MNRAS, in press (astro-ph/0201230) Wang, Q., Immler, S., & Pietsch, W. 1999, ApJ, 523, 121 Warwick, R. S., Done, C., & Smith, D. A. 1995, MNRAS, 275, 1003 Wu, H., Xue, S. J., Xia, X. Y., Deng, Z. G., & Mao, S. 2002, ApJ, in press (astro-ph/0204359) Yaqoob, T., Serlemitsos, P. J., Ptak, A., Mushotzky, R., Kunieda, H., & Terashima, Y. 1995, ApJ, 455, 508 Zezas, A. L., Georgantopoulos, I., & Ward, M. J. 1998, MNRAS, 301, 915 Zezas, A. L., Georgantopoulos, I., & Ward, M. J. 1999, MNRAS, 308, 302
[rrrclcccrlrr]{} 1 & 01 52 49.7 & $-$13 42 11 & 39.1 & NGC 720 & $-$5.0 & & & 22.88 & 1 & 22.5 & 31.1\
2 & 01 53 02.9 & $-$13 46 53 & 39.7 & NGC 720 & $-$5.0 & & & 22.88 & 1 & 17.3 & 31.1\
3 & 02 22 45.8 & $+$42 25 55 & 39.2 & NGC 891 & $+$3.0 & & $+$1.0 & 9.60 & 2 & 15.7 & 37.7\
4 & 02 40 25.8 & $-$08 24 30 & 40.1 & NGC 1042 & $+$6.0 & & $+$0.0 & 18.31 & 1 & 8.5 & 24.9\
5 & 02 43 38.3 & $+$01 24 13 & 39.1 & NGC 1073 & $+$5.0 & & $-$0.1 & 16.15 & 1 & 8.0 & 23.0\
6 & 03 17 13.6 & $-$41 10 32 & 39.2 & NGC 1291 & $+$0.0 & & $-$0.9 & 8.60 & 2 & 10.3 & 24.4\
7 & 03 18 19.9 & $-$66 29 07 & 39.8 & NGC 1313 & $+$7.0 & & $+$0.0 & 3.70 & 2 & 0.9 & 9.8\
8 & 03 18 22.2 & $-$66 36 01 & 39.4 & NGC 1313 & $+$7.0 & & $+$0.0 & 3.70 & 2 & 6.7 & 9.8\
9 & 03 21 59.8 & $-$37 05 06 & 40.0 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 77.3 & 83.6\
10 & 03 22 40.4 & $-$37 16 41 & 39.7 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 29.4 & 83.6\
11 & 03 22 51.3 & $-$37 09 47 & 39.5 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 23.0 & 83.6\
12 & 03 22 57.3 & $-$37 15 57 & 39.4 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 32.5 & 83.6\
13 & 03 23 14.4 & $-$37 16 51 & 39.4 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 54.7 & 83.6\
14 & 03 23 28.8 & $-$37 10 09 & 39.5 & NGC 1316 & $-$2.0 & $-$1.0 & $-$0.8 & 23.91 & 1 & 67.3 & 83.6\
15 & 03 33 11.9 & $-$36 11 35 & 39.8 & NGC 1365 & $+$3.0 & 1.8 & $+$0.9 & 22.16 & 1 & 38.5 & 72.3\
16 & 03 33 13.0 & $-$36 11 53 & 39.8 & NGC 1365 & $+$3.0 & 1.8 & $+$0.9 & 22.16 & 1 & 38.5 & 72.3\
17 & 03 38 31.8 & $-$35 26 01 & 39.4 & NGC 1399 & $-$5.0 & & & 19.29 & 1 & 6.2 & 38.8\
18 & 03 38 32.7 & $-$35 27 01 & 39.7 & NGC 1399 & $-$5.0 & & & 19.29 & 1 & 4.2 & 38.8\
19 & 03 38 41.2 & $-$35 31 35 & 40.4 & NGC 1399 & $-$5.0 & & & 19.29 & 1 & 29.4 & 38.8\
20 & 03 38 51.7 & $-$35 26 43 & 39.6 & NGC 1399 & $-$5.0 & & & 19.29 & 1 & 26.0 & 38.8\
21 & 03 40 12.3 & $-$35 37 36 & 39.9 & NGC 1427A & $+$10.0 & & $-$0.2 & 26.99 & 1 & 5.0 & 18.4\
22 & 03 45 55.7 & $+$68 04 58 & 39.2 & IC 342 & $+$6.0 & & $+$0.5 & 3.90 & 2 & 5.8 & 24.2\
23 & 04 15 51.9 & $-$55 49 01 & 39.4 & NGC 1553 & $-$2.0 & $-$1.0 & $-$1.3 & 14.40 & 1 & 14.2 & 18.7\
24 & 04 19 55.6 & $-$54 56 37 & 39.7 & NGC 1566 & $+$4.0 & 1.0 & $+$0.3 & 19.95 & 1 & 4.4 & 48.3\
25 & 04 45 30.9 & $-$59 14 55 & 39.1 & NGC 1672 & $+$3.0 & 2.0 & $+$0.8 & 18.00 & 1 & 7.6 & 34.6\
26 & 04 45 33.6 & $-$59 14 41 & 39.6 & NGC 1672 & $+$3.0 & 2.0 & $+$0.8 & 18.00 & 1 & 5.9 & 34.6\
27 & 04 45 53.2 & $-$59 14 57 & 39.5 & NGC 1672 & $+$3.0 & 2.0 & $+$0.8 & 18.00 & 1 & 7.4 & 34.6\
28 & 05 05 11.1 & $-$37 58 48 & 39.3 & NGC 1792 & $+$4.0 & & $+$0.8 & 16.33 & 1 & 3.8 & 24.9\
29 & 05 41 43.3 & $+$69 20 46 & 40.5 & NGC 1961 & $+$5.0 & $-$1.0 & $+$0.6 & 52.40 & 1 & 41.6 & 69.7\
30 & 06 15 15.3 & $+$71 02 04 & 40.4 & Mrk 3 & $-$2.0 & 2.0 & $+$0.8 & 53.31 & 1 & 26.1 & 28.2\
31 & 08 19 30.2 & $+$70 42 18 & 40.2 & Holmberg II & $+$10.0 & & $-$0.5 & 4.50 & 2 & 2.7 & 10.4\
32 & 09 10 19.9 & $+$07 06 00 & 39.2 & NGC 2775 & $+$2.0 & & $-$0.1 & 18.05 & 1 & 19.4 & 22.4\
33 & 09 10 27.0 & $+$06 59 10 & 39.5 & NGC 2775 & $+$2.0 & & $-$0.1 & 18.05 & 1 & 18.6 & 22.4\
34 & 09 57 54.4 & $+$69 03 43 & 40.0 & NGC 3031 & $+$2.0 & 1.8 & $-$0.8 & 3.55 & 3 & 13.0 & 27.8\
35 & 10 23 26.0 & $+$19 56 20 & 39.2 & NGC 3226 & $-$5.0 & $-$1.0 & & 17.63 & 1 & 12.5 & 16.2\
36 & 10 27 37.6 & $-$43 53 08 & 39.8 & NGC 3256 & $+$99.0 & & $+$1.7 & 37.08 & 1 & 30.0 & 41.0\
37 & 10 37 39.6 & $-$27 05 23 & 40.9 & IC 2597 & $-$4.0 & & & 40.09 & 1 & 21.7 & 30.0\
38 & 10 38 43.4 & $+$53 31 06 & 39.5 & NGC 3310 & $+$4.0 & & $+$0.8 & 18.70 & 2 & 5.7 & 16.8\
39 & 11 20 15.5 & $+$13 35 25 & 39.5 & NGC 3628 & $+$3.0 & & $+$0.8 & 7.70 & 2 & 0.5 & 33.1\
40 & 11 50 57.9 & $-$28 44 02 & 39.9 & NGC 3923 & $-$5.0 & & & 22.24 & 1 & 28.8 & 38.1\
41 & 12 01 40.4 & $+$62 19 58 & 39.6 & UGC 7009 & $+$10.0 & & & 14.93 & 1 & 2.4 & 7.0\
42 & 12 05 32.6 & $+$50 32 46 & 39.5 & NGC 4088 & $+$4.0 & & $+$0.8 & 17.00 & 2 & 2.5 & 28.5\
43 & 12 10 07.9 & $+$39 23 12 & 39.3 & NGC 4151 & $+$2.0 & 1.5 & & 20.30 & 2 & 29.6 & 37.3\
44 & 12 10 22.6 & $+$39 23 16 & 39.0 & NGC 4151 & $+$2.0 & 1.5 & & 20.30 & 2 & 13.8 & 37.3\
45 & 12 15 15.7 & $+$33 10 20 & 39.0 & NGC 4203 & $-$3.0 & $-$1.0 & $-$0.4 & 14.48 & 1 & 11.2 & 14.3\
46 & 12 18 56.1 & $+$14 24 18 & 40.5 & NGC 4254 & $+$5.0 & & $+$0.5 & 32.09 & 1 & 16.9 & 50.1\
47 & 12 22 43.0 & $+$15 44 00 & 39.5 & NGC 4321 & $+$4.0 & & $+$0.3 & 21.15 & 1 & 37.7 & 45.6\
48 & 12 22 48.4 & $+$15 43 09 & 39.2 & NGC 4321 & $+$4.0 & & $+$0.3 & 21.15 & 1 & 39.6 & 45.6\
49 & 12 23 08.6 & $+$15 51 23 & 39.3 & NGC 4321 & $+$4.0 & & $+$0.3 & 21.15 & 1 & 23.3 & 45.6\
50 & 12 25 05.1 & $+$12 53 06 & 40.0 & NGC 4374 & $-$5.0 & & $-$1.2 & 16.80 & 2 & 1.8 & 31.6\
51 & 12 25 15.4 & $+$12 56 02 & 39.6 & NGC 4374 & $-$5.0 & & $-$1.2 & 16.80 & 2 & 19.5 & 31.6\
52 & 12 25 16.8 & $-$39 43 26 & 40.0 & NGC 4373 & $-$2.9 & & & 45.28 & 1 & 29.1 & 44.6\
53 & 12 26 01.8 & $+$33 31 34 & 39.0 & NGC 4395 & $+$9.0 & 1.8 & $-$0.4 & 3.60 & 2 & 2.9 & 13.8\
54 & 12 27 57.5 & $+$13 02 30 & 39.9 & NGC 4438 & $+$0.0 & $-$1.0 & $-$0.1 & 16.80 & 2 & 17.0 & 41.6\
55 & 12 29 13.1 & $+$07 57 40 & 39.1 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 42.0 & 50.0\
56 & 12 29 22.3 & $+$07 53 31 & 39.4 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 43.0 & 50.0\
57 & 12 29 23.9 & $+$07 54 00 & 39.8 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 40.0 & 50.0\
58 & 12 29 27.9 & $+$08 06 34 & 39.4 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 39.3 & 50.0\
59 & 12 29 35.1 & $+$08 09 37 & 39.3 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 49.1 & 50.0\
60 & 12 29 39.9 & $+$07 53 32 & 39.9 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 32.4 & 50.0\
61 & 12 30 06.7 & $+$08 02 05 & 39.2 & NGC 4472 & $-$5.0 & 2.0 & & 16.80 & 2 & 26.5 & 50.0\
62 & 12 30 30.8 & $+$41 41 45 & 39.3 & NGC 4485 & $+$10.0 & & & 9.30 & 2 & 0.8 & 6.2\
63 & 12 35 19.2 & $+$12 33 17 & 39.5 & NGC 4552 & $-$5.0 & & & 16.80 & 2 & 24.7 & 25.1\
64 & 12 35 29.4 & $+$12 31 10 & 39.3 & NGC 4552 & $-$5.0 & & & 16.80 & 2 & 16.7 & 25.1\
65 & 12 35 51.7 & $+$27 56 04 & 39.9 & NGC 4559 & $+$6.0 & & $-$0.0 & 9.70 & 2 & 5.8 & 30.2\
66 & 12 35 58.7 & $+$27 57 42 & 39.7 & NGC 4559 & $+$6.0 & & $-$0.0 & 9.70 & 2 & 0.6 & 30.2\
67 & 12 36 17.4 & $+$25 58 56 & 40.1 & NGC 4565 & $+$3.0 & 1.9 & $+$0.0 & 16.36 & 1 & 3.5 & 75.4\
68 & 12 41 55.5 & $+$32 32 14 & 39.1 & NGC 4631 & $+$7.0 & & $+$0.6 & 6.90 & 2 & 5.2 & 31.1\
69 & 12 43 36.8 & $+$11 30 06 & 39.4 & NGC 4649 & $-$5.0 & & & 14.85 & 1 & 12.9 & 32.0\
70 & 12 44 07.2 & $+$11 35 25 & 39.3 & NGC 4649 & $-$5.0 & & & 14.85 & 1 & 30.4 & 32.0\
71 & 12 44 09.2 & $+$11 33 36 & 39.7 & NGC 4649 & $-$5.0 & & & 14.85 & 1 & 30.7 & 32.0\
72 & 12 59 00.9 & $+$34 50 42 & 39.8 & NGC 4861 & $+$9.0 & & & 17.80 & 2 & 5.8 & 20.6\
73 & 12 59 02.0 & $+$34 51 11 & 40.3 & NGC 4861 & $+$9.0 & & & 17.80 & 2 & 3.0 & 20.6\
74 & 13 15 19.6 & $+$42 02 58 & 39.6 & NGC 5055 & $+$4.0 & $-$1.0 & $+$0.4 & 7.20 & 2 & 11.7 & 26.4\
75 & 13 25 07.4 & $-$43 04 06 & 39.3 & NGC 5128 & $-$2.0 & 2.0 & $+$0.3 & 4.90 & 2 & 7.2 & 36.6\
76 & 13 25 19.9 & $-$43 03 13 & 39.9 & NGC 5128 & $-$2.0 & 2.0 & $+$0.3 & 4.90 & 2 & 4.0 & 36.6\
77 & 13 29 38.7 & $+$58 25 06 & 39.5 & NGC 5204 & $+$9.0 & & $+$0.0 & 4.80 & 2 & 0.4 & 7.0\
78 & 13 29 40.2 & $+$47 12 41 & 39.0 & NGC 5194 & $+$4.0 & 2.0 & $+$0.3 & 8.40 & 2 & 5.8 & 25.1\
79 & 13 29 43.6 & $+$47 11 36 & 39.1 & NGC 5194 & $+$4.0 & 2.0 & $+$0.3 & 8.40 & 2 & 4.1 & 25.1\
80 & 13 30 01.3 & $+$47 13 42 & 39.4 & NGC 5194 & $+$4.0 & 2.0 & $+$0.3 & 8.40 & 2 & 5.7 & 25.1\
81 & 13 30 07.2 & $+$47 11 02 & 39.3 & NGC 5194 & $+$4.0 & 2.0 & $+$0.3 & 8.40 & 2 & 6.1 & 25.1\
82 & 13 37 20.1 & $-$29 53 46 & 39.2 & NGC 5236 & $+$5.0 & & $+$0.3 & 4.70 & 2 & 6.3 & 17.6\
83 & 14 04 14.3 & $+$54 26 05 & 39.1 & NGC 5457 & $+$6.0 & & $+$0.3 & 5.40 & 2 & 16.3 & 45.3\
84 & 14 53 44.7 & $+$03 33 30 & 39.4 & NGC 5775 & $+$5.0 & & $+$1.1 & 22.41 & 1 & 21.8 & 27.2\
85 & 20 35 00.3 & $+$60 09 05 & 39.0 & NGC 6946 & $+$6.0 & & $+$0.6 & 5.50 & 2 & 1.7 & 18.4\
86 & 22 35 48.3 & $-$26 01 26 & 39.4 & NGC 7314 & $+$4.0 & 1.9 & $+$0.3 & 18.96 & 1 & 9.4 & 25.2\
87 & 23 18 55.9 & $-$42 13 53 & 39.8 & NGC 7590 & $+$4.0 & 2.0 & & 21.28 & 1 & 2.7 & 16.7\
[rlcrrr]{} 1 & rh600472n00 & 0.50$\pm$0.02 & 3.9 & 20.2 & 1.38\
2 & rh600472n00 & 1.77$\pm$0.08 & 13.3 & 20.2 & 4.89\
3 & rh500501n00 & 2.13$\pm$0.13 & 10.7 & 20.9 & 1.52\
& rh600690n00 & 1.41$\pm$0.05 & 12.8 & & 1.01\
4 & rh600469a01 & 6.04$\pm$0.36 & 17.2 & 20.5 & 12.51\
5 & rh600999n00 & 0.65$\pm$0.05 & 4.2 & 20.6 & 1.13\
6 & rh600828n00 & 3.81$\pm$0.30 & 14.6 & 20.3 & 1.59\
7 & rh400065n00 & 68.50$\pm$3.29 & 83.6 & 20.6 & 6.11\
& rh500403n00 & 63.20$\pm$2.02 & 149.5 & & 5.64\
& rh500403a01 & 29.70$\pm$0.85 & 98.2 & & 2.65\
& rh500404n00 & 64.80$\pm$1.43 & 187.1 & & 5.78\
& rh500404a01 & 36.90$\pm$1.26 & 88.3 & & 3.29\
& rh500492n00 & 62.90$\pm$1.53 & 182.1 & & 5.61\
& rh500550n00 & 28.20$\pm$0.93 & 84.5 & & 2.52\
& rh600505n00 & 49.20$\pm$1.36 & 161.8 & & 4.39\
& rh600505a01 & 62.20$\pm$1.62 & 161.8 & & 5.55\
8 & rh400065n00 & 18.70$\pm$1.62 & 43.9 & 20.6 & 1.67\
& rh500403n00 & 25.30$\pm$1.25 & 73.6 & & 2.26\
& rh500403a01 & 11.50$\pm$0.50 & 49.0 & & 1.03\
& rh500404n00 & 17.50$\pm$0.69 & 65.2 & & 1.56\
& rh500404a01 & 15.00$\pm$0.74 & 42.4 & & 1.34\
& rh500492n00 & 16.40$\pm$0.74 & 76.6 & & 1.46\
& rh500550n00 & 15.70$\pm$0.67 & 58.9 & & 1.40\
& rh600505n00 & 8.45$\pm$0.46 & 33.2 & & 0.75\
9 & rh600255a01 & 3.30$\pm$0.23 & 9.6 & 20.3 & 10.39\
10 & rh600255a01 & 1.51$\pm$0.11 & 9.7 & 20.3 & 4.76\
11 & rh600255n00 & 0.97$\pm$0.11 & 3.7 & 20.3 & 3.05\
& rh600255a01 & 1.01$\pm$0.08 & 5.5 & & 3.18\
12 & rh600255a01 & 0.75$\pm$0.06 & 3.1 & 20.3 & 2.36\
13 & rh600255a01 & 0.80$\pm$0.06 & 3.9 & 20.3 & 2.53\
14 & rh600255a01 & 1.12$\pm$0.09 & 3.3 & 20.3 & 3.53\
15 & rh701297n00 & 2.41$\pm$0.30 & 5.4 & 20.1 & 6.11\
& rh701297a02 & 2.64$\pm$0.32 & 4.2 & & 6.69\
16 & rh701297n00 & 2.56$\pm$0.30 & 6.6 & 20.1 & 6.48\
& rh701297a02 & 2.14$\pm$0.28 & 6.7 & & 5.42\
17 & rh600940n00 & 1.08$\pm$0.03 & 3.1 & 20.1 & 2.06\
& rh600831a01 & 1.21$\pm$0.03 & 5.0 & & 2.31\
18 & rh600831n00 & 2.53$\pm$0.06 & 3.6 & 20.1 & 4.82\
& rh600831a01 & 2.34$\pm$0.06 & 3.8 & & 4.46\
& rh600940n00 & 2.76$\pm$0.07 & 3.5 & & 5.26\
19 & rh600220n00 & 12.00$\pm$0.96 & 26.8 & 20.1 & 22.88\
& rh600256n00 & 9.94$\pm$0.88 & 22.3 & & 18.95\
& rh600831n00 & 3.86$\pm$0.14 & 22.6 & & 7.36\
& rh600831a01 & 7.75$\pm$0.21 & 53.7 & & 14.78\
& rh600940n00 & 3.70$\pm$0.13 & 25.7 & & 7.05\
20 & rh600940n00 & 1.67$\pm$0.07 & 7.5 & 20.1 & 3.18\
& rh600831a01 & 2.33$\pm$0.08 & 14.2 & & 4.44\
21 & rh600940n00 & 1.96$\pm$0.08 & 3.5 & 20.1 & 7.34\
22 & rh600022n00 & 7.38$\pm$0.47 & 26.7 & 21.5 & 1.52\
23 & rh600479n00 & 2.54$\pm$0.14 & 16.5 & 20.2 & 2.76\
& rh600490a01 & 1.88$\pm$0.12 & 11.5 & & 2.04\
24 & rh600860n00 & 2.05$\pm$0.15 & 4.4 & 20.1 & 4.18\
& rh600962n00 & 2.39$\pm$0.10 & 8.4 & & 4.87\
25 & rh601011n00 & 0.63$\pm$0.04 & 3.0 & 20.4 & 1.17\
26 & rh601011n00 & 1.92$\pm$0.12 & 7.0 & 20.4 & 3.57\
& rh701022n00 & 0.82$\pm$0.06 & 4.5 & & 1.53\
27 & rh601011n00 & 1.63$\pm$0.10 & 8.0 & 20.4 & 3.03\
& rh701022n00 & 1.86$\pm$0.14 & 6.8 & & 3.46\
28 & rh600433n00 & 1.34$\pm$0.08 & 6.3 & 20.4 & 2.11\
29 & rh600499n00 & 1.62$\pm$0.07 & 14.3 & 20.9 & 35.23\
30 & rh700339n00 & 1.20$\pm$0.12 & 4.9 & 20.9 & 27.10\
31 & rh600745n00 & 118.00$\pm$3.63 & 136.5 & 20.5 & 15.10\
32 & rh600826a01 & 0.77$\pm$0.05 & 3.8 & 20.6 & 1.66\
33 & rh600826a01 & 1.35$\pm$0.08 & 7.8 & 20.6 & 2.90\
34 & rh600247n00 & 45.40$\pm$1.27 & 71.3 & 20.6 & 3.79\
& rh600247a01 & 65.10$\pm$1.70 & 84.4 & & 5.44\
& rh600739n00 & 35.50$\pm$1.27 & 49.0 & & 2.96\
& rh600740n00 & 57.40$\pm$1.66 & 65.7 & & 4.79\
& rh600881n00 & 47.50$\pm$1.74 & 60.3 & & 3.97\
& rh600882n00 & 61.60$\pm$1.76 & 67.4 & & 5.14\
& rh600882a01 & 109.00$\pm$4.54 & 56.3 & & 9.10\
& rh601001n00 & 112.00$\pm$2.39 & 126.4 & & 9.36\
& rh601002n00 & 39.10$\pm$1.33 & 48.6 & & 3.27\
& rh601095n00 & 41.30$\pm$1.63 & 31.4 & & 3.45\
35 & rh701299n00 & 0.88$\pm$0.06 & 4.7 & 20.3 & 1.55\
36 & rh701983n00 & 0.59$\pm$0.03 & 4.9 & 21.0 & 6.73\
37 & rh800741n00 & 7.07$\pm$0.35 & 33.2 & 20.7 & 79.12\
38 & rh600685a01 & 1.89$\pm$0.10 & 8.7 & 20.1 & 3.27\
39 & rh700009n00 & 10.40$\pm$0.72 & 42.6 & 20.3 & 3.52\
40 & rh600679n00 & 1.78$\pm$0.10 & 11.2 & 20.8 & 6.45\
& rh600679a01 & 2.15$\pm$0.16 & 10.4 & & 7.79\
41 & rh500499n00 & 3.08$\pm$0.17 & 8.9 & 20.3 & 3.83\
42 & rh500391n00 & 1.91$\pm$0.18 & 4.6 & 20.3 & 3.08\
43 & rh701707n00 & 0.88$\pm$0.03 & 9.2 & 20.3 & 2.02\
44 & rh701707n00 & 0.47$\pm$0.02 & 4.9 & 20.3 & 1.09\
45 & rh600221n00 & 0.98$\pm$0.08 & 4.1 & 20.1 & 1.02\
46 & rh600972n00 & 5.08$\pm$0.32 & 17.9 & 20.4 & 31.32\
47 & rh500542n00 & 1.03$\pm$0.08 & 4.9 & 20.4 & 2.68\
& rh600731n00 & 1.11$\pm$0.07 & 7.1 & & 2.89\
48 & rh600731n00 & 0.62$\pm$0.04 & 3.4 & 20.4 & 1.62\
49 & rh500542n00 & 0.77$\pm$0.06 & 3.5 & 20.4 & 2.00\
50 & rh600493n00\
& rh700192n00 & 6.56$\pm$0.38 & 3.1 & 20.4 & 10.97\
51 & rh600493n00 & 2.13$\pm$0.15 & 8.6 & 20.4 & 3.56\
52 & rh600976n00 & 0.65$\pm$0.04 & 3.6 & 20.8 & 9.94\
53 & rh702725n00 & 15.20$\pm$1.00 & 48.3 & 20.1 & 1.01\
54 & rh600608n00 & 5.13$\pm$0.33 & 18.5 & 20.4 & 8.62\
& rh701657n00 & 3.95$\pm$0.52 & 6.1 & & 6.64\
55 & rh600216a01 & 0.88$\pm$0.07 & 3.5 & 20.2 & 1.34\
56 & rh600216a01 & 1.69$\pm$0.12 & 5.8 & 20.2 & 2.56\
57 & rh600216n00 & 4.29$\pm$0.47 & 7.4 & 20.2 & 6.48\
& rh600216a01 & 4.52$\pm$0.26 & 16.7 & & 6.83\
58 & rh600216a01 & 1.70$\pm$0.12 & 6.3 & 20.2 & 2.57\
59 & rh600216a01 & 1.47$\pm$0.12 & 4.1 & 20.2 & 2.22\
60 & rh600216n00 & 5.63$\pm$0.65 & 11.3 & 20.2 & 8.52\
& rh600216a01 & 5.13$\pm$0.28 & 21.3 & & 7.76\
61 & rh600216a01 & 1.13$\pm$0.08 & 3.0 & 20.2 & 1.71\
62 & rh600697n00 & 4.45$\pm$0.51 & 4.4 & 20.3 & 2.10\
& rh600855n00 & 2.52$\pm$0.15 & 11.7 & & 1.19\
& rh600855a01 & 1.87$\pm$0.14 & 9.8 & & 0.88\
63 & rh600491n00\
& rh600491a01 & 2.03$\pm$0.14 & 8.5 & 20.4 & 3.38\
64 & rh600491n00 & 1.10$\pm$0.10 & 3.0 & 20.4 & 1.83\
65 & rh600861n00 & 15.20$\pm$0.63 & 91.8 & 20.2 & 7.49\
66 & rh600861n00 & 9.18$\pm$0.31 & 48.9 & 20.2 & 4.52\
67 & rh600691n00 & 8.63$\pm$0.97 & 16.0 & 20.1 & 11.76\
68 & rh600193a00 & 2.40$\pm$0.30 & 7.0 & 20.1 & 0.58\
& rh600193a01 & 2.70$\pm$0.22 & 8.7 & & 0.65\
& rh600193a02 & 5.36$\pm$0.47 & 19.2 & & 1.29\
& rh600934n00 & 3.74$\pm$0.24 & 16.1 & & 0.90\
69 & rh600480a01 & 2.17$\pm$0.21 & 4.4 & 20.3 & 2.73\
& rh600620a01 & 2.21$\pm$0.16 & 7.4 & & 2.78\
70 & rh600620a01 & 1.43$\pm$0.12 & 4.4 & 20.3 & 1.80\
71 & rh600480a01 & 2.56$\pm$0.29 & 9.4 & 20.3 & 3.22\
& rh600620a01 & 4.35$\pm$0.31 & 14.1 & & 5.47\
72 & rh600273n00 & 3.58$\pm$0.42 & 8.3 & 20.1 & 5.67\
& rh600273a01 & 2.28$\pm$0.28 & 8.8 & & 3.61\
73 & rh600273n00 & 2.18$\pm$0.25 & 8.6 & 20.1 & 3.46\
& rh600273a01 & 11.50$\pm$0.90 & 39.3 & & 18.23\
74 & rh600742n00 & 13.80$\pm$0.87 & 41.1 & 20.1 & 3.65\
& rh601110n00 & 7.80$\pm$0.64 & 23.6 & & 2.06\
75 & rh150004n00 & 5.54$\pm$0.36 & 20.0 & 20.9 & 1.06\
& rh701924n00 & 11.00$\pm$1.17 & 18.3 & & 2.11\
& rh701925n00 & 9.32$\pm$0.95 & 15.3 & & 1.79\
& rh701926n00 & 5.32$\pm$0.63 & 7.3 & & 1.02\
& rh701928n00 & 7.95$\pm$1.04 & 8.3 & & 1.53\
& rh704206n00 & 5.49$\pm$0.35 & 18.3 & & 1.05\
76 & rh701924n00 & 32.20$\pm$2.20 & 45.9 & 20.9 & 6.20\
& rh701925n00 & 37.00$\pm$2.40 & 60.1 & & 7.12\
& rh701926n00 & 32.40$\pm$2.10 & 48.0 & & 6.23\
& rh701927n00 & 24.60$\pm$2.12 & 31.8 & & 4.73\
& rh701928n00 & 33.70$\pm$2.55 & 34.6 & & 6.48\
77 & rh702723n00 & 23.10$\pm$1.06 & 61.2 & 20.1 & 2.75\
& rh702723a01 & 28.30$\pm$1.28 & 61.8 & & 3.37\
78 & rh600062a01 & 2.73$\pm$0.30 & 5.1 & 20.2 & 1.02\
& rh600601n00 & 2.55$\pm$0.15 & 9.6 & & 0.95\
79 & rh600062a01 & 2.70$\pm$0.27 & 6.3 & 20.2 & 1.00\
& rh600601n00 & 0.90$\pm$0.05 & 4.6 & & 0.33\
& rh601115n00 & 3.02$\pm$0.32 & 7.0 & & 1.13\
80 & wh600062 & 5.50$\pm$0.51 & 11.2 & 20.2 & 2.06\
& rh600062a00 & 5.37$\pm$0.49 & 11.0 & & 2.01\
& rh600062a01 & 6.42$\pm$0.55 & 11.5 & & 2.39\
& rh600601n00 & 4.27$\pm$0.20 & 17.2 & & 1.59\
81 & wh600062 & 3.91$\pm$0.42 & 9.3 & 20.2 & 1.46\
& rh600062a00 & 3.90$\pm$0.42 & 9.4 & & 1.46\
& rh600062a01 & 4.86$\pm$0.47 & 12.4 & & 1.82\
& rh600062a03 & 3.49$\pm$0.36 & 9.6 & & 1.31\
& rh600601n00 & 5.82$\pm$0.28 & 27.2 & & 2.18\
82 & rh600024n00 & 5.92$\pm$0.35 & 22.9 & 20.6 & 0.85\
& rh600024a01 & 10.30$\pm$0.51 & 40.7 & & 1.48\
83 & rh600092n00 & 5.65$\pm$0.41 & 13.6 & 20.1 & 0.82\
& rh600383n00 & 8.37$\pm$0.41 & 22.2 & & 1.22\
& rh600820n00 & 8.52$\pm$0.23 & 38.1 & & 1.24\
& rh600820a01 & 4.57$\pm$0.17 & 19.1 & & 0.67\
84 & rh600964n00 & 0.78$\pm$0.05 & 5.1 & 20.5 & 2.15\
85 & rh600501n00 & 3.15$\pm$0.14 & 25.4 & 21.3 & 1.06\
& rh600718n00 & 1.31$\pm$0.09 & 5.0 & & 0.44\
86 & rh701300n00 & 1.36$\pm$0.11 & 6.4 & 20.2 & 2.55\
87 & rh702055n00 & 2.29$\pm$0.23 & 3.4 & 20.3 & 5.76\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Signal recovery from nonlinear measurements involves solving an iterative optimization problem. In this paper, we present a framework to optimize the sensing parameters to improve the quality of the signal recovered by the given iterative method. In particular, we learn illumination patterns to recover signals from coded diffraction patterns using a fixed-cost alternating minimization-based phase retrieval method. Coded diffraction phase retrieval is a physically realistic system in which the signal is first modulated by a sequence of codes before the sensor records its Fourier amplitude. We represent the phase retrieval method as an unrolled network with a fixed number of layers and minimize the recovery error by optimizing over the measurement parameters. Since the number of iterations/layers are fixed, the recovery incurs a fixed cost. We present extensive simulation results on a variety of datasets under different conditions and a comparison with existing methods. Our results demonstrate that the proposed method provides near-perfect reconstruction using patterns learned with a small number of training images. Our proposed method provides significant improvements over existing methods both in terms of accuracy and speed.'
author:
- |
Zikui Cai\
ECE Department\
UC Riverside\
`[email protected]` Rakib Hyder\
ECE Department\
UC Riverside\
`[email protected]` M. Salman Asif\
ECE Department\
UC Riverside\
`[email protected]`
bibliography:
- 'learnPR.bib'
title: Learning Illumination Patterns for Coded Diffraction Phase Retrieval
---
Introduction
============
The problem of signal recovery from nonlinear measurements arises in various imaging and signal processing tasks [@shechtman2015phase; @maiden2009improved; @millane1990phase; @rodenburg2008ptychography; @candes2015phasediff]. Conventional methods for solving such inverse problems use an iterative method to recover the signal from given measurements. In this paper, we present a framework to optimize over the measurement parameters to improve the quality of signals recovered by the given iterative method. In particular, we learn illumination patterns to recover the signal from coded diffraction patterns (CDP) using a fixed-cost alternating minimization method.
Coded diffraction imaging is a specific instance of Fourier phase retrieval problems. Phase retrieval refers to a broad class of nonlinear inverse problems where we seek to recover a complex- (or real-) valued signal from its phase-less (or sign-less) measurements [@fienup1978reconstruction; @candes2013phaselift; @fienup1982phase; @jaganathan2015phase; @shechtman2015phase; @muminov2020small]. In practice, these problems often arise in coherent optical imaging where an image sensor records the intensity of the Fourier measurements of the object of interest. In coded diffraction imaging, the signal of interest gets modulated by a sequence of known illumination patterns/masks before observing the Fourier intensity at the sensor [@candes2013phaselift; @jaganathan2015phase; @shechtman2015phase]. Applications include X-ray crystallography [@watson1953structure; @millane1990phase; @harrison1993phase], astronomy [@fienup1987phase; @gonsalves2014perspectives], microscopy [@misell1973method; @miao2008extending; @rodenburg2008ptychography; @tian2014multiplexed], speech processing and acoustics [@rabiner1993fundamentals; @balan2006signal; @balan2010signal; @jaganathan2016stft], and quantum mechanics [@corbett2006pauli; @reichenbach1998philosophic].
We can model the sensor measurements for coded diffraction imaging as follows. Let us denote the signal of interest as $x \in \mathbb{R}^n$ or $\mathbb{C}^n$ that is modulated by $T$ illumination patterns $D = \{d_1, \ldots, d_T\}$, where $d_t \in \mathbb{R}^n$ or $\mathbb{C}^n$. The amplitude of sensor measurements for $t$th illumination pattern can be written as $$y_t = |\mathcal{F} (d_t \odot x)|,
\label{eq:yt}$$ where $\mathcal{F}$ denotes the Fourier transform operator and $\odot$ denotes an element-wise product. We note that real sensor measurements are proportional to the intensity of the incoming signal (i.e., square of the Fourier transform). In practice, however, solving the inverse problem with (non-square) amplitude measurements provides better results; therefore, we use the amplitude measurements throughout the paper.
![The block diagram of our proposed framework to learn illumination patterns while solving coded diffraction imaging using $K$ iterations of a phase retrieval algorithm. The iterative algorithm is represented as an unrolled network with $K$ layers. Steps at every iteration are fixed and depicted as an unrolled network (details can be found in Algorithm \[algo:test\]). Phase retrieval algorithm uses the measurements $Y = \{y_t\}$ and illumination patterns $D = \{d_t\}$ to provide an estimate $x^K$ after $K$ iterations. Our goal is to learn the illumination patterns $D$ to minimize the error between the estimated $x^K$ and the ground truth.[]{data-label="fig:intro"}](fig/intro_v7.pdf){width="1\columnwidth"}
To recover the signal $x$ from the nonlinear measurements, we can solve the following optimization problem: $$\label{eq:PR}
\min_x \; \sum_{t=1}^T \|y_t - |\mathcal{F}(d_t \odot x)| \|_2^2.$$ In recent years, a number of iterative algorithms have been proposed for solving the problem in , which includes lifting-based convex methods, alternating minimization-based nonconvex methods, and greedy methods [@candes2013phaselift; @gross2017improved; @netrapalli2013phase; @hyder2019alternating; @Jagatap2017].
Our goal is to learn a set of illumination patterns to optimize the recovery of an alternating minimization (AltMin) algorithm for solving the problem in . The AltMin method can be viewed as an unrolled gradient descent network, as shown in Fig. \[fig:intro\], where we fix the steps at every iteration and the total number of iterations for AltMin. One forward pass through the unrolled network is equivalent to $K$ iterations of the AltMin algorithm. To minimize the computational complexity of the recovery algorithm, we keep the total number of iterations very small (e.g., $K=50$). At the training stage, we optimize over the illumination patterns to minimize the error between the AltMin outputs after $K$ iterations and the ground truth training images. At the test time, we solve the problem in using $K$ AltMin iteration with the learned illumination patterns (equivalent to one forward pass). We evaluated our method on different image datasets and compared against existing methods for coded diffraction imaging. We demonstrate that our proposed method of designing illumination patterns for a fixed-cost algorithm outperforms existing methods both in terms of accuracy and speed.
Related Work
============
**Phase Retrieval and Coded Diffraction Patterns:** Fourier phase retrieval problem arises in a number of imaging systems because standard image sensors can only record intensity of the observed measurements. This problem has been extensively studied over last five decades in optics, signal processing, and optimization [@gerchberg1972practical; @fienup1982phase; @millane1990phase; @rodenburg2008ptychography; @maiden2009improved; @shechtman2015phase; @hand2018phase; @nayer2019phaseless]. Coded diffraction imaging is a physically realistic setup in which we can first modulate by signal of interest and then collect the intensity measurements [@candes2015phasediff; @chandra2017phasepack]. The modulation can be performed using a spatial light modulator or custom transparencies [@miao2008extending; @rodenburg2008ptychography; @ptychTCI]. The recovery problems involves solving a phase retrieval problem; the presence of modulation patterns makes this a more tractable problem compared to classical Fourier phase retrieval [@candes2015phasediff].
The algorithms for solving phase retrieval problem can be broadly divided into non-convex and convex methods. Classical algorithms for phase retrieval rely on solving the underlying non-convex problem using alternating minimization [@fienup1982phase; @bauschke2002phase; @netrapalli2013phase]. Amplitude flow [@wang2016sparse; @wang2016solving], Wirtinger flow [@candes2015phase; @zhang2016reshaped; @chen2015solving; @cai2016optimal], alternating minimization (AltMin) [@netrapalli2013phase; @zhang2016reshaped; @Jagatap2017; @hyder2019alternating] are recent methods that solve the non-convex problem. Convex methods usually lift the nonconvex problem of signal recovery from quadratic measurements into a convex problem of low-rank matrix recovery from linear measurements [@candes2013phaselift; @soltani2016fast; @fazel2008compressed]. The PhaseLift algorithm [@candes2013phaselift] and its variations [@gross2017improved; @candes2015phasediff] can be considered under this class. We can also incorporate prior knowledge about the signal structure (e.g., sparsity, support, or positivity) in the recovery process constraints [@ohlsson2012cprl; @li2013sparse; @bahmani2015efficient; @jaganathan2012recovery; @netrapalli2013phase; @cai2016optimal; @wang2016sparse].
**Data-Driven Approaches for Phase Retrieval:** A number of papers have recently explored the idea of replacing the classical (hand-designed) signal priors with deep generative priors for solving inverse problems [@bora2017compressed; @hand2016compressed; @ulyanov2018deep; @van2018compressed]. Some of the generative prior-based approaches for phase retrieval are presented in [@hyder2019alternating; @jagatap2019algorithmic; @hand2018phase; @shamshad2018robust; @metzler2020deep].
Another growing trend is to apply deep learning to solve inverse problems (including phase retrieval) in an end-to-end manner, where deep networks are trained to learn a mapping from sensor measurements to the signal of interest using a large number of measurement-signal pairs. A few examples demonstrating the benefit of the data-driven approaches include robust phase retrieval [@metzler2018prdeep], Fourier ptychographic microscopy [@kellman2019data], holographic image reconstruction [@rivenson2018phase], and correlography for non-line-of-sight imaging [@metzler2020deep2].
While our method is partially driven by data, our goal is not to learn a signal prior or a mapping from measurements to signal. We use data to learn the illumination patterns for a fixed recovery algorithm. The number of training images required by our method is extremely small (32 or 128 images only). Furthermore, the patterns we learn on one class of images provide good results on other types of images (see Table \[table:generalzation\]). Apart from the great flexibility, our method uses a well-defined AltMin routine, where we know exact steps for every iteration as opposed to the black-box deep models.
**Unrolled Network for Inverse Problem:** Iterative methods for solving the inverse problems, such as AltMin or other first-order methods, can be represented as unrolled networks. Every layer of such a network performs the same steps as a single iteration of the original method [@kellman2019physics; @diamond2017unrolled; @gregor2010learning; @wang2016proximal; @hammernik2018learning; @sun2016deep; @kamilov2016learning; @bostan2018learning; @monga2019algorithm; @liang2019deep]. Some parameters of the iterative steps can be learned from data (e.g., step size, denoiser, or threshold parameters) but the basic structure and physical forward model are kept intact.
**Learn to Sense:** Deep learning methods have also been recently used to design the sensing system; especially in the context of compressive sensing and computational imaging [@mousavi2017learning; @wu2019learning; @Bergman:2020:DeepLiDAR; @wang2020learning]. The main objective in these methods is similar to ours, which is to select sensor parameters to recover best possible signal/image from the sensor measurements. The sensor parameters may involve selection of samples/frames, design of sampling waveforms, or illumination patters as we discuss in this paper. In principle, the sensor can be treated as the first layer of the network with some physical constraints on the parameters [@kellman2019physics]. In contrast to most of the existing methods that learn a deep network to solve the inverse problem, our method uses a predefined iterative method as an unrolled network while learning the illumination patterns using a small number of training images.
Proposed Method
===============
We use $N$ training images ($x_1,\ldots, x_N$) to learn $T$ illumination patterns that provide best reconstruction using a predefined (iterative) phase retrieval algorithm. Furthermore, to ensure that the illumination patterns are physically realizable, we constrain their values to be in the range $[0,1]$. We use a sigmoid function over unconstrained parameters $\Theta = \{\theta_1, \ldots, \theta_T\}$ to define the illumination patterns; that is, $d_t = \text{sigmoid}(\theta_t)$ for all $t=1,\ldots,T$.
Our proposed method for learning illumination patterns can be divided into two parts: The first (inner) part involves solving the phase retrieval problem with given coded diffraction patterns using AltMin as an unrolled network (see block diagram in Fig. \[fig:intro\]); Second part is updating the illumination patterns based on backpropagating the image reconstruction loss. These two parts together yield optimal image reconstruction and illumination patterns. Pseudocodes for both parts are listed in Algorithms \[algo:train\],\[algo:test\].
Training set $X$ with $N$ images $X=\{x_1, \ldots, x_N\}$.
Initialize the optimization variables for $T$ patterns as $\Theta = \{\theta_1,\ldots, \theta_T \}$ from a uniform distribution. Generate illumination patterns in the range \[0,1\] as $d_t = \text{sigmoid}(\theta_t)$ for all $t$. $Y^n = \{y_1^n, \ldots, y_T^n ~|~y_t^n=|\mathcal{F} (d_t \odot x_n)| \}$ $x_n^K(\Theta) \gets \texttt{solveCDP($Y^n$,$D$)}$ $L_\Theta = \sum_{n=1}^{N} \| x_n - x_n^K(\Theta)\|_2^2$ $\Theta \leftarrow \Theta - \beta \nabla_{\Theta}L_{\Theta}$ Optimal illumination patterns ${D} = \{d_1, \ldots, d_T ~|~ d_t = \text{sigmoid}(\theta_t)\}$.
Measurements $Y = \{y_1, \ldots, y_t\}$ and illumination patterns $D =\{d_1,\ldots, d_T\}$.
Initialize estimate $x^0 = 0$. $p_t^{k-1} \gets \text{sign}(\mathcal{F}(d_t \odot x^{k-1})$ $\nabla_x L_{x,p} = \frac{2}{T}\sum_{t=1}^T [|d_t|^2\odot x^{k-1} - d_t^* \odot \mathcal{F}^*(p_t^{k-1}\odot y_t)]$ $x^k \leftarrow x^{k-1} - \alpha\nabla_{x}L_{x,p} $ Estimated signal $x^K$
**Phase retrieval as alternating minimization (AltMin):** Given measurements $Y = \{y_1,\ldots, y_T\}$ and illumination patterns $D = \{d_1,\ldots,d_T\}$, we seek to solve the CDP phase retrieval problem by minimizing the loss function defined in as $$\label{eq:Lx}
L_{x} = \frac{1}{T}\sum_{t=1}^{T} \| y_t - |\mathcal{F} (d_t \odot x)| \|_2^2.$$ Even though the loss function in is nonconvex and nonsmooth with respect to $x$, we can minimize it using the well-known alternating minimization (AltMin) with a single gradient descent step [@netrapalli2013phase; @zhang2016reshaped]. We define a new variable for the estimated phase of linear measurements as $p_t = \text{phase}(\mathcal{F} (d_t \odot x))$ and reformulate the loss function in into $$\label{eq:Lx_p}
L_{x,p} = \frac{1}{T}\sum_{t=1}^{T} \| p_t \odot y_t - \mathcal{F} (d_t \odot x) \|_2^2.$$ The gradient step can be computed as $$\nabla_{x}L_{x,p} = \frac{2}{T}\sum_{t=1}^{T} |d_t|^2 \odot x - d_t^* \odot \mathcal{F}^*(p_t \odot y_t). \label{eqn:grad_x}$$
We initialize the image estimate as a zero vector and update the estimate at every iteration as $ x^k = x^{k-1} - \alpha_{k-1}\nabla_{x}L_{x,p},$ where $\alpha_{k-1}$ denotes the step size. In our implementation, we used a fixed step size $\alpha$ for all iterations. We use $K$ iterations of the gradient descent in the unrolled network and denote the final estimate as $x^K$. We summarize the steps of CDP phase retrieval routine as `solveCDP` in Algorithm \[algo:test\].
**Learning illumination patterns:** To learn a set of illumination patterns that provide the best reconstruction with the predefined iterative method (or the unrolled network), we seek to minimize the difference between the original training images and their estimates. In this regard, we minimize the following quadratic loss function with respect to $\Theta$: $$\label{eq:L_theta}
L_\Theta = \sum_{n=1}^N \|x_n - x_n^K(\Theta)\|_2^2,$$ where $x_n^K(\Theta)$ denotes the `solveCDP` estimate of $n$th training image for the given values of $\Theta$. Note that for a given values of $\Theta$, we can define illumination patterns as $d_t = \text{sigmoid}(\theta_t)$ and sensor measurements for $x_n$ as $y^n_t = |\mathcal{F}(d_t\odot x_n)|$ for $t=1,\ldots, T$ and $n = 1,\ldots, N$. We use Adam optimizer in PyTorch [@kingma2014adam; @paszke2019pytorch] to minimize the loss function in . A summary of the algorithm for learning the illumination patterns is also listed in Algorithm \[algo:train\].
Experiments
===========
**Datasets.** \[section:dataset\] We used MNIST digits, Fashion MNIST (F. MNIST), CIFAR10, SVHN and CelebA datasets for training and testing in our experiments. We use 128 images from each of the datasets for training and another 1000 images for testing. To make the tiny-image datasets uniform, we reshaped all of them to $32\times32$ size with grayscale values. Images in CelebA dataset have $218\times178$ pixels, we first converted all the images to grayscale, cropped $178\times178$ region in the center, and resized to $200\times200$. In Fig. \[fig:classical\], we evaluated the performance of our method on images used in [@metzler2018prdeep].
**Measurements.** We use the amplitude of the 2D Fourier transform of the images modulated with $T$ illumination patterns as the measurements. Unless otherwise mentioned, we used noiseless measurements. We report results for measurements with Gaussian and Poisson noise in Fig. \[fig:test\_noise\].
**Computing Infrastructure.** Intel Core i7-8700 CPU and NVIDIA TITAN Xp GPU.
Setup and hyper-parameter search
--------------------------------
The hyper-parameters include the number of iterations ($K$), step size $\alpha$, and the number of training samples $N$. We set the default value of $K=50$, but we will show later that $K$ can be adjusted as a trade-off between better reconstruction quality and shorter run time. We tested all methods for $T = \{2,4,8\}$ to evaluate three cases where signal recovery is hard, moderate, and easy, respectively. Through grid search, we found that it provides the best results over all datasets when $\alpha*T = 4$. We also studied the effect of the number of training images and found that illumination patterns learned on $32$ randomly selected images provide good recovery over the entire dataset. The test accuracy improves slightly as we increase the number of training samples. To be safe, we used 128 training images in all our experiments.
Comparison between random and learned illumination patterns
-----------------------------------------------------------
To demonstrate the advantages of our learned illumination patterns, we compare the performance of learned and random illumination patterns on five different datasets. We learn a set of $T=\{2,4,8\}$ illumination patterns on 128 training images from a dataset and test them on 1000 test images from the same dataset. For random patterns, we draw $T$ independent patterns from Uniform(0,1) distribution and test their performance on the same 1000 samples that we used for the learned case. We repeat this process 30 times and choose the best result to compare with the results for the learned illumination patterns. The average PSNR over all 1000 test image reconstructions is presented in Table \[table:compare\_learned\_random\], which shows that the learned illumination patterns perform significantly better than the random patterns for all values of $T$. In addition to that, we can observe a transition in the performance for $T=4$, where random patterns provide poor quality reconstructions and learned patterns provide very high quality reconstructions.
To highlight this effect, we show a small set of reconstructed images and histograms of PSNRs of all the reconstructed images from learned and random illumination patterns in Fig. \[fig:compare\] for $T=4$ patterns. The result suggests that the learned illumination patterns demonstrate consistently better performance compared to random illumination patterns.
0.15in
------------ -------- --------- -------- --------- -------- ---------
(l)[2-7]{} Random Learned Random Learned Random Learned
MNIST 13.59 27.77 28.61 101.50 46.43 111.79
F. MNIST 16.26 25.17 30.05 92.44 52.05 108.85
CIFAR10 13.40 22.93 27.51 82.02 60.64 104.61
SVHN 12.24 20.27 26.24 84.43 65.33 109.25
12.67 18.49 21.67 78.98 41.49 90.29
------------ -------- --------- -------- --------- -------- ---------
-0.1in
[0.24]{} {width="1\linewidth"}
[0.24]{} {width="1\linewidth"}
[0.24]{} {width="1\linewidth"}
[0.24]{} {width="1\linewidth"}
Effect of number of iterations/layers (K)
-----------------------------------------
Figure \[fig:diff\_k\] shows the performance of the learned and random illumination patterns as we increase $K$ to 200 at test time using the patterns learned for $K=50$. We observed that, with the learned patterns, the image reconstruction process converges faster and is more stable (smaller variance) than the case with random patterns. The red curve in Fig. \[fig:diff\_k\] has a steeper slope and narrower shades. Besides the default setting for $K=50$, we also learn the illumination patterns for different values of $K$. Figure \[fig:diff\_k\_train\_test\] shows that we can recover images in a small number of iterations if we use learned illumination patterns. We also observe that we can perform better if we use more iterations in testing than in training. We have chosen $K=50$ for most of the experiments as a trade-off between computational cost and reconstruction performance.
[0.24]{} ![Comparison of the reconstruction quality with random and learned illumination patterns for different values of $K=1,\ldots,200$. We plot the average PSNR in bright color and the PSNR of randomly selected 100 samples in light shadows. **Learned** represents the reconstruction PSNR with learned illumination patterns (shown in red), and **Random** represents PSNR for random illumination patterns (shown in blue). The number of illumination patterns is $T=4$. Random illumination patterns are selected best out of 30 trials. The learned illumination patterns are trained on 128 training images and number of iterations $K=50$ during training.[]{data-label="fig:diff_k"}](fig/fig_diff_k_mnist.pdf "fig:"){width="1\linewidth"}
[0.24]{} ![Comparison of the reconstruction quality with random and learned illumination patterns for different values of $K=1,\ldots,200$. We plot the average PSNR in bright color and the PSNR of randomly selected 100 samples in light shadows. **Learned** represents the reconstruction PSNR with learned illumination patterns (shown in red), and **Random** represents PSNR for random illumination patterns (shown in blue). The number of illumination patterns is $T=4$. Random illumination patterns are selected best out of 30 trials. The learned illumination patterns are trained on 128 training images and number of iterations $K=50$ during training.[]{data-label="fig:diff_k"}](fig/fig_diff_k_fmnist.pdf "fig:"){width="1\linewidth"}
[0.24]{} ![Comparison of the reconstruction quality with random and learned illumination patterns for different values of $K=1,\ldots,200$. We plot the average PSNR in bright color and the PSNR of randomly selected 100 samples in light shadows. **Learned** represents the reconstruction PSNR with learned illumination patterns (shown in red), and **Random** represents PSNR for random illumination patterns (shown in blue). The number of illumination patterns is $T=4$. Random illumination patterns are selected best out of 30 trials. The learned illumination patterns are trained on 128 training images and number of iterations $K=50$ during training.[]{data-label="fig:diff_k"}](fig/fig_diff_k_cifar.pdf "fig:"){width="100.00000%"}
[0.24]{} ![Comparison of the reconstruction quality with random and learned illumination patterns for different values of $K=1,\ldots,200$. We plot the average PSNR in bright color and the PSNR of randomly selected 100 samples in light shadows. **Learned** represents the reconstruction PSNR with learned illumination patterns (shown in red), and **Random** represents PSNR for random illumination patterns (shown in blue). The number of illumination patterns is $T=4$. Random illumination patterns are selected best out of 30 trials. The learned illumination patterns are trained on 128 training images and number of iterations $K=50$ during training.[]{data-label="fig:diff_k"}](fig/fig_diff_k_celeba.pdf "fig:"){width="100.00000%"}
[0.3]{} {width="1\linewidth"}
[0.3]{} {width="1\linewidth"}
[0.3]{} {width="100.00000%"}
Generalization of learned patterns on different datasets
--------------------------------------------------------
To explore the generalizability of our learned illumination patterns, we use patterns learned on one dataset to recover images from another. The results are shown in Table. \[table:generalzation\]. As we can see in the table, the diagonal numbers are generally the best, and off-diagonal numbers are generally better than the ones with random illumination patterns.
We have also tested the learned illumination patterns on several classical images. Some results are shown in Fig. \[fig:classical\]. We used illumination patterns learned on 128 celebA images, but we can see that the learned illumination patterns perform better than the randomly chosen illumination patterns for classical images which indicates some generalizability of our learned illumination patterns.
\[table:generalzation\]
------------- ------------- ------------ ------------ ------------ ------------ ------------- ------------- ------------- ------------- ------------
(l)[2-11]{} MNIST FMNIST CIFAR SVHN [ Rand]{} MNIST FMNIST CIFAR SVHN [ Rand]{}
MNIST [ 101.50]{} [ 83.13]{} [ 78.04]{} [ 54.83]{} [ 28.61]{} [ 111.79]{} [ 92.66]{} [ 82.75]{} [ 74.86]{} [ 46.43]{}
FMNIST [ 64.47]{} [ 92.44]{} [ 85.14]{} [ 76.41]{} [ 30.05]{} [ 82.37]{} [ 108.85]{} [ 85.80]{} [ 93.09]{} [ 52.05]{}
CIFAR [ 31.77]{} [ 46.79]{} [ 82.02]{} [ 68.12]{} [ 27.51]{} [ 52.75]{} [ 84.94]{} [ 104.61]{} [ 87.85]{} [ 60.64]{}
SVHN [ 40.04]{} [ 55.31]{} [ 87.77]{} [ 84.43]{} [ 26.24]{} [ 65.88]{} [ 100.00]{} [ 104.69]{} [ 109.25]{} [ 65.33]{}
------------- ------------- ------------ ------------ ------------ ------------ ------------- ------------- ------------- ------------- ------------
![**First Row:** Ground truth images from image processing standard test datasets. **Second Row:** Reconstruction using random illumination patterns with uniform random distribution \[0, 1\] (we selected $T=4$ patterns that provided best results on celebA test images in **30 trials**). PSNR numbers are shown on the top of reconstructed images. **Third Row:** Reconstruction using the patterns trained on celebA dataset. Each image has $200\times 200$ pixels and the number of illumination patterns is $T=4$.[]{data-label="fig:classical"}](fig/fig_classical.pdf){width="60.00000%"}
Robustness to noise
-------------------
To investigate the robustness of our method to noise, we train our illumination patterns on noiseless measurements obtained from the training datasets. We then added Gaussian and Poisson noise at different levels to the measurements from the test datasets. Poisson noise or shot noise is the most common in the imaging systems. We model the Poisson noise following the approach in [@metzler2018prdeep]. Let us denote the $i^{th}$ element of measurement vector corresponding to $t^{th}$ illumination pattern, $y_t$ as $$y_t(i)=|z_t(i)|+ \eta_t (i) \;\;\;\forall i=1,2,\ldots, m$$ where $ \eta_t (i) \sim \mathcal{N}(0,\lambda|z_t(i)|)$ and $z_t=\mathcal{F}(d_t\odot x)$. We varied $\lambda$ to generate noise at different signal-to-noise ratios. Poisson noise affects larger values in measurements with higher strength than the smaller values. Since the sensors can measure only positive measurements, we kept the measurements positive by applying ReLU function after noise addition. We expect the reconstruction to be affected by noise as we did not use any denoiser. We observe the effect of noise in Figure \[fig:test\_noise\]. Even though noise affects the reconstructions, we can get reasonable reconstruction up to a certain level of noise. The relationship between noise level and reconstruction performance also indicates that our phase retrieval system is quite stable.
[0.35]{} ![Reconstruction quality of the test images vs noise level of the measurements for different datasets. Here we show shaded error bar of $\pm 0.25\sigma$ for each dataset. We learn the illumination patterns ($T=4$) on 128 noiseless training images of corresponding datasets.[]{data-label="fig:test_noise"}](fig/noise/noise_curve_gaussian.pdf "fig:"){width="100.00000%"}
[0.35]{} ![Reconstruction quality of the test images vs noise level of the measurements for different datasets. Here we show shaded error bar of $\pm 0.25\sigma$ for each dataset. We learn the illumination patterns ($T=4$) on 128 noiseless training images of corresponding datasets.[]{data-label="fig:test_noise"}](fig/noise/noise_curve_poisson.pdf "fig:"){width="100.00000%"}
0
Adding a denoiser
-----------------
The unrolling network enables us to use plug-n-play denoiser for additional regularization [@any; @paper; @here?; @venkatakrishnan2013plug; @?]. From the noise curves in Fig. \[fig:test\_noise\], we observe that at increasingly high noise level, our method show linearly low reconstruction performance. In case of high noise level, if we use a denoiser, it may potentially give better performance than the case when we do not use a denoiser. We investigate such behaviour by adding a convolutional autoencoder after every layer of the unrolling network. We demonstrate some reconstruction results in Fig \[fig:denoiser\_rec\]. We can observe that the reconstructions are better when we add a denoiser for a very high noise level (10 dB SNR Poisson noise). However, due to the generation capacity of the denoiser, it is not expected to give us better performance at low noise level. For 10 dB SNR Poisson noise, the reconstruction quality without denoiser is 16.41 dB for CIFAR10, 17.58 dB for SVHN and 20.71 dB for Fashion MNIST. These values are average reconstruction performance over 1000 test images of the corresponding datasets. Whereas with an autoencoder (denoiser) trained on corresponding training datasets, we can reach 19 dB for CIFAR10, 21.81 dB for SVHN and 22.56 dB for Fashion MNIST.
[0.32]{} ![Denoiser reconstruction performance for high noise level. Illumination Patterns are learned for clean images. We have added 10 dB Poisson noise to the test images. (i) The ground truth, (ii) reconstruction without using denoiser during testing and (iii) reconstruction using denoiser while testing. The results are for measurements using 4 illumination patterns. []{data-label="fig:denoiser_rec"}](fig/denoiser/fashion_denoiser.pdf "fig:"){width="100.00000%"}
[0.32]{} ![Denoiser reconstruction performance for high noise level. Illumination Patterns are learned for clean images. We have added 10 dB Poisson noise to the test images. (i) The ground truth, (ii) reconstruction without using denoiser during testing and (iii) reconstruction using denoiser while testing. The results are for measurements using 4 illumination patterns. []{data-label="fig:denoiser_rec"}](fig/denoiser/cifar_denoiser.pdf "fig:"){width="100.00000%"}
[0.32]{} ![Denoiser reconstruction performance for high noise level. Illumination Patterns are learned for clean images. We have added 10 dB Poisson noise to the test images. (i) The ground truth, (ii) reconstruction without using denoiser during testing and (iii) reconstruction using denoiser while testing. The results are for measurements using 4 illumination patterns. []{data-label="fig:denoiser_rec"}](fig/denoiser/svhn_denoiser.pdf "fig:"){width="100.00000%"}
0.15in
------------- ------------ ---------- ----------- ---------- ----------- ---------- ----------- ---------- ----------- ----------
PSNR Time PSNR Time PSNR Time PSNR Time PSNR Time
Fienup 13.17 0.31 23.28 0.30 32.79 0.29 40.57 0.29 31.57 5.42
GS 12.96 0.31 23.28 0.30 32.71 0.29 40.99 0.29 31.69 5.22
WirtFlow 10.74 0.07 13.72 0.06 16.28 0.05 19.65 0.04 16.93 1.26
AmpFlow 17.09 0.07 32.83 0.05 42.47 0.05 47.43 0.04 37.76 1.24
Kaczmarz 11.50 0.05 13.46 0.04 15.31 0.03 18.39 0.03 16.63 0.94
Deep Models 31.73 8.41 22.33 8.45 25.63 8.40 27.81 8.25 22.31 10.55
Ours **101.50** **0.02** **92.44** **0.02** **82.02** **0.02** **84.43** **0.02** **78.98** **0.04**
------------- ------------ ---------- ----------- ---------- ----------- ---------- ----------- ---------- ----------- ----------
: Reconstruction PSNR and run time (sec) of our and existing methods on different datasets.[]{data-label="table:compare_existing_methods"}
Comparision with existing methods
---------------------------------
We compare our method with various existing methods on different datasets and show the reconstruction PSNR (dB) and run time (seconds) per image in Table. \[table:compare\_existing\_methods\]. The description of the datasets can be found in section \[section:dataset\]. We compare with Fineup [@fienup1978reconstruction], Gerchberg-Saxton [@gerchberg1972practical], Wirtinger Flow [@candes2015phase], Amplitude Flow [@chen2015solving], Kaczmarz [@wei2015solving] and deep generative models [@metzler2020deep]. For [@fienup1978reconstruction; @gerchberg1972practical; @candes2015phase; @chen2015solving; @wei2015solving], we use the code collected in PhasePack [@chandra2017phasepack]. In PhasePack, we restricted all the illumination patterns in the range of $[0,1]$ and set the maximum number of iterations to be $50$ which is the same as our default setting. We note that all of the methods in PhasePack use optimal spectral initialization; in contrast, we only use zero initialization. For deep generative models, we use a modified version of the publicly available code for [@metzler2020deep]. The code only provided pretrained DCGAN models for MNIST and F. MNIST; therefore, we trained our DCGAN models on the other datasets. This method is noticeably time-consuming because it optimizes over the latent vector for the deep model and uses 2000 iterations for each image where each iteration requires a forward and backward pass through the deep model. For all of the methods, we selected the best PSNR from 5 trials and report the average time. For our method, we report the run time with batch size equal to 1, which aligns with other methods. We observe from the table that our method with learned patterns performs significantly better than all the other algorithms in terms of both reconstruction quality and run time.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The power spectrum of finite-temperature quantum electromagnetic fluctuations produced by elementary charge carriers under the influence of external electric field is investigated. It is found that under the combined action of the photon heat bath and the external field, the low-frequency asymptotic of the power spectrum is modified both qualitatively and quantitatively. The new term in the power spectrum is inversely proportional to frequency, but is odd with respect to it. It comes from the connected part of the correlation function, and is related to the temperature and external field corrections to the photon and charge carrier propagators. In application to the case of a biased conducting sample, this term gives rise to a contribution to the voltage power spectrum which is proportional to the absolute system temperature, the charge carrier mobility, the bias voltage squared, and a factor describing dependence of the noise intensity on the sample geometry. It is verified that the derived expression is in agreement with the experimental data on $1/f$-noise measurements in metal films. It is shown also that the obtained result provides a natural resolution to the problem of divergence of the total noise power.'
author:
- 'Kirill A. Kazakov'
title: 'Quantum fluctuations of Coulomb Potential as a Source of Flicker Noise. The Influence of Heat Bath'
---
Introduction
============
The origin of flicker noise [@buck] observed in virtually all conducting media remains an open issue in condensed matter theory. Despite numerous models suggested since its discovery eighty years ago there is presently no consistent theory which would explain the main characteristic properties of this omnipresent noise. The power spectral density of flicker noise is proportional to $f^{-\gamma},$ where $f$ is frequency, and the exponent $\gamma$ is about unity (usually $\gamma$ takes values $0.9$ to $1.5$). One of the essential difficulties for theoretical explanation is the fact that experiments show no limits for this dispersion law, neither lower or upper. Although flicker noise dominates only at sufficiently low frequencies, the $1/f$-component is detected in the whole measured band up to $10^6 {\rm Hz}.$ It is also experimentally established that the noise power spectrum of a sample is proportional to the applied bias squared, and roughly inversely proportional to its volume.
Ubiquity of flicker noise and universality of its properties suggest existence of a simple reason for its occurrence. It is natural to expect this reason to have a quantum origin. Although some of the models suggested so far do consider various quantum effects as underlying mechanisms of flicker noise (such as, for instance, trapping of charge carriers), it may well be that its origin is to be sought at the most fundamental level. Namely, it is plausible that the phenomenon of flicker noise has its roots in the very quantum nature of interaction of elementary charges with electromagnetic field. From this point of view, the problem was attacked by Handel [@handel1], who suggested that the observed flicker noise is related to the spectrum of low-energy photons emitted in any scattering process, which, according to Handel, has the $1/f$ profile and reflects the well-know property of bremsstrahlung, namely, the infrared divergence of the cross-section considered as a function of the energy loss. Later, the argument was modified and the so-called coherent quantum $1/f$ effect described [@handel2] on the basis of quantum electrodynamic results of Kibble and Zwanziger [@zwan; @kibble]. Although Handel’s theory was severely criticized in many respects [@tremblay; @kampen], it has found support in independent investigations of Refs. [@vliet; @ziel]. Handel’s approach is based on consideration of [*current*]{} fluctuations. An essentially different quantum approach to the problem was recently proposed [@kazakov1; @kazakov2], in which flicker noise is treated as a [*voltage*]{} fluctuation originating from quantum fluctuations of individual electric fields of charge carriers. In this case, the noise power spectrum is found by evaluating the two-point correlation function of the Coulomb potential of an elementary particle, dispersion of this function being related to the quantum spreading of the particle wave packet. However, the magnitude of noise induced by an external electric field, given by this theory, turned out to be too small to explain the observed noise level.
The purpose of this paper is to show that the value of induced quantum noise is actually significantly higher than the previously calculated. It turns out that in evaluating the effect of external electric field it is essential to take into account statistical properties of electromagnetic field. The point is that simultaneous account of the effects of photon heat bath and external field leads to appearance of a contribution of new type. Namely, the new term in the noise power spectrum is [*odd*]{} with respect to frequency, and hence corresponds to the part of the correlation function, which is odd with respect to the difference of its time arguments. The underlying reason that makes the appearance of this term possible is the inhomogeneity in time of fluctuations produced by individual charge carriers. As a consequence of this inhomogeneity, the correlation function of each individual contribution to the Coulomb field fluctuation depends on both time arguments separately, rather than on their difference only, so that symmetry under interchanging of the arguments does not forbid appearance of the odd contribution. The time homogeneity is restored only after summing up all independent contributions. The new contribution stems from the connected part of the correlation function of particle’s field, rather than from the disconnected one that was in focus of Ref. [@kazakov2], and is related to the temperature and external field corrections to the photon and charge carrier propagators, in contrast to considerations of Ref. [@kazakov2] where thermal bath and external field affected only real particle states.
The paper is organized as follows. In Sec. \[prelim1\] we briefly discuss the role of the heat bath effects in evaluating the mean value and the correlation function of electromagnetic field produced by elementary particles, and identify the contributions relevant in the low-frequency regime. The power spectral functions of the Coulomb field and voltage fluctuations are defined in Sec. \[prelim2\], and written in the form convenient for explicit calculations which are carried out in Sec. \[calcul\]. The low-frequency asymptotic of the power spectrum of Coulomb potential fluctuations in the absence of external electric field is evaluated in Sec. \[zerofield\], and is found to exhibit an inverse frequency dependence. However, the $1/f$ part of the spectrum is cancelled in the expression for the voltage spectral density. The non-vanishing $1/f$-contribution to the voltage power spectrum is obtained in Sec. \[nonzerofield\] upon account of the influence of external homogeneous electric field on the virtual charger particle propagation. Application of the obtained results to solids and comparison with experimental data is given in Sec. \[comparison\]. Gauge independence of our treatment of electromagnetic fluctuations is proved in the Appendix.
Preliminaries
=============
Heat bath contribution to particle propagators {#prelim1}
----------------------------------------------
Consider a quantized system of charged particles interacting with electromagnetic field. Let $T$ be the absolute temperature of the system. We are interested in the influence of finite temperature on quantum properties of the electromagnetic field produced by charged particles. Specifically, finite-temperature correlations in the values of the particle’s Coulomb fields will be investigated. To the leading order in the electromagnetic coupling, these correlations are a single-particle effect, in the sense that in this case only fields produced by one and the same particle correlate. In what follows, we thus confine ourselves to systems which allow perturbative treatment of charge carrier collisions, either in terms of original particles, or in terms of quasi-particles (e.g., conduction electrons in metals). In the latter case, the particle’s mass and energy-momentum relation should be replaced by the effective ones.
In this section, we shall discuss some general features of the temperature effect, related to the heat bath influence on virtual propagation of the electromagnetic and charged field quanta. Evidently, the heat bath has no effect on the mean electromagnetic field of a charged particle. Indeed, under above assumptions, this field can be represented as the amplitude of one-photon emission in a transition between free charged particle states, contracted with the photon propagator. Since the 4-vector of momentum transfer to a free massive particle, $p,$ is always spacelike, $p^2<0,$ distribution of real photons appearing in the definition of the photon propagator is immaterial in the calculation of the field. As a consequence, quantities built from the mean field, such as the disconnected part of the correlation function, are not affected by the photon heat bath (details of evaluation of this part see in Refs. [@kazakov1; @kazakov2]). Things change, however, when the connected part of the two-point correlation function of electric potential is considered. It is defined by the following symmetric expression $$\begin{aligned}
\label{corr}&&
C^{\rm con}_{00}(x;x') = \frac{1}{2}\langle {\rm
in}|\hat{A}_0(x)\hat{A}_0(x') + \hat{A}_0(x')\hat{A}_0(x) |{\rm
in}\rangle\,,\end{aligned}$$ where $x$ and $x'$ are the spacetime coordinates of two observation points, $\hat{A}_0$ is the scalar potential Heisenberg operator, and $|{\rm in }\rangle$ denotes the given [*in*]{} state of the system “charged particle + electromagnetic field.” In the two-photon processes, photon momenta are allowed to take on lightlike directions, and hence the photon heat bath does contribute to the function $C^{\rm con}_{00}(x;x').$
As is well-known, the ordinary Feynman rules of the S-matrix theory are not generally applicable for the calculation of [*in-in*]{} expectation values, and must be modified, e.g., according to Schwinger and Keldysh [@keldysh]. This complication was overcome in [@kazakov2] by rewriting Eq. (\[corr\]) in the form $$\begin{aligned}
\label{corr2}&&
C^{\rm con}_{00}(x;x') = {\rm Re}\langle {\rm out}|T
\{\hat{A}_0(x)\hat{A}_0(x')\}|{\rm in}\rangle\,,\end{aligned}$$ which allows the use of the S-matrix rules. This transformation uses equivalence of the one-particle [*in*]{} and [*out*]{} states (and Hermiticity of the electromagnetic field operator). It is applicable to the present case as well, despite the fact that now the [*in*]{} state is not one-particle, because we are not going to take into account scattering processes in the heat bath itself. This can also be shown directly at the diagrammatic level following the route of transformations taken in Ref. [@kazakov], which is formally the same in zero- and nonzero-temperature cases. Thus, in order to calculate the [*in-in*]{} expectation value (\[corr\]) taking into account the heat bath effect, the standard finite-temperature-field-theory techniques can be used [@landsman]. Below we employ a version of the real time formalism, developed in [@niemi], which is especially convenient in actual calculations since the momentum space propagators in this formulation do not involve the step function.
The real time formulation involves doubling of all fields, which will be specified by a two-valued lower index. According to the diagrammatic rules derived in [@niemi], the photon propagator has the following matrix structure[^1] $$\begin{aligned}
\label{phiphot}
\mathfrak{D}_{\mu\nu}(x) &=& \int\frac{d^4 k}{(2\pi)^4}
\mathfrak{D}_{\mu\nu}(k)e^{-ikx}\,, \quad \mathfrak{D}_{\mu\nu}(k)
= 4\pi\eta_{\mu\nu}\left(
\begin{array}{cc}
D_{11}(k)&D_{12}(k)\\
D_{21}(k)&D_{22}(k)
\end{array}\right)\,,\end{aligned}$$ where $$D_{11}(k) = - D^*_{22}(k) = \frac{1}{k^2 + i0} -
\frac{2\pi i \delta(k^2)}{e^{\beta|k_0|} - 1}\,, \quad D_{12}(k) =
D_{21}(k) = - \frac{2\pi
i\delta(k^2)e^{\beta|k_0|/2}}{e^{\beta|k_0|} - 1}\,,$$ $\beta =
1/T$ being the inverse absolute temperature of the system. In applications to the problem of $1/f$-noise considered below, the value of the product $\beta |k_0|$ turns out to be very small. For instance, even for frequencies as large as $10^6 {\rm Hz},$ and temperatures as small as $1^{\circ}{\rm K}, $ it is less than $10^{-27}\cdot 10^6/(10^{-16}) = 10^{-5}$ (the factors $10^{-27}$ and $10^{-16}$ are contributed by the Planck and Boltzmann constants, respectively), so the denominators in the above expressions can be replaced by $\beta |k_0|,$ implying that the second term in $D_{11}$ dominates. On the other hand, the temperature effect on the propagation of massive particles is much less prominent. For instance, in the case of conduction electrons in a crystal (this case will be used throughout as a standard example), the particle energy $\varepsilon$ is of the order $(\hbar/d)^2/m,$ where $m$ is the particle mass, and $d$ is the lattice spacing. Taking $d\sim 10^{-8}{\rm cm},$ $m\sim
10^{-27}{\rm g},$ we find that $\varepsilon/T \sim 10^{+5},$ so that the temperature contributions can be completely neglected (for fermions as well as for bosons), and the propagator taken in the simple diagonal form $$\begin{aligned}
\label{phiprop}
\mathfrak{D}^{\phi}(x) &=& \int\frac{d^4 k}{(2\pi)^4}
\mathfrak{D}^{\phi}(k)e^{-ikx}\,, \quad \mathfrak{D}^{\phi}(k) =
\left(
\begin{array}{cc}
D^{\phi}_{11}(k)& 0\\
0& D^{\phi}_{22}(k)
\end{array}\right)\,, \\
D^{\phi}_{11}(k) &=& - D^{\phi *}_{22}(k) = (m^2 - k^2 -
i0)^{-1}\,.\nonumber\end{aligned}$$ This is for a scalar particle described by the action $$\begin{aligned}
\label{action}
S_{\phi} &=& {\displaystyle\int} d^4 x
\left\{(\partial_{\mu}\phi^* + i e A_{\mu}\phi^*)
(\partial^{\mu}\phi - ie A^{\mu}\phi) - m^2 \phi^*\phi\right\}\,,\end{aligned}$$ where $e$ is the particle charge. Account of particle spin, though adds some extra algebra, does not change the long-range properties of its field, so that following Ref. [@kazakov2] we work with the simplest case of zero-spin particles. The matrix propagators are multiplied in the interaction vertices, generated by the triple and higher order terms in the Lagrangian, with an additional minus sign for the product of 2-components, as in the Schwinger-Keldysh techniques. The arguments of the Green functions are treated as 1-component fields. Finally, external particle lines represent normalized particle amplitudes or their conjugates, according to whether the particle is incoming or outgoing, just like in the conventional techniques.
As was mentioned above, the photon propagator is dominated by the temperature contribution (as long as the range of momentum integration contains lightlike directions, see discussion in Sec. \[zerofield\]), while in the massive particle propagator this contribution is negligible. It is important, on the other hand, that the heat bath affects significantly the real particle propagation, i.e., external matter lines in the diagrams. Bilinears of the particle amplitudes representing these lines are expressed eventually via statistical distribution function (see Sec. \[nonzerofield\] for details). Thus, apart from explicit $T$-dependence coming from the photon propagator, the correlation function also depends on temperature implicitly through the particle statistical distribution.
Power spectral densities of potential and voltage fluctuations {#prelim2}
--------------------------------------------------------------
Connected contribution to the power spectral density of electric potential fluctuations is obtained by Fourier transforming Eq. (\[corr\]) with respect to the difference of the time instants $t,t'$: $$\begin{aligned}
\label{corrf}&&
C(\bm{x},\bm{x}',t',\omega) = \int\limits_{-\infty}^{+\infty}d\tau
C^{\rm con}_{00}(\bm{x},t'+\tau;\bm{x}',t')e^{-i\omega\tau}\,.\end{aligned}$$ The upper and lower indices in the notation of the correlation function are suppressed in the left hand side, for brevity. We are interested ultimately in the power spectrum of voltage fluctuations, $C_U,$ measured between two observation points $\bm{x},\bm{x}'$ (e.g., two leads attached to a conducting sample). The connected contribution to the voltage correlation function is given by $$\begin{aligned}
\label{corru}&&
C_U(\bm{x},\bm{x}',t,t') = \frac{1}{2}\langle{\rm
in}|\hat{U}(t)\hat{U}(t') + \hat{U}(t')\hat{U}(t)|{\rm
in}\rangle\,,\end{aligned}$$ where $\hat{U}(t) = \hat{A}_0(\bm{x},t) -
\hat{A}_0(\bm{x}',t)$ is the operator of voltage between the two points. This function is separately symmetric with respect to the interchanges $\bm{x}\leftrightarrow \bm{x}',$ and $t\leftrightarrow t',$ unlike the function $C^{\rm
con}_{00}(x;x')$ which is only symmetric under $x\leftrightarrow
x'.$ Substituting the definition of $\hat{U}(t)$ in Eq. (\[corru\]), the former can be expressed via the latter $$\begin{aligned}
C_U(\bm{x},\bm{x}',t,t') &=& C^{\rm con}_{00}(\bm{x},t;\bm{x},t')
+ C^{\rm con}_{00}(\bm{x}',t;\bm{x}',t') \nonumber \\ &-&
\left[C^{\rm con}_{00}(\bm{x},t;\bm{x}',t') + C^{\rm
con}_{00}(\bm{x}',t;\bm{x},t')\right]\,.\end{aligned}$$ Accordingly, the power spectral density of voltage fluctuations, defined by $$\begin{aligned}
\label{corruf}&&
C_U(\bm{x},\bm{x}',t',\omega) =
\int\limits_{-\infty}^{+\infty}d\tau
C_U(\bm{x},t'+\tau,\bm{x}',t')e^{-i\omega\tau}\,,\end{aligned}$$ is expressed through that of potential fluctuations as $$\begin{aligned}
\label{corrfup}
C_U(\bm{x},\bm{x}',t',\omega) &=& C(\bm{x},\bm{x},t',\omega) +
C(\bm{x}',\bm{x}',t',\omega) - \left[C(\bm{x},\bm{x}',t',\omega) +
C(\bm{x}',\bm{x},t',\omega)\right]\,.\end{aligned}$$ Although $C_U(\bm{x},\bm{x}',t,t')$ is symmetric with respect to the interchange $t\leftrightarrow t',$ it depends on both time arguments separately, and therefore $C_U(\bm{x},\bm{x}',t,\omega)$ does not have to be an even function of $\omega.$
When calculating the power spectrum of potential fluctuations according to Eqs. (\[corr2\]), (\[corrf\]), it is convenient to perform the Fourier transformation under the sign “[Re]{}” in Eq. (\[corr2\]). For this purpose we introduce the Fourier transform of the two-point Green function: $$\begin{aligned}
\label{fourierg}&&
G(\bm{x},\bm{x}',t',\omega) = \int\limits_{-\infty}^{+\infty}d\tau
G(\bm{x},t'+\tau;\bm{x}',t') e^{-i\omega\tau}\,, \quad G(x;x') =
\langle {\rm out}|T \{\hat{A}_0(x)\hat{A}_0(x')\}|{\rm
in}\rangle\,, \nonumber \\\end{aligned}$$ with the help of which the power spectral density of potential fluctuations can be written as $$\begin{aligned}
\label{corrg}
C(\bm{x},\bm{x}',t',\omega) &=& \frac{1}{2}{\rm Re
}\left\{G(\bm{x},\bm{x}',t',\omega) +
G(\bm{x},\bm{x}',t',-\omega)\right\} \nonumber\\ &+&
\frac{i}{2}{\rm Im }\left\{G(\bm{x},\bm{x}',t',\omega) -
G(\bm{x},\bm{x}',t',-\omega)\right\}\,.\end{aligned}$$ We see that contributions to the function $C(\bm{x},\bm{x}',t',\omega),$ and hence to the voltage power spectrum, are either real even, or imaginary odd functions of frequency.
Evaluation of low-frequency asymptotic of spectral density {#calcul}
==========================================================
Power spectrum in the absence of external electric field {#zerofield}
--------------------------------------------------------
The tree contribution to the right hand side of Eq. (\[corr2\]) when the influence of external electric field is neglected is shown in Fig. \[fig1\]. Repeating the argument of Ref. [@kazakov2], it is not difficult to show that the leading contribution is contained in the diagrams \[fig1\](a), \[fig1\](b). Distinguishing their contributions by the corresponding Latin subscript, we have $$\begin{aligned}
\label{diagen}
G_a(x,x') &=& ie^2\iint d^4 z d^4 z'
\left\{\mathfrak{D}(x,z)\left[\phi_0(z)
\stackrel{\leftrightarrow}{\partial_{0}}
\mathfrak{D}^{\phi}(z,z')\stackrel{\leftrightarrow}{\partial_{0}^{\,\prime}}
\phi^*_0(z')\right]\mathfrak{D}(z',x')\right\}_{11}\,, \\
G_b(x,x') &=& G_a(x',x)\,,\nonumber\end{aligned}$$ where $$\begin{aligned}
\varphi\stackrel{\leftrightarrow}{\partial_{0}}\psi &=&
\varphi\partial_{0}\psi - \psi\partial_{0}\varphi\,, \nonumber\end{aligned}$$ and $\phi_0$ is the given particle state. Going over to momentum space with the help of Eqs. (\[phiphot\]), (\[phiprop\]), introducing the spectral function for $G_a(x,x')$ according to Eq. (\[fourierg\]), and writing the matrix product longhand yields $$\begin{aligned}
\label{diagenk1}
G_a(\bm{x},\bm{x}',t',\omega) &=& (4\pi e)^2\iint \frac{d^3
\bm{q}}{(2\pi)^3} \frac{d^3
\bm{p}}{(2\pi)^3}\frac{a(\bm{q})a^*(\bm{q} +
\bm{p})}{\sqrt{2\varepsilon_{\bm q}2\varepsilon_{{\bm q} +
\bm{p}}}} e^{ip^0(t' - t_0) - i\bm{p}\bm{x}'}
J_a(p,q,\bm{x}-\bm{x}',\omega)\,, \nonumber \\
p^0 &=& \varepsilon_{\bm{q} + \bm{p}} - \varepsilon_{\bm{q}}\,,
\quad \varepsilon_{\bm{q}} = +\sqrt{\bm{q}^2 + m^2}\,,\end{aligned}$$ where $$\begin{aligned}
\label{diagenk3}&&
J_a(p,q,\bm{x}-\bm{x}',\omega) = - i\int \frac{d^3
\bm{k}}{(2\pi)^3} e^{i\bm{k}(\bm{x}'-\bm{x})}(2q_0 + k_0)(2q_0 +
k_0 + p_0)\nonumber\\&&\times
\left[D_{11}(k)D^{\phi}_{11}(q+k)D_{11}(k-p) +
D_{12}(k)D^{\phi}_{22}(q+k)D_{21}(k-p) \right]_{k_0=\omega}\,.\end{aligned}$$ Here $q_{\mu}$ is the particle 4-momentum, $a(\bm{q})$ its momentum wave function at some time instant $t_0,$ normalized by $$\begin{aligned}
\label{norm}
\int\frac{d^3 \bm{q}}{(2\pi)^3}|a(\bm{q})|^2 = 1\,,\end{aligned}$$and it is taken into account that $D^{\phi}_{12} = 0\,.$
The second term in the square brackets in Eq. (\[diagenk3\]) can be neglected. Indeed, in view of the factor $D_{12}(k)$ which is proportional to $\delta(k^2),$ and the condition $k_0 = \omega,$ the momentum $k$ contributes only a tiny value to the argument of the factor $D_{21};$ for electrons in a crystal, for instance, the ratio $|\bm{k}|/|\bm{p}|$ is of the order $(\hbar
\omega/c)/(\hbar/d) = \omega d/c \approx \omega \cdot
10^{-18}\,{\rm s}.$ Therefore, this factor can be written simply as $D_{21}(p)\sim \delta(p^2).$ On the other hand, momentum transfer $p$ to the massive particle is spacelike, and hence the argument of the delta-function is always nonzero. Furthermore, using explicit expression for the photon propagator, the first term in the square brackets reads $$\begin{aligned}
&&
\hspace{-0,5cm}\left\{\frac{1}{k^2 + i0} - \frac{2\pi i
\delta(k^2)}{e^{\beta|k_0|} - 1} \right\}D^{\phi}_{11}
\left\{\frac{1}{(k-p)^2 + i0} - \frac{2\pi i
\delta((k-p)^2)}{e^{\beta|k_0-p_0|} - 1} \right\} = \frac{1}{[k^2
+ i0]}D^{\phi}_{11}\frac{1}{[(k-p)^2 + i0]} \nonumber\\&& -
\frac{2\pi i \delta(k^2)}{e^{\beta|k_0|} - 1}D^{\phi}_{11}
\frac{1}{(k-p)^2 + i0} - \frac{1}{k^2 + i0}D^{\phi}_{11}\frac{2\pi
i \delta((k-p)^2)}{e^{\beta|k_0-p_0|} - 1} + \frac{2\pi i
\delta(k^2)}{e^{\beta|k_0|} - 1}D^{\phi}_{11}\frac{2\pi i
\delta((k-p)^2)}{e^{\beta|k_0-p_0|} - 1}\,.\nonumber\end{aligned}$$ As before, the last term on the right hand side can be neglected, while the first term, describing the zero-temperature contribution, has already been considered in Ref. [@kazakov2]. Furthermore, $k_0$ enters the temperature exponent in the third term in the combination $(k_0 - p_0),$ and, therefore, this term does not contribute to the leading order in $1/\omega,$ because in practice $|\omega|\ll p_0.$ Indeed, estimating the energy transfer as $p_0 \approx (\bm{p}\bm{q})/m,$ and taking $|\bm{p}| \sim 1/l,$ where $l$ is the characteristic sample length, one finds for our standard example $|\omega/p_0|\sim 10^{-8}|\omega|l,$ where $l,\omega$ are supposed to be expressed in the $CGS$ system of units. Even for $l$ as large as $1\,{\rm cm},$ this ratio is very small for all practically relevant frequencies. Thus, the contribution of the second term only remains to be considered. In view of the factor $\delta(k^2),$ the pole of the function $D_{11}(q+k)$ in this term does not contribute. This is again a consequence of the requirement that the momentum transfer to a massive particle on-shell be spacelike: conditions $k^2 = 0,$ $q^2
= m^2,$ and $(q+k)^2 = m^2$ cannot be satisfied altogether. Hence, the scalar particle propagator can be written simply as $D_{11}(q+k) = - 1/(2qk),$ or, assuming that the particle is nonrelativistic, $|\bm{q}|\ll m,$ as $$\begin{aligned}
\label{prsingular}
D^{\phi}_{11}(q+k) = -\frac{1}{2m\omega}\,.\end{aligned}$$ Finally, neglecting $k$ in comparison with $p$ in the factor $1/[(k-p)^2 + i0],$ using $p^2<0$ to omit $i0,$ and retaining only terms singular in $\omega,$ we find $$\begin{aligned}
\label{diagenk4}
J_a(p,q,\bm{x}-\bm{x}',\omega) &=& \frac{2\pi (2m +
p_0)}{\omega\beta |\omega| p^2}\int \frac{d^3 \bm{k}}{(2\pi)^3}
e^{i\bm{k}(\bm{x}'-\bm{x})}\delta(\omega^2 - \bm{k}^2 )
\nonumber\\
&=& \frac{(2m + p_0)}{2\pi \omega\beta p^2} + O(\omega)\,.\end{aligned}$$ The reason why $p_0$ has been kept along with $m$ in the numerator will become clear soon. As to the contribution of the diagram \[fig1\](b), changing $\bm{k} \to \bm{k} + \bm{p}$ in Eq. (\[diagenk3\]), and then $\bm{q}\to \bm{q} - \bm{p},$ $\bm{p}\to -\bm{p}$ in Eq. (\[diagenk1\]) shows that $$G_b(\bm{x},\bm{x}',t',\omega) = G^*_a(\bm{x},\bm{x}',t',-\omega)\,.$$ Thus, the total contribution to the function $G(\bm{x},\bm{x}',t',\omega)$ is $$\begin{aligned}
\label{gab}
G(\bm{x},\bm{x}',t',\omega) &=& G_a(\bm{x},\bm{x}',t',\omega) +
G^*_a(\bm{x},\bm{x}',t',-\omega)\,.\end{aligned}$$ It is seen from this relation and Eq. (\[diagenk4\]) that only imaginary part of $G_a$ gives rise to a nonzero contribution to the total Green function $G,$ and this part corresponds to the term proportional to $p_0$ on the right hand side of Eq. (\[diagenk4\]). Thus, $$\begin{aligned}
\label{diagenk5}
G(\bm{x},\bm{x}',t',\omega) &=& - \frac{8\pi e^2}{m^2\beta
\omega}\iint \frac{d^3 \bm{q}}{(2\pi)^3} \frac{d^3
\bm{p}}{(2\pi)^3}a(\bm{q})a^*(\bm{q}+\bm{p})\frac{(\bm{p}\bm{q})}{\bm{p}^2}e^{
ip_0(t' - t_0) - i\bm{p}\bm{x}'}\,,\end{aligned}$$ where, in the denominator, the energies $\varepsilon_{\bm{q}},\varepsilon_{\bm{q}+\bm{p}}$ have been replaced by $m,$ and $p^0$ neglected in comparison with $|\bm{p}|$ on account of the condition $|\bm{q}|\ll m,$ while in the numerator, $p^0$ has been replaced by its leading long-range term, $(\bm{p}\bm{q})/m.$ It is instructive to see what the right hand side of Eq. (\[diagenk5\]) becomes in a particulary simple model case when the amplitude $a(\bm{q})$ can be written as $$\begin{aligned}
\label{abrel}
a(\bm{q}) = b(\bm{q})e^{-i\bm{q}\bm{x}_0}\,,\end{aligned}$$where $b(\bm{q})$ is a real function of the particle momentum. In this case, $\bm{x}_0$ is easily identified as the mean particle position, and hence, the function $b(\bm{q})$ describes the momentum space profile of the particle wave packet. After extraction of the position-dependent phase factor, the amplitude becomes a relatively slowly varying function of the particle momentum, therefore, to the leading order of the long-range expansion, $b(\bm{q} + \bm{p})$ in the integrand of Eq. (\[diagenk5\]) can be replaced by $b(\bm{q})$ $$\begin{aligned}
\label{diagenk51}
G(\bm{x},\bm{x}',t',\omega) &=& - \frac{8\pi e^2}{m^2\beta
\omega}\iint \frac{d^3 \bm{q}}{(2\pi)^3} \frac{d^3
\bm{p}}{(2\pi)^3}|b(\bm{q})|^2\frac{(\bm{p}\bm{q})}{\bm{p}^2}e^{
ip_0(t' - t_0) + i\bm{p}(\bm{x}'-\bm{x}_0)}\,,\end{aligned}$$ In the exponent, $p_0$ generally cannot be replaced by $(\bm{p}\bm{q})/m,$ because the subleading term of its long-range expansion, $\bm{p}^2/(2m),$ although small compared to $(\bm{p}\bm{q})/m,$ can change the phase significantly, provided that the difference $(t'-t_0)$ is sufficiently large, as is the case if the particle collisions are neglected. However, taking into account the latter reduces this difference to the particle mean free time, $\tau_f,$ so that the product $p_0(t'-t_0)$ can be completely neglected. Indeed, $\tau_f$ can be estimated roughly as $d/(|\bm{q}|/m),$ and hence, $p_0(t'-t_0)\sim
|\bm{p}||\bm{q}|\tau_f/m \sim d|\bm{p}| \sim
d/|\bm{x}'-\bm{x}_0|\ll 1.$ Then the $\bm{p}$ integral can be evaluated with the help of the formula $$\int\frac{d^3\bm{p}}{(2\pi)^3}e^{i(\bm{p}\bm{x})}\frac{4\pi\bm{p}}{\bm{p}^2}
= \frac{i \bm{x}}{|\bm{x}|^3}\,,$$ yielding $$\begin{aligned}
\label{gzero}
G(\bm{x},\bm{x}',t',\omega) &=& - 2i\frac{e^2}{m^2\beta
\omega}\frac{(\overline{\bm{q}}\bm{r}')}{r'^3}\,, \quad \bm{r}' =
\bm{x}' - \bm{x}_0\,,\end{aligned}$$ where the overline denotes $\bm{q}$-averaging over the given particle state. In the absence of external electric field, $\overline{\bm{q}}$ is zero, and therefore so is the right hand side of Eq. (\[gzero\]). As we will see in the next section, the same result is obtained in the general case without the use of the model decomposition (\[abrel\]). But even for nonzero $\overline{\bm{q}},$ the voltage power spectrum calculated from $G$ given by Eq. (\[gzero\]) turns out to be zero. This is verified directly by substituting expression (\[gzero\]) into Eqs. (\[corrg\]), (\[corrfup\]). The reason for nullification of the voltage power spectrum is easily identified – it is the consequence of the fact that the function $G_a(\bm{x},\bm{x}',t',\omega)$ is independent of the $\bm{x}$ coordinate. The $\bm{x}$-dependence has been lost upon extracting the low-frequency asymptotic of $J_a$ in Eq. (\[diagenk4\]).
The influence of external electric field and particle collisions {#nonzerofield}
----------------------------------------------------------------
Let us now consider corrections to the power spectrum due to constant homogeneous external electric field, taking into account also the influence of particle collisions. These corrections are twofold. First of all, the field affects the particle wave function $a(\bm{q}),$ which is symbolized in Fig. \[fig2\] by inserting the vertices of particle-field interaction into the two external solid lines. This is only a schematic picture, because the effect of a constant homogeneous field on the free particle states cannot be treated perturbatively. The latter circumstance, however, is not important in view of the particle collisions which prevent the particle from gaining too much momentum from the field, thus cutting down its effect (particle collisions are symbolized in Fig. \[fig2\] by a virtual photon interchange between particles). Account of these two factors is accomplished by replacing the particle momentum probability distribution, $|a(\bm{q})|^2,$ by the statistical distribution function, obtained as a solution of the kinetic equation in the presence of external electric field. This point will be discussed in more detail later in this section.
Next, the particle propagator is also modified by the external field. It is not difficult to see that for a sufficiently small field strength, $\bm{E},$ this modification can be treated perturbatively. It can be recalled that in coordinate space, it amounts to multiplying the zero-field propagator $D^{\phi}(z,z')$ by[^2] $\exp\left\{ie(\bm{E},\bm{z}+\bm{z}')(z'_0-z_0)/2\right\}.$ Although the low-frequency limit is determined by the large-time behavior of the quantities involved, implying that $(z'_0-z_0)\sim
1/\omega$ is large, the exponent can be made as small as desired for any given $\omega$ by taking $|\bm{E}|$ sufficiently small. This also will be clear from the explicit calculations to follow. The lowest order correction to the correlation function is represented by diagrams with a single insertion of the particle-field interaction vertex into the internal solid line, as shown in Fig. \[fig3\]. As in the preceding section, these diagrams vanish unless all interaction vertices are 1-type, on account of momentum conservation in the vertices together with the mass shell conditions for the massive particle. The contribution of the diagram \[fig3\](a) to the spectral density of the two-point Green function $$\begin{aligned}
\label{diagenk1e}
G^E_a(\bm{x},\bm{x}',t',\omega) = && (4\pi e)^2\iiint \frac{d^3
\bm{q}}{(2\pi)^3} \frac{d^3 \bm{p}}{(2\pi)^3}\frac{d^3
\bm{k}_1}{(2\pi)^3}\frac{a(\bm{q})a^*(\bm{q} +
\bm{p})}{\sqrt{2\varepsilon_{\bm q}2\varepsilon_{{\bm q} +
\bm{p}}}} \nonumber\\&& \times e^{ip^0(t' - t_0) - i(\bm{p} -
\bm{k}_1)\bm{x}'}\varphi(\bm{k}_1)
J^{E}_a(p,q,\bm{x}-\bm{x}',\omega)\,,\end{aligned}$$ where $$\begin{aligned}
\label{diagenk3e}&&
J^E_a(p,q,\bm{x}-\bm{x}',\omega) = - \left.ie\int \frac{d^3
\bm{k}}{(2\pi)^3} e^{i\bm{k}(\bm{x}'-\bm{x})}(2q_0 + k_0)(2q_0 +
2k_0)(2q_0 + k_0 + p_0)\right.\nonumber\\&& \left.
D_{11}(k)D^{\phi}_{11}(q+k)D^{\phi}_{11}(q+k+k_1)D_{11}(k+k_1-p)
\right|_{k_0=\omega}\,, \quad k_1 = (0,\bm{k}_1)\,,\end{aligned}$$ and $\varphi(\bm{k})$ is the Fourier transform of the external field potential, $\varphi(\bm{x}) = -(\bm{E},\bm{x}).$ We do not include an arbitrary constant in this expression because it is clear in advance that it cannot affect the final result. It is proved in the Appendix that the correlation function is actually invariant under the most general gauge variations of the electromagnetic potential. Substituting $$\varphi(\bm{k}) = - i(2\pi)^3 \left(\bm{E}\frac{\partial}{\partial
\bm{k}}\right)\delta^{(3)}(\bm{k})$$ into Eq. (\[diagenk1e\]) and integrating by parts, the $\bm{k}_1$ integral in Eq. (\[diagenk1e\]) is brought to the form $$\begin{aligned}
\label{k1int}&&
i\int d^3\bm{k}_1 \delta^{(3)}(\bm{k})
\left(\bm{E}\frac{\partial}{\partial
\bm{k}_1}\right)\left[D^{\phi}_{11}(q+k+k_1)D_{11}(k+k_1-p)e^{-
i(\bm{p} - \bm{k}_1)\bm{x}'}\right]\,,\end{aligned}$$ where only terms involving $\bm{k}_1$ are retained. As we have seen, the singularity at $\omega = 0$ in the function $G_a(\bm{x},\bm{x}',t',\omega)$ comes from integration over small $\bm{k},$ and this singularity is now strengthened by the extra particle propagator and the differentiation with respect to $\bm{k}_1.$[^3] Consider first the case when the $\bm{k}_1$-derivative acts on the last two factors in the square brackets. Since these depend on the difference $(\bm{p}-\bm{k}_1),$ changing $\partial/\partial \bm{k}_1 \to -
\partial/\partial \bm{p},$ and then integrating by parts with respect to $\bm{p}$ in Eq. (\[diagenk1e\]), this derivative is rendered to act on terms independent of $\bm{k}$[^4]. Thus, of the three factors in the square brackets, only the first is to be differentiated in effect, thus reducing the expression (\[k1int\]) to $$\frac{2i(\bm{E},\bm{q} + \bm{k})}{[m^2 - (q+k)]^2}
D_{11}(k-p)e^{-i(\bm{p}\bm{x}')}\,.$$ Taking into account also the vertex factor $e(2q_0 + 2k_0),$ we see that the first order correction to the charged particle propagator due to constant homogeneous external electric field is obtained by inserting the factor $$\frac{4ie(q_0+k_0)(\bm{E},\bm{q} + \bm{k})}{[m^2 - (q+k)]^2}
\equiv \varkappa$$ into the integrand in Eq. (\[diagenk3\]). The rest of the calculation repeats the steps of Sec. \[zerofield\]. Substituting explicit expressions for the photon propagators, one sees that only the part proportional to $\delta(k^2)$ is to be retained in the expression for $D_{11}(k),$ while the corresponding part in $D_{11}(k-p)$ is to be omitted. One consequence of this observation is that the factor $\varkappa$ simplifies to $$\frac{ie(\bm{E}\bm{q})}{m\omega^2}\,,$$ where we have taken into account that $k_0 = \omega \ll q_0,$ and $|\bm{k}|
= |\omega| \ll |\bm{q}|$ \[indeed, taking our standard example of electron in a crystal, the ratio $|\omega|/|\bm{q}|$ is, in the ordinary units, $(\hbar|\omega|/c)/|\bm{q}| \sim |\omega|d/c\sim
|\omega|10^{-18},$ with $\omega$ expressed in Hz\]. Another consequence is that the contributions of diagrams \[fig3\](a), \[fig3\](b) are related, as before, by[^5] $$G^E_b(\bm{x},\bm{x}',t',\omega) =
G^{E*}_a(\bm{x},\bm{x}',t',-\omega)\,.$$ Furthermore, Eq. (\[diagenk4\]) is now replaced by $$\begin{aligned}
\label{diagenk5e}
J^E_a(p,q,\bm{x}-\bm{x}',\omega) &=& \varkappa\frac{2\pi (2m +
p_0)}{\omega\beta |\omega| p^2}\int \frac{d^3 \bm{k}}{(2\pi)^3}
e^{i\bm{k}(\bm{x}'-\bm{x})}\delta(\omega^2 - \bm{k}^2 )
\nonumber\\
&=& \frac{ie(\bm{E}\bm{q})}{m\omega^2}\frac{(2m + p_0)}{2\pi
\omega\beta p^2}\left[1 - \frac{\omega^2}{6}(\bm{x}'-\bm{x})^2
\right] + O(\omega)\,.\end{aligned}$$ An important difference in comparison with the result of the preceding section is that because of the extra imaginary unit brought in by the factor $\varkappa,$ the term proportional to $p_0$ in the last formula gives rise now to a purely real contribution upon substituting $J^E_a$ in Eq. (\[diagenk1e\]), and hence is cancelled[^6] in the sum of diagrams \[fig3\](a), \[fig3\](b). On the other hand, the terms independent of $p_0$ survive, and lead to the following expression for the first order correction to the spectral density of the two-point Green function $$\begin{aligned}
\label{gefree}
G^E(\bm{x},\bm{x}',t',\omega) = - \frac{16\pi i
e^3}{m\beta\omega^3}\hspace{-0,4cm}&&\left[1 -
\frac{\omega^2}{6}(\bm{x}'-\bm{x})^2 \right]\nonumber\\&& \times
{\rm Re} \iint \frac{d^3 \bm{q}}{(2\pi)^3}\frac{d^3
\bm{p}}{(2\pi)^3} a(\bm{q})a^*(\bm{q}+\bm{p})
\frac{(\bm{E}\bm{q})}{\bm{p}^2} e^{ip_0(t'-t_0) -
i(\bm{p}\bm{x}')}\,.\end{aligned}$$ In a many-particle system, this result is to be expressed through the one-particle density matrix, which is accomplished by replacing $a^*(\bm{q}')a(\bm{q}) \to
\varrho_0(\bm{q}',\bm{q}),$ where $\varrho_0$ is the momentum space density matrix at the time instant $t_0.$ We recall that $t_0$ denotes the instant at which the particle state $a(\bm{q})$ is prepared. It can be identified, for instance, as the moment the charge carrier enters the sample, or escapes from a surface trap, etc. The factor $e^{ip_0(t'-t_0)}$ in the integrand then realizes time evolution of the density matrix from $t_0$ to $t'.$ Since $p_0 = (\bm{q} + \bm{p})^2/2m - \bm{q}^2/2m,$ this is a free evolution. The fact that the evolution of charge carriers in solids is not actually free on macroscopic scales is not important for the present consideration in which appearance of the $1/\omega$-singularity is related to the effects of the medium on the field propagators. In this respect, it is essentially different from considerations of Ref. [@kazakov2], in which time evolution of the particle wave packet was the central issue, and, in particular, the requirement that the particle collisions be elastic was important. To take into account particle collisions in the present case, it is sufficient to consider them as instantaneous. Then the interval $(t_0,t')$ is divided into a sequence of short time intervals of order $\tau_f$ (the particle mean free time), on each of which the density matrix evolves freely as in Eq. (\[gefree\]), and changes abruptly at the collision instants. Going through this sequence the density matrix tends to the stationary statistical distribution function, $\varrho(\bm{q}',\bm{q}),$ which is independent of the initial particle state. What is important here is the sign of the difference $(t'-t_0).$ Recall that $t'$ is a fixed time instant to count off the time interval $\tau$ with respect to which the correlation function is Fourier-transformed, and that each particle has its own $t_0.$ This means that for a given $\omega,$ the system is observed during the time interval $(t' - \Delta t,
t' + \Delta t),$ where $\Delta t\sim 1/\omega,$ and $t_0$s are distributed uniformly over this interval. The density matrix evolves forward (backward) in time, if $t'>t_0$ ($t'<t_0$). But time reversal involves inversion of particle momentum, and therefore, the reciprocal contributions to the function $G^E(\bm{x},\bm{x}',t',\omega)$ have opposite signs. To be more specific, let $(t'-t_0)>0.$ Then the exponent $e^{ip_0(t'-t_0)}$ realizes a forward evolution of the density matrix, so that the integral in Eq. (\[gefree\]) takes eventually the form $$\begin{aligned}
\label{contrf}
\iint \frac{d^3 \bm{q}}{(2\pi)^3}\frac{d^3 \bm{p}}{(2\pi)^3}
\varrho(\bm{q}+\bm{p},\bm{q}) \frac{(\bm{E}\bm{q})}{\bm{p}^2} e^{-
i(\bm{p}\bm{x}')}\,.\end{aligned}$$ On the other hand, if $(t'-t_0)<0,$ then the density matrix evolves backward. In momentum space, the initial state of the reversed motion is represented by the amplitude $\tilde{a}(\bm{q}) = a^*(-\bm{q}).$ Taking complex conjugate of the integral in Eq. (\[gefree\]) (which does not change the value of $G^E$ in view of the sign “Re”), and changing the integration variables $\bm{q}\to - \bm{q},$ $\bm{p}\to - \bm{p}$ gives in this case $$- \iint \frac{d^3 \bm{q}}{(2\pi)^3}\frac{d^3 \bm{p}}{(2\pi)^3}
\tilde{a}(\bm{q})\tilde{a}^*(\bm{q}+\bm{p})
\frac{(\bm{E}\bm{q})}{\bm{p}^2} e^{ip_0(t_0-t') -
i(\bm{p}\bm{x}')}\,.$$ After replacement $\tilde{a}^*(\bm{q}+\bm{p})\tilde{a}(\bm{q}) \to
\tilde{\varrho}(\bm{q}+\bm{p},\bm{q}),$ where $\tilde{\varrho}$ plays the role of the momentum density matrix at the moment $t',$ the exponent $e^{ip_0(t_0-t')}$ governs forward evolution of this state on the interval $(t',t_0),$ so that the above expression takes the form $$- \iint \frac{d^3 \bm{q}}{(2\pi)^3}\frac{d^3 \bm{p}}{(2\pi)^3}
\varrho(\bm{q}+\bm{p},\bm{q}) \frac{(\bm{E}\bm{q})}{\bm{p}^2} e^{-
i(\bm{p}\bm{x}')}\,.$$ The density matrix here is the same as in (\[contrf\]), because the statistical distribution is independent of the initial state. We see that reciprocal contributions to the function $G^E(\bm{x},\bm{x}',t',\omega)$ cancel each other when summed over all particles in the system. Thus, we arrive at the important conclusion that the total noise intensity is independent of the number of particles, and remains at the level of individual contribution. As was shown in Ref. [@kazakov2], this conclusion is also true of the disconnected part of the correlation function, though by virtue of quite different reasons.
It is customary to further express the function $\varrho(\bm{q}',\bm{q})$ via the real mixed distribution function, $n(\bm{r},\bm{q}),$ according to $$\begin{aligned}
\varrho(\bm{q}+\bm{p},\bm{q}) = \int d^3\bm{r}e^{i(\bm{p}\bm{r})}
n\left(\bm{r},\bm{q} + \frac{\bm{p}}{2}\right)\,. \nonumber\end{aligned}$$ Probability distributions for the particle position in a sample and for its momentum can be obtained by integrating $n(\bm{r},\bm{p})$ over all $\bm{p}$ and the sample volume, respectively. Using this in the expression (\[contrf\]), and substituting the latter into Eq. (\[gefree\]) yields $$\begin{aligned}
\label{diagenk6e}
G^E(\bm{x},\bm{x}',t',\omega) = - \frac{16\pi i
e^3}{m\beta\omega^3}\hspace{-0,4cm}&&\left[1 -
\frac{\omega^2}{6}(\bm{x}'-\bm{x})^2 \right]\nonumber\\&& \times
{\rm Re} \iiint \frac{d^3 \bm{q}}{(2\pi)^3}\frac{d^3
\bm{p}}{(2\pi)^3} d^3\bm{r} \frac{(\bm{E}\bm{q})}{\bm{p}^2}
n\left(\bm{r},\bm{q} + \frac{\bm{p}}{2}\right)e^{i\bm{p}(\bm{r} -
\bm{x}')}\,.\end{aligned}$$ As we know, the first term in the square brackets in this formula doest not contribute to the voltage power spectrum, and can be omitted. Furthermore, after shifting $\bm{q}\to \bm{q}
- \bm{p}/2,$ and omitting the imaginary term proportional to $(\bm{E},\bm{p}),$ the triple integral becomes purely real, so the symbol “${\rm Re}$” can be omitted. Integrating then over $\bm{p}$ with the help of the formula $$\int\frac{d^3\bm{p}}{(2\pi)^3}e^{i(\bm{p}\bm{x})}\frac{4\pi}{\bm{p}^2}
= \frac{1}{|\bm{x}|}\,,$$ we thus obtain $$\begin{aligned}
\label{diagenk7e}
G^E(\bm{x},\bm{x}',t',\omega) = \frac{2i
e^3(\bm{x}'-\bm{x})^2}{3m\beta\omega} \iint \frac{d^3
\bm{q}}{(2\pi)^3}d^3\bm{r} (\bm{E}\bm{q})
\frac{n(\bm{r},\bm{q})}{|\bm{r}-\bm{x}'|}\,.\end{aligned}$$ Equation (\[corrg\]) shows that the spectral density of the correlation function is given by the same expression (\[diagenk7e\]), so that the power spectrum of voltage fluctuations is, according to Eq. (\[corrfup\]), $$\begin{aligned}
\label{main}
C_U(\bm{x},\bm{x}',t',\omega) = - \frac{2i
e^3(\bm{x}'-\bm{x})^2}{3\beta\omega \Omega} \int d^3\bm{r}
(\bm{E}\overline{\bm{v}}(\bm{r}))\left( \frac{1}{|\bm{r}-\bm{x}|}
+ \frac{1}{|\bm{r}-\bm{x}'|}\right)\,,\end{aligned}$$ where $$\overline{\bm{v}}(\bm{r}) = \Omega\int\frac{d^3
\bm{q}}{(2\pi)^3} \frac{\bm{q}}{m}n(\bm{r},\bm{q})$$ is the local drift velocity of the charge carriers, $\Omega$ denoting the sample volume. For a crystal in the homogeneous external field, $\overline{\bm{v}}$ is a function of the crystalline direction, $$\overline{v}_i = \mu_{ik}E_k\,, \quad i,k=1,2,3,$$ where $\mu_{ik}$ is the charge carrier mobility tensor. With the help of this formula Eq. (\[main\]) can be rewritten as $$\begin{aligned}
\label{mainhom}
C_U(\bm{x},\bm{x}',t',\omega) = - i\eta\,\frac{U^2_0}{\omega}\,,
\quad \eta \equiv \frac{2e^3\mu g}{3\beta}\,,\end{aligned}$$ where $$\mu=\mu_{ik}n_i n_k\,, \quad \bm{n} =
\frac{\bm{E}}{|\bm{E}|}\,,$$ $U_0 = |\bm{E}||\bm{x} - \bm{x}'|$ is the bias applied to the sample (it is assumed that $\bm{E}
\parallel (\bm{x} - \bm{x}'),$ as is usually the case), and $g$ is a geometrical factor $$\begin{aligned}
\label{gfactor}
g \equiv \frac{1}{\Omega}\int\limits_{\Omega} d^3\bm{r}\left(
\frac{1}{|\bm{r}-\bm{x}|} + \frac{1}{|\bm{r}-\bm{x}'|}\right)\,.\end{aligned}$$ If Fourier transformation is defined in a purely real form, i.e., as a decomposition in $\cos(\omega \tau),$ $\sin(\omega\tau),$ rather than in $e^{i\omega\tau},$ then the spectral density is also real: $$\begin{aligned}
\label{mainhomr}
C_U(\bm{x},\bm{x}',t',\omega) = \eta\,\frac{U^2_0}{\omega}\,,
\quad \eta \equiv \frac{2e^3\mu g}{3\beta}\,,\end{aligned}$$
We mention for future reference that if the sample is an elongated parallelepiped with the leads attached to its ends (as is usually the case in practice), then the $g$-factor can be evaluated as $$\begin{aligned}
\label{gfactorapprox}
g \approx \frac{2}{l w h}\int\limits_{w}^{l} \frac{whd x}{x} =
\frac{2}{l}\ln \frac{l}{w}\,,\end{aligned}$$ where $l,w,h$ denote the sample length, width and thickness, respectively, and it is assumed that $h< w\ll l.$ We note also that in the ordinary units, the dimensionless factor $\eta$ reads $$\eta = \frac{2e^3\mu g}{3\beta \hbar^2 c^3} =
\frac{2\alpha^2}{3ec}g\mu T\,,$$ $\alpha$ being the fine structure constant. In particular, in the $CGS$ system of units, $$\begin{aligned}
\label{etaapprox}
\eta \approx 3.4\cdot 10^{-22} g\mu T\,,\end{aligned}$$ where the absolute temperature $T$ is to be expressed in $^{\circ}{\rm K}.$
On the unboundedness of $1/f$-spectrum {#unbound}
--------------------------------------
In this section, a special feature of the derived expression for the power spectrum, namely, its oddness in frequency, will be discussed in connection with the problem of observed absence of frequency limits of the $1/f$-law. As was mentioned in Introduction, $1/f$ noise has been detected in a very wide frequency band $\sim 10^{-6}\,{\rm Hz}$ to $10^6\,{\rm Hz}.$ This fact represents one of the essential difficulties for theoretical explanation, because all physical mechanisms underlying existing models of flicker noise work in much narrower subbands, and none of the models suggested so far has been able to explain the observed plenum of the $1/f$-spectrum.
On the other hand, existence of bounds on this spectrum is generally believed to be necessary in order to guarantee finiteness of the total noise power. There is a well-known argument [@flinn] according to which these limits are actually unnecessary when the flicker noise exponent $\gamma$ is strictly equal to unity, because the logarithmic divergence of the total power is not a problem in this case in view of the existence of natural frequency cutoffs such as the inverse Planck time and lifetime of Universe. However, this reasoning does not work for $\gamma \ne 1,$ in which case divergence is a power of the cutoff. At the same time, the results obtained above reconcile unboundedness of $1/f$-spectrum with the requirements of stationarity and finiteness of the total noise power in a quite natural way. Indeed, using Eq. (\[mainhom\]) we find $$\int\limits_{-\infty}^{+\infty}\frac{d\omega}{2\pi} C_U(\bm{x},\bm{x}',t',\omega)
e^{i\omega\tau} = \eta
U^2_0\int\limits_{-\infty}^{+\infty}\frac{d\omega}{2\pi}
\frac{\sin(\omega \tau)}{\omega} = \frac{\tau}{|\tau|}\frac{\eta
U^2_0}{2}\,.$$ More generally, if the spectrum $C_U\sim
1/f^{\gamma}$ is continued to negative $f$’s as an odd function, then for any $0<\gamma<2$ the integral $$\int\limits_{-\infty}^{+\infty}d\omega C_Ue^{i\omega\tau} \sim
\int\limits_{0}^{+\infty}df \frac{\sin(2\pi f \tau)}{f^{\gamma}}$$ is convergent in both limits $f \to 0$ and $f \to \infty.$ In particular, the singular contribution to the voltage variance (i.e. to the quantity $C_U(t,t')|_{\tau = 0}$) vanishes.
Since appearance of odd contributions to the power spectrum is somewhat unusual in macroscopic fluctuation theory, let us discuss it in more detail. Under stationary external conditions, the voltage noise power spectrum (to be denoted below simply as $C_U(t,t'),$ with the spatial arguments suppressed, for brevity) must be independent of $t'.$ This is an expression of the noise stationarity, or, using a term more suitable for the subsequent discussion, time homogeneity with respect to the macroscopic system. It is usually realized as the requirement that $C_U(t,t')$ be a function of the difference $t-t' \equiv \tau.$ Since $C_U(t,t')$ is also symmetric with respect to the interchange $t
\leftrightarrow t',$ an immediate consequence of this is that it is actually a function of $|\tau|,$ and hence the spectral density is a real even function of frequency. It is important, on the other hand, that time homogeneity is not necessarily exhibited by individual contributions to the total voltage fluctuation, whatever mechanism of flicker noise generation be. In particular, this property evidently does not take place at the microscopic level, i.e., with respect to elementary processes such as charge carrier trapping, surface or grain boundary scattering, etc. Stationarity of the macroscopic process emerges usually upon summation over a large number of individual contributions, so that this microscopic inhomogeneity turns out to be inconsequential. However, this summation is not the only way to obtain a stationary correlation function symmetric in $t,t'.$ Another possibility, which is realized in the present paper, is that flicker noise may be a one-particle phenomenon, in the sense that the entire effect can be ascribed to elementary fluctuations produced by single charge carriers. In this case the function $C_U(t,t')$ does not have to depend solely on $|\tau|,$ and as the explicit calculations of Sec. \[calcul\] show, it actually does not. As was mentioned above, elementary processes are inhomogeneous in time, and hence the symmetry with respect to $t \leftrightarrow
t'$ imposes no restriction on the $\tau$-dependence of the correlation function. The only remaining requirement, namely reality of the correlation function, implies that contributions to the spectral density must be real even, [*or*]{} imaginary odd functions of frequency \[Cf. Eq. (\[corrg\])\]. These two cases correspond to the Fourier decomposition of the function $C_U(t'+\tau,t')$ in $\cos(\omega\tau)$ and $\sin(\omega\tau),$ respectively, and describe the parts symmetric and antisymmetric with respect to the difference of its time arguments. Finally, transition to the statistical distribution removes the $t'$-dependence of the power spectrum \[Cf. transition from Eq. (\[gefree\]) to Eq. (\[diagenk6e\])\]. This restores macroscopic time homogeneity of the correlation function, but leaves the possibility of being odd with respect to the difference of its time arguments. In other words, dependence of the power spectrum on $t'$ shows itself only at microscopic scales, while macroscopically fluctuations look as if they were homogeneous in time.
The $1/f$-spectrum derived in the previous section has no lower frequency cutoff. As to the upper bound, it is given by the condition $f\ll T$ \[see Sec. \[prelim1\]\], or in the ordinary units, $f\ll k T/\hbar \approx 10^{11}T\,{\rm Hz},$ with $T$ expressed in $^{\circ}{\rm K}.$ We see that from the practical point of view, the obtained spectrum has no upper cutoff either.
Validity of Eq. (\[mainhomr\]) and comparison with experimental data {#comparison}
--------------------------------------------------------------------
Let us next discuss the range of applicability of the obtained results. The validity of the perturbative treatment of the external field imposes a very strong bound on the field strength. For this purpose we first collect all characteristic factors that have appeared in the course of extracting the $1/\omega$-asymptotic of the voltage power spectrum. As we have seen, insertion of the vertex describing interaction of the virtual charged particle with external field amounts to multiplying the zero-field diagram by the factor, in the ordinary units, $e(\bm{E}\bm{q})/(m\omega^2 \hbar)\,,$ which is eventually promoted by the $\bm{q}$-integration into $e\mu\bm{E}^2/(\omega^2\hbar).$ Furthermore, the leading non-vanishing contribution to the voltage correlation function has been obtained after expanding $e^{i\bm{k}(\bm{x} - \bm{x}')}$ in the integrand of the $\bm{k}$-integral, which brought in a factor $\omega^2(\bm{x} - \bm{x}')^2/c^2.$ Thus, the overall factor is $e\mu\bm{E}^2(\bm{x} - \bm{x}')^2/(c^2\hbar) = e\mu
U^2_0/(c^2\hbar).$ This is small provided $\mu U^2_0\ll c^2\hbar/e
\approx 10^{3}$ units $CGS,$ which is a quite soft requirement met in virtually all flicker noise measurements (Cf. examples below). The problem with this estimation, however, is that at higher orders, the scalar product $(\bm{E}\bm{q})$ is to be estimated as $|\bm{E}||\bm{q}|,$ because $\overline{(\bm{E}\bm{q})^2}$ is of the order $(|\bm{E}||\bm{q}|)^2\,,$ rather than $(\bm{E}\overline{\bm{q}})^2\,.$ As a result, the requirement that the factor $e(\bm{E}\bm{q})/(m\omega^2 \hbar)$ be small leads to the following upper bound on the electric field strength for a given frequency $\omega,$ in the $CGS$ system of units, $|\bm{E}|
\ll dm\omega^2/e\approx 10^{-25}\omega^2.$ At the same time, the values $|\bm{E}|\sim 1$ are quite normal in flicker noise measurements. In other words, from the point of view of the developed theory, the experimentally relevant regime is identified as the strong field limit. Yet the use of Eq. (\[mainhomr\]) in this limit can be justified to a certain extent by recalling that the perturbative expansion is in reality an asymptotic expansion, and hence the fact that Eq. (\[mainhomr\]) gives the [*first*]{} non-vanishing term of the voltage noise power spectrum implies that the question of validity of the perturbative expansion is actually a question of whether or not it is legitimate to use this expansion to obtain higher order corrections to Eq. (\[mainhomr\]). A rigorous justification is a difficult task because it requires the use of non-perturbative methods. Thus, this issue is left open until careful investigation of the strong field limit. One of the possible ways this problem can hopefully be resolved is a partial summation of the perturbation series, followed by an analytical continuation with respect to $\varkappa.$
After this discouraging observation of strong divergence of the asymptotic series, the more striking turns out to be the fact that Eq. (\[mainhomr\]) is in a general agreement, qualitative and even quantitative, with the existing experimental data. First of all, the spectral density is quadratic in the applied bias. This is perhaps the most solidly established property of flicker noise. Second, the noise level is inversely proportional to the sample size. As to the dependence of flicker noise amplitude on sample dimensions, agreement in the literature is not that good. Experiments are usually arranged so as to prove one of the two main competing points of view on the flicker noise origin, namely wether it is a bulk or surface effect. Although this issue is far from being resolved, there is no doubt that the noise level increases with decreasing sample size. Third, it is generally agreed that, with other things being equal, the flicker noise is more intensive in semiconductors than in metals, and this is again in conformity with Eq. (\[mainhomr\]), because charge carrier mobility is higher in semiconductors than in metals, usually by several orders. Unfortunately, determination of mobility in semiconductors (or semimetals) is a difficult problem, both theoretically and experimentally, and different experiments often give significantly different results. By this reason, the subsequent quantitative consideration will be carried out for metals only. Even in this case careful estimation of the noise level takes some effort. This is because electron mobilities in thin metal films commonly used in flicker noise measurements differ essentially from the corresponding bulk values, varying non-monotonically with the film thickness, and exhibiting complicated temperature dependence. Thus, the thicker the film, the more reliable comparison of theoretical and experimental results. Fortunately, the modern instrumentation allows measurements in sufficiently thick samples, electrical transport in which has bulk properties (usually, effects related to film thickness become important for $h$ less than a few hundred nanometers). As is well known, temperature dependence of the electron mobility in this case is well approximated by the $1/T$ law. Theoretically, this approximation is valid for $T$ higher than the Debay characteristic temperature, but in most cases it is practically applicable already for $T\gtrsim 50^{\circ}{\rm K}.$ Thus, it follows from Eq. (\[mainhomr\]) that the flicker noise level in thick samples is temperature independent. This conclusion is confirmed, e.g., by the results of Ref. [@massiha] where $1/f$ noise was measured in $2.44\,{\rm \mu m}$ thick metal films, which is quite sufficient for bulk treatment of the sample conduction. According to Fig. 5 of Ref. [@massiha], the flicker noise level is constant for $T\gtrsim 50^{\circ}{\rm K}$ indeed. Unfortunately, the authors of [@massiha] did not specify the metals used in their experiments, which makes further comparison with Eq. (\[mainhom\]) impossible.
In order to compare the absolute value of the noise spectral density given by Eq. (\[mainhomr\]) with experimental data, we use the results of the classic paper [@voss1] where flicker noise in thin metal films was investigated. The information provided in this paper is sufficient for estimation of the noise intensity in the gold film shown in Fig. 2 of [@voss1]. This was an elongated sample with $h = 25\,{\rm nm},$ $w = 8\,{\rm\mu
m},$ $l = 625\,{\rm\mu m},$ biased at $U_0 = 0.81\,{\rm V},$ and operated at about $40^{\circ}{\rm K}$ above room temperature. Substituting the sample dimensions in Eq. (\[gfactorapprox\]) gives $g = 140\,{\rm cm}^{-1}.$ Estimation of the electron mobility is more subtle. As was mentioned above, charge carrier mobility in thin films strongly deviates from its bulk value, and this deviation is the main source of uncertainty in evaluating the noise level. In the case under consideration, $\mu$ is isotropic and can be found using the relation $\mu = \sigma/en,$ where $\sigma$ is the electrical conductivity of gold, and $n = 5.9\cdot
10^{22}\,{\rm cm}^{-3}$ is the free electron concentration. The bulk conductivity of gold at $T = 330^{\circ}{\rm K}$ is equal to $4.0\cdot 10^{7}\,{\rm \Omega^{-1} m^{-1}},$ but in thin films the value of $\sigma$ is strongly affected by the grain boundary and surface scattering, surface roughness and other factors. The relevant value of conductivity can be calculated indirectly using the $I-V$ characteristic of the given gold sample, shown in Fig. 3 of [@voss1]. According to this figure, the sample resistance was about $100\, \Omega.$ Taking into account the sample dimensions given above, this implies that $\sigma =1.2\cdot
10^{6}\,{\rm \Omega^{-1} m^{-1}}.$ It should be mentioned that this value is approximately six times lower than that obtained in more recent studies of electrical transport in thin films. For instance, according to Ref. [@bieri] conductivity of a $25\,{\rm nm}$ thick, $15\,{\rm \mu m}$ wide gold film obtained by a laser-improved deposition of nanoparticle suspension, is $7.1
\cdot 10^{6}\,{\rm \Omega^{-1} m^{-1}}.$ The same value can be obtained also indirectly using the data given in Refs. [@chen; @pov]. According to [@chen], the conductivity of gold is $75\%$ to $85\%$ of its bulk value for $h = 100\,{\rm
nm},$ depending on the choice of the substrate, and decreases below that value approximately linearly with decreasing thickness. On the other hand, according to Ref. [@pov] conductivity drops to about $3\cdot 10^{5}\,{\rm \Omega^{-1} m^{-1}}$ for $h =
5\,{\rm nm}.$ One readily finds from this that for $h = 25\,{\rm
nm},$ $\sigma = (6.5 \div 7.5) \cdot 10^{6}\,{\rm \Omega^{-1}
m^{-1}}.$ Presumably, this difference in the values of conductivity is to be attributed to the quality of film deposition. In the case of $\sigma =1.2\cdot 10^{6}\,{\rm
\Omega^{-1} m^{-1}},$ the electron mobility equals to $\mu =
1.3\,{\rm cm^2/Vs},$ and then Eq. (\[etaapprox\]) gives $\eta =
6.0\cdot 10^{-15}\,.$ Substituting this together with the bias value given above in Eq. (\[mainhomr\]), and setting $\omega =
2\pi f$ yields $C_U = 6.3\cdot 10^{-16}\,{\rm V^2/Hz}$ for the frequency $f = 1\,{\rm Hz},$ which is to be compared with the experimental value $C_U \approx 10^{-15}\,{\rm V^2/Hz}\,.$
Discussion and conclusions
==========================
We have shown that the combined action of the temperature and external field effects results in appearance of a principally new contribution to the power spectral density of quantum electromagnetic fluctuations, given by Eqs. (\[main\]), (\[mainhom\]). The power spectrum is thus modified both qualitatively and quantitatively. Being odd with respect to frequency, the new term in the power spectrum describes correlations in the values of voltage measured at two time instants, which are finite for all times. The underlying reason that makes the appearance of the new term possible (apart from the two factors mentioned in the beginning of this paragraph) is the inhomogeneity in time of fluctuations produced by individual charge carriers. As discussed in Sec. \[unbound\], oddness of the found $1/f$-contribution gives a natural explanation to the observed unboundedness of flicker noise spectrum. Although the obtained result is valid, strictly speaking, only for very weak fields, we have seen in Sec. \[comparison\] that it is in qualitative and quantitative agreement with experimental data even beyond its formal range of applicability.
Next, an important qualitative difference of the present considerations from those of Ref. [@kazakov1] is to be emphasized. As we have seen in Sec. \[calcul\], frequency dependence of the power spectral function is determined completely by internal structure of the Feynman diagrams representing the connected part of the correlation function. In other words, dispersion of the correlation function, considered in the present paper, is related to the properties of virtual quanta propagation, and not to the time evolution of the charge carrier wave function. This is in contrast to considerations of Ref. [@kazakov1] where the $1/f$ asymptotic of the power spectrum was related to the spreading of the particle wave packet, and was derived by evaluating the disconnected part of the correlation function.
Finally, regarding discussion of Sec. \[comparison\] it should be stressed that for the purpose of experimental verification of Eq. (\[mainhom\]) only the genuine $1/f$ noise data was used, i.e., the data that fits the law $f^{-\gamma}$ in which $\gamma =
1,$ within experimental error. Otherwise the comparison would be meaningless, even for $f\approx 1\,{\rm Hz}.$ Large deviations of $\gamma$ from unity, observed in some thin films, are presumably due to back reaction of the conducting medium on the electromagnetic field produced by the charge carriers. This issue will be considered elsewhere.
Gauge independence of the correlation function
==============================================
Consider the theory of interacting scalar and electromagnetic fields described by the action $$S = S_{\phi} + S_{A}\,,$$ where $S_{\phi}$ is given by Eq. (\[action\]), and $$S_{A} = - \frac{1}{4}{\displaystyle\int} d^4 x F_{\mu\nu} F^{\mu\nu}
+ S_{gf}\,, \quad S_{gf} = \frac{1}{2\alpha}{\displaystyle\int}
d^4 x~(\partial_{\mu}A^{\mu})^2\,,\quad F_{\mu\nu} =
\partial_{\mu} A_{\nu} - \partial_{\nu} A_{\mu}\,.$$ For arbitrary constant parameter $\alpha,$ the gauge fixing term describes the generalized Lorentz gauge. Let us introduce the generating functional of Green functions $$\begin{aligned}
\label{gener1}
Z[J,\eta,\eta^*] = \int dA d\phi d\phi^* \exp\left\{i\left(S +
\int d^4x[J^{\mu}A_{\mu} + \eta^*\phi + \eta\phi^*]
\right)\right\}\,,\end{aligned}$$ where $J,\eta,\eta^*$ denote sources for the fields $A,\phi^*,\phi,$ respectively. Vanishing of $Z$ under the gauge variation of the functional integral variables $$\delta A_{\mu} =
\partial_{\mu}\xi(x)\,, \quad \delta\phi = ie\xi(x)\phi\,,
\quad \delta\phi^* = -ie\xi(x)\phi^*,$$ with $\xi(x)$ a small gauge function, leads to the Ward identity $$\begin{aligned}
\label{ward}
- i\partial_{\mu} J^{\mu}(y)Z +
\frac{\Box}{\alpha}\partial_{\mu}\frac{\delta Z}{\delta
J_{\mu}(y)} + ie\eta^*(y)\frac{\delta Z}{\delta \eta^*(y)} -
ie\eta(y)\frac{\delta Z}{\delta \eta(y)} = 0\,.\end{aligned}$$ Since we are interested in the connected contribution to the correlation function, we rewrite this identity for the generating functional of connected Green functions, $W = - i\ln
Z,$ $$\begin{aligned}
\label{wardc}
- \partial_{\mu} J^{\mu}(y) +
\frac{\Box}{\alpha}\partial_{\mu}\frac{\delta W}{\delta
J_{\mu}(y)} + ie\eta^*(y)\frac{\delta W}{\delta \eta^*(y)} -
ie\eta(y)\frac{\delta W}{\delta \eta(y)} = 0\,.\end{aligned}$$ The consequence of this equation we need is obtained by functional differentiation with respect to $\eta,\eta^*,$ and twice with respect to $J,$ with all the sources set equal to zero afterwards, $$\begin{aligned}
&&
\frac{\Box^y}{\alpha}\partial^y_{\mu}\frac{\delta^5 W}{\delta
J_{\mu}(y)\delta J_{\alpha}(x)\delta
J_{\beta}(x')\delta\eta(z)\delta\eta^*(z')} + ie\delta^{(4)}(y -
z')\frac{\delta^4 W}{\delta J_{\alpha}(x)\delta
J_{\beta}(x')\delta\eta(z)\delta \eta^*(y)} \nonumber\\&& -
ie\delta^{(4)}(y - z)\frac{\delta^4 W}{\delta J_{\alpha}(x)\delta
J_{\beta}(x')\delta \eta(y)\delta\eta^*(z')} = 0\,. \nonumber\end{aligned}$$ Fourier transform of this identity with respect to $y$ reads $$\begin{aligned}
\label{wardc1}&&
\frac{k^2_1}{\alpha}k_{1\mu}\int d^4 y e^{-ik_1 y}\frac{\delta^5
W}{\delta J_{\mu}(y)\delta J_{\alpha}(x)\delta
J_{\beta}(x')\delta\eta(z)\delta\eta^*(z')} \nonumber\\ &&=
e(e^{-ik_1 z'} - e^{-ik_1 z}) \frac{\delta^4 W}{\delta
J_{\alpha}(x)\delta J_{\beta}(x')\delta\eta(z)\delta
\eta^*(z')}\,.\end{aligned}$$ The argument of the Fourier transform is purposely denoted here by $k_1$ to stress that the left hand side of this equation corresponds to the variation of the Green function we dealt with in Sec. \[calcul\], under gauge variation of the external field. Indeed, the longitudinal part of the photon propagator in the generalized Lorentz gauge has the form $$\begin{aligned}
\label{long}
D^l_{\mu\nu}(k) = - \alpha\frac{k_{\mu}k_{\nu}}{k^4}\,.\end{aligned}$$ Therefore, contraction with the factor $k^2_1
k_{1\mu}/\alpha$ is equivalent to amputation of the photon propagator attached to the $y$ vertex, followed by contraction of this vertex with $k_{1\mu}.$ Exactly the same result is obtained under a gauge variation of the external field coming into this vertex. The only difference with the Green function we considered in Sec. \[calcul\] is that the external scalar lines in Eq. (\[wardc1\]) are the particle propagators. To promote them into particle amplitudes, according to the standard rules, Eq. (\[wardc1\]) is to be Fourier transformed with respect to the variables $z,z',$ and then multiplied by $a(\bm{q})a^*(\bm{q}')(m^2 - q^2)(m^2 - q'^2),$ where the arguments $q,q'$ of the Fourier transformations with respect to $z,z'$ are to be taken eventually on the mass shell. But these operations give zero identically when applied to the right hand side of Eq. (\[wardc1\]), because each of the factors $e^{-ik_1
z'},$ $e^{-ik_1 z}$ makes the corresponding particle propagator nonsingular on the mass shell. For instance, the first term in Eq. (\[wardc1\]) gives rise to the contribution of the form $(m^2 - q'^2)D^{\phi}(q' + k_1)$ times terms nonsingular on the mass shell. For $k_1 \ne 0,$ the function $D^{\phi}(q' + k_1)$ is also nonsingular at $q'^2=m^2,$ and hence this contribution vanishes on the mass shell.
Thus, the correlation function is invariant under the gauge transformations of the external field, which are part of the gauge freedom in the theory. The other part is related to the explicit dependence of the photon propagator on the choice of the gauge conditions used to fix the gauge invariance of the action. As is well known, it is the longitudinal part of the propagator that depends on the gauge, and the most general Lorentz-invariant form of this part is given by Eq. (\[long\]) in which $\alpha$ is to be regarded as an arbitrary function of $k^2.$ It is not difficult to see that variations of $\alpha(k^2)$ do not affect the observable quantities. Recall, first of all, that we are interested ultimately in the fluctuations of gauge-invariant quantities such as the electric field strength. The $\alpha$-independence of these quantities is a direct consequence of their gauge invariance, because variations of $\alpha(k^2)$ give rise to terms that are pure gradients with respect to the spacetime arguments $x,x',$ as is easily verified by substituting the expression (\[long\]) in place of one or two photon propagators in Eq. (\[diagen\]). Then, if the vector potential contribution to the field strength is negligible, as is the case in our nonrelativistic calculation (recall the condition $|\bm{q}|\ll m$ used throughout), the voltage correlation function can be found by integrating the correlation function for the field strength with respect to $\bm{x},\bm{x}'$ using the relation $\bm{E} = - \bm{\nabla}A_0.$
Thus, gauge-independence of our results expressed by Eqs. (\[main\]), (\[mainhom\]) is proved.
See, for instance, Buckingham M 1983 [*Noise in Electronic Devices and Systems*]{} (Chichester: Ellis Horwood). For recent reviews of the problem see
Wong H 2003 [*Microelectron. Reliab.*]{} [**43**]{} 585
Raychaudhuri A K 2002 [*Current Opinion in Solid State & Materials Science*]{} [**60**]{} 67
Milotti E 2002 [*E-print archive physics*]{}/0204033 and references therein. General mathematical description of $1/f$ noise can be found in
Kaulakys B, Gontis V and Alaburda M 2005 [*Phys. Rev.*]{} E [**71**]{} 051105, which also contains an extensive bibliography. An up-to-date bibliographic list on $1/f$-noise can be found at http://www.nslij-genetics.org/wli/1fnoise
Handel P H 1975 [*Phys. Rev. Lett.*]{} [**34**]{} 1492
Handel P H 1980 [*Phys. Rev.*]{} A [**22**]{} 745. The complete reference list on Handel’s theory is too extensive to cite here. A fairly complete bibliography on the quantum theory approach to $1/f$-noise can be found at http://www.umsl.edu/ handel/QuantumBib.html
Handel P H 1994 [*IEEE Trans. on Electron. Devices*]{} [**41**]{} 2023
Handel P H 1996 [*Phys. Stat. Sol.*]{} (b) [**194**]{} 393
Handel P H 1999 in [*Wiley Encyclopedia of Electrical and Electronics Engineering*]{} [**14**]{} ed Webster J G (John Wiley [&]{} Sons) p 428
Zwanziger D 1975 [*Phys. Rev.*]{} D [**11**]{} 3481
Kibble T W B 1968 [*Phys. Rev.*]{} [**173**]{} 1527
Tremblay A-M 1978 [*PhD thesis*]{} Massachusetts Institute of Technology
Nieuwenhuizen Th M, Frenkel D and van Kampen N G 1987 [*Phys. Rev.*]{} A [**35**]{} 2750
Van Vliet C M 1990 Physica A [**165**]{} 101
Van Vliet C M 1990 Physica A [**165**]{} 126
van der Ziel A 1988 Unified Presentation of 1/f Noise in Electronic Devices; Fundamental 1/f Noise Sources [*Proc. IEEE*]{} [**76**]{} 233
van der Ziel A 1988 [*J. Appl. Phys.*]{} [**63**]{} 2456
Kazakov K A 2006 [*Int. J. Mod. Phys.*]{} B [**20**]{} 233
Kazakov K A 2006 [*Journal of Physics A: Mathematical and General*]{} [**39**]{} 7125
Schwinger J 1961 [*J. Math. Phys.*]{} [**2**]{} 407
Schwinger J Particles, Sources and Fields (Addison-Wesley, Reading, Mass., 1970)
Keldysh L V 1964 [*Zh. Eksp. Teor. Fiz.*]{} [**47**]{} 1515 \[1965 [*Sov. Phys. JETP*]{} [**20**]{} 1018\]
Kazakov K A 2005 [*Phys. Rev. D*]{} [**71**]{} 113012
Landsman N P and van Weert Ch G 1987 [*Phys. Reports*]{} [**145**]{} 141
Niemi A J and Semenoff G W 1984 [*Ann. Phys.*]{} [**152**]{} 105
Niemi A J and Semenoff G W 1984 [*Nucl. Phys.*]{} B [**230**]{} \[FS10\] 181
Flinn I 1968 [*Nature*]{} [**219**]{} 1356
Massiha G H and Rawat K S 2002 [*Journal of Industrial Technology*]{} [**18**]{} 1
Voss R F and Clarke J 1976 [*Phys. Rev.*]{} B [**13**]{} 556
Chen G et al 2005 [*Appl. Phys.*]{} A [**80**]{} 659
Povilus A 2003 Electronic properties of metals and semiconductors, [*Michigan Univ. Report*]{} N 441
Bieri N R et al 2004 [*Superlattices and Microstructures*]{} [**35**]{} 437
**Figure captions**
Fig.1: Feynman diagrams representing connected part of correlation function. Wavy lines denote photon propagators, solid lines massive particle. $q$ and $p$ are the particle 4-momentum and 4-momentum transfer, respectively.
Fig.2: Symbolic diagrammatic picture of the effect of particle collisions and external electric field (dashed line) on the particle wave function.
Fig.3: Feynman diagrams describing the first order external field correction to the particle propagator.
![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig1"}](fig1a.eps "fig:") ![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig1"}](fig1b.eps "fig:") ![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig1"}](fig1c.eps "fig:")
![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig2"}](fig2.eps)
![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig3"}](fig3a.eps "fig:") ![Kazakov, Quantum fluctuations of Coulomb potential as a source of flicker noise. The influence of heat bath[]{data-label="fig3"}](fig3b.eps "fig:")
[^1]: The photon propagator is written here in the Feynman gauge. The question of gauge independence of the correlation function is considered in the Appendix.
[^2]: This phase factor incorporates contributions involving only vertices linear in the photon field. This is sufficient for the subsequent discussion concerned with the leading correction to the propagator. Inclusion of the other interaction vertex adds terms of higher orders in $\bm{E}$ to the exponent.
[^3]: Independently of the proof given in the Appendix, it can be noticed that the constant term in the potential does not involve the $\bm{k}_1$-differentiation, and hence does not contribute to the leading singularity for $\omega
\to 0$ anyway.
[^4]: Moreover, if the function $a(\bm{q})$ is decomposed as in Eq. (\[abrel\]), and the potential is chosen respectively to vanish at the point $\bm{x}_0,$ i.e., $\varphi(\bm{x}) = -(\bm{E},\bm{x} - \bm{x}_0),$ then the $\bm{p}$-derivative acts on the product of $e^{i\bm{p}\bm{x}_0}$ with the slowly varying factors $b(\bm{q}+\bm{p}),$ $\varepsilon_{\bm{q}+\bm{p}},$ $p_0.$ Differentiation of the exponent gives $\bm{x}_0$ which just cancels the purposely chosen constant term in the potential, while the result of differentiation of the remaining factors can be neglected to the leading order of the long-range expansion. This observation can be useful in assessing the higher order corrections to the correlation function.
[^5]: It is convenient to prove this relation before the integration over $\bm{k}_1$ using the following sequence of substitutions: $\bm{k}\to \bm{k} +
\bm{p} - \bm{k}_1\,,$ $\bm{q} \to \bm{q} - \bm{p}\,,$ and then $\bm{p} \to - \bm{p}\,.$ The extra factor $(-1)$ coming from complex conjugation of the imaginary unit in the factor $\varkappa$ is compensated by that from the integration by parts with respect to $\bm{k}_1.$ Alternatively, this can be proved directly in coordinate space (after extracting the relevant contribution) by substituting $\exp\left\{ie(\bm{E},\bm{z}+\bm{z}')(z'_0-z_0)/2\right\}D^{\phi}(z,z')$ for the particle propagator, interchanging the integration variables $z,z',$ and taking into account symmetry of the functions $D(x,z),$ $D^{\phi}(z,z')\,.$
[^6]: Here $\bm{p}$ is neglected in comparison with $\bm{q}.$ Otherwise, there is a residual term proportional to $p_0$ in the total of the two diagrams. The usual manipulation with the integration variables $\bm{q}\to \bm{q} -
\bm{p},$ $\bm{p}\to-\bm{p}$ shows that this term can be obtained by substituting $(\bm{E}\bm{q})\to (\bm{E}\bm{p}).$ Its ratio to the main contribution considered in the text is $|\bm{p}|p_0/m|\bm{q}| \sim \bm{p^2}/m^2.$ For electrons in a sample of characteristic size $l\,{\rm cm},$ this is, in the ordinary units, $(\hbar/lmc)^2 \sim 10^{-20}/l^2.$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a version of the stochastic maximum principle (SMP) for ergodic control problems. In particular we give necessary (and sufficient) conditions for optimality for controlled dissipative systems in finite dimensions. The strategy we employ is mainly built on duality techniques. We are able to construct a dual process for all positive times via the analysis of a suitable class of perturbed linearized forward equations. We show that such a process is the unique bounded solution to a Backward SDE on infinite horizon from which we can write a version of the SMP.'
address:
- 'Dipartimento di Matematica, Sapienza Università di Roma. Piazzale Aldo Moro 5, 00185 Roma, Italia'
- 'Dipartimento di Matematica e Applicazioni, Università di Milano-Bicocca. via Cozzi 55, 20125 Milano, Italia'
- 'Institute of Information Theory and Automation, Czech Academy of Sciences, Pod Vodárenskou věž'' i 4, Praha 8, 182 08, Czech Republic'
author:
- Carlo Orrieri
- Gianmario Tessitore
- Petr Veverka
bibliography:
- 'mybib.bib'
title: Ergodic maximum principle for stochastic systems
---
Introduction
============
We consider an optimal control problem with the following controlled *dissipative stochastic state equation* $$\left\lbrace\begin{array}{@{}l@{}}
dX_t = b(X_t,u_t)dt + \sigma(X_t,u_t)dW_t, \qquad t\geq 0,\\
X_0 = x,
\end{array}\right.$$ and an *ergodic cost functional* (e.g. a functional that depends only on the asymptotic behaviour of the state and of the control) such as: $$\begin{aligned}
J^{\inf} (u(\cdot)) &= \liminf_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T f(X_t,u_t)dt, \label{cost inf}\\
J^{\sup} (u(\cdot)) &= \limsup_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T f(X_t,u_t)dt. \label{cost sup}\end{aligned}$$ In the above the state $X$ is a ${\mathbb{R}}^n$-valued process and $(W_t)_{t\geq 0}$ is a $d$-dimensional Wiener process. Moreover the drift $b$ and diffusion $\sigma$ satisfy a joint monotonicity condition. Finally the control process $(u_t)$ is progressively measurable and takes values in a non-empty convex subset $U\subset {\mathbb{R}}^l$.
We refer to this setting as *ergodic control problem*. The choice of the functionals refers to “minmin” and “minmax” formulation. Our aim is to find a correct formulation of the stochastic maximum principle (SMP) in the sense of Pontryagin, by means of which we have at our disposal some necessary (and sufficient) condition for optimality.
Alternatively, under stronger regularity assumptions, one can use the dynamic programming and derive the Hamilton-Jacobi-Bellman equation whose solution gives the optimal cost and the optimal feedback control. In finite dimensions, the first result was obtained in the paper by Mandl [@mandl1964control], later generalized by Bokar and Gosh in [@borkar1988ergodic]. For further generalizations of such an HJB approach, both in finite and infinite dimensional framework both by analytic and by probabilistic tools see e.g. [@goldys1999ergodic; @fuhrman2009ergodic; @hu2015probabilistic; @debussche2011ergodic; @richou2009ergodic; @bensoussan1987equations; @arisawa1998ergodic] We also refer to [@goldys2011stochastic] for a survey on recent results obtained in this direction.
Nevertheless, it is by now well known that, even if it only provides necessary ( only under strong convexity requirements also sufficient) optimality conditions, the SMP normally requires much less regularity and structural condition allowing for instance to easily include the case of control dependent diffusion. The first general formulation of the SMP for finite horizon controlled stochastic systems in finite dimensions was obtained by Peng in [@peng1990general]. After this seminal paper, many directions have been followed by many authors. For what concerns ergodic costs, though, the theory is not yet fully developed. As far as we know, the only version of necessary and sufficient condition for optimality goes back to the paper by Kushner [@kushner1978optimality] in 1978, in which no backward stochastic equation appeared. In that framework the author adopted a martingale solution approach and considered only Markov feedback controls. The system is also assumed to be stable for each control. Under these assumptions, for each stationary Markov control there exists a unique invariant measure $\mu_u(\cdot)$ such that the initial cost functional can be rewritten in the following way $$\lim_{T \to \infty} {\mathbb{E}}_{x_0}^u \frac{1}{T} \int_0^T f(X_t,u(X_t)) dt = \int f(x,u(x)) \mu_u(dx).$$ Using this formulation, Kushner derived a necessary and sufficient condition for $u(\cdot)$ to be optimal, which he called a “dynamic programming like” condition. Let us also mention a recent preprint [@cohen2015classical] in which the authors give some sufficient condition for optimality, studying the adjoint Backward SDE, as well as Feller property and exponential ergodicity of the controlled process. As in the present paper the adjoint BSDE is multidimensional in an infinite horizon. The point is that the approach chosen in [@cohen2015classical] to prove well posedness of such an equation relies on Girsanov argument and seems to work under commutativity requirements that are satisfied when $n=1$ or when $\sigma$ is constant. Also see [@guatteri2009ergodic] for infinite horizon multidimensional BSDEs in the context of linear quadratic stationary optimal control.
Our formulation is fairly general. We do not impose the existence of a limit in the formulation of the cost functional and we consider general progressive controls. Moreover, notice that the convexity assumption on the control actions is a natural choice for the ergodic control problems. Indeed, due to the dissipativity of the system, a spike variation argument is not sufficient to extract useful information on the behaviour of the system at infinity. In the present paper we deduce a version of the maximum principle written in terms of the unique bounded solution to a multidimensional backward SDE on infinite horizon $$-dp_t = \left[ D_xb(X_t,u_t)^*p_t + D_x\sigma(X_t,u_t)^*q_t - D_xf(X_t,u_t)\right]dt - q_tdW_t.$$ As far as we know, a well-posedness result for backward equations of this form is new. The major difficulty to overcome is the lack of integrability in time of the forcing term of the equation. Due to the hypothesis on the state equation we can guarantee that $$\sup_{t \geq 0} \left( {\mathbb{E}}\abs{D_x f(X_t,u_t)}^r \right)^{1/r} < \infty;\qquad \text{for some }r > 1.$$
Similar equations are studied in the formulation of the SMP for discounted cost functionals in infinite horizon, see e.g. [@maslowski2014sufficient], [@orrieri2015necessary]. In that case, though, the spaces in which one is looking for a solution are weighted $L^2$-spaces, allowing the solution to explode at infinity in a controlled way. Here, due to the stability of the system, we expect the solution to be bounded up to infinity.
The strategy we employ is mainly built on duality techniques. Via the analysis of a suitable class of perturbed linearized forward eqautions, see equation below, we are able, exploiting their dissipativity, to construct an adjoint process for all positive times. We introduce then a well-suited family of truncated equations and we show the consistency of the family with respect to the varying finite horizon $T>0$, as $T \to \infty$.
We also propose a second version of maximum principle involving a family of backward equations on finite time horizon $T$ with terminal condition $p^T_T = 0$ that could be verifiable in certain cases, see Remark \[thm:necessary truncated\] below.
Once we have a necessary condition for optimality, it is natural to ask also for a sufficient counterpart of it. As in the classical setting, an extra convexity assumption on the Hamiltonian of the system guarantees the required sufficiency.
The paper is structured as follows. In Section 2 we fix the notation and we discuss the main assumptions on the state equation and on the control actions. In Section 3 we study the convex perturbation of the optimal control and we expand the optimal trajectory and cost functional with respect to the perturbation. Section 4 is the core of the paper. Here we introduce the adjoint equation and we present a well-posedness result for it. The main results concerning the necessary and sufficient versions of the SMP are contained in Section 5 and 6.
Preliminaries and assumptions
=============================
Let $(\Omega, {\mathcal{F}}, {\mathbb{P}})$ be a complete probability space and $(W_t)_{t\geq 0}$ a standard $d$-dimensional Brownian motion. Throughout the paper we use the natural filtration $(\mathcal{F}_t)_{t\geq 0}$ associated to $W$, augmented in the usual way with the family of ${\mathbb{P}}$-null sets of $\mathcal{F}$. By $|\cdot|$ we denote the Euclidean norm on ${\mathbb{R}}^n$ and $\norm{\cdot}_2$ denotes the Hilbert-Schmidt norm on ${\mathbb{R}}^{n \times n}$.
For any $p \geq 1$ and $T>0$ we define
- $L^p(\Omega\times [0,T]; {\mathbb{R}}^n)$, the set of all $(\mathcal{F}_t)$-progressive processes with values in ${\mathbb{R}}^n$ such that $$\norm{X}_{L^p(\Omega\times[0,T];{\mathbb{R}}^n)} = \left( {\mathbb{E}}\int_0^T \abs{X_t}^p dt\right)^{1/p} < \infty;$$
- $L^p({\mathbb{R}}_+; L^q(\Omega;{\mathbb{R}}^n))$ the set of all $(\mathcal{F}_t)$-progressive processes with values in ${\mathbb{R}}^n$ with $1\leq q < +\infty$ such that $$\norm{X}_{L^p({\mathbb{R}}_+; L^q(\Omega;{\mathbb{R}}^n))}^p = \int^{\infty}_0 {\left(}{\mathbb{E}}\abs{X_t}^q {\right)}^{\frac{p}{q}}dt < \infty,\quad \text{for } 1\leq p < +\infty,$$ and $$\norm{X}_{L^{\infty}({\mathbb{R}}_+; L^q(\Omega;{\mathbb{R}}^n))} = \sup_{t \geq 0} {\left(}{\mathbb{E}}\abs{X_t}^q {\right)}^{\frac{1}{q}} < \infty.$$
The aim of this work is to give some necessary (and sufficient) condition for optimality of a controlled system of the form $$\label{SDE}
\left\lbrace\begin{array}{@{}l@{}}
dX_t = b(X_t,u_t)dt + \sigma(X_t,u_t)dW_t, \qquad t\geq 0,\\
X_0 = x,
\end{array}\right.$$when a cost functional of ergodic type has to be minimized. The form of the cost functional slightly differs when considering a $\liminf$ or a $\limsup$ formulation. We define a truncated cost functional in the following form $$\label{cost truncated}
J_T (u(\cdot)) = {\mathbb{E}}\int_0^T f(X_t,u_t)dt.$$ Let us denote the two forms in the following way $$\label{eq:cost_inf}
J^{\inf} (u(\cdot)) = \liminf_{T \to \infty} \frac{1}{T} J_T(u(\cdot)) = \liminf_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T f(X_t,u_t)dt,$$ $$\label{eq:cost_sup}
J^{\sup} (u(\cdot)) = \limsup_{T \to \infty} \frac{1}{T} J_T(u(\cdot)) = \limsup_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T f(X_t,u_t)dt.$$ An control process $\bar{u}(\cdot)$ is said to be optimal either if $$\label{ergodic cost}
J^{\inf} (\bar{u}(\cdot)) = \inf_{u(\cdot) \in \mathcal{U}} J(u(\cdot)) \qquad \text{ or } \qquad J^{\sup} (\bar{u}(\cdot)) = \inf_{u(\cdot) \in \mathcal{U}} J(u(\cdot)),$$ where $\mathcal{U}$ indicates a class of admissible controls. Now we give some assumptions on the state equation and on the control actions.
\[Hyp\] Assumptions involve three constants $m\geq 0$ and $p>(4m+2){\color{red}\vee 4}$ and $k>(p-1)/2$ that we fix now and for the rest of the paper.
- - - ([**[Controls]{}**]{}) $U$ is a closed convex subset of ${\mathbb{R}}^l$. Moreover $u$ is a progressively measurable $U$-valued process. We say that $u$ is an *admissible control* if it satisfies: $$\sup_{t \geq 0}{\mathbb{E}}\abs{u_t}^p < +\infty.$$
- ([**[Polynomial growth]{}**]{}) The vector field $b: {\mathbb{R}}^n \times U \rightarrow {\mathbb{R}}^n$ is $\mathcal{B}({\mathbb{R}}^n)\otimes \mathcal{B}(U)$-measurable and $\mathcal{C}^2$ with respect to $x$ and $u$. There exists $C_1 >0$ such that $$\abs{D_ub(x,u)} \leq C_1, \qquad x \in {\mathbb{R}}^n, u \in U.$$ Moreover: $$\label{eq:grad b}
\sup_{u \in U}\sup_{x \in \mathbb{R}^n} \frac{\abs{D^{\beta}_x b(x,u)} }{1+ \abs{x}^{2m+1-\beta} + \abs{u}^{1-\beta}} < +\infty,\qquad \beta =0,1.$$
- ([**[Polynomial growth]{}**]{}) The mapping $\sigma: {\mathbb{R}}^n \times U \rightarrow {\mathbb{R}}^{n\times d} $ is measurable with respect to $\mathcal{B}({\mathbb{R}}^n) \otimes \mathcal{B}(U)$. There exists $C_2 >0$ such that $$\norm{D_u\sigma(x,u)}_2 \leq C_2, \qquad x \in {\mathbb{R}}^n, u \in U.$$ Moreover it is $\mathcal{C}^2$ with respect to $x$, $u$ and: $$\label{eq:grad sigma}
\sup_{u\in U}\sup_{x \in \mathbb{R}^n} \frac{\norm{D^{\beta}_x \sigma(x,u)}_2 }{1+ \abs{x}^{m-\beta} + \abs{u}^{1-\beta}} < +\infty,\qquad \beta =0,1.$$
- ([**[Joint dissipativity]{}**]{}) There is $c_p <0$ such that $$\label{eq:joint dissipativity}
\braket{D_xb(x,u)y,y} + k\norm{D_x\sigma(x,u)y}^2_2 \leq c_p \abs{y}^2, \qquad x,y \in \mathbb{R}^n, u \in U.$$
- ([**[Cost]{}**]{}) The function $f: {\mathbb{R}}^n \times U \rightarrow {\mathbb{R}}$ is $\mathcal{B}({\mathbb{R}}^n)\otimes \mathcal{B}(U)$-measurable, bounded from below by a constant $f_0$, it is differentiable in $x$ and $u$ and $$\abs{D_xf(x,u)} + \abs{D_uf(x,u)}\leq C(1 + \abs{x} + \abs{u}),$$ for some $C >0$.
We refer to [@cerrai2001second] and [@orrieri2015necessary] for a discussion on the joint monotonicity and on the relation between the growth of $b$ and $\sigma$. Concerning (H5), here we limit ourselves to linear growth for simplicity. A general polynomial growth can be easily achieved.
The choice of $p > 4m+2$ in (H1)-(H4) comes from the interplay between the dissipative behaviour of the system and polynomial growth of the coefficients. Actually this bound can be easily derived from the maximal moment of the state process that we need to estimate in the proofs (see Proposition \[l:convergence X tilde\]). The condition for $k$ is then the natural one.
We can state the following
\[t.existence.SDE\] Assume that Hypothesis \[Hyp\] holds true. Then, for every $x \in {\mathbb{R}}^n$ and every admissible control $u(\cdot)$, equation admits a unique progressively measurable solution for each admissible control. Moreover, the following estimate holds $$\label{eq:estimate.state}
{\mathbb{E}}\abs{X_t}^{p} \leq e^{-p\beta t}\abs{x}^{p} + K \sup_{t \geq 0}{\mathbb{E}}\abs{u_t}^{p},$$ for some positive constants $K = K(p,c_p)$ and $\beta$.
Define $\tilde{X}_t := e^{\beta t}X_t$ for a positive $\beta$. Then $\tilde{X}$ solves $$\label{eq:SDE tilde}
\left\lbrace\begin{array}{@{}l@{}}
d\tilde{X}_t = \beta\tilde{X}_t + e^{\beta t} b( e^{-\beta t} \tilde{X}_t,u_t)dt
+ e^{\beta t}\sigma ( e^{-\beta t} \tilde{X}_t,u_t)dW_t, \qquad \forall t\geq 0,\\
\tilde{X}_0 = x.
\end{array}\right.$$ If we call $\tilde{b}_t(x,u) = e^{\beta t} b\bigl( e^{-\beta t} x, u \bigr)$ and $\tilde{\sigma}_t(x,u) = e^{\beta t} \sigma \bigl( e^{-\beta t} x,u \bigr)$ then also $\tilde{b_t}, \tilde{\sigma_t}$ satisfy Hypothesis \[Hyp\]. In particular the joint dissipativity holds with the same constant $$\braket{\tilde{b}_t(x,u) - \tilde{b}_t(y,u), x-y} + \frac{p-1}{2}\norm{\tilde{\sigma}_t(x,u) - \tilde{\sigma}_t(y,u)}^2_2
\leq c_p |x-y|^2.$$ Let $p \geq 2$, denote $p = 2q$ and $\tilde a = \tilde \sigma(x,u)^*\tilde \sigma(x,u) $ (we omit the time dependence $\tilde \sigma = \tilde \sigma_t$ when it is clear). We apply the Itô formula to the function $f(x) = \abs{x}^{2q}$ to get $$\begin{split}
{\mathbb{E}}&\abs{\tilde{X}_t}^{2q} = |x|^{2q}
+ 2q{\mathbb{E}}\int^t_0 \abs{\tilde{X}_s}^{2(q-1)}\left(\braket{\tilde{X}_s, \tilde{b}( \tilde X_s,u_s)}
+ \frac{1}{2} \norm{\tilde\sigma(\tilde X_s,u_s)}_2^2 \right) ds \\
&\quad+ 2q\beta{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2q}ds + 2q(q-1){\mathbb{E}}\int^t_0 \abs{\tilde{X}_s}^{2(q-2)}Tr{\left\lbrace}\tilde{a}_s {\left(}\tilde{X}_s \otimes \tilde{X}_s {\right)}{\right\rbrace}ds \\
&\leq |x|^{2q} + 2q{\mathbb{E}}\int^t_0 \abs{\tilde{X}_s}^{2(q-1)}\left(\braket{\tilde{X}_s, \tilde{b}( \tilde X_s,u_s)} +(q- \frac{1}{2}) \norm{\tilde\sigma(\tilde X_s,u_s)}_2^2 \right) ds + 2q\beta{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2q}ds\\
&\leq |x|^{2q} + 2q{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2(q-1)}\left(\braket{\tilde{X}_s, \tilde{b}(\tilde X_s,u_s) - \tilde{b}(0,u_s)} + (q -\frac{1}{2})(1+\varepsilon)\norm{\tilde\sigma(\tilde X_s,u_s) - \tilde\sigma(0,u_s)}_2^2 \right) ds \\
\end{split}$$ $$\begin{split}
&\quad+2q{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2(q-1)}\left(\braket{\tilde{X}_s, \tilde{b}(0,u_s)}
+ c_\varepsilon\norm{\tilde\sigma(0,u_s)}_2^2 \right) ds + 2q\beta{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2q}ds\\
&\leq |x|^{2q} + 2q\left( c_{r} + \beta + \frac{\delta}{2}\right){\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2q}ds \qquad \qquad \qquad \qquad \qquad \qquad(\text{ with } r = 2q(1+\varepsilon) - \varepsilon)\\
&\quad+2q{\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2(q-1)}\left(\frac{1}{2\delta}\abs{\tilde{b}(0,u_s)}^2 + c_\varepsilon\norm{\tilde\sigma(0,u_s)}_2^2 \right) ds \\
&\leq |x|^{2q} + 2q\left( c_{r} + \beta + \frac{\delta}{2} + c_\delta\delta^{q/(q-1)} \right){\mathbb{E}}\int_0^t\abs{\tilde{X}_s}^{2q}ds\\
&\quad+ 2q{\mathbb{E}}\int_0^t e^{-2q\beta t}\left(\frac{1}{2^q \delta^{q+1}q} + \frac{c_\varepsilon}{\delta^q} \right)\abs{u_s}^{2q}ds, \\
\end{split}$$ where we employed joint dissipativity for the process $\tilde X$, we repeatedly used weighted Young inequality and in the end the growth condition on the coefficients. Choosing $\beta$ and $\delta$ small enough, thanks to Hypothesis (H1) we end up with the following estimate $$\begin{split}
{\mathbb{E}}\abs{X_t}^{2q} &\leq e^{-2q\beta t}\abs{x}^{2q} + C \int_0^t e^{-2q\beta(t-s)} {\mathbb{E}}\abs{u_s}^{2q}ds \\
&\leq e^{-2q\beta t}\abs{x}^{2q} + C \sup_{t \geq 0}{\mathbb{E}}\abs{u_t}^{2q}.
\end{split}$$ Notice that, taking the supremum on both sides we also have that $$\sup_{t \geq 0}{\mathbb{E}}\abs{X_t}^{2q} \leq C(\abs{x}^{2q} + 1),$$ and the claim is proved.
Perturbation of the controls
============================
When considering ergodic control problems we can not expect to gain information by the use of local in time perturbations of the optimal control.
More precisely, let ${u}^i(\cdot)$, $i=1,2$ are admissible controls with $u^1_t= {u}^2_t $ for all $t>T_0$. If one denotes by ${X}^i$, $i=1,2$ the corresponding states then by the dissipativity assumption (H4) one gets $\mathbb{E}|X^1_t-X^2_t|^2\rightarrow 0$ for $t>T_0$ exponentially fast (let us say with exponential decay $\varepsilon$). Consequently (assume for a moment that $f$ is Lipschitz) $$\begin{split}
| J(u^1(\cdot)) - J({u}^2(\cdot))| &=
\lim_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T \left| f({X}^1_t,{u}^1_t) - f({X}^2_t,{u}^2_t)\right|dt \\
&=
\lim_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_{T_0}^T \left| f({X}^1_t,{u}^1_t) - f({X}^2_t,{u}^1_t)\right|dt \\
&\leq C \lim_{T \to \infty} \frac{1}{T}\int_{T_0}^T e^{-\varepsilon t} dt = 0,
\end{split}$$
This is the reason for considering the perturbations which act on the system up to infinity. Notice that it is crucial to require that $U$ is convex.
Let then $ \bar{u}(\cdot)$ be an optimal control for the ergodic control problem and denote the corresponding state process as $\bar{X}$. For $\theta \in (0,1]$ and $u(\cdot)$ admissible control define $u^{\theta}$ as a convex combination by $u^{\theta}(\cdot):
=
(1-\theta) \bar{u}(\cdot) + \theta u(\cdot)= \bar{u}(\cdot) + \theta v(\cdot)$, where $v(\cdot) := u(\cdot) - \bar{u}(\cdot)$. Then $u^{\theta}(\cdot)$ is admissible and the corresponding state is denoted by $X^{\theta}$.
\[lemma:x epsilon\] Under Hypothesis \[Hyp\] the following holds $$\operatorname{sup}_{t \geq 0} {\mathbb{E}}\abs{ X^{\theta}_t - \bar{X}_t }^p \leq C \theta^2 \operatorname{sup}_{t \geq 0} {\mathbb{E}}\abs{ v_t }^p.$$ Where $C$ only depends on the constants appearing in Hypothesis \[Hyp\].
Denote $\Delta X^{\theta}_t := X^{\theta}_t - \bar{X}_t$ and write the corresponding equation $$\Delta X^{\theta}_t = \int_0^t \left[ b(X^{\theta}_s,u^{\theta}_s) - b(\bar{X}_s, \bar{u}_s) \right]ds + \int_0^t \left[ \sigma(X^{\theta}_s,u^{\theta}_s) - \sigma(\bar{X}_s, \bar{u}_s) \right]dW_s.$$ Following the technique developed in the proof of Theorem \[t.existence.SDE\] we define $\Delta \tilde{X}^\theta_t := e^{\beta t}\Delta X^\theta_t$ for a positive $\beta$. Then the Itô formula gives $$\begin{split}
&{\mathbb{E}}\abs{ \Delta \tilde X^{\theta}_t}^{2q}\\
&\leq 2q{\mathbb{E}}\int^t_0 \abs{\Delta \tilde X^{\theta}_s}^{2(q-1)}{\left[}\braket{\tilde b(\tilde X^{\theta}_s, u^\theta_s) - \tilde b(\tilde{\bar{X}}_s,u^\theta_s) ,\Delta \tilde X^{\theta}_s} + (q-\frac{1}{2})(1+\varepsilon)\norm{\tilde \sigma(\tilde X^{\theta}_s,u^{\theta}_s) - \tilde \sigma(\tilde{\bar{X}}_s, u^\theta_s)}_2^2{\right]}ds\\
&\quad+ 2q \beta{\mathbb{E}}\int^t_0 \abs{\Delta \tilde X^{\theta}_s}^{2q}ds + 2q {\mathbb{E}}\int^t_0 \abs{\Delta \tilde X^{\theta}_s}^{2(q-1)}\braket{ \int^1_0 D_u \tilde b( \tilde{\bar{X}}_s, \bar{u}_s + \lambda \theta v_s )\theta v_s d\lambda , \Delta \tilde X^{\theta}_s} ds \\
&\quad+ 2qc_\varepsilon{\mathbb{E}}\int^t_0 \abs{\Delta\tilde X^{\theta}_s}^{2(q-1)} \norm{\int_0^1 D_u\tilde \sigma(\tilde{\bar X}_s,\bar u_s + \lambda \theta v_s)\theta v_s d\lambda}_2^2 ds \\
&\leq 2q(c_r + \beta + \frac{\delta}{2})\int^t_0 \abs{ \Delta \tilde X^{\theta}_s }^{2q}ds + 2q\theta^2{\mathbb{E}}\int_0^t \abs{ \Delta \tilde X^{\theta}_s }^{2(q-1)}\frac{e^{2\beta s}}{2\delta}\abs{v_s}^2ds \quad \qquad(\text{ with } r = 2q(1+\varepsilon) - \varepsilon)\\
&\quad+ 2qc_\varepsilon\theta^2{\mathbb{E}}\int_0^t \abs{ \Delta \tilde X^{\theta}_s }^{2(q-1)}e^{2\beta s}\abs{v_s}^2ds \\
&\leq 2q\left(c_r + \beta + \frac{\delta}{2} + c_\delta \theta^2 \delta^{q/(q-1)}\right)\int^t_0 \abs{ \Delta \tilde X^{\theta}_s }^{2q}ds\\
&\quad+ 2q\theta^2{\mathbb{E}}\int_0^t e^{2q\beta s}\left( \frac{1}{2^q\delta^{q+1}}+ \frac{c_\varepsilon}{\delta^q} \right)\abs{v_s}^{2q} ds.
\end{split}$$ Where we used the joint dissipativity and weighted Young inequality, for every $\delta >0$. Choosing $\beta$, $\delta$ small enough, from the boundedness of $\sup_{s\geq 0}{\mathbb{E}}\abs{v_s}^{2q}$ we get $${\mathbb{E}}\abs{ \Delta X^{\theta}_t}^{2q} \leq C\theta^2 {\mathbb{E}}\int_0^t e^{-2q\beta (t-s)}ds.$$ The result follows by taking the supremum in time and finally by sending $\theta \rightarrow 0_+$.
Now we introduce the first variation equation of the system. Notice that in the equation appears the derivative of the coefficients with respect to the control, which are bounded due to our assumptions. $$\label{eq:first:variation}
\left\lbrace\begin{array}{@{}l@{}}
dY_t = \left[ D_xb(\bar{X}_t,\bar{u}_t)Y_t + D_ub(\bar{X}_t,\bar{u}_t)v_t \right]dt + \left[ D_x\sigma(\bar{X}_t,\bar{u}_t)Y_t + D_u\sigma(\bar{X}_t,\bar{u}_t)v_t \right]dW_t,\\
Y_0 = 0,
\end{array}\right.$$
\[lemma:y\] Under Hypothesis \[Hyp\], the first variation equation admits a unique adapted solution. Moreover the following estimate holds true $$\label{eq:pth moment first variation}
{\mathbb{E}}{\left|}Y_t {\right|}^p \leq K \sup_{s \in [0,t]}{\mathbb{E}}\abs{v_s}^p.$$ where again $K$ only depends on the constants appearing in Hypothesis \[Hyp\].
In particular, $\sup_{t \geq 0}{\mathbb{E}}{\left|}Y_t {\right|}^p \leq K \sup_{t \geq 0}{\mathbb{E}}{\left|}v_t {\right|}^p <+\infty$.
The proof goes through by the same technique adopted in Theorem \[t.existence.SDE\]. What is crucial here is the uniform boundedness of $D_u b(x,u)$ and $D_u\sigma(x,u)$, along with the assumption (H1) on admissible controls.
The following lemma is fundamental in order to obtain the right expansion of the cost functional with respect to the control.
\[l:convergence X tilde\] Under our assumptions the process $\hat{X}^{\theta}$ defined as $$\hat{X}^{\theta}_t = \dfrac{X^{\theta}_t - \bar{X}_t}{\theta} - Y_t,$$ satisfies $$\lim_{\theta \rightarrow 0_+} \sup_{t \geq 0}{\mathbb{E}}\abs{ \hat{X}^{\theta}_t}^2 = 0.$$
The equation for $\hat{X}^{\theta}$ reads $$\begin{split}
d\hat{X}^{\theta}_t &=
\frac{1}{\theta}\left[ b( X^{\theta}_t,u^{\theta}_t) - b(\bar{X}_t,\bar{u}_t) - \theta D_x b(\bar{X}_t,\bar{u}_t) Y_t - \theta D_u b(\bar{X}_t,\bar{u}_t) v_t \right]dt \nonumber \\
& \quad + \frac{1}{\theta}\left[ \sigma( X^{\theta}_t,u^{\theta}_t) - \sigma(\bar{X}_t,\bar{u}_t) - \theta D_x \sigma(\bar{X}_t,\bar{u}_t) Y_t - \theta D_u \sigma(\bar{X}_t,\bar{u}_t) v_t \right]dW_t \nonumber \\
&= \frac{1}{\theta}\left[ b{\left(}\bar{X}_t + \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t+\theta v_t{\right)}- b(\bar{X}_t,\bar{u}_t ) - \theta D_x b(\bar{X}_t,\bar{u}_t) Y_t - \theta D_u b(\bar{X}_t,\bar{u}_t) v_t\right]dt \nonumber \\
&\quad + \frac{1}{\theta}\left[ \sigma {\left(}\bar{X}_t + \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t+\theta v_t {\right)}- \sigma(\bar{X}_t,\bar{u}_t) - \theta D_x \sigma(\bar{X}_t,\bar{u}_t) Y_t - \theta D_u \sigma(\bar{X}_t,\bar{u}_t) v_t \right]dW_t,
\end{split}$$ with $\hat{X}^{\theta}_0 = 0$ as initial condition. Further, by Taylor expansion we have that $$\begin{split}
d\hat{X}^{\theta}_t &=
\int^1_0 D_x b {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}\hat{X}^{\theta}_t d\lambda dt \nonumber\\ & \quad
+ \int^1_0 \left[ D_x b {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}- D_x b(\bar{X}_t,\bar{u}_t) \right] Y_t d\lambda dt \nonumber \\
& \quad + \int^1_0 \left[ D_u b {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}- D_u b(\bar{X}_t,\bar{u}_t) \right] v_t d\lambda dt \nonumber \\
& \quad + \int^1_0 D_x \sigma {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}\hat{X}^{\theta}_t d\lambda dW_t \nonumber\\ & \quad + \int^1_0 \left[ D_x \sigma {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}- D_x \sigma(\bar{X}_t,\bar{u}_t) \right] Y_t d\lambda dW_t \nonumber \\
& \quad + \int^1_0 \left[ D_u \sigma {\left(}\bar{X}_t + \lambda \theta (Y_t + \hat{X}^{\theta}_t),\bar{u}_t + \lambda \theta v_t {\right)}- D_u \sigma (\bar{X}_t,\bar{u}_t) \right] v_t d\lambda dW_t.
\end{split}$$ To keep the notation simple, we rewrite the above equation as $$d\hat{X}^{\theta}_t =
{\left(}A^x_t \hat{X}^{\theta}_t + A^y_t Y_t + A^v_t v_t {\right)}dt + {\left(}B^x_t \hat{X}^{\theta}_t + B^y_t Y_t + B^v_t v_t {\right)}dW_t,$$ where we have kept the order of the terms from the previous equation.\
Now apply the Itô formula to $e^{\beta t}\big| \hat{X}^{\theta}_t \big|^2$ to get $$\label{eq:x tilde Ito}
\begin{split}
{\mathbb{E}}\big( e^{\beta t}\big| \hat{X}^{\theta}_t \big|^2 \big) &=
2{\mathbb{E}}\int^t_0 e^{\beta s}
\big< A^x_s \hat{X}^{\theta}_s + A^y_s Y_s + A^v_s v_s, \hat{X}^{\theta}_s\big> ds \\
& + {\mathbb{E}}\int^t_0 e^{\beta s} \norm{ B^x_s \hat{X}^{\theta}_s + B^y_s Y_s + B^v_s v_s}_2^2 ds + \beta {\mathbb{E}}\int^t_0 e^{\beta s}\big| \hat{X}^{\theta}_s \big|^2 ds.
\end{split}$$ By the joint dissipativity assumption (H4) in Hypothesis \[Hyp\] we have $$2\big< A^x_s \hat{X}^{\theta}_s, \hat{X}^{\theta}_s \big> +2 k \norm{ B^x_s \hat{X}^{\theta}_s }^2 +\beta | \hat{X}^{\theta}_s|^2 <0,$$ for some $k>1/2$ and $\beta$ small enough.
Thus, repeating the same computations as in the proof of Theorem \[t.existence.SDE\], we get the following intermediate estimate $$\label{eq:x tilde prefinal}
{\mathbb{E}}\big| \hat{X}^{\theta}_t \big|^2 \leq
C \int^t_0 e^{-\beta(t-s)} {\mathbb{E}}{\left(}\abs{A^y_s Y_s}^2 + \abs{A^v_s v_s}^2 + \abs{B^y_s Y_s}^2 + \abs{B^v_s v_s}^2 {\right)}ds.$$ Now we show how to treat the first term in . The estimate of the remaining ones goes along similar lines.
We fix $\alpha$ with $p/(p-2)< \alpha <p/(4m)$, if $m\geq 1$, or $\alpha=2$, if $m=0$. Recall that $p> 4m+2$ and notice that, this way, denoting by $\alpha'$ the conjugate of $\alpha$ (that is $1/\alpha+1/\alpha'=1$) then $2 \alpha '<p$ $4m\alpha <p$ and $2\alpha <p$. First we start by observing that by Hölder inequality and by we have that for any $\alpha>1$. $$\label{eq:x tilde loc lipsch1}
\begin{split}
&\int^t_0 e^{-\beta(t-s)} {\mathbb{E}}\abs{A^y_s Y_s}^2 ds \\
&= \int^t_0 e^{-\beta(t-s)} {\mathbb{E}}{\left|}\int^1_0 \left[ D_x b {\left(}\bar{X}_s + \lambda \theta (Y_s + \hat{X}^{\theta}_s),\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s) \right] Y_s d\lambda {\right|}^2 ds \\
&\leq \int^t_0\!\! e^{-\beta(t-s)}{\left(}\int^1_0
{\mathbb{E}}{\left|}D_x b {\left(}\bar{X}_s + \lambda \theta (Y_s + \hat{X}^{\theta}_s),\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s){\right|}^{2\alpha}d\lambda {\right)}^{\frac{1}{\alpha}} \!\!\! \cdot {\left(}{\mathbb{E}}|Y_s|^{2\alpha'} {\right)}^{\frac{1}{\alpha'}} ds
\end{split}$$ Since $2\alpha'<p$, using Lemma \[lemma:y\] to estimate $\sup_{s\in \mathbb{R}^+} {\mathbb{E}}|Y_s|^{2\alpha'} $: $$\begin{split}
&\int^t_0 e^{-\beta(t-s)} {\mathbb{E}}\abs{A^y_s Y_s}^2 ds \\
&\leq C \int^t_0 e^{-\beta(t-s)} {\left(}\int^1_0
{\mathbb{E}}{\left|}D_x b {\left(}\bar{X}_s + \lambda \theta (Y_s + \hat{X}^{\theta}_s),\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s){\right|}^{2\alpha} d\lambda{\right)}^{\frac{1}{\alpha}} ds.\\
&\leq C \int^t_0 e^{-\beta(t-s)} {\left(}\int^1_0
{\mathbb{E}}{\left|}D_x b {\left(}\bar{X}_s + \lambda \theta (Y_s + \hat{X}^{\theta}_s),\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,,\bar{u}_s + \lambda \theta v_s){\right|}^{2\alpha} \!\! d\lambda {\right)}^{\frac{1}{\alpha}} \!\!\!\! ds\\ & \quad + C \int^t_0 e^{-\beta(t-s)} {\left(}\int^1_0
{\mathbb{E}}{\left|}D_x b {\left(}\bar{X}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s){\right|}^{2\alpha} \!\! d\lambda {\right)}^{\frac{1}{\alpha}} \!\!\!\! ds.
\end{split}$$ We prove convergence of the first term, being the second similar (and easier).
Due to Hypothesis \[Hyp\], the gradients $D_xb$ are locally Lipschitz functions with respect to $x$, so that for all $R > 0$ there exists $C_R$ such that $D_xb$ is Lipschitz with constant $C_R$ in the ball of radius $R$. For each $t$ and $\theta$ we define the sets $$A_{t,\theta}(R) = \lbrace w \in \Omega: \abs{\bar X_t} > R \rbrace \cup \lbrace w \in \Omega: \abs{X^\theta_t} > R \rbrace.$$ By Chebyshev inequality we know that $$\label{stimadiA}
{\mathbb{P}}(A_{t,\theta}(R)) \leq \frac{{\mathbb{E}}\abs{\bar X_t}^2}{R^2} + \frac{{\mathbb{E}}\abs{X^\theta_t}^2}{R^2} \leq \frac{C}{R^2}, \qquad \forall\, t, \forall \, \theta.$$ Denoting for simplicity $X^{\lambda}_s=\bar{X}_s+\lambda \theta (Y_s+\hat{X}_s))=(1-\lambda)\bar{X}_s+\lambda X^{\theta}_s$ we have $$\label{eq:x tilde loc lipsch2}
\begin{split}
\int^t_0 & e^{-\beta(t-s)} {\left(}\int^1_0
{\mathbb{E}}{\left|}D_x b {\left(}X^{\lambda}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s + \lambda \theta v_s){\right|}^{2\alpha} \!\! d\lambda {\right)}^{\frac{1}{\alpha}} \!\!ds \\
&\leq C \int^t_0 e^{-\beta(t-s)} {\left(}\int^1_0
\int_{A_{s,\theta}(R)} {\left|}D_x b {\left(}X^{\lambda}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s + \lambda \theta v_s){\right|}^{2\alpha}\!\! d{\mathbb{P}}d\lambda \right)^{\frac{1}{\alpha}}\!\!\! ds\\
&\quad + C \int^t_0 e^{-\beta(t-s)} {\left(}\int_0^1 \int_{A_{s,\theta}^c(R)} {\left|}D_x b {\left(}X^{\lambda}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s + \lambda \theta v_s) {\right|}^{2\alpha}\!\! d{\mathbb{P}}d\lambda
{\right)}^{\frac{1}{\alpha}} \!\! ds \\
&\leq C\int^t_0 e^{-\beta(t-s)} \left( \int_0^1 {\mathbb{P}}(A_{s,\theta}(R))^{\frac{\delta}{1+\delta}} {\mathbb{E}}{\left|}D_x b {\left(}X^{\lambda}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s + \lambda \theta v_s){\right|}^{2\alpha(1+\delta)}\!\!\! d\lambda \right)^{\frac{1}{\alpha(1+\delta)}} \!\!\! \!\!\! ds \\
&\quad + C\int^t_0 e^{-\beta(t-s)} C_R^{\frac{1}{\alpha}} \left( {\mathbb{E}}\abs{X^\theta_s - \bar X_s}^{2\alpha} \right)^{\frac{1}{\alpha}} ds,
\end{split}$$ where $\delta>0$ is such that $4m\alpha(1+\delta)\leq p$.\
Fixed $\varepsilon >0$ we know by that there exists $R$ large enough so that ${\mathbb{P}}(A_{s,\theta}(R)) \leq \varepsilon$. Moreover by Hypothesis \[Hyp\], Theorem \[t.existence.SDE\] and Lemma \[lemma:x epsilon\]: $${\mathbb{E}}{\left|}D_x b {\left(}X^{\lambda}_s,\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s + \lambda \theta v_s){\right|}^{2\alpha(1+\delta)}\leq
C \left( \mathbb{E} | X^{\lambda}_s |^{4m\alpha(1+\delta)}+ \mathbb{E} | \bar{X}_s |^{4m\alpha(1+\delta)}\right)\leq C,$$ (if $m=0$ the above relation is straight forward). Thus the first of the two integrals in the last two lines in (\[eq:x tilde loc lipsch2\]) can be estimated, for $R$ large enough and all $\theta$, $\lambda$ in $[0,1]$, by $C \varepsilon ^{\delta/[\alpha(1+\delta)^2]}$.
Moreover, due to Lemma \[lemma:x epsilon\] we have that $
\sup_{t \geq 0} {\mathbb{E}}\abs{ X^{\theta}_t - \bar{X}_t }^p \rightarrow 0$ as $\theta\rightarrow 0$.
Combining the two estimates above we have: $$\sup_{t \geq 0} \int_0^t e^{-\beta(t-s)}{\mathbb{E}}\abs{A^y_sY_s}^2 ds \rightarrow 0 \hbox{ as }\theta\rightarrow 0.$$ Repeating the argument for all the terms in we get the required result.
Notice that we estimate only the second moment of the error term, uniformly in time. Nevertheless, estimate of higer moments of the the state and first variation process are needed in order to complete the proof. More precisely, we can tune the value of $\alpha$ in in order to minimize the maximal moment of the state equation we need to control. Indeed, the growth of the first term is $${\mathbb{E}}{\left|}D_x b {\left(}\bar{X}_s + \lambda \theta (Y_s + \hat{X}^{\theta}_s),\bar{u}_s + \lambda \theta v_s {\right)}- D_x b(\bar{X}_s,\bar{u}_s){\right|}^{2\alpha}
\leq C {\mathbb{E}}\abs{\bar{X}_s}^{4m\alpha}.$$ So that, $4m\alpha = 2\alpha' = 2 \frac{\alpha}{\alpha - 1}$, from which $\alpha = \frac{2m+1}{2m}$. The maximal moment is then $p = 4m\alpha = 2(2m+1)$, which is the one appearing in Hypothesis \[Hyp\].
Perturbation of the cost
------------------------
Due to the hypotheses on the admissible controls and the estimate the cost is well posed: $$\liminf_{T \to \infty} \frac{1}{T}{\mathbb{E}}\int_0^T f(X_t,u_t)dt \leq K \left[1+ \sup_{t \geq 0}{\mathbb{E}}\abs{X_t}^2 + \sup_{t \geq 0}{\mathbb{E}}\abs{u_t}^2\right] < \infty.$$ The same is true for the $\limsup$ formulation. The expansion of the functional with respect to a convex perturbation of the control is given in the following
\[l:derivative\_J\] Let $\bar{u}$ be an optimal control and let $u$ be any admissible control. Letting $v=u-\bar{u}$ and using the above notation the following holds: $$\label{eq:Gateaux J inf}
\lim_{\theta \rightarrow 0_+}\frac{J^{\inf}(\bar{u}(\cdot) + \theta v(\cdot)) - J^{\inf}(\bar{u}(\cdot))}{\theta} \leq \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \left[\braket{ D_xf(\bar{X}_t,\bar{u}_t), Y_t}_{\mathbb{R}^n}+ \braket{ D_uf(\bar{X}_t,\bar{u}_t), v_t }_{\mathbb{R}^l} \right]dt,$$ and $$\label{eq:Gateaux J sup}
\lim_{\theta \rightarrow 0_+}\frac{J^{\sup}(\bar{u}(\cdot) + \theta v(\cdot)) - J^{\sup}(\bar{u}(\cdot))}{\theta} \leq \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \left[\braket{ D_xf(\bar{X}_t,\bar{u}_t), Y_t}_{\mathbb{R}^n}+ \braket{ D_uf(\bar{X}_t,\bar{u}_t), v_t }_{\mathbb{R}^l} \right]dt.$$
We prove the first relation. The proof of the second one goes along the same lines. Let us compute
$$\begin{split}
&\frac{J_T(\bar{u}(\cdot) + \theta v(\cdot)) - J_T(\bar{u}(\cdot))}{\theta} = \dfrac{1}{\theta} {\mathbb{E}}\int_0^T \bigr[ f(X^{\theta}_t,\bar{u}_t + \theta v_t) - f(\bar{X}_t,\bar{u}_t) \bigl]dt \\
&= {\mathbb{E}}\int_0^T \int_0^1 D_xf\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big) \big(\hat{X}^{\theta}_t + Y_t\big) d\lambda dt \\
&\quad+ {\mathbb{E}}\int_0^T \int_0^1 D_u f\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big) v_t d\lambda dt. \\
\end{split}$$
Passing to the ergodic $\liminf$ cost functional we have that $$\begin{split}
& \frac{J^{\inf}(\bar{u}(\cdot) + \theta v(\cdot)) - J^{\inf}(\bar{u}(\cdot))}{\theta} = \frac{1}{\theta}\left[\liminf_{T \to \infty} \frac{1}{T}J_T(\bar{u}(\cdot) + \theta v(\cdot)) - \liminf_{T \to \infty}\frac{1}{T} J_T(\bar{u}(\cdot))\right]\\
&\quad \leq \limsup_{T \to \infty} \frac{1}{T} \left[ \frac{J_T(\bar{u}(\cdot) + \theta v(\cdot)) - J_T(\bar{u}(\cdot))}{\theta}\right]\\
&\quad= \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \int_0^1 \big< D_x f\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big), \hat{X}^{\theta}_t + Y_t \big>d\lambda dt \\
&\quad \quad+ \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \int_0^1 \big< D_u f\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big), v_t \big>_U d\lambda dt,
\end{split}$$ where we used that $\limsup (a_n) - \limsup (b_n) \leq \limsup (a_n - b_n)$, for $(a_n)_{n \geq 1}$ and $(b_n)_{n \geq 1}$ two general real sequences. The extra term can be estimated by $$\begin{split}
&\limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \int_0^1 \big< D_x f\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big), \hat{X}^{\theta}_t \big> d\lambda dt \\
&\qquad \leq \limsup_{T \to \infty} \frac{1}{T}\int_0^T \int_0^1({\mathbb{E}}\abs{D_x f\big(\bar{X}_t + \lambda( X^{\theta}_t - \bar{X}_t),\bar{u}_t+ \lambda\theta v_t\big)}^2)^{1/2} ({\mathbb{E}}\abs{\hat{X}^{\theta}_t}^2)^{1/2} d\lambda dt, \\
\end{split}$$ which converges to zero, uniformly in $T$, as $\theta \to 0_+$. In fact, this follows from the linear growth of $D_x f(\cdot)$, the a priori estimates on $X_t$ and Lemma \[l:convergence X tilde\]. The conclusion now easily follows.
The adjoint equation
====================
In this section we introduce the dual equation associated to the system, which is an infinite horizon Backward SDE in ${\mathbb{R}}^n$. Different approaches have been developed in the literature to study this class of equations. Here we present a duality method built on construction of a family of truncated equations and associated with a consistency argument. More precisely, the infinite horizon backward equation has the form $$\label{eq:BSDE}
-dp_t = \left[ D_xb(X_t,u_t)^*p_t + \sum_{i=1}^dD_x \sigma^i(X_t,u_t)^*q^i_t - D_xf(X_t,u_t)\right]dt - \sum_{i=1}^dq^i_tdW^i_t,$$ where, fixed any orthonormal basis $(e_i)_{i=1,..d}$ in $\mathbb{R}^d$ we set $W^i_s=<e_i,W_s>$ and $\sigma^i(x,u)= \sigma(x,u)e_i$ moreover we denote by $(\cdot)^*$ the transposition operation in $\mathcal{L}(\mathbb{R}^n)$. For every $T >0$ fixed, its solution has to be understood as $$\label{eq:BSDE_int}
\begin{split}
p_t &= p_T + \int_t^T \left[ D_xb(X_s,u_s)^*p_s +\sum_{i=1}^d D_x\sigma^i(X_s,u_s)^*q^i_s + D_xf(X_s,u_s)\right]ds - \sum_{i=1}^d \int_t^T q^i_sdW^i_s.
\end{split}$$ where $p$ and $q^i$, $i=1,...,d$ take values in $\mathbb{R}^n$. Due to Hypothesis \[Hyp\] and estimate the forcing term in the driver is no better than bounded, so that $D_xf(X_s,u_s) \in L^{\infty}{\left(}{\mathbb{R}}_+; L^2(\Omega;{\mathbb{R}}^n) {\right)}$. Therefore we cannot expect the solution of to be integrable up to infinity but only that $p \in L^{\infty}{\left(}{\mathbb{R}}_+; L^2(\Omega;{\mathbb{R}}^n) {\right)}$. Up to the authors„1¤7 knowledge, there is not a general wellposendess result for such multidimensional BSDE’s. Partial results have been obtained in [@cohen2015classical] by a Girsanov argument that seems to work only if one knows a-priori that $\sum_{i=1}^d D_x\sigma^i(X_s,u_s)^*q^i_s$ can be written as $\sum_{i=1}^d q^i_s f^i$ for suitable adapted real process $(f_i)_{i=1,...,d}$. In particular this is the case when $n = 1$ or the noise is additive.
Here the solution will be obtained via the introduction of a family of time truncations: $$\label{eq:BSDE_trunc}
\left\lbrace\begin{array}{@{}l@{}}
-dp^{T,\nu}_t = \left[ D_xb(X_t,u_t)^*p^{T,\nu}_t + \sum_{i=1}^dD_x\sigma^i(X_t,u_t)^*q^{i,T,\nu}_t - D_xf(X_t,u_t)\right]dt -
\sum_{i=1}^dq^{i,T,\nu}_t dW^i_t,\\
p_T^{T,\nu} = \nu.
\end{array}\right.$$ which will be estimated by duality. For the approximating equation a wellposedness result has been already adressed in [@orrieri2015necessary].\
To shorten the notation in the following paragraphs, let us denote $$\Lambda_t := D_xb(X_t,u_t), \quad \Gamma^i_t := D_x\sigma^i(X_t,u_t), \quad \Psi_t := D_xf(X_t,u_t);$$ moreover when $\nu=0$ the solution of equation will be denoted by $(p^T,q^{i,T})$.
\[t:truncatedBSDEs\] For all $T\geq 0$ and all $\nu\in L^2(\Omega,\mathcal{F}_T,\mathbb{P};\mathbb{R}^n)$ there exists a unique $(d+1)$-tuple of $\mathbb{R}^n$-valued, adapted processes $(p^{T,\nu },q^{1,T,\nu},...,q^{d,T,\nu})$ such that $p^{T,\nu }$ has continuous trajectories, $\sup_{t\in [0,T]}\mathbb{E}|p^{T,\nu }_t|^2 +
\sum_{i=1}^d \mathbb{E}\int_0^T |q^{i,T,\nu}_t|^2 dt <\infty$ and, $\mathbb{P}$-almost surely, for all $t\in [0,T]$ it holds: $$p^{T,\nu}_t=\nu+\int_t^T \Lambda^*_s p^{T,\nu}_sds+\sum_{i=1}^d\int_t^T (\Gamma^{i}_s)^* q^{i,T,\nu}ds+\int_t^T\Psi_sds+\sum_{i=1}^d \int_t^T q^{i,T,\nu}_s dW^i_s$$.
Consider now the following affine forward SDE with general forcing term $(\gamma,\rho^i)_{i=1,..,d}$ with $\gamma$ and $\rho^i$, $i=1,..,d$ in $L^2([0,T];L^2(\Omega;{\mathbb{R}}^{n}))$ and initial condition $\eta \in L^2(\Omega,\mathcal{F}_t;{\mathbb{R}}^n)$: $$\label{eq:duality:general}
\left\lbrace\begin{array}{@{}l@{}}
d{\mathcal{Y}}_s^{t,\eta,\gamma,\rho} = \Lambda_s {\mathcal{Y}}_s^{t,\eta,\gamma,\rho}dt + \sum_{i=1}^d \Gamma^i_t{\mathcal{Y}}_s^{t,\eta,\gamma,\rho}dW^i_t +\gamma_sds +\sum_{i=1}^d \rho^i_s dW^i_s,\ s \geq t,\\
{\mathcal{Y}}_t^{t,\eta,\gamma,\rho} = \eta.
\end{array}\right.$$ Then by the same technique we adopted in the proof of Theorem \[t.existence.SDE\], the above equation admits a unique adapted solution and $$\label{stimadiY}
{\mathbb{E}}\abs{{\mathcal{Y}}_r^{t,\eta,\gamma,\rho}}^2 \leq e^{-2\beta(r-t)}{\mathbb{E}}\abs{\eta}^2 + K \int_t^r e^{-2\beta(r-s)}{\mathbb{E}}\left[|\gamma_s|^2+\abs{\rho^1_s}^2+....+\abs{\rho^d_s}^2\right] ds.$$ When $\gamma\equiv 0$ then the solution to the above equation will be denoted by ${\mathcal{Y}}^{t,\eta,\rho}$ and when $\rho\equiv 0$ as well, it will be denoted by ${\mathcal{Y}}^{t,\eta,\gamma}$.
The next result is proven in [@orrieri2015necessary] by computing the Itô formula the differential of the product $d\braket{Y^{t,\eta,\gamma,\rho}_s,p^{T,\nu}_s}$
\[l:dualityFO\] Given $(\rho^i)_{i=1,..,d}$ with $\gamma, \rho^i \in L^2([0,T];L^2(\Omega;{\mathbb{R}}^{n}))$, $\eta \in L^2(\Omega,\mathcal{F}_t;{\mathbb{R}}^n)$, $\nu \in L^2(\Omega,\mathcal{F}_T;{\mathbb{R}}^n)$ it holds: $$\label{general_duality}
{\mathbb{E}}\int_t^T \braket{p^{T,\nu}_s, \gamma_s}ds+
\sum_{i=1}^d {\mathbb{E}}\int_t^T \braket{q^{i,T,\nu}_s, \rho^i_s}ds + {\mathbb{E}}\braket{p^{T,\nu}_t,\eta} = {\mathbb{E}}\int_t^T \braket{{\mathcal{Y}}_s^{t,\eta,\gamma,\rho}, \Psi_s}ds + {\mathbb{E}}\braket{\nu, {\mathcal{Y}}_T^{t,\eta,\gamma,\rho}}.$$
In the following, relation will be the main instrument to get information on the behaviour of the BSDE. We will specifically choose the values of $t,\eta,\rho$ according to our needs.\
We are now in a position to define the solution to the infinite horizon multidimensional BSDE and prove its existence and uniqueness
A solution to equation is a $(d+1)$-tuple of $\mathbb{R}^n$-valued, adapted processes $(p_t,q^1_t,...,q^d_t)_{t\in [0,\infty[}$ such that, for all $T>0$ and all $i=1,...,d$ it holds $ \mathbb{E}\int_0^T |q^{i}_t|^2 dt <\infty$. Moreover $p$ has continuous trajectories and $\sup_{t\in [0,\infty)}\mathbb{E}|p_t|^2<\infty$. Finally, for all $0\leq t \leq T$, holds $\mathbb{P}$-almost surely.
The main result of this section is the following
\[t.existence\_BSDE\] Let Hypothesis \[Hyp\] holds true. Then equation admits a unique solution $(p^{\infty},$ $q^{1,\infty},...,q^{d,\infty})$.
$ $
*Existence:* Let in $\nu \equiv 0,\ \gamma \equiv 0,\ \rho \equiv 0$, $\eta \in L^2(\Omega,\mathcal{F}_t;{\mathbb{R}}^n)$ then $${\mathbb{E}}\braket{p^T_t,\eta} = {\mathbb{E}}\int_t^T \braket{{\mathcal{Y}}^{t,\eta}_s ,\Psi_s } ds.$$ Since $\Psi \in L^{\infty}{\left(}{\mathbb{R}}_+; L^2(\Omega;{\mathbb{R}}^n) {\right)}$ by we deduce that $${\mathbb{E}}\int_t^T \braket{{\mathcal{Y}}^{t,\eta}_s, \Psi_s } ds \rightarrow {\mathbb{E}}\int_t^{\infty} \braket{{\mathcal{Y}}^{t,\eta}_s, \Psi_s} ds$$ and that the right hand side is a bounded linear operator from $L^2(\Omega,\mathcal{F}_t;{\mathbb{R}}^n) \to {\mathbb{R}}$. Hence, by Riesz representation theorem there exists an element $P(t)\in L^2(\Omega,\mathcal{F}_t;{\mathbb{R}}^n)$ such that $$\label{eq:Riesz:p_infty}
{\mathbb{E}}\braket{P(t),\eta} = {\mathbb{E}}\int_t^\infty \braket{{\mathcal{Y}}^{t,\eta}_s, \Psi_s } ds.$$ Moreover $p^T(t)\rightharpoonup P(t)$ in $L^2(\Omega,\mathcal{F}_t,\mathbb{P};\mathbb{R}^n)$ and $
{\mathbb{E}}\abs{P(t)}^2 \leq \beta^{-1} \sup_{s\in [0,\infty[} (\mathbb{E}|\Psi_s|^2)^{1/2}$ for all $t>0$.
Let now for all $N\in \mathbb{N}$, $(\tilde{p}^N_t,\tilde{q}^{1,N}_t,...,\tilde{q}^{d,N}_t)_{t\in [0,N]}$ be the solution of equation with $T=N$ and $\nu=P(N)$.
We claim that, for all $N,M\in \mathbb{N}$ with $0\leq N\leq M$ and all $t\leq N$ it holds $$\label{eq:consistency}\tilde{p}^N(t)=\tilde{p}^M(t), \hbox{ $\mathbb{P}$-a.s.}$$ By definition and Lemma \[l:dualityFO\] we deduce that for all $\eta\in L^2(\Omega,\mathcal{F}_t,\mathbb{P};\mathbb{R}^n)$ $${\mathbb{E}}\braket{\tilde{p}^{N}_t, \eta} = {\mathbb{E}}\int_t^N \braket{{\mathcal{Y}}_s^{t,\eta}, \Psi_s}ds + {\mathbb{E}}\braket{P(N), {\mathcal{Y}}_N^{t,\eta}}.$$ Plugging with $t=N$ in the above relation we have $${\mathbb{E}}\braket{\tilde{p}^{N}_t, \eta} = {\mathbb{E}}\int_t^N \braket{{\mathcal{Y}}_s^{t,\eta}, \Psi_s}ds + {\mathbb{E}}\int_N^{\infty} \braket{{\mathcal{Y}}_s^{N,{\mathcal{Y}}_N^{t,\eta}}, \Psi_s}ds.$$ and finally, observing that by uniqueness of the solution to equation ${\mathcal{Y}}_s^{N,{\mathcal{Y}}_N^{t,\eta}}={\mathcal{Y}}_s^{t,\eta}$ $\mathbb{P}$-a.s. we conclude $${\mathbb{E}}\braket{\tilde{p}^{N}_t, \eta} = {\mathbb{E}}\int_t^{\infty} \braket{{\mathcal{Y}}_s^{t,\eta}, \Psi_s}ds=\mathbb{E}\braket{\eta, P(t)}.$$ and our claim is proved since the right hand side does not depend on $N$. We also remark that by the above identity we deduce that $$\sup_{t\in [0,N]} |\tilde{p}^N_t|^2 \leq \beta^{-1} \sup_{s\in [0,\infty[} (\mathbb{E}|\psi_s|^2)^{1/2},$$ and that the right hand side does not depend neither on $t$ nor on $N$.
Now we define $$p^{\infty}_t=\sum_{N=1}^{\infty} \tilde{p}^N_t I_{[N-1,N[}(t),\quad
q^{i,\infty}_t=\sum_{N=1}^{\infty} \tilde{q}^{i,N}_t I_{[N-1,N[}(t),$$ and claim that it is the desired solution. Indeed it satisfies the desired integrability and adaptedness conditions. Moreover fixed $0\leq t\leq T$ then $$\begin{split}p^{\infty}_t-p^{\infty}_T &=
[ p^{\infty}_t-p^{\infty}_{\lfloor t\rfloor+1}]+ [ p^{\infty}_{\lfloor T\rfloor }-p^{\infty}_T]+
\sum_{n=\lfloor t\rfloor+1}^{\lfloor T\rfloor -1}[ p^{\infty}_n-p^{\infty}_{n+1}] \\
&=\big[ \tilde{p}^{\lfloor t\rfloor+1}_t-\tilde{p}^{\lfloor t\rfloor+2}_{\lfloor t\rfloor+1}\big]+ \big[ \tilde{p}^{\lfloor T\rfloor+1}_{\lfloor T\rfloor }-\tilde{p}^{\lfloor T\rfloor+1}_T\big]+
\sum_{n=\lfloor t\rfloor+1}^{\lfloor T\rfloor -1}[ \tilde{p}^{n+1}_n-\tilde{p}^{n+2}_{n+1}]
\\
&=\big[ \tilde{p}^{\lfloor t\rfloor+1}_t-\tilde{p}^{\lfloor t\rfloor+1}_{\lfloor t\rfloor+1}\big]+ \big[ \tilde{p}^{\lfloor T\rfloor+1}_{\lfloor T\rfloor }-\tilde{p}^{\lfloor T\rfloor+1}_T\big]+
\sum_{n=\lfloor t\rfloor+1}^{\lfloor T\rfloor -1}[ \tilde{p}^{n+1}_n-\tilde{p}^{n+1}_{n+1}],
\end{split}$$ where in the last equality we have exploited where it was needed. Now recalling that $(\tilde{p}^N_t,\tilde{q}^{1,N}_t,...,\tilde{q}^{d,N}_t)_{t\in [0,N]}$ solves equation and the definition of $({p}^{\infty},{q}^{1,\infty},...,\tilde{q}^{d,\infty})$ the above equality can be rewritten as $$\begin{split}p^{\infty}_t-p^{\infty}_T &=\int_t^{\lfloor t\rfloor+1}\!\!\!\! \Lambda^*_s p^{\infty}_sds+\sum_{i=1}^d\int_t^{\lfloor t\rfloor+1}\!\!\!\! (\Gamma^{i}_s)^* q^{i,\infty}_sds+\int_t^{\lfloor t\rfloor+1}\!\!\!\!\Psi_sds+\sum_{i=1}^d\int_t^{\lfloor t\rfloor+1}\!\!\!\! q^{i,\infty}_s dW^i_s
\\
& \quad + \sum_{n=\lfloor t\rfloor+1}^{\lfloor T\rfloor-1}
\left[
\int_n^{n+1} \!\!\!\!\Lambda^*_s p^{\infty}_sds+\sum_{i=1}^d\int_n^{n+1}\!\!\!\! (\Gamma^{i}_s)^* q^{i,\infty}_sds+\int_n^{n+1}\!\!\!\!\Psi_sds+\sum_{i=1}^d\int_n^{n+1} \!\!\!\! q^{i,\infty}_s dW^i_s \right]\\
&\quad +\int_{\lfloor T\rfloor}^T\!\! \Lambda^*_s p^{\infty}_sds+\sum_{i=1}^d\int_{\lfloor T\rfloor}^T\!\! (\Gamma^{i}_s)^* q^{i,\infty}_sds+\int_{\lfloor T\rfloor}^T\!\!\Psi_sds+\sum_{i=1}^d\int_{\lfloor T\rfloor}^T\!\! q^{i,\infty}_s dW^i_s\\
&= \int_t^T\!\!\Lambda^*_s p^{\infty}_sds+\sum_{i=1}^d\int_t^T\!\! (\Gamma^{i})^*_s q^{i,\infty}_sds+\int_t^T\!\!\Psi_sds+\sum_{i=1}^d\int_t^T\!\! q^{i,\infty}_s dW^i_s
\end{split}$$ and this completes the proof of existence of a solution to equation .
$ $
*Uniqueness:* Let $(p_t,q^1_t,...,q^d_t)_{t\geq 0}$ be a solution to equation . We choose $\rho\in L^2(\Omega\times [0,\infty[;\mathbb{R}^n)$ with support in the finite interval $[0,T]$ ($\rho_r = 0$, if $r \geq T$) and $\eta \in L^2(\Omega, \mathcal{F}_t,\mathbb{P}; \mathbb{R}^n)$. Noticing that $(p_t,q^1_t,...,q^d_t)_{t\geq 0}$ is, in particular a solution to equation in $[0,T]$ with $\nu=p_T$ by Lemma \[l:dualityFO\] we get: $${\mathbb{E}}\int_t^T \braket{{\mathcal{Y}}_s^{t,\eta,\rho}, \Psi_s}ds + {\mathbb{E}}\braket{p_T, {\mathcal{Y}}_T^{t,\eta,\rho}} = \sum_{i=1}^d{\mathbb{E}}\int_t^T \braket{ q^i_s, \rho_s}ds + {\mathbb{E}}\braket{\eta, \tilde p_t} .$$ We notice that since $\rho_t=0$ for $t>T$ then by we have that ${\mathbb{E}}\abs{{\mathcal{Y}}_s^{t,\eta,\rho}}^2\leq Ce^{-2\beta(s-t)}$ for a suitable $C$. So letting $T\rightarrow \infty$ in the above equality we get (recall that $\sup_{t\geq 0}{\mathbb{E}}|p_t|^2< \infty $ by definition of solution): $$\label{eq:dualityIO}
{\mathbb{E}}\int_t^{\infty} \braket{{\mathcal{Y}}_s^{t,\eta,\rho}, \Psi_s}ds= \sum_{i=1}^d{\mathbb{E}}\int_t^T \braket{ q^i_s, \rho_s}ds + {\mathbb{E}}\braket{\eta, \tilde p_t}$$ and this completes the proof of uniqueness due to the arbitrariness of $t,T,\rho$ and $\eta$.
As a by-product of the above proof we have the following infinite-horizon version of the duality relation:
\[cor:dualityIO\] Let $(p_t,q^1_t,...,q^d_t)_{t\geq 0}$ be a solution to equation . Fix $\rho\in L^2(\Omega\times [0,\infty[;\mathbb{R}^n)$ with support in $[0,T]$, $ t\in [0,T)$ and $\eta\in L^2(\Omega, \mathcal{F}_t,\mathbb{P};\mathbb{R}^n)$ then holds.
Necessary Ergodic SMP
=====================
We give two versions of the SMP in its necessary form. The first is based on the well-posedness result for the infinite horizon BSPDE. The second one is written in terms of the family of truncated backward equations introduced in the previous section. The Hamiltonian associated to the system is $$H(x,u,p,q^1,...q^d) = \braket{b(x,u),p} + \sum_{i=1}^d\braket{\sigma^i(x,u),q^i} + f(x,u).$$ We are now in a condition to formulate a necessary condition corresponding to the ergodic control problem.
\[thm:necessary inf horizon\] Suppose that $(\bar{X},\bar{u})$ is an optimal pair for the control problem $J^{\inf} $ or $J^{\sup}$ and let $(p^\infty, q^\infty) = (p^\infty,q^{\infty,1},...,q^{\infty,d})$ be the solution of equation . Then under Hypothesis 1, the following variational inequality holds: $$\begin{split}
0&\leq \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \braket{ D_u H{\left(}\bar{X}_t, \bar{u}_t, p^\infty_t, q^\infty_t {\right)}, u_t-\bar{u}_t}_{{\mathbb{R}}^l} dt,
\end{split}$$ where $H(x,u,p,q)$ is the Hamiltonian of the system, and $u(\cdot)$ is an arbitrary admissible control.
Let $v(\cdot)= u(\cdot)-\bar{u}(\cdot)$ and let $Y_t$ be the solution to equation . Lemma \[l:dualityFO\] with $t=0$, $\eta=0$, $\nu=p^{\infty}_T$, $\gamma= D_u b(\bar{X},\bar{u})$, $\rho^{i}=D_u \sigma(\bar{X},\bar{u})v$ yields $$\begin{split}
& {\mathbb{E}}\int_0^T \braket{D_x f(\bar{X}_t,\bar{u}_t),Y_t}dt \\
\quad &= {\mathbb{E}}\braket{p^\infty_T,Y_T} + {\mathbb{E}}\int_0^T \braket{p^\infty_t,D_u b(\bar{X}_t,\bar{u}_t)v_t} dt + {\mathbb{E}}\int_0^T \braket{q^\infty_t,D_u \sigma(\bar{X}_t,\bar{u}_t)v_t} dt.
\end{split}$$ So that, from Lemma \[l:derivative\_J\] and the relation above, we have $$\begin{split}
0& \leq \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \left[\braket{ D_xf(\bar{X}_t,\bar{u}_t), Y_t}_{\mathbb{R}^n}+ \braket{ D_uf(\bar{X}_t,\bar{u}_t), v_t }_{\mathbb{R}^l} \right]dt.\\
& \leq -\limsup_{T \to \infty} \frac{1}{T}{\mathbb{E}}\braket{Y_T,p^\infty_T} + \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \Big[ \braket{ D_uH(\bar X_t, \bar u_t, p^\infty_t, q^\infty_t), v_t}_{{\mathbb{R}}^l} \Big] dt .
\end{split}$$ Recalling that $sup_{t\geq 0}\mathbb{E}|p^{\infty}_t|^2<+\infty$ by definition of solution to equation and $sup_{t \geq 0}\mathbb{E}|Y_t|^2<+\infty$ by we can conclude that $$0 \leq \limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \Big[ \braket{ D_uH(\bar X_t, \bar u_t, p^\infty_t, q^\infty_t), v_t}_{{\mathbb{R}}^l} \Big] dt$$ and the claim is proved
\[thm:necessary truncated\] Similarly we can prove a truncated version of the stochastic maximum principle that involves the solution $(p^T,q^{1,T}, ... q^{d,T})$ of equation with $\eta=0$. Indeed if $(\bar{X},\bar{u})$ is an optimal pair for the control problem , then under Hypothesis 1 the following variational inequality holds $$0\leq\limsup_{T \to \infty} \frac{1}{T} {\mathbb{E}}\int_0^T \braket{ D_u H{\left(}\bar{X}_t, \bar{u}_t, p^T_t, q^T_t {\right)}, u_t-\bar{u}_t}_{{\mathbb{R}}^l} dt,$$ where $H(x,u,p,q)$ is the Hamiltonian of the system and $u(\cdot)$ is an arbitrary admissible control.
Let $v_t = u_t-\bar{u}_t$, for every $u_t$ admissible. The result easily follows combining Lemma \[l:derivative\_J\] with a duality argument. Precisely, choose $\eta = 0$, $\nu_t = D_u b{\left(}\bar{X}_t,\bar{u}_t {\right)}v_t$ and $\Psi_t = D_x f(\bar{X}_t,\bar{u}_t)$ in the general formula .
Sufficient SMP
==============
In this part we prove that under some additional convexity assumption on the Hamiltonian function $H$, the variational inequality obtained in Theorem \[thm:necessary inf horizon\] (the same hold also for Theorem \[thm:necessary truncated\]) is sufficient for optimality.
\[thm:sufficient infinite\] Let $u^*(\cdot) \in \mathcal{U}_{ad}$ be an admissible control, $X^*$ be the corresponding state process and $p^*$ the first adjoint process on infinite time horizon solving for the couple $(u^*,X^*)$. Further, let $(x,u) \mapsto H(x,u,p^*_t,q^*_t)$ be a convex function $d {\mathbb{P}}\times dt-$a.e. and the following minimality condition holds $$\label{eq:suf minimality}
\limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0 \braket{ D_u H(X^*_t,u^*_t,p^*_t,q^*_t)
, u_t - u^*_t}_{{\mathbb{R}}^l} dt \geq 0,$$ for every $u(\cdot) \in \mathcal{U}_{ad}$. Then $u^*(\cdot)$ is optimal both for $\liminf$ and $\limsup$ formulations of the ergodic control problem.
Let $u(\cdot) \in \mathcal{U}_{ad}$ be arbitrary but fixed. Then the goal is to show that the difference $J(u^*(\cdot)) - J(u(\cdot)) $ is non-positive. Using the sub additivity of the $\limsup$ we have $$\begin{aligned}
J(u^*(\cdot)) &- J(u(\cdot)) \leq \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\left[ f{\left(}X^*_t, u^*_t {\right)}- f{\left(}X_t, u_t {\right)}\right] dt \nonumber \\
&\quad = \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\left[ H ( X^*_t, u^*_t, p^{*}_t, q^{*}_t ) - H( X_t, u_t, p^{*}_t, q^{*}_t) \right] dt
\nonumber \\
& \quad + \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\braket{ b {\left(}X_t, u_t {\right)}- b {\left(}X^*_t, u^*_t {\right)}, p^{*}_t} dt \\
& \quad + \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\braket{ \sigma {\left(}X_t, u_t {\right)}- \sigma {\left(}X^*_t, u^*_t {\right)}, q^{*}_t} dt = I_1 + I_2 + I_3. \end{aligned}$$
Now, due to convexity of $H$, the term $I_1$ can be estimated from above as follows
$$\begin{aligned}
I_1 & \leq \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\braket{ D_x H(X^*_t, u^*_t, p^{*}_t, q^{*}_t ), X^*_t - X_t} dt \nonumber \\
& \quad + \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\braket{ D_u H(X^*_t, u^*_t, p^{*}_t, q^{*}_t), u^*_t - u_t}_U dt \nonumber \\
& \leq \limsup_{T \rightarrow +\infty} \frac{1}{T} {\mathbb{E}}\int^T_0
\braket{ D_x H(X^*_t, u^*_t, p^{*}_t, q^{*}_t), X^*_t - X_t } dt, \end{aligned}$$
where in the last step we have used the minimality condition . Next, $$\lim_{T \to \infty} \frac{1}{T} {\mathbb{E}}\braket{p^*_T,X^*_T - X_T} = 0,$$ due to the fact that $p^*,X,X^* \in L^{\infty}{\left(}{\mathbb{R}}_+; L^2(\Omega;H) {\right)}$.
By applying the Itô formula to $\braket{p^{*,T}_T, X^*_T - X_T}_H$ and putting all the terms together we arrive at $$J(u(\cdot)) - J(u^*(\cdot)) \leq 0.$$ The above inequality means that $u^*(\cdot)$ is optimal control.
The form of minimality condition is related to our definition of the Hamiltonian. In fact, one could introduce an another sign convention for $H$, namely $H(x,u,p,q) = \braket{b(t,x,u),p} + \sum_{i=1}^d \braket{\sigma^i(x,u),q^i} - f(x,u)$ which would lead to the corresponding modification in the driver of the first adjoint equation, concavity assumption (instead of convexity) on $H$ in $(x,u)$ and the opposite inequality in . All these changes would lead to the maximality condition usually considered with stochastic maximum principle.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present population synthesis modeling of the X-ray background with genetic algorithm – based optimization method. In our models the best fit could be achieved for lower values of high-energy exponential cut-off ($\approx 170~keV$) and larger amount of the highly obscured ($\log N_{H}=25.5$) AGNs.'
address: |
Department of Astronomy, Odessa National University,\
T.G.Shevchenko park, 65014, Odessa, Ukraine,\
[email protected]
author:
- 'Alexander V. Halevin'
title: 'Population synthesis modeling of the X-ray background with genetic algorithm – based optimization method'
---
Introduction
============
The cosmic X-ray background (XRB) above $\sim1$ keV is known to be produced by integrated emission of discrete sources [@Mush; @Has02; @Mo]. The resent best reviews in this area are [@Gil03] and [@Has03]. XRB synthesis models are usually based on the so-called unification scheme for AGN [@Ant], which explains the different observational appearances by the orientation of accretion disk and molecular torus surrounding the nucleus. The intersection of the line of sight with the torus determines a type 2 AGNs, and the direct observation of the nucleus identifies a type 1 AGNs.
The last works in the area of population synthesis modeling used assumption about extra quantity of absorbed AGNs at high redshifts [@Wil; @Gil99a; @Pom; @Gil01; @Mor]. These models give good approach to the exist observations in the energy range of $1\div100$ kev.
However, there are many unresolved yet problems exist, like as a role of the soft X-ray excess in AGN spectra at high redshifts, behaviour of the AGN luminosity function [@Has03], possible flattening of the spectra of AGNs at high redshifts and a role of the high luminosity obscured AGNs. Furthermore, we do not have yet detailed information about an exponential cut-off at high energies.
Here we present population synthesis models of the X-ray background in order to investigate some additional conditions to obtain the best fit.
Fitting method
==============
In our work we have tested one of the newest effective fitting techniques: a genetic algorithm – based optimization [@Char]. This method implements the Darvin’s natural selection law for the mathematical problems.
The main steps of this technique are:
1\. Constructing a random initial population of the sets of the model parameters and evaluating the fitness of its members. 2. Constructing a new population by breeding selected individuals from the old population. 3. Evaluating the fitness of each member of the new population. 4. Replacing the old population by the new population. 5. Test convergence: unless fittest phenotype matches target phenotype within tolerance, goto 2.
The breeding consists of the several steps such as crossover (as a result the offspring obtains the properties of the both its parents) and mutation (which allows to probe the alternative sets of the model parameters. This option especially important, when population members become practically identical).
This method already was implemented as one of the XSpec fitting methods and now often used for different complicated fitting problems.
Description of the model
========================
As for any population synthesis model, the resulting spectrum was calculated as a mix of the AGN spectra, which are typical for the different classes of AGNs, integrated over redshift and luminosity. To avoid a contribution of the Galactic diffuse radiation component, we have modeled the energy range above 1.5 keV only.
Thus, following [@Com], for AGN spectra we have the next expressions:
$$F(Quasars~1)\propto E^{-\alpha_{h}}*\exp(-\frac{E}{E_{c}})$$
$$F(Seyferts~1)\propto E^{-\alpha_{h}}*\exp(-\frac{E}{E_{c}}) +
F_{refl}(E)$$
$$F(Type~2)\propto F(Type~1)*\exp(-\sigma_{E}*N_{H})$$
where the hard energy index $\alpha_{h}=0.9$. For the exponential cut-off as a base value we have used $E_{c}=320~keV$, but in different models we have tried another values.
The term $F_{refl}(E)$ represents the Compton reflection component by the accretion disk and has been evaluated following [@Mag] assuming inclination angle of $60^{\circ}$.
For obscured (type 2) AGNs, photoelectric absorption cross sections for given hydrogen column $N_{H}$ were calculated following [@Morr]. As the base distribution of equivalent hydrogen column densities $N_{H}$ was taken from [@Ris] (RMS hereafter). Following [@Mai], the local ratio of absorbed and unabsorbed AGNs is $R=4.0\pm0.9$. To evaluate this ratio with redshift, we have used the next formula
$$\begin{aligned}
R(z)=\left(R(0) + \Delta
R(\infty)*\left(1-\frac{1}{1+z/z_{r}}\right)\right)\times\\ \quad
\times\exp(-z/z_{e}) \end{aligned}$$
which in contrast to the “power law & constant” form of [@Gil01] has continuous first derivative and is convenient to represent the ratio dependence from the redshift. Here $(R(0)+\Delta R(\infty))$ is the ratio at “infinity” and $z_{r}$ is a distance for R(z) to get value $R(0)+\frac{\Delta
R(\infty)}{2}$. $z_{e}$ is an exponential cut-off scale, the same as in [@Pom].
During integration for the X-ray luminosity function we have used expression in the form presented by [@Miy00]. For all our models the Hubble constant given by $H_{0}=50~km~s^{-1}~Mpc^{-1}$.
We also computed contribution of the clusters of galaxies to the overall background spectrum, adopting the temperature distribution taken from [@Dav] and cluster luminosity function of [@Ebel].
Galactic hydrogen column was taken as $N_{H}^{gal}=10^{20}~
cm^{-2}$.
To fit our models we have used kindly provided data from different missions like as HEAO-1 A4 LED and A2 HED detectors [@Grub92; @Grub99; @Bol], ROSAT PSPC and ASCA GIS [@Miy98; @Miy03].
The main parameters which describe our implementation of the genetic algorithm are population number (100 members for the model with 3 free parameters and 200 members for the case of 4 free parameters), variable mutation rate (initial value is 0.005) and high selection pressure regime for a breeding of the population members. Furthermore, we have used so-called “elitism”, when the fittest generation member is copied without alteration into the next generation. A good algorithm convergence we achieve after 40–50 iterations.
Results and discussion
======================
In the present work we realized several kinds of models. A simple model $A$ was calculated using equation 4 with $\Delta R(\infty)$, $z_{r}$ and $z_{e}$ as free parameters. The best fit results presented in the Table 1. The corresponded XRB fit and $R(z)$ curve one can see in Fig.1 and Fig.2.
For the model A we have the same rapid increase of the $R(z)$ with redshift, as in work of [@Gil01] and significantly much more slower exponential decay than found by [@Pom]. We consider that such rapid increase of $R(z)$ can not be real because it makes an observer as a particular person. The most probable mistake could be using ’as is’ the distribution of equivalent hydrogen column densities of RMS. In spite of very detailed calculations of the selection effects, their sample based on optical identifications using emission lines and, hence, can lose some highly obscured AGNs. In reality the distribution could shifts towards high values of hydrogen column densities.
In order to investigate this possibility in our model B we have used as an additional parameter the quantity of AGNs with $\log{N_{H}}=25.5$. In the Table 1 the last parameter $n_{25.5}$ was calculated as a fraction of all population of type 2 AGNs in units of the RMS fractional population value (they have used $n_{25.5}\approx0.26$). Then new local ratio of absorbed and unabsorbed AGNs is calculated as $4.0\times(1-n_{25.5}^{old}+n_{25.5}^{new})$. You can see that our new value of $n_{25.5}$ is distinctly higher than used before. For new model values the XRB spectrum is better fitted below 50 keV but is worse above 50 keV. At the same time the $R(z)$ function becomes more smooth. The distribution of hydrogen column densities for the different cases one can see in the Fig. 3.
The main problem of our fitting curves is high bias to the data above 50 keV. In reality if we analyse the data up to 400 keV it is obvious that we do not have clear exponential cut-off profile. Assuming that data below 100 keV represents the ’true’ exponential cut-off we have tried to make the model C with $E_{C}=160~keV$. One can see that in this case we achieve practically ideal approach to the observations, although it demands to increase the $n_{25.5}$ parameter up to 0.93.
Table 1. The best fit parameters for different models.\
model $\Delta R(\infty)$ $z_{r}$ $z_{e}$ $n_{25.5}$
------- -------------------- --------- --------- ------------
A 3.34 0.11 2.49 -
B 1.21 3.01 4.46 0.47
C 2.64 0.47 2.96 0.93
\
In spite of approximate character of our models we have to conclude the presence of a large amount of undetected highly obscured AGNs (with $\log N_{H}=25.5$). I our future models we are going to use more precise simulations of the spectra of absorbed AGNs and the latest results obtained for the properties of the luminosity function.\
[*Acknowledgements. Author is thankful to Duane Gruber, Elihu Boldt, Takamitsu Miyaji and Keith Gendreau for providing of the spectral data from different missions and/or useful comments. Special thanks to G$\rm \ddot u$nther Hasinger and Roberto Gilli for their exciting article “The Cosmic Reality Check”.\
*]{}
[9]{} Antonucci R.R.J., 1993, ARA&A, 31, 473 Boldt E., 2003, private communication Charbonneau P., 1995, AJ Suppl., 101, 309 Comastri A., Setti G., Zamorani G., & Hasinger G., 1995, A&A, 296, 1 David L.P. et al., 1993, ApJ 412, 479 Ebeling H, Edge A., Fabian A., Allen S., Crawford C. & Bohringer H., 1997, ApJ 479, L101 Gilli R., 2003, astro-ph/0303115 Gilli R., Comastri A., Brunetti G. & Setti G., 1999, New Astronomy, 4, 45 Gilli R., Risaliti G. & Salvati M., 1999, A&A, 347, 424 Gilli R, Salvati M & Hasinger G., 2001, A&A, 366, 407 Gruber D., 1992, In the Proceedings of “The X-ray background”, eds. X.Barcons & A.C.Fabian (Cambridge: Cambridge Univ. Press), p.44. Gruber D., Matteson J., Petersen L. & Jung G., 1999, ApJ, 520, 124 Hasinger G., 2002, astro-ph/0202430 Hasinger G. and the CDF-S team, 2003, astro-ph/0302574 Iwasawa K. & Taniguchi Y., 1993, ApJ 413, L15 Loeb A. & Barkana R., 2000, ARA&A, 39, 19 Magdziarz P. & Zdziarski A., 1995, MNRAS 273, 837 Maiolino R. & Rieke G.H.,1995, ApJ 454, 95 Markevitch et al. 2003, ApJ, 583, 70 Miyaji T., 2003, private communication Miyaji T., Hasinger G. & Schmidt M., 2000, A&A, 353, 25 Miyaji T., Ishisaki, Y., Ogasaka, Y., Ueda, Y., Freyberg, M. J., Hasinger, G. & Tanaka, Y., 1998, A&A, 334, L13 Moran E., Filippenko A. & Chornock R., 2002, ApJ, 579, 71 Moretti A, Campana S., Lazzati D., & Tagliaferri G., 2003, astro-ph/0301555 Morrison R. & McCammon D., 1983, ApJ 270, 119 Mushotzky R., Cowie L., Barger A., & Arnaud K., 2000, Nature, 404, 459 Nandra K., George L., Mushotzky R., Turner T., & Yaqoob T., 1997, ApJ 488, L91 Pompilio F., La Franca F. & Matt G., 2000, A&A, 353, 440 Risaliti G., Maiolino R. & Salvati M., 1999, ApJ 522, 157 Wilman R., & Fabian A., 1999, MNRAS, 309, 862
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'fft.bib'
---
[ **Circalunar variations of the night sky brightness – an FFT\
perspective on the impact of light pollution** ]{}
Johannes Puschnig,$^{1,*}$ Stefan Wallner,$^{2,3}$ Thomas Posch$^{2}$\
e-mail: [email protected]\
1 Universität Bonn, Argelander-Institut für Astronomie, Auf dem Hügel 71, D-53121 Bonn, Germany\
2 Universität Wien, Institut für Astrophysik, Türkenschanzstra[ß]{}e 17, A-1180 Wien, Austria\
3 ICA, Slovak Academy of Sciences, Dubravska cesta 9, 84503 Bratislava, Slovak Republic\
Accepted for publication in MNRAS on Dec 11, 2019.\
Abstract {#abstract .unnumbered}
========
Circa-monthly activity conducted by moonlight is observed in many species on Earth. Given the vast amount of artificial light at night (ALAN) that pollutes large areas around the globe, the synchronization to the circalunar cycle is often strongly perturbed. Using two-year data from a network of 23 photometers (Sky Quality Meters; SQM) in Austria (latitude $\sim$48$^{\degree}$), we quantify how light pollution impacts the recognition of the circalunar periodicity. We do so via frequency analysis of nightly mean sky brightnesses using Fast Fourier Transforms. A very tight linear relation between the mean zenithal night sky brightness (NSB) given in $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ and the amplitude of the circalunar signal is found, indicating that for sites with a mean zenithal NSB brighter than 16.5 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ the lunar rhythm practically vanishes. This finding implies that the circalunar rhythm is still detectable (within the broad bandpass of the SQM) at most places around the globe, but its amplitude against the light polluted sky is strongly reduced. We find that the circalunar contrast in zenith is reduced compared to ALAN-free sites by factors of $\mathsf{\sfrac{1}{9}}$ in the state capital of Linz ($\sim$200,000 inhabitants) and $\mathsf{\sfrac{1}{3}}$ in small towns, e.g. Freistadt and Mattighofen, with less than 10,000 inhabitants. Only two of our sites, both situated in national parks (Bodinggraben and Zöblboden), show natural circalunar amplitudes. At our urban sites we further detect a strong seasonal signal that is linked to the amplification of anthropogenic skyglow during the winter months due to climatological conditions.
Introduction {#sec:intro}
============
Impact of moonlight on animals, plants and humans
-------------------------------------------------
The Moon’s synodic period of 29.5 days is its orbital time around the Earth required to show the exact same lunar phase, i.e. for example the time span between two consecutive full moons. The corresponding circalunar oscillation of the Moon’s illumination impacts many types of life on Earth, in particular in the context of reproduction cycles. Scientific work on this topic dates back to the early 20$\mathsf{^{th}}$ century (e.g. [@Fox1924]) and it was likely already recognized by fishermen in the antiquity – due to practical implications – that the size of (edible) gonads of sea urchins varies over the lunar month [@Raible2017].
Later studies revealed that also predator-prey interactions change with moon phase and illumination [@Clarke1983; @Shimose2013], giving advantages to either side, depending on the context. More recently, [@Fallows2016] studied interactions between white sharks (*Carcharodon carcharias*) and Cape fur seals (*Arctocephalus pusillus pusillus*). They found that the shark attack frequency and seal capture success was significantly higher at sunrise during periods of low lunar illumination. In other species, the lunar cycle may control activity [@Kolb1992], foraging, habitat use and communication. See [@Kronfeld-Schor2013] for a recent review on those topics. And for the golden rabbitfish (*Siganus guttatus*), [@Takemura2006] found a direct influence of moonlight intensity on changes in melatonin production.
Some animals are driven by the Moon in their orientation [@Papi1963; @Frisch1993], in particular [@Dacke2011] found that dung beetles (*carabaeus lamarcki*) use the polarization pattern around the Moon as a compass for maintaining their travel direction.
[@Buenning1968] studied how different types of plants react on moonlight. They revealed that plants may undergo leaf movements such that the intensity of lunar illumination is reduced and disturbing effects caused by moonlight are eliminated. It was further shown that illumination by moonlight may even promote flowering, e.g. [@Ben-Attia2016] found flowering patterns in the cactus *Cereus peruvianus* with a period of $\sim$29.5 days and a correlation between moon phase and number and proportion of flowers in bloom. In aquatic systems, [@Zantke2013] first established that the marine worm *Platynereis dumerilii* possesses an endogenous circalunar clock and [@Last2016] was one of the first to search for impacts on aquatic ecosystems and especially noticed a vertical migration of zooplankton which takes place in winter when the Moon is above the horizon at the Arctic, fjord or other sea areas.
Besides animals and plants, the impact of moonlight on humans is still under debate, see for example [@Zimecki2006] for a review. However, many authors find evidence that women of reproductive age do follow the circalunar rhythm [@Reinberg2016], especially the ovulation seems to accumulate around new moon [@Law1986]. What about the effect of lunar illumination on human sleep? On one hand, [@Cajochen2013] find that at full moon the electroencephalogram delta activity during the deep sleep phase is 30 percent decreased and that the sleep duration is reduced by 20 minutes, but on the other hand, [@Cordi2014] find no such significant effects depending on lunar cycle. However, [@Cajochen2014reply] pointed out that the volunteers tested by [@Cordi2014] were not synchronised with respect to their own natural sleep timing, which may have led to low signal-to-noise in their result.
Given the plethora of studies about how moonlight impacts various species on Earth, our work on how artificial light at night (ALAN) impacts the recognizability of the lunar cycle, seems to be a timely matter, as ALAN thus presumably perturbs those species in manifold ways (reproduction, predator-prey interaction, activity, orientation, and so forth) as well.
Moonlight versus light pollution
--------------------------------
Given the fact that ALAN is ever increasing on a global scale [@Kyba2015], effects of artificial light on organisms and ecosystems have gained increasingly more attention in recent years [@Hoelker2010; @Gaston2015; @Manfrin2017]. It has become clear that ALAN indeed has multifaceted consequences for flora and fauna, see e.g. [@Schroer2016] for a review.
However, despite the fact that the influence of moonlight on beings on our planet is well documented (see previous section), and that ALAN’s impact on various organisms and ecosystems have been demonstrated by many authors, to date no studies were performed to investigate ALAN’s impact on processes that rely on synchronization with moonlight. The reason for this knowledge gap is that the community is lacking a quantification of the strength of the circalunar rhythm in dependence of the level of light pollution [@Davies2013]. With this work, we aim to provide such a quantification, i.e. a simple empirical relation between the mean night sky brightness and the amplitude of the circalunar rhythm, allowing the knowledge gap to be filled in the near future.
Locations and methods of our measurements {#sec:locmeth}
=========================================
The present paper is primarily based on zenithal night sky brightness (NSB) measurements taken with Sky Quality Meters (model SQM-LE). These are photodiode based devices with an optical element on the front that narrows down the field of view to a Gaussian-like cone with a full-with-at-half-maximum of $\sim$20$^\degree$. Its effective bandpass ranges from approximately 400–650nm (see Figure \[fig:sqmtrans\]). Technical characterisation and testing was done by [@Cinzano2005] and details about the absolute radiometric calibration are found in [@Bara2019].
Our SQM measurements have been carried out in the Austrian county of Upper Austria at 23 sites, distributed over the whole area of this county and ranging from its capital city Linz – which has very bright skies – to very remote locations such as Krippenstein on the Dachstein plateau ($\sim$2000m above sea level). See Table \[tab:sqmsites\] for station codes and geographic coordinates. Measurements are taken in an automated way, with SQM devices located in weather-proof housings. The network of SQMs is run by the provincial government of Upper Austria. Starting with six devices in 2014, it has grown to 23 SQMs by 2016. A detailed description of the individual sites, their exact locations and light pollution levels are found in [@Posch2018].
Measurements are taken every minute, thus a huge amount of data is generated every night. However, the present data analysis is based on the [*mean* nocturnal NSBs]{} ([<NSB>]{}) and is restricted to data obtained during the years 2016–2017. For each night and SQM site, we calculate [<NSB>]{} as arithmetic means of the minute-by-minute SQM readings for each night. The data series is further constrained for solar elevations below -15 degrees, in the same way as described by [@Posch2018]. The contribution of scattered sunlight to the night sky brightness is negligible below this altitude. Our [<NSB>]{} measurements range from 17.3–21.0 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$.
We stress that we do not apply any further filtering, i.e. data obtained when the sky was cloudy are included.
------ --------------------------- ------------ ---------------- -------------------------
Code Name Latitude N Longitude E Elevation [\[]{}m[\]]{}
(above sea level)
*urban*
LSM Linz, Schlossmuseum N 48 18 19 E 14 16 58 287
LGO Linz, Göthestra[ß]{}e N 48 18 19 E 14 18 30 259
LSW Linz, Sternwarte N 48 17 36 E 14 16 6 336
STY Steyr N 48 2 57 E 14 26 32 307
STW Steyregg-Weih N 48 17 19 E 14 21 13 331
TRA Traun N 48 14 8 E 14 15 11 269
WEL Wels, Rathaus N 48 9 23 E 14 1 29 317
*intermediate*
BRA Braunau N 48 15 40 E 13 2 41 351
GRI Grieskirchen N 48 14 4 E 13 49 33 336
FRE Freistadt N 48 30 33 E 14 30 7 512
MAT Mattighofen N 48 5 50 E 13 9 6 454
PAS Pasching N 48 15 31 E 14 12 36 292
VOE Vöcklabruck N 48 0 21 E 13 38 43 434
*rural*
BOD Nationalpark Bodinggraben N 47 47 31 E 14 23 38 641
FEU Feuerkogel N 47 48 57 E 13 43 15 1628
GIS Giselawarte N 48 23 3 E 14 15 11 902
GRU Grünbach N 48 31 50 E 14 34 30 918
KID Kirchschlag-Davidschlag N 48 26 31 E 14 16 26 813
KRI Krippenstein N 47 31 23 E 13 41 36 2067
LOS Losenstein, Hohe Dirn N 47 54 22 E 14 24 40 982
MUN Münzkirchen N 48 28 45 E 13 33 29 486
ULI Ulrichsberg, Schöneben N 48 42 20 E 13 56 44 935
ZOE Nationalpark Zöbloden N 47 50 18 E 14 26 28 899
------ --------------------------- ------------ ---------------- -------------------------
{width="\textwidth"}
Synthetic models of ground illumination by the Moon and total zenithal night sky brightness {#sec:models}
===========================================================================================
In order to study ALAN’s impact on the lunar rhythm, we first want to understand the amplitude of the *naturally* occuring oscillation of moonlight at our sites, i.e. without any contribution of anthropogenic light at night. We do so using two models. The first one, describing the lunar variation of *ground illumination* in units of *lux* and the second one, describing the naturally occuring variation of zenithal night sky brightness in units of $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, which may be approximately converted to luminance in units of $\mathsf{cd\ m^{-2}}$ [@Bara2016; @Bara2017], using Equation \[eq:unihedron\]. Despite the fact that this formula is widely used to estimate luminances from SQM magnitudes, it was originally derived by [@Garstang1986] for Johnson V-band magnitudes, and is thus just an approximation.
Luminance \[cd/m\^2\] = 10.8 10\^4 10\^[(-0.4 \[mag arcsec\^[-2]{}\])]{} \[eq:unihedron\]
Simplified model of ground illumination by the Moon {#sec:illuminancemodel}
---------------------------------------------------
### General solution for all possible values
To obtain insight into the contribution of moonlight to the total ground illumination, we make use of the moonlight model by [@Seidelmann1992]. The model does not take into account contributions of the skyglow, stars or airglow and thus depend only on two parameters: the mean altitude of the Moon (0-90$^{\circ}$, where 90$^{\circ}$ is the zenith) and the Moon’s phase angle (0-180$^{\circ}$, where 180$^{\circ}$ is full moon). Note that the parallax value is neglected due to its very small contribution. We show the whole parameter space covered by the model in Figure \[fig:moonmodel\].
Unsurprisingly, the amount of ground illuminance is highest when the full moon can be observed in the zenith. It is recognized that with increasing phase angle and altitude the illumination does not increase linear, but rather exponential, owed to the fact that the transmittance of the atmosphere is proportional to $e^{-\frac{\tau}{cos(z)}}$, with $\tau$ being the optical depth and $z$ the zenith distance. Some values which underline this can be found in Table \[tab:moonmodel\]. In fact, for zenithal positions, there is a factor $\sim$8.4 in ground illumination between full and half moon. The large difference is caused by *coherent backscattering* or *opposition surge* [@Hapke1998]. In this phenomenon portions of waves traveling along same paths but in opposite directions, interfere constructively with each other, causing a peak at zero phase (full moon). In case of full moon the illuminance varies by a factor of $\sim$166 between zenithal and horizontal positions. Please note that horizontal position, i.e. altitude of zero degrees, is valid for the center of the lunar disk. Hence, a slight ground illumination is visible and not zero. The resulting values can be considered as theoretical only, since every other possible light emitting source is neglected, even stars or the Milky Way are not taken into account. However, the model shows how the lunar position and its phase angle are associated with the consequential ground illumination.
![Results of the illuminance model as developed by [@Seidelmann1992] for all phase angles and altitudes of the Moon.[]{data-label="fig:moonmodel"}](illuminancemodel_grid.png){width="0.7\columnwidth"}
ALT \[$^{\circ}$\] PA \[$^{\circ}$\] $\mathsf{G_{ILL}}$ \[$\mathsf{mlux}$\]
-------------------- ------------------- ----------------------------------------
0 90 0.192
0 180 1.622
45 90 19.474
45 180 164.059
90 90 31.949
90 180 269.153
: Resulting values of the theoretical Moon model for half and full moons. ALT is the lunar altitude, PA is the lunar phase angle and $\mathsf{G_{ILL}}$ is the ground illumination caused only by moonlight.
\[tab:moonmodel\]
### Application of the illuminance model to one of our sites {#sec:appliillumodel}
We now apply the model to one of our sites, namely VOE. To do so, we first calculate for the years 2016–2017 the altitude and phase of the Moon *at midnight* and then derive the corresponding illuminance through the model. The result is shown in Figure \[fig:LUXVOE\]. A strong seasonal variation is recognized in the model, mainly caused by the changing altitude of the ecliptic (and thus the Moon) between summer and winter, leading to a natural variation of the moon illuminance by a factor of $\sim$3.6 (peak-to-peak for full moon).
![Ground illuminance (the Moon’s contribution to it) evaluated for the location of Vöcklabruck, Upper Austria, over the years 2016–2017. The high-frequency lunar cycle is modulated with a seasonal variation caused by the seasonal change of the Moon’s altitude.[]{data-label="fig:LUXVOE"}](LUXVoecklabruckOriginal.png){width="0.7\columnwidth"}
Zenithal night sky model for the SQM band {#sec:skymodel}
-----------------------------------------
### The Cerro Paranal Advanced Sky Model
Using the *Cerro Paranal Advanced Sky Model* (Skycalc), we are able to compare our zenithal SQM measurements to a synthetic sky model that is cloud-free and takes into account several sources of light such as scattered moonlight, starlight, molecular emission of the lower atmosphere and the airglow (upper atmosphere).
Skycalc was published by [@Noll2012] and [@Jones2013], as part of an Austrian in-kind contribution to the European Southern Observatory (ESO), e.g. ESO’s exposure time calculator is based on it. Skycalc’s current version (2.0.4) also comes with a Python-based command line interface[^1]. However, the model used for our study is based on Skycalc 1.4.4, which was available through a web interface only. The main input parameters are zenith distance (or airmass) of the observation, precipitable water vapor (PWV) and monthly averaged solar flux. For the moon radiance component, the separation of Sun and Moon as seen from Earth, the Moon-target separation, moon altitude over horizon and the Moon-Earth distance are needed.
We have decided to make some simplifications, allowing us to evaluate the model on a 2-dimensional parameter grid with vectors of (Sun-Moon-separation, moon altitude) only. This is reasonable in our case, because the measurement devices we are using, the Sky Quality Meters of type SQM-LE, are equipped with a front lens that narrows down the field of view to a roughly 20 degree wide cone, pointed towards zenith. Hence, we only need to consider zenithal night sky brightness. The two input parameters Moon-target separation and moon altitude can thus be simplified to one parameter, with the former one being the moon zenith distance. We have further decided to evaluate the model for a fixed PWV value of 5mm, a monthly averaged solar flux of 130sfu and for a fixed mean Moon-Earth distance. These simplifications have practically no influence on our results, since ALAN’s contribution to our SQM measurements is magnitudes larger than the natural variation caused by phenomena such as PWV or solar flux. However, variations due to moon phase and height are fully treated by our gridded model evaluation for the zenith. Since the natural, cloudless sky brightness changes smoothly, a grid spacing of one degree in both parameters (Sun-Moon-separation, moon altitude) was found to be sufficient.
The result is a synthetic (cloud-free) night sky spectrum for the target location in units of $\mathsf{photons\,s^{-1}\,m^{-2}\,\upmu m^{-1}\,arcsec^{-2}}$, i.e. spectral radiance. We first convert to $\mathsf{{erg\,s^{-1}\,cm^{-2}\,{\text{\normalfont\AA}}^{-1}\,arcsec^{-2}}}$ and then multiply with the SQM transmission curve as published in [@Cinzano2005] and shown in Figure \[fig:sqmtrans\], together with transmission curves of Bessel BVR filters. The radiance within the SQM band is then calculated via integration over the wavelength axis. Using a zeropoint (ZP) of -12.92 (Puschnig et al. in prep), we finally convert to $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ via equation \[eq:radiance2mag\].
NSB \[[mag\_[SQM]{}arcsec\^[-2]{}]{}\] = -2.5 log\_[10]{}(radiance \[erg s\^[-1]{} cm\^[-2]{} arcsec\^[-2]{}\]) + ZP \[eq:radiance2mag\]
Note that Equation \[eq:radiance2mag\] results to a modeled zenithal sky brightness of 21.87$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ for new moon, which is in agreement with SQM observations of remote, rural sites [@Posch2018; @Bertolo2019; @Bara2019]. We further stress that the exact absolute value is not critical for the scientific results of the paper.
### Application of the Cerro Paranal Advanced Sky Model to one of our sites {#sec:appliskymodel}
Analogous to Section \[sec:appliillumodel\], we now evaluate the Skycalc model for one of our sites (VOE) that we use as a proxy for our network. We do so for the years 2016–2017. However, in this case we calculate for each night the mean NSB ([<NSB>]{}) rather than just the value at midnight. The result is shown in Figure \[fig:SYNVOE\]. As for the illuminance model, a strong seasonal variation is recognized, that is mainly caused by the changing altitude of the Moon. However, the relative change between brightest and darkest [<NSB>]{}(18.04 vs. 18.69 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$) corresponds to a (linear-scale) factor of 1.8 only, which is half the amplitude that is seen in the illuminance model. The main reason for this discrepancy lies in the fact that we evaluate the illuminance model at midnight only, i.e. at the point of maximum illumination, whereas for Skycalc we calculate the nightly means within dark-time limits. Given the fact that summer nights are shorter, the relative contribution of the full moon to the zenithal [<NSB>]{} is thus higher in summer than in winter and the dynamic range of the seasonal variation shrinks in that case.
![Skycalc sky model for the zenith, evaluated for one of our sites (VOE) for the years 2016–2017. The y-axis was limited to show only the bright peaks around full moon. Beside the circa-monthly oscillation, a seasonal variation is recognised, caused by changing altitude of the Moon between summer and winter.[]{data-label="fig:SYNVOE"}](SYN2016-2017original.png){width="0.7\columnwidth"}
Data analysis
=============
Fourier analysis {#sec:fft}
----------------
Using our 2-year data of nightly mean [<NSB>]{} values, we aim to reveal the amplitude of the circalunar periodicity (and other periodic signals that might exist). Numerous implementations of discrete Fourier transforms exist, but probably the most common one is the Fast Fourier Transform (FFT), which we also use for the analysis of our SQM data. In particular, we apply the FFT algorithm as implemented in `numpy` [@van2011numpy], the fundamental package for scientific computing with Python. In the following, we describe some general properties of the FFT and highlight common obstacles of the method and how we treat them.
### Which unit to choose? {#sec:fftunit}
The SQM delivers data in units of $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$. However, conversions into linear units are available, see Equation \[eq:unihedron\]. Thus, we investigate how well the amplitude of periodic variations such as the circalunar cycle caused by varying moon phases or the seasonal rhythm, that is driven by variations of the moon zenith distances between summer and winter, are recovered in frequency space after application of the FFT in dependence of the data input unit.
For a linear input, the increase in zenithal night sky brightness from new moon to full moon starts very shallow, almost unrecognizable. Then, few days before full moon, the increase accelerates to finally form a sharp peak in the (time,[<NSB>]{}) plane (see Figure \[fig:ifft\_cdm\]). On the contrary, in logarithmic units such as $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ the phase of the shallow increase is more pronounced (stretched in time) and the later steep increase somewhat compressed. Thus, the apparent course in the (time,[<NSB>]{}) plane is rather smooth compared to linear units (see Figure \[fig:ifft\]).
Since the FFT of a signal that is spread out in time delivers a compact result in frequency space and vice versa, logarithmic units such as $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ are preferred, because the smoother course leads to a better definition of the circalunar cycle in frequency space, i.e. a single peak.
### The role of gaps in [<NSB>]{} measurements {#sec:fftgaps}
Gaps in SQM data series can occur due to several reasons, be it hardware- or software-failures or even meteorological conditions. In this section, we investigate how gaps in the time series affect the ability of the FFT algorithm to correctly reproduce amplitudes of time-dependent variations. We do so by introducing single and double gaps of varying lengths into our synthetic sky model. The recovered amplitudes of the circalunar and seasonal rhythm are then evaluated against the gap fraction, i.e. the fraction of data points on the equal-distant input time grid without measurement.
We stress that under all circumstances gaps should not be left as such, but replaced with some form of interpolation or reasonable value. We made tests using linear, quadratic and cubic spline interpolations, as well as using the mean of the remaining data as fill value. We find that the latter one is the most robust and preferred method. Although cubic spline interpolation gives *slightly* better results for small gaps, i.e. when the gaps size is much smaller than the periodicity of the desired signal, it can cause unforeseen results for larger gaps.
Our results are summarized in table \[tab:gaps\], which shows that the single- and double gap tests give very similar results: gap fractions of 1%, 5%, 7.5% and 10% recover the amplitude of the circalunar variation (A) at levels of 99%, 92–95%, 91% and 87–88% respectively. Even gap fractions of 20% recover 75–77% of the true amplitude.
The seasonal variation (S) is less affected, because of its longer periodic time. It can be accurately derived even for gap fractions of 20%.
However, as described by [@Munteanu2016], for even larger periods of missing data, one should consider to perform spectral analysis using other algorithms such as the *Z transform* or the *Lomb-Scargle algorithm*. They might be able to reproduce the amplitude for cases where gaps make up more than 50 percent of the time series.
Additionally, the presence of gaps in the time series leads to an increase of ‘frequency noise’ in the amplitude spectrum, limiting the chance to detect low-amplitude variations at certain frequencies. As shown in Table \[tab:gaps\], the noise roughly doubles between 0 and 5% gap fractions, but stays almost constant from thereon.
----- ------- ------- ------- -- ------- ------- -------
A S N A S N
(2) (3) (4) (5) (6) (7)
0 1.508 0.073 0.006 1.508 0.073 0.006
1 1.489 0.077 0.008 1.488 0.075 0.009
5 1.415 0.078 0.013 1.438 0.083 0.011
7.5 1.397 0.070 0.012 1.403 0.078 0.014
10 1.356 0.081 0.012 1.349 0.075 0.015
20 1.217 0.072 0.014 1.205 0.074 0.015
30 1.037 0.060 0.015 1.067 0.067 0.015
50 0.754 0.036 0.013 0.765 0.010 0.015
----- ------- ------- ------- -- ------- ------- -------
: We test the influence of single and double gaps on the ability to recover signal amplitudes from synthetic sky model data. The gap fraction in percent is given in *column 1* and the recovered amplitudes of the circalunar and the seasonal variation are shown in *column 2 and 3* for the single gap case, and *column 5 and 6* for the double gap test. FFT Noise is given in *column 4 and 7*. The unit for all measurements is $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$.[]{data-label="tab:gaps"}
### Importance of the length of the data series for FFTs {#sec:fftdatalength}
The length of a time series as well as its sampling rate are of importance for FFT studies, because they define the frequency resolution in the final amplitude spectrum.
Our sampling rate $\mathsf{f_s}$ is one measurement per night or $\mathsf{f_s=\,1\,d^{-1}}$. The distance $\mathsf{\Updelta t}$ between two data points is: $\mathsf{\Updelta t\,=\,\frac{1}{f_s}\,=\,1\,d}$. The final frequency range is thus limited to the interval $\mathsf{[-\frac{f_s}{2};+\frac{f_s}{2}]}$, i.e. $\mathsf{-0.5...0.5\,d^{-1}}$. The highest measurable frequency is $\mathsf{0.5\,d^{-1}}$ or two days periodic time. The number of discrete points $\mathsf{N}$ in the final frequency domain equals the number in time domain. Thus, the distance $\mathsf{\Updelta \nu}$ in frequency space is $\mathsf{\Updelta \nu\,=\,\frac{f_s}{N}}$, which shows that the frequency resolution is controlled by the sampling rate $\mathsf{f_s}$ and the length of the time series $\mathsf{N}$. Since our sampling rate is fixed, the number of data points is the main quantity that drives the frequency resolution in our final amplitude spectrum.
For example, if we wanted to *detect* the frequency of the lunar synodic month (without prior knowledge), with an accuracy of 0.5d, we would need at least 1711 data points, equivalent to 4.68 years at a sampling rate of one measurement per night.
However, with prior knowledge of the period - in our case 29.5d for the synodic month - one can adjust the time axis such that the final discrete amplitude spectrum covers the corresponding frequency, i.e. $\mathsf{\frac{1}{29.5}\,=\,0.0339\,d^{-1}}$. This is achieved by limiting the data points such that no discontinuities occur at the edges of the time series, i.e. spectral leakage (see next paragraph) is eliminated. That way, even time series of only one year recover the amplitude on levels better than 90 percent.
### Avoiding spectral leakage {#sec:spectralleakage}
The ability of the FFT algorithm (and any other discrete Fourier transforms) to recover amplitudes is limited due to the fact that the duration of the observation is finite. This means that the input signal is factual a product with a rectangular window. The discrete spectrum of any finite signal is thus spread out over multiple frequency components and the amplitude is not fully recovered anymore. This behaviour is called *spectral leakage*.
However, the effect may be reduced by 1) avoiding discontinuities of the input signal or 2) gradually decreasing the amplitude of the signal towards the edges of the measurement series. The first method requires prior knowledge of the periodic time and phase of the signal of interest (which in most applications is not fulfilled). The latter case can be achieved by multiplying the input time series with a window function, e.g. a Hanning window, before the FFT is applied.
Since the periodicity of the synodic month is known, we could test both scenarios using sinusoidal and our synthetic sky time series as input. Although the application of a Hanning window improved amplitude recovery in cases where discontinuities occurred at the edges, we find that an continuous input time series (e.g. from first new moon 2016 to last new moon in 2017) gives best results with only negligible amounts of spectral leakage and amplitude recovery at levels better than 95 percent. Thus, for our main science case, i.e. studying the circalunar rhythm, a continuous time series as input for the FFT is recommended and all our FFT analysis was performed using data from new moon 2016/01/09 to new moon 2017/12/18 only.
Results
=======
FFT analysis of the illuminance and Skycalc models {#sec:fftmodel}
--------------------------------------------------
As a reference for our measurement sites, we perform an FFT analysis of our model data using the considerations from Section \[sec:fft\]. To do so, for the Skycalc model we first calculate the nightly averages, i.e. [<NSB>]{}, as described in Section \[sec:locmeth\]. The resulting FFT amplitude spectra are shown in Figures \[fig:LUXVOEfft\] and \[fig:SYNVOEfft\].
An analysis of both models reveals significantly strong amplitudes at the frequency of $\mathsf{\frac{1}{29.5}\,=\,0.0339\,d^{-1}}$, i.e. the fundamental of the circalunar rhythm. Also its harmonics are identified at multiples of that frequency. However, the noise level in the illuminance model is higher because of its peak-like input (see Section \[sec:fftunit\]).
We also note that the zero-frequency amplitude is not comparable between the models. While for the Skycalc model, the zero-frequency is in fact a representation of the mean [<NSB>]{}, this is not true for the illuminance model, because for the FFT analysis all input data with illuminance levels of 0lux is considered as gaps and are thus filled with the mean in order to improve recognition of the circalunar rhythm as explained in Section \[sec:fftgaps\].
It is also recognized that the seasonal variation seen in the amplitude spectrum of the illuminance model is stronger than in the Skycalc model. This is caused because we only consider midnight values for the illuminance model, while for the Skycalc model nightly averages are calculated. Hence, the contribution of the peak NSB at midnight is smoothed out in time due to averaging, while it is fully captured for midnight data. The averaging effect is even stronger in winter when the Moon’s contribution to the zenithal [<NSB>]{} is highest, because then the nights are longer. The seasonal signal is thus equalized throughout the year.
![FFT amplitude spectrum of the illuminance model as described in Section \[sec:illuminancemodel\]. The shown frequency range is limited to values between 0 and 0.2. Note that the unit of the y-axis is *milli-lux*. Labeled amplitudes (*from left to right* correspond to the mean of all input data, the seasonal variation and the circalunar cycle.[]{data-label="fig:LUXVOEfft"}](LUXVoecklabruckFFT.png){width="0.7\columnwidth"}
![FFT amplitude spectrum of the Skycalc sky model as described in Section \[sec:skymodel\]. The shown frequency range is limited to values between 0 and 0.2. Labeled amplitudes (*from left to right* correspond to the mean [<NSB>]{} of the input data, the seasonal variation and the circalunar cycle.[]{data-label="fig:SYNVOEfft"}](SYN2016-2017fft.png){width="0.7\columnwidth"}
Identifying the synodical month in the [<NSB>]{} data and quantifying its amplitude
-----------------------------------------------------------------------------------------
From our SQM data, we first calculate the nightly averages, i.e. [<NSB>]{}, as described in section \[sec:locmeth\] and then perform an FFT analysis using the considerations from section \[sec:fft\]. That way, we can clearly detect the circalunar rhythm at all our sites (see Table \[tab:fftresult\]), be it rural or urban. However, the amplitude decreases from 1.55 to 0.33$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, corresponding to a factor of three on a linear scale, from the darkest to the brightest sites. A comparison between amplitude spectra of a typical urban and rural site is shown in Figure \[fig:lgo\_vs\_kri\].
It is also recognized that at rural sites, harmonics of the main variation at a frequency $\mathsf{0.0339\,d^{-1}}$ can be identified up to 3rd order. In contrast, at urban sites, they perish in frequency noise, which increases with light pollution. In fact, the FFT noise roughly shows a bimodal distribution between rural and urban sites, as seen in Figure \[fig:fftnoise\]. This is, because day-to-day variations are more pronounced at urban, light-polluted sites due to backscattering of ALAN at clouds [@Kyba2011; @Kyba2012; @Puschnig2014a; @Puschnig2014b; @Aube2016]. Thus, the observed night sky brightness typically jumps between two preferred values; see e.g. [@Posch2018 Figures A1a, A1b and 7]. Since the noise as calculated here is dominated by high frequencies, i.e. roughly 64 percent of the frequencies correspond to periodic times of equal or less than five days, the same bimodality is seen here.
Another interesting feature seen in Figure \[fig:fftnoise\] is that the scatter of the noise level increases along the [<NSB>]{} axis, i.e. darker sites show larger variance in the noise. The cause of this trend might be explained by findings of [@Kocifaj2014], who showed that the amplification factor due to clouds decreases with increasing city size and thus the level of light pollution.
[ccccccc]{} COD & $\mathsf{{<NSB>}_W}$ & $\mathsf{{<NSB>}_{mag}}$ & A & CLC & S & N\
(1) & (2) & (3) & (4) & (5) & (6) & (7)\
\
LGO & 18.25 & 17.31 & 0.34 & 36.8 & 0.73 & 0.071\
LSM & 15.32 & 17.51 & 0.33 & 34.3 & 0.77 & 0.070\
WEL & 12.74 & 17.70 & 0.36 & 39.3 & 0.86 & 0.067\
LSW & 12.39 & 17.73 & 0.41 & 45.9 & 0.58 & 0.069\
STW & 10.31 & 17.93 & 0.45 & 50.0 & 0.49 & 0.069\
TRA & 8.90 & 18.09 & 0.46 & 52.8 & 0.55 & 0.061\
STY & 7.54 & 18.27 & 0.50 & 58.5 & 0.69 & 0.068\
PAS & 6.33 & 18.46 & 0.53 & 62.9 & 0.57 & 0.054\
BRA & 5.88 & 18.54 & 0.66 & 83.7 & 0.67 & 0.057\
GRI & 5.56 & 18.60 & 0.57 & 69.0 & 0.92 & 0.057\
VOE & 5.56 & 18.62 & 0.64 & 82.0 & 0.74 & 0.060\
FRE & 5.56 & 18.60 & 0.71 & 92.3 & 0.76 & 0.070\
MAT & 3.67 & 19.05 & 0.83 & 114.8 & 0.65 & 0.062\
MUN & 1.91 & 19.76 & 1.10 & 175.4 & 0.56 & 0.047\
GIS & 1.44 & 20.07 & 1.02 & 155.9 & 0.06 & 0.050\
ULI & 1.38 & 20.13 & 1.16 & 191.1 & 0.50 & 0.057\
KID & 1.12 & 20.37 & 1.19 & 204.8 & 0.09 & 0.040\
GRU & 1.07 & 20.39 & 1.34 & 243.6 & 0.38 & 0.036\
FEU & 0.84 & 20.66 & 1.22 & 207.6 & 0.07 & 0.057\
LOS & 0.83 & 20.86 & 1.30 & 243.6 & 0.01 & 0.059\
KRI & 0.73 & 20.81 & 1.16 & 191.1 & 0.07 & 0.049\
ZOE & 0.70 & 21.01 & 1.55 & 324.6 & 0.10 & 0.056\
BOD & 0.62 & 21.04 & 1.51 & 301.8 & 0.10 & 0.064\
SYN & 0.73 & 20.81 & 1.51 & 301.8 & 0.08 & 0.004
{width="\textwidth"}
![FFT noise increases with the level of light pollution.[]{data-label="fig:fftnoise"}](OOE_noise.png){width="0.5\columnwidth"}
Relation between the circalunar amplitude and zenithal [<NSB>]{} for mid latitudes
----------------------------------------------------------------------------------------
Previous observations have already qualitatively shown, that the circalunar rhythm steadily fades away with an increasing level of anthropogenic light at night. Using our FFT analysis, we can now quantify how ALAN affects the lunar cycle’s degree of recognition. We do so by plotting the mean nightsky brightness, given by the amplitude at zero-frequency, against the amplitude of the circalunar rhythm. The result is shown in the left panel of Figure \[fig:circalunarrelation\]. A linear fit (see Equation \[equ:nsbamprelation\]) is found to be appropriate, with a scatter of only 0.062 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$. Extrapolation leads us to an [<NSB>]{} level of 16.5$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, the level at which the circalunar rhythm practically vanishes (at zenith) and is indistinguishable from the polluted nightsky. $$\label{equ:nsbamprelation}
\mathsf{
A\,=\,0.322\,{<NSB>}\,-\,5.324
}$$ We may also convert the amplitudes in linear units, e.g. to luminance (see Equation \[eq:unihedron\]), and then express the amplitude of the circalunar cycle as multiple of [<NSB>]{}, which we denote as *circalunar contrast* (CLC). As shown in the right panel of Figure \[fig:circalunarrelation\], the contrast between the average [<NSB>]{} level and the darkening/brightening due to new/full moon ranges between 30 and 300 percent for our urban and rural sites respectively.
{width="\textwidth"}
Searching for other than lunar frequencies in the [<NSB>]{} data
----------------------------------------------------------------------
As seen in Sections \[sec:models\] and \[sec:fftmodel\] a naturally occuring seasonal variation caused by the changing height of the ecliptic (and thus the Moon) is expected. In our measurements, however, for the darkest, rural stations (FEU, KRI, LOS, ZOE, BOD), we do not recover any such seasonal variation at a significant level (compare Table \[tab:fftpeaks\]). This is mainly because our [<NSB>]{} data includes overcast skies that weaken the amplitude of the signal in an unforeseeable way, and given the fact that we cover only two years, the signal easily vanishes.
On the other side, we do detect a very strong seasonal variation at all urban and also at most intermediate stations (compare Table \[tab:fftpeaks\]). However, the amplitude of that signal is too large to be caused by the Moon. This oscillation might be the result of climatological effects that enhance ALAN during winters, i.e. combined effects of increased surface albedo and lower vegetation state [@Wallner2019]. The observed seasonal variation might also be related to the aerosol optical depth (AOD). As demonstrated by [@Aube2015], the zenith radiance can increase several tens of times when optical depth is significantly lowered. However, disentangling the contribution of the several effects (surface albedo, vegetation, AOD) is beyond the scope of this paper and would require ancillary data products.
At urban sites (LGO, LSM, WEL, LSW, STY, PAS, BRA, GRI, VOE, MAT), we further detect a significant ($>$3 sigma) signal with a periodic time of $\sim$100 days. The cause of this roughly quaterly variation is unclear and was not reported previously. In our Table \[tab:fftpeaks\] we denote it as *unknown*.
We also search for weekly variations that might be associated with an increased human nighttime activity on weekends. Although we do see at our urban stations amplitudes on the order of 0.15–0.20 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, these are not significant ($<$3 sigma). Furthermore, its frequency is very close to that of the third lunar harmonic.
![The amplitude of the seasonal oscillation against [<NSB>]{}, both derived through FFT analysis of nightly [<NSB>]{} values. Each point corresponds to one of our SQM stations, with error bars indicating the $\pm 1\upsigma$ errors found from the noise in the FFT amplitude spectrum. The red point shows the location of our synthetic Skycalc sky model, indicating that for our SQM network the amplitude of the naturally occuring seasonal variation due to the Moon (its seasonal change of altitude) has an amplitude of less than 0.1$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, which is too low to be detectable within the frequency noise as we see in our FFT amplitude spectra. Thus, no clear trend or relation is identified. The strong seasonal amplitudes seen at urban sites are thus caused by other seasonal effects such as changing albedo or vegetation state between winter and summer (Puschnig et al. in prep.). []{data-label="fig:seasonalrelation"}](OOE_seasonal_relation.png){width="0.5\columnwidth"}
Discussion and Conclusion
=========================
Prior to a discussion of possible limitations of our method and implications of our findings in a wider context, we first compare our results to those of [@Bara2016], who performed an SQM based FFT analysis of 14 Galician stations using a 1-year dataset. [@Bara2016] have previously defined the *moonlight factor* (y) as the ratio of *power densities* between the fundamental circalunar frequency and the zero frequency. Note that they used double-sided power density spectra, while we use single-sided amplitude spectra. The power spectra of [@Bara2016] thus show negative and positive frequency components $k>0$ with heights of $\sfrac{A_k^2}{4}$ compared to our amplitudes (A) in Table \[tab:fftresult\] and at a height of $A_0^2$ for the zero-frequency component ($\mathsf{{<NSB>}_{mag}}$). They reported ranges for the moonlight factor between 0.2–0.3$\times10^{-3}$ for urban sites, 1.5–2.4$\times10^{-3}$ for dark rural sites and values in-between for intermediate regions. A compilation of the moonlight factor calculated for our sites is shown in Table \[tab:moonlightfactor\]. We find that the moonlight factor derived through our FFT methodology and data is almost a factor of 2 lower than those reported in [@Bara2016]. Although it is expected that the Moon’s impact on zenithal NSBs in Galicia is stronger than in Upper Austria, because the Galician network’s geographic latitude is lower by $\sim$5$^{\degree}$ and thus the ecliptic and Moon closer to zenith, a factor of 2 seems to be relatively high. We argue that the relatively large difference is mainly caused because [@Bara2016] take into account NSB measurements obtained at midnight only rather than averaging over dark time hours as we do. As a result – in analogy to the explanation in Section \[sec:fftmodel\] – for nightly averages peak values that occur at midnight are smoothed in time, which is also the reason why we do not capture the seasonal variation (compare Figure \[fig:seasonalrelation\]). On the other hand, averaging over dark times has the advantage of providing a better definition and thus less scatter of the amplitudes over several months, because a single nightly measurement may easily be affected by short-term perturbations such as cloud cover, which may impact NSB measurements in a complex way: Considering that backscattering of moonlight on scattered clouds (in zenith) exists in analogy to previous findings of [@Kyba2011; @Kyba2012; @Puschnig2014a; @Puschnig2014b], clouds may enhance the zenithal NSB on the one hand, but it may also lower the zenithal NSB on the other hand, e.g for fully overcast skies, similar to previous findings of [@Posch2018; @Jechow2019].
However, recognition of the circalunar amplitude as performed by [@Bara2016] is affected by spectral leakage that lowers the amplitude (compared to our approach) due to smearing out of the signal over multiple frequencies, as explained in Section \[sec:spectralleakage\]. Additionally, [@Bara2016] use only 1-year of input data, leading to lower frequency resolution and thus even more leakage, as explained in Section \[sec:fftdatalength\].
We conclude that a combination of these effects causes the factor 2 discrepancy, but we stress that the linear relation in Figure \[fig:circalunarrelation\], its intercept mainly, would not change by a factor 2 for geographic locations similar to those of the Galician network. This is shown in the following.
In principle, it is expected that the relation in Figure \[fig:circalunarrelation\] shifts towards lower circalunar amplitudes for geographic latitudes north of 48$^\degree$ and towards higher amplitudes otherwise. In order to asses how much the relation’s intercept may shift, we perform FFT analysis of 2-year Skycalc models calculated for different latitudes. For the first test, we choose a latitude of N 36$^{\degree}$ (e.g. Gibraltar). The derived amplitude of the circalunar cycle is 1.61$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$, which is only $\sim$0.1$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ higher than for the synthetic models calculated for our network. For the second test, we choose a latitude of N 60$^{\degree}$ (e.g. slightly north of Stockholm). However, we caution that for latitudes above 51.5$^{\degree}$, during summer there are no dark times with the sun being more than 15$^{\degree}$ below the horizon. As a result, gaps in the input data limit the recoverability of the circalunar amplitude, as explained in Section \[sec:fftgaps\]. In that case we thus find an amplitude of 1.02$\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ only, which is mainly due to the large summer gaps accounting for roughly 25 percent of the input data. From these tests we conclude that our relation between the circalunar amplitude and [<NSB>]{}is at least valid for latitudes between $\sim$40–50$^\degree$.
Our linear relation between zenithal [<NSB>]{} and the circalunar amplitude in Figure \[fig:circalunarrelation\] implies that the circa-monthly variation of moonlight is still traceable over large areas, not only in Upper Austria, but also in many other countries (compare light pollution atlas by [@Falchi2016]). The circalunar rhythm is thus expected to practically vanish due to ALAN only in the innermost parts of major cities where the zenithal NSB may exceed 16.5 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$. However, at the same time, it is recognized that the contribution of the Moon to the total zenithal NSB (i.e. contrast) is largely reduced due to ALAN, namely by a factor of $\sim\sfrac{1}{9}$ for urban areas (e.g. Linz with $\sim$200,000 inhabitants), a factor of $\sim\sfrac{1}{3}$ for small towns with less than 10,000 inhabitants (e.g. Freistadt or Mattighofen) and still up to $\sim\sfrac{1}{2}$ for one of our rural stations (e.g. Münzkirchen, a village with less than 2,600 inhabitants). Only two of our sites, both situated in national parks (Bodinggraben and Zöblboden), show natural circalunar amplitudes.
Finally, we discuss implications of the SQM’s spectral bandpass that covers a range of $\sim$300–680$\upmu$m. Although this is very close to the so called *photosynthetically active radiation*, i.e. the spectral range between 400–700$\upmu$m to which photosynthetic organisms are sensitive, we caution that within that spectral range, Chlorophyll – the most abundant plant pigment – has a sensitivity curve that is very different from the SQM’s sensitivity curve shown in Figure \[fig:sqmtrans\]. Rather than being mostly sensitive to green photons, Chlorophyll is mostly sensitive to red and blue photons. The exact degradation of the circalunar rhythm as recognized by photosynthetic organisms may thus be different from what we observe using SQMs. However, other species such as e.g. ocean fish, that have maximum spectral sensitivities between 500 and 550$\upmu$m [@Marshall2017] may recognize the degradation of the circalunar cycle exactly as described in our paper. How they are affected by a decreased amplitude of the circa-monthly signal is yet to be shown in future (chrono)biological studies.
----- ----------------- ---------------
COD [<NSB>]{} $y\times10^3$
(1) (2) (3)
*urban*
LGO 17.31 0.10
LSM 17.51 0.09
WEL 17.7 0.10
LSW 17.73 0.13
STW 17.93 0.16
TRA 18.09 0.16
STY 18.27 0.19
*intermediate*
PAS 18.46 0.21
BRA 18.54 0.32
GRI 18.6 0.23
VOE 18.62 0.30
FRE 18.6 0.36
MAT 19.05 0.47
*rural*
MUN 19.76 0.77
GIS 20.07 0.65
ULI 20.13 0.83
KID 20.37 0.85
GRU 20.39 1.08
FEU 20.66 0.87
LOS 20.86 0.97
KRI 20.81 0.78
ZOE 21.01 1.36
BOD 21.04 1.29
SYN 20.81 1.32
----- ----------------- ---------------
: Station codes and zero-frequency [<NSB>]{} in units of $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ are shown in columns 1 and 2. Calculations of the moonlight factor (y) as defined in [@Bara2016] are given in column 3[]{data-label="tab:moonlightfactor"}
Summary
=======
We studied the circalunar periodicity via FFT analysis of night-time averages of zenithal NSB measurements obtained during the years 2016 and 2017 through a network of 23 SQMs located in Upper Austria at a latitude of $\sim$48$^\degree$. Models of the sky [@Noll2012; @Jones2013] and the Moon [@Seidelmann1992] were used as a reference of ideal conditions and to develop an optimal methodology for the recognition of the circalunar periodicity. The following conclusions are drawn from our study:
- A tight linear relation between [<NSB>]{} given in $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ and the circalunar amplitude is found (see Figure \[fig:circalunarrelation\]). This relation indicates that for sites with [<NSB>]{} brighter than 16.5 $\mathsf{{mag_{SQM}\,arcsec^{-2}}}$ the lunar rhythm practically vanishes. This finding implies that the circalunar rhythm is still detectable (within the broad bandpass of the SQM) at most places around the globe.
- However, the circalunar contrast in zenith is largely reduced compared to ALAN-free sites. In the state capital of Linz ($\sim$200,000 inhabitants) the Moon’s contribution to zenithal [<NSB>]{} is reduced by a factor of $\mathsf{\sfrac{1}{9}}$. For small towns (e.g. Freistadt or Mattighofen) with less than 10,000 inhabitants, we find that the circalunar contrast in zenith is lowered by a factor of $\mathsf{\sfrac{1}{3}}$ due to ALAN and even at one of our rural sites, Münzkirchen, a village with less than 2,600 inhabitants, the circalunar zenithal contrast is reduced to a level of $\mathsf{\sfrac{1}{2}}$ compared to ALAN-free conditions.
- Only two of our sites, both situated in national parks (Bodinggraben and Zöblboden), show natural circalunar amplitudes.
- At our urban sites we further detect a strong seasonal signal that is linked to the amplification of anthropogenic skyglow during the winter months due to combined effects of enhanced albedo (due to snow) and a lower vegetation state (Puschnig et al. in prep).
- At urban sites we further detect a significant ($>$3 sigma) signal with a periodic time of $\sim$100 days. The cause of this roughly quaterly variation is unclear and was not reported previously.
Acknowledgements {#acknowledgements .unnumbered}
================
We dedicate this work to Thomas Posch, our most valued colleague, mentor and dearest friend, who passed away during the development phase of this manuscript. This paper would have never been written without Thomas, as it was him who first introduced Johannes and Stefan to this stunning field of research many years ago. He will be missed by us, not only because of his expertise and keen mind, but also because of his cordiality.
We acknowledge support by the provincial government of Upper Austria, in particular we thank Heribert Kaineder and Martin Waslmaier for their dedication and tireless efforts to establish a network of SQMs in Upper Austria, and share the data with the public.
We are grateful to the referee for his/her constructive input.
This research further made use of SciPy [@jones2001scipy] and NumPy [@van2011numpy], two `Python` packages that make life as a scientist easier.
![The panels show cutouts of the FFT amplitude spectra ranging from $\mathsf{0}$ to $\mathsf{0.1\,d^{-1}}$ for 23 locations in Upper Austria, sorted by increasing [<NSB>]{} values, i.e. decreasing light pollution. The amplitude spectra were obtained from a two year long dataset (2016-2017) of nightly [<NSB>]{} values. For comparison, an amplitude spectrum of a cloud-free synthetic sky model as described in section \[sec:skymodel\] is shown in the bottom right panel (SYN). The peak at zero frequency, i.e. [<NSB>]{}, is indicated by a black dotted line, the seasonal variation $\mathsf{\frac{1}{365}\ d^{-1}}$ is marked with a green dash-dotted line, the expected lunar synodic frequency of $\mathsf{\frac{1}{29.5}\ d^{-1}}$ and its first harmonic are marked with blue, dashed lines and the mean noise level is shown as horizontal dashed line.[]{data-label="fig:fft"}](OOE_fft.png){width="85.00000%"}
---------------- ------------------ -------------- -------------- ---------------- ------------------------ -- ---------------- ------------------ -------------- -------------- ---------------- ------------------------
$\mathsf{COD}$ $\mathsf{\upnu}$ $\mathsf{T}$ $\mathsf{A}$ $\mathsf{S/N}$ $\mathsf{note}$ $\mathsf{COD}$ $\mathsf{\upnu}$ $\mathsf{T}$ $\mathsf{A}$ $\mathsf{S/N}$ $\mathsf{note}$
(1) (2) (3) (4) (5) (6) (1) (2) (3) (4) (5) (6)
LGO 0.00e+00 inf 17.31 244.0 mean BRA 0.00e+00 inf 18.54 326.3 mean
LGO 2.83e-03 353.0 0.73 10.3 seasonal BRA 2.82e-03 354.0 0.67 11.8 seasonal
LGO 9.92e-03 100.9 0.24 3.4 unknown BRA 5.65e-03 177.0 0.18 3.2 seasonal (1st harm.)
LGO 3.40e-02 29.4 0.34 4.7 circalunar BRA 9.89e-03 101.1 0.19 3.3 unknown
LGO 6.09e-02 16.4 0.24 3.3 BRA 3.39e-02 29.5 0.66 11.7 circalunar
BRA 6.07e-02 16.5 0.17 3.1
LSM 0.00e+00 inf 17.51 250.6 mean BRA 6.78e-02 14.8 0.30 5.3 circalunar (1st harm.)
LSM 2.82e-03 354.0 0.77 11.1 seasonal BRA 7.06e-02 14.2 0.19 3.3
LSM 5.65e-03 177.0 0.24 3.4 seasonal (1st harm.) BRA 7.77e-02 12.9 0.19 3.3
LSM 9.89e-03 101.1 0.24 3.5 unknown BRA 1.03e-01 9.7 0.17 3.0 circalunar (2nd harm.)
LSM 3.39e-02 29.5 0.33 4.8 circalunar
LSM 6.07e-02 16.5 0.24 3.4 FRE 0.00e+00 inf 18.60 267.6 mean
FRE 2.82e-03 354.0 0.76 10.9 seasonal
WEL 0.00e+00 inf 17.70 263.2 mean FRE 3.39e-02 29.5 0.71 10.2 circalunar
WEL 2.82e-03 354.0 0.86 12.8 seasonal FRE 6.78e-02 14.8 0.27 3.9 circalunar (1st harm.)
WEL 5.65e-03 177.0 0.22 3.3 seasonal (1st harm.) FRE 7.06e-02 14.2 0.23 3.3
WEL 9.89e-03 101.1 0.21 3.1 unknown
WEL 3.39e-02 29.5 0.36 5.4 circalunar GRI 0.00e+00 inf 18.60 324.6 mean
WEL 6.07e-02 16.5 0.21 3.1 GRI 1.42e-03 706.0 0.18 3.1
GRI 2.83e-03 353.0 0.92 16.1 seasonal
LSW 0.00e+00 inf 17.73 256.2 mean GRI 5.67e-03 176.5 0.20 3.5 seasonal (1st harm.)
LSW 2.82e-03 354.0 0.58 8.4 seasonal GRI 9.92e-03 100.9 0.19 3.2 unknown
LSW 9.89e-03 101.1 0.24 3.5 unknown GRI 2.69e-02 37.2 0.17 3.0
LSW 3.39e-02 29.5 0.41 6.0 circalunar GRI 3.40e-02 29.4 0.57 10.0 circalunar
LSW 6.07e-02 16.5 0.23 3.3 GRI 6.09e-02 16.4 0.18 3.2
LSW 7.06e-02 14.2 0.21 3.1 GRI 6.80e-02 14.7 0.27 4.7 circalunar (1st harm.)
GRI 1.03e-01 9.7 0.17 3.0 circalunar (2nd harm.)
STW 0.00e+00 inf 17.93 259.4 mean
STW 2.83e-03 353.0 0.49 7.0 seasonal VOE 0.00e+00 inf 18.62 309.0 mean
STW 3.40e-02 29.4 0.45 6.5 circalunar VOE 2.82e-03 354.0 0.74 12.2 seasonal
STW 6.09e-02 16.4 0.23 3.3 VOE 9.89e-03 101.1 0.20 3.3 unknown
STW 7.08e-02 14.1 0.24 3.5 VOE 2.12e-02 47.2 0.18 3.0
VOE 3.39e-02 29.5 0.64 10.6 circalunar
TRA 0.00e+00 inf 18.09 296.7 mean VOE 6.78e-02 14.8 0.32 5.3 circalunar (1st harm.)
TRA 2.82e-03 354.0 0.55 9.1 seasonal VOE 7.06e-02 14.2 0.23 3.8
TRA 3.39e-02 29.5 0.46 7.5 circalunar VOE 7.77e-02 12.9 0.24 3.9
TRA 6.07e-02 16.5 0.20 3.3 VOE 1.03e-01 9.7 0.18 3.0 circalunar (2nd harm.)
TRA 6.78e-02 14.8 0.21 3.4 circalunar (1st harm.)
TRA 7.06e-02 14.2 0.21 3.5 MAT 0.00e+00 inf 19.05 305.3 mean
MAT 2.83e-03 353.0 0.65 10.4 seasonal
STY 0.00e+00 inf 18.27 268.6 mean MAT 9.92e-03 100.9 0.25 3.9 unknown
STY 2.82e-03 354.0 0.69 10.1 seasonal MAT 3.40e-02 29.4 0.83 13.3 circalunar
STY 9.89e-03 101.1 0.23 3.3 unknown MAT 6.52e-02 15.3 0.24 3.9
STY 3.39e-02 29.5 0.50 7.4 circalunar MAT 6.80e-02 14.7 0.32 5.1 circalunar (1st harm.)
STY 6.07e-02 16.5 0.21 3.1 MAT 7.08e-02 14.1 0.22 3.6
STY 6.78e-02 14.8 0.23 3.5 circalunar (1st harm.) MAT 7.79e-02 12.8 0.23 3.8
STY 7.06e-02 14.2 0.21 3.1 MAT 1.78e-01 5.6 0.19 3.1
STY 7.77e-02 12.9 0.24 3.6
MUN 0.00e+00 inf 19.76 421.4 mean
PAS 0.00e+00 inf 18.46 344.1 mean MUN 2.82e-03 354.0 0.56 12.0 seasonal
PAS 2.82e-03 354.0 0.57 10.7 seasonal MUN 3.39e-02 29.5 1.10 23.4 circalunar
PAS 5.65e-03 177.0 0.16 3.0 seasonal (1st harm.) MUN 3.67e-02 27.2 0.14 3.0
PAS 9.89e-03 101.1 0.17 3.2 unknown MUN 6.50e-02 15.4 0.16 3.5
PAS 3.39e-02 29.5 0.53 9.9 circalunar MUN 6.78e-02 14.8 0.45 9.6 circalunar (1st harm.)
PAS 6.07e-02 16.5 0.17 3.2 MUN 7.06e-02 14.2 0.24 5.1
PAS 6.50e-02 15.4 0.17 3.1 MUN 1.02e-01 9.8 0.19 4.1 circalunar (2nd harm.)
PAS 6.78e-02 14.8 0.24 4.5 circalunar (1st harm.)
PAS 7.06e-02 14.2 0.18 3.4
---------------- ------------------ -------------- -------------- ---------------- ------------------------ -- ---------------- ------------------ -------------- -------------- ---------------- ------------------------
---------------- ------------------ -------------- -------------- ---------------- ------------------------ -- ---------------- ------------------ -------------- -------------- ---------------- ------------------------
$\mathsf{COD}$ $\mathsf{\upnu}$ $\mathsf{T}$ $\mathsf{A}$ $\mathsf{S/N}$ $\mathsf{note}$ $\mathsf{COD}$ $\mathsf{\upnu}$ $\mathsf{T}$ $\mathsf{A}$ $\mathsf{S/N}$ $\mathsf{note}$
(1) (2) (3) (4) (5) (6) (1) (2) (3) (4) (5) (6)
GIS 0.00e+00 inf 20.07 400.5 mean FEU 0.00e+00 inf 20.66 362.6 mean
GIS 1.42e-03 706.0 0.21 4.2 FEU 3.11e-02 32.2 0.20 3.5
GIS 5.67e-03 176.5 0.18 3.6 seasonal (1st harm.) FEU 3.39e-02 29.5 1.22 21.5 circalunar
GIS 7.08e-03 141.2 0.21 4.1 FEU 3.67e-02 27.2 0.19 3.4
GIS 8.50e-03 117.7 0.23 4.7 FEU 6.78e-02 14.8 0.60 10.5 circalunar (1st harm.)
GIS 1.13e-02 88.2 0.16 3.3 FEU 7.06e-02 14.2 0.23 4.0
GIS 1.27e-02 78.4 0.18 3.7 FEU 1.02e-01 9.8 0.24 4.2 circalunar (2nd harm.)
GIS 1.42e-02 70.6 0.18 3.6 FEU 1.24e-01 8.0 0.19 3.3
GIS 1.56e-02 64.2 0.18 3.7
GIS 1.84e-02 54.3 0.17 3.3 KRI 0.00e+00 inf 20.81 428.6 mean
GIS 2.55e-02 39.2 0.18 3.6 KRI 3.12e-02 32.1 0.28 5.8
GIS 3.12e-02 32.1 0.18 3.6 KRI 3.40e-02 29.4 1.16 23.9 circalunar
GIS 3.40e-02 29.4 1.02 20.3 circalunar KRI 6.09e-02 16.4 0.16 3.2
GIS 6.80e-02 14.7 0.47 9.4 circalunar (1st harm.) KRI 6.80e-02 14.7 0.51 10.6 circalunar (1st harm.)
GIS 7.08e-02 14.1 0.30 6.0 KRI 7.08e-02 14.1 0.16 3.2
GIS 1.02e-01 9.8 0.19 3.7 circalunar (2nd harm.) KRI 9.49e-02 10.5 0.15 3.1
KRI 1.02e-01 9.8 0.19 3.9 circalunar (2nd harm.)
ULI 0.00e+00 inf 20.13 353.8 mean KRI 1.29e-01 7.8 0.16 3.3
ULI 2.83e-03 353.0 0.50 8.8 seasonal
ULI 8.50e-03 117.7 0.19 3.3 LOS 0.00e+00 inf 20.86 355.1 mean
ULI 1.13e-02 88.2 0.18 3.1 LOS 1.27e-02 78.7 0.20 3.4
ULI 1.56e-02 64.2 0.21 3.7 LOS 2.40e-02 41.6 0.21 3.5
ULI 2.12e-02 47.1 0.25 4.3 LOS 3.11e-02 32.2 0.24 4.0
ULI 2.83e-02 35.3 0.18 3.1 LOS 3.39e-02 29.5 1.30 22.0 circalunar
ULI 3.12e-02 32.1 0.29 5.2 LOS 3.67e-02 27.2 0.20 3.5
ULI 3.40e-02 29.4 1.16 20.4 circalunar LOS 6.78e-02 14.8 0.56 9.5 circalunar (1st harm.)
ULI 3.97e-02 25.2 0.21 3.8 LOS 7.06e-02 14.2 0.31 5.2
ULI 6.52e-02 15.3 0.19 3.4
ULI 6.66e-02 15.0 0.18 3.1 ZOE 0.00e+00 inf 21.01 378.2 mean
ULI 6.80e-02 14.7 0.45 7.9 circalunar (1st harm.) ZOE 1.55e-02 64.4 0.17 3.1
ULI 7.08e-02 14.1 0.20 3.5 ZOE 1.98e-02 50.6 0.25 4.5
ZOE 3.39e-02 29.5 1.55 28.0 circalunar
KID 0.00e+00 inf 20.37 506.3 mean ZOE 3.67e-02 27.2 0.39 7.1
KID 7.06e-03 141.6 0.14 3.6 ZOE 3.81e-02 26.2 0.20 3.7
KID 9.89e-03 101.1 0.14 3.5 unknown ZOE 3.95e-02 25.3 0.19 3.5
KID 1.27e-02 78.7 0.14 3.5 ZOE 6.78e-02 14.8 0.62 11.2 circalunar (1st harm.)
KID 1.55e-02 64.4 0.14 3.4 ZOE 7.06e-02 14.2 0.28 5.1
KID 1.84e-02 54.5 0.13 3.3
KID 2.54e-02 39.3 0.17 4.3 BOD 0.00e+00 inf 21.04 329.6 mean
KID 3.11e-02 32.2 0.14 3.4 BOD 5.62e-03 178.0 0.20 3.1
KID 3.39e-02 29.5 1.19 29.6 circalunar BOD 9.36e-03 106.8 0.23 3.5
KID 3.67e-02 27.2 0.20 5.0 BOD 2.81e-02 35.6 0.20 3.1
KID 6.50e-02 15.4 0.12 3.1 BOD 3.18e-02 31.4 0.23 3.6
KID 6.78e-02 14.8 0.52 13.0 circalunar (1st harm.) BOD 3.37e-02 29.7 1.51 23.6 circalunar
KID 7.06e-02 14.2 0.25 6.3 BOD 3.56e-02 28.1 0.48 7.6
KID 1.02e-01 9.8 0.20 4.9 circalunar (2nd harm.) BOD 3.93e-02 25.4 0.34 5.3
KID 1.34e-01 7.5 0.12 3.0 circalunar (3rd harm.) BOD 6.74e-02 14.8 0.56 8.7 circalunar (1st harm.)
BOD 6.93e-02 14.4 0.29 4.5
GRU 0.00e+00 inf 20.39 568.6 mean BOD 7.49e-02 13.3 0.20 3.1
GRU 2.83e-03 353.0 0.38 10.6 seasonal BOD 9.74e-02 10.3 0.23 3.6
GRU 3.12e-02 32.1 0.17 4.6 BOD 1.01e-01 9.9 0.20 3.1 circalunar (2nd harm.)
GRU 3.26e-02 30.7 0.13 3.6
GRU 3.40e-02 29.4 1.34 37.3 circalunar
GRU 3.68e-02 27.2 0.28 7.7
GRU 6.52e-02 15.3 0.19 5.3
GRU 6.80e-02 14.7 0.53 14.8 circalunar (1st harm.)
GRU 7.08e-02 14.1 0.23 6.4
GRU 1.02e-01 9.8 0.15 4.1 circalunar (2nd harm.)
GRU 1.05e-01 9.5 0.11 3.1
---------------- ------------------ -------------- -------------- ---------------- ------------------------ -- ---------------- ------------------ -------------- -------------- ---------------- ------------------------
![The *gray shaded areas* in the panels show [<NSB>]{} time series for 2016 and 2017 obtained from 23 SQM stations in Upper Austria and from a cloud-free synthetic sky model (SYN) as described in section \[sec:skymodel\]. On top of that, the result of an inverse FFT (iFFT) of identified frequency components is shown as *blue line*. The iFFT frequencies correspond to the mean brightness level, the circalunar rhythm plus its first two harmonics as well as a yearly cycle, i.e. bright winters and dark summers. This cycle is driven by an increase of overcast skies during the winter months and an amplification of light pollution by clouds. Data gaps that were set to the mean [<NSB>]{} value are marked in *red*.[]{data-label="fig:ifft"}](OOE_ifft.png){width="85.00000%"}
![The *gray shaded areas* in the panels show [<NSB>]{} time series for 2016 and 2017 obtained from 23 SQM stations in Upper Austria and from a cloud-free synthetic sky model (SYN) as described in section \[sec:skymodel\]. On top of that, the result of an inverse FFT (iFFT) of identified frequency components is shown as *blue line*. The iFFT frequencies correspond to the mean brightness level, the circalunar rhythm plus its first two harmonics as well as a yearly cycle, i.e. bright winters and dark summers. This cycle is driven by an increase of overcast skies during the winter months and an amplification of light pollution by clouds. Data gaps that were set to the mean [<NSB>]{} value are marked in *red*.[]{data-label="fig:ifft_cdm"}](OOE_ifft_cdm.png){width="85.00000%"}
[^1]: <https://www.eso.org/observing/etc/doc/skycalc/helpskycalccli.html>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have performed the first measurement of the angular power spectrum in the large-scale diffuse emission at energies from 1-50 GeV. We compared results from data and a simulated model in order to identify significant differences in anisotropy properties. We found angular power above the photon noise level in the data at multipoles greater than $\sim 100$ for energies $1 \lesssim E \lesssim 10$ GeV. The excess power in the data suggests a contribution from a point source population not present in the model.'
author:
- 'G. Gómez-Vargas, on behalf of the Fermi-LAT collaboration, and E. Komatsu\'
title: 'Measurement of anisotropies in the large-scale diffuse gamma-ray emission'
---
\[1999/12/01 v1.4c Il Nuovo Cimento\]
Introduction
============
The Fermi Gamma-Ray Telescope, launched on June 11th 2008 from Cape Canaveral, performs gamma-ray measurements over the whole celestial sphere. Its main scientific instrument the Large Area Telescope (LAT) measures the tracks of the electron and positron that result when an incident gamma-ray undergoes pair conversion in a thin, tungsten foil, and measures the energy of the subsequent electromagnetic shower that develops in the telescope’s calorimeter. Some *Fermi*-LAT specifications are: Energy range from 20MeV to $\sim$300GeV, angular resolution $\sim$0.1 deg above 10 GeV, field of view (FoV) $\sim$2.4sr, and uniform sky exposure of $\sim$30 minutes every 3 hours. Detailed descriptions of the *Fermi*-LAT telescope can be found in [@Atwood:2009ez]. One of the key science targets of the Fermi mission is diffuse gamma-ray emission. Its main component is correlated with Milky Way structures, the galactic emission, arising from interactions of high-energy cosmic rays with the interstellar medium and the interstellar radiation field. A fainter component considered to have an isotropic or nearly isotropic distribution on the sky, the so-called extragalactic emission, has been observed. This observation is based on the modelization of galactic component, detected *Fermi*-LAT sources and the solar gamma-ray emission [@Abdo:2010nz]. Also there is a contribution from populations of sources, of various kinds, including blazars, pulsars, SNR, and possibly dark matter (DM) structures, not yet detected due to *Fermi*-LAT spatial resolution and photon statistics. The angular distribution of photons in the diffuse gamma-ray background contains information about the presence and nature of these unresolved source populations (USP). Fluctuations on small scales may originate from USP if they are different from those expected from the Poisson noise due to finite statistics. Recent studies have predicted the contributions to the angular power spectrum (APS) from extragalactic and galactic DM annihilation or decay, e.g. [@Ando:2005xg; @Ando:2006cr; @SiegalGaskins:2008ge; @SiegalGaskins:2009ux; @Ando:2009fp; @Fornasa:2009qh; @Zavala:2010pw; @Taoso:2008qz]. A detailed *Fermi*-LAT sensitivity study of anisotropies from DM annihilation has been presented in [@Cuoco:2010jb]. I present the results of an anisotropy analysis of the diffuse emission measured by the *Fermi*-LAT. We calculate the angular power spectrum of the emission from $\sim 22$ months of Fermi data and of the emission from a simulated model (galactic diffuse emission, 11-month sources from Fermi catalog and isotropic emission), and compare the results from the data and model in order to identify significant differences in anisotropy properties.
The angular power spectrum (APS) as a metric for anisotropy
===========================================================
We consider the APS $C_l$ of intensity fluctuations,
$$\label{delta}
\delta I(\psi)=\frac{I(\psi)-\langle I\rangle}{\langle I\rangle},$$
where $I(\psi)$ is the intensity in the direction $\psi$. The APS is given by $C_l=\langle \vert a_{lm}\vert^ 2\rangle$, where $a_{lm}$ are determined by expanding (\[delta\]) in spherical harmonics, $\delta I(\psi)=\sum_{l,m}{a_{lm}Y_{lm}}.$
The $1-\sigma$ statistical uncertainty in the measured APS is given by $$\delta C_l = \sqrt{\frac{2}{(2l+1)\Delta l f_{sky}}}\left(C_l+\frac{C_N}{W_l^ 2} \right)$$ where $W_l=$exp$(-l^2\sigma^2_b/2)$ is the window function of a Gaussian beam of width $\sigma_b$. $f_{sky}$ is the fraction of the sky observed and $\Delta l$ multipole bins. The noise power spectrum $C_N$ is the Poisson noise, $C_N=(4\pi f_{sky})/N_{\gamma}$, where $N_{\gamma}$ is the number of photons observed. Predicted values of $C_l$ at $l=100$ of various USP cover a large range, e.g., $\sim 1\times 10^{-4}$ for blazars [@Ando:2006mt], $\sim 1\times 10^{-7}$ for starforming galaxies [@Ando:2009nk] , and $\sim 1\times 10^{-4}$ to $1$ for DM [@Ando:2005xg; @Ando:2006cr; @SiegalGaskins:2008ge; @SiegalGaskins:2009ux; @Ando:2009fp; @Fornasa:2009qh; @Zavala:2010pw; @Taoso:2008qz].
Method
======
Select regions of the sky which are relatively clean
[-]{}
mask sources in the 11-month catalog within a 2 deg radius
mask the galactic plane $\vert b\vert<30$deg
Calculate angular power spectrum of the data in several energy bins using the HEALPix package [@Gorski:2004by].
Focus on multipoles greater than 100 (angular scales $\lesssim 1-2^{\circ}$), because the contamination from Galactic diffuse is likely to be small.
Compare results from data and simulated model to identify significant differences in anisotropy properties.
Error bars on points indicate $1-\sigma$ statistical uncertainty in the measurement; systematic uncertainties are NOT included.
Data from $\sim 22$ months of diffuse class events in the energy range 1GeV to 50GeV were analyzed. We used P6\_V3 instrument response, for data and simulations. Maps have been binned into order 9 HEALPix. The simulated data are produced using the `gtobssim` routine, part of the Fermi Science Tools package. We used current background models released by the Fermi collaboration[^1] and 1-year point source catalog[^2].
Results and conclusions
========================
Plots of Fig. 1. show the APS of the data and the default model (Galactic diffuse model + 11 month source catalog + isotropic) in different energy ranges. These figures show at what energy ranges and multipole ranges the APS of the data and the model differ, as well as where each of these is consistent with the photon noise level. We have found that at multipoles greater than $\sim 100$ the excess power in the data suggest a contribution from a point source population not present in the model. Also, at large angular scales ($l<100$) angular power above the noise is seen in the data and model, probably due to contamination from the galactic diffuse. Due to decreasing photon statistics, the amplitude of anisotropies detectable by this analysis decreases with increasing energy. For this reason, the non-detection of power above the noise level at $10$-$50$ GeV does not exclude the presence of anisotropies at the level of those detected at $1$-$10$ GeV.
\[plots\] {width="45.00000%"}{width="45.00000%"}{width="45.00000%"}{width="45.00000%"}{width="45.00000%"}
The *Fermi*-LAT Collaboration acknowledges support from a number of agencies and institutes for both development and the operation of the *Fermi*-LAT as well as scientific data analysis. These include NASA and DOE in the United States, CEA/Irfu and IN2P3/CNRS in France, ASI and INFN in Italy, MEXT, KEK, and JAXA in Japan, and the K. A. Wallenberg Foundation, the Swedish Research Council and the National Space Board in Sweden. Additional support from INAF in Italy and CNES in France for science analysis during the operations phase is also gratefully acknowledged. G.G.V thanks the support of the Spanish MICINN’s Consolider-Ingenio 2010 Programme under grant MultiDark CSD2009-00064, MICINN under grant FPA2009-08958, the Community of Madrid under grant HEPHACOS S2009/ESP-1473, and the European Union under the Marie Curie-ITN program PITN-GA-2009-237920.
[0]{} Atwood W. B. [*et al.*]{} \[*Fermi*-LAT Collaboration\] Astrophys. J. [**697**]{} (2009) 1071 Abdo A. A. [*et al.*]{} \[The *Fermi*-LAT collaboration\] *Phys. Rev. Lett.* [**104**]{} (2010) 101101 Ando S. and Komatsu E. *Phys. Rev. D* [**73**]{} (2006) 023521 Ando S., Komatsu E., Narumoto T. and Totani T. *Phys. Rev. D* [**75**]{} (2007) 063519 Siegal-Gaskins J. M. *JCAP* [**0810**]{} (2008) 040 Siegal-Gaskins J. M. and Pavlidou V. *Phys. Rev. Lett.* [**102**]{} (2009) 241301 Ando S. *Phys. Rev. D* [**80**]{} (2009) 023520 Fornasa M., Pieri L., Bertone G. and Branchini E. *Phys. Rev. D* [**80**]{} (2009) 023518 Zavala J., Springel V. and Boylan-Kolchin M., arXiv:1001.3307 \[astro-ph.CO\].
Taoso M., Ando S., Bertone G. and Profumo S., *Phys. Rev. D* [**79**]{} (2009) 043521
Cuoco A., Sellerholm A., Conrad J. and Hannestad S. \[arXiv:1005.0843\].
Ando S., Komatsu E., Narumoto T. and Totani T. *Mon. Not. Roy. Astron. Soc.* [**376**]{} (2007) 1635 Ando S. and Pavlidou V. *Mon. Not. Roy. Astron. Soc.* [**400**]{} (2009) 2122
Gorski K. M., Hivon E., Banday A. J., Wandelt B. D., Hansen F. K., Reinecke M. and Bartelman M. *Astrophys. J.* [**622**]{} (2005) 759
[^1]: http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html
[^2]: http://fermi.gsfc.nasa.gov/ssc/data/access/lat/1yr\_catalog/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Ge hut clusters, which appear during heteroepitaxy of Ge on Si(001), are prototypical examples of islands formed through strain. Experimentally, complete facets are observed to form rapidly, though the mechanism is unknown. We model the growth of new faces on Ge hut clusters, using linear scaling DFT. We build realistic small huts on substrates, and show that the growth of {105} facets proceeds from top-to-bottom, even for these small huts. The growth of the facet is driven by the reconstruction on the {105} facet, and nucleates at the boundaries of the facet. We thus resolve any ambiguities in kinetic models of hut cluster growth.'
author:
- Sergiu Arapan
- 'David R. Bowler'
- Tsuyoshi Miyazaki
title: 'A linear scaling DFT study of the growth of a new {105} facet layer on a Ge hut cluster'
---
Heteroepitaxy and strained growth have a long history, and are emerging as important techniques to improve the performance of microelectronic devices. In particular, the formation and stability of islands has been studied as a potential source of quantum dots, while nanowires are being pursued as potential channels for field-effect transistors [@Aqua2013]. Germanium on silicon is a prototypical system, following a well-known progression from two dimensional growth, through the emergence of hut clusters and their transition into domes, but the structure and growth of these islands is poorly understood [@Aqua2013; @Mo1990; @Voigtlander2001]. Substantial efforts have been made to achieve a narrow distribution of these 3D islands by using kinetically self-limiting growth [@Jesson1998; @Kastner1999; @McKay2008]. By assuming that islands grow by adding new facet layers, various kinetic models provide a fair description of experimental observations but cannot describe the actual mechanism of the facet growth. Direct STM imaging of 3D islands lacks any direct evidence for the facet nucleation sites and also cannot reveal the microscopic picture of the growth process.
However, an atomistic description of the Ge(105) surface may be provided by simulations. The first models for the Ge(105) surface reconstruction, as seen on hut cluster facets [@Mo1990], were given by density functional theory (DFT) calculations [@Fujikawa2002; @Raiteri2002]. These showed that the basic unit of the reconstruction which is responsible for the stability of the surface was a U-shaped set of six atoms (known as Uss, and discussed further below). Experimental and forcefield modelling [@Montalenti2004] of the transition from the hut cluster to Ge domes showed that facets appear to nucleate at the apex, and that accumulation of partial facets near the top of large huts can drive the transition. DFT modelling of the growth of the Ge(105) surface [@Cereda2007] showed how new layers can nucleate, and drew on step-flow arguments to suggest reasons for the growth of facets from top to bottom. Realistic modelling of the growth of facets, however, is challenging. The reconstruction involves charge transfer between buckled Ge dimers [@Fujikawa2002], which requires a quantum mechanical method for accurate modelling. Moreover, just as on the Si(001) surface, there are at least two ways to arrange the buckled dimers on the Ge(105) surface [@Miyazaki2007]. Conventional DFT does not allow simulations of realistic models of hut clusters (which will necessarily contain more than 10,000 atoms), while classical force-fields cannot account for quantum mechanical effects. We have shown that linear scaling DFT methods [@Bowler:2012zt] can be applied to model systems of over 2,000,000 atoms [@Bowler:2010uq]. We have applied these methods to models of small hut clusters (up to 23,000 atoms) [@Miyazaki:2008tt], showing that the transition from two-dimensional to three-dimensional growth is due to thermodynamic stability.
![(colour online) $a)$ A $16\times 26$ Ge hut (light- and dark-grey spheres) on a $24\times 36$ Si substrate (blue spheres) with two Ge wetting layers (grey spheres). Dark-grey spheres show atom positions of the outermost layer of small {105} facets. b) Top view of the left half of the left {105} facet of the hut shown above. Dark-grey spheres show positions of the {105} facet layer atoms (large spheres) as well as atoms of the (105) surface (smaller spheres) that form the Uss (marked with a grey box). c) Top view of the left half of the left {105} facet of the $16\times 28$ hut with additional facet layers (A-dimers and B-dimers are shown as yellow and orange spheres, respectively) and a close up view of a Uss. Atoms at positions close to that of atoms of the Uss of the previous facet (Fig. \[fig1\]b) are shown as dark-grey spheres.[]{data-label="fig1"}](Fig1 "fig:"){width="8cm"}\
In this work, we use linear scaling DFT methods to model the stability of dimers and groups of dimers, and the formation of a new layer, on the facet of a realistic Ge hut cluster. We use the approach we have developed to model hut cluster structure[@Miyazaki:2008tt] to generate the starting hut cluster. We notate the size of the substrate and the hut in terms of multiples of the cubic silicon (or germanium) bulk unit cell, as $L_{1}\times L_{2}$. The starting system, shown in Fig. \[fig1\]a, consists of an elongated $16\times 26$ Ge hut on a $24\times 36$ Si substrate, with two layers of Ge as a wetting layer. Both top and bottom of the substrate is reconstructed in the $p(2\times 2)$ reconstruction of Si(001) and Ge(001). The system consists of 2,457 Ge atoms in the hut, 3,456 Ge atoms in the wetting layer and 13,824 Si atoms in the substrate, giving a total of 19,737 atoms. To create the positions for dimers in the new facet, we also built a $16\times 28$ hut cluster, adding new facets on both ends of the hut (in this case with 19,973 atoms). The size of the simulation cell was $24\times 36 \times 7.5$, giving a vacuum gap of $\sim 20$Å. The simulations used linear scaling DFT[@RevModPhys.71.1085; @Bowler:2012zt] as implemented in the <span style="font-variant:small-caps;">Conquest</span> code[@Bowler:2010uq; @Miyazaki:2004ee; @Bowler:2002pt], in the non-self-consistent Harris-Foulkes regime. We used pseudo-atomic orbitals[@Torralba:2008wm] with single zeta basis functions in the local density approximation, which has been shown to model the Ge(105) surface and hut cluster facets well[@Miyazaki:2008tt; @Miyazaki2007; @Arita2014]. The charge density grid used was equivalent to a 60Ha energy cutoff, and structural relaxation was performed via quenched molecular dynamics by using the FIRE method [@Bitzek2006] with some modifications to account for slow convergence in the case of large systems [@Arita2014]. The structure of the {105} surface, particularly on the facet of a hut cluster, is based around arrangements of buckled Ge dimers. We show the top two layers of atoms for part of a facet on our hut cluster model in Fig. \[fig1\]b. The key building block of the surface is known as the Uss (U-shaped structure) and consists of three Ge dimers: an example is highlighted in a shaded rectangle in Fig. \[fig1\]b, and is enlarged in Fig. \[fig1\]c (where the three constituent dimers are the two labelled A and B and the dimer parallel to A). We consider growth of the facet in terms of dimers, both because the Uss is built from dimers, and because the dimer is a very stable unit (ad-atoms diffuse quickly over the surface[@PhysRevB.70.245315], and form dimers[@Cereda2007]).
The growth of a new facet will involve deposition of new dimers on the surface, continuing the {105} reconstruction. Fig. \[fig1\]c shows the structure of a complete new facet, with the added dimers shown as yellow and orange spheres. As can be seen in the shaded rectangle, two new dimers combine with dimers on the surface to create a new layer of Uss features, though it is important to note that this is altered at the edges of the facet (we return to this point later). When considering the new dimers individually, however, it becomes clear that there are two types: A (coloured yellow); and B (coloured orange). The A dimers are more stable than the B dimers on the perfect facet, as they can form relaxed bonds to the existing substrate atoms. To form fully relaxed bonds, a B dimer must bond to an A dimer—that is, a dimer in the *new* facet—as can be seen in the magnified view. While B dimers can bond to the existing surface atoms, this leads to strained bonds and distorted bond angles. The stable dimers formed from Ge adatoms during the growth of the perfect (105) surface[@Cereda2007] correspond to A dimers.
![(colour online) $a$) An energy map of single Ge ad-dimers: each pair of spheres shows the location of a single ad-dimer whose colour represents the ad-dimer energy. Small black spheres show the location of the third dimer that makes the Uss (this helps to identify B-dimers at the head of the Uss). Labels denote pairs of dimers on the same terrace. $b$) Local atomic structure (light green and dark grey spheres) of selected ad-dimers (yellow spheres). Dark grey spheres show optimised positions of atoms that form Uss of the underlying (105) surface. Bonds that exceed 2.8Å are shown as thin tubes. []{data-label="fig2"}](Fig2 "fig:"){width="8cm"}\
To understand the growth of a new facet on a realistic model of a hut cluster, we break down the problem: first, we will consider individual dimers on the facet; then we will examine interactions between pairs of dimers; finally, we will extend our studies to three dimers and then complete rows. We show the energy for single dimers added to the small facet of the hut, relative to the energy of a dimer on the wetting layer, in Fig. \[fig2\]$a$ (studies of growth on the long facet will be presented in a future publication). In the centre of the facet, it is clear that A-type dimers are more stable than B-type dimers, though the overall energy is slightly higher than on the wetting layer. The structures of two typical dimers are shown in Fig. \[fig2\]$b$, where the stretched bond for the B-type dimer is shown as a thin bond. In general, an A-type dimer inserts into the bonds of a single Uss on the surface, allowing it to make strong, relaxed bonds, while a B-type dimer bridges between two Uss, causing distorted, less stable bonds.
At the edges of the facet, however, the stability of dimers is very dependent on the local environment. At the base of the facet, the first row of B-type dimers are very stable, due to the interface with the wetting layer. There are several B-type dimers near the edge of the facet which are also stable, in particular B(12), which is illustrated in Fig. \[fig2\]$b$. The stability of this dimer comes from the structure of the edge between facets, where there are broken bonds which can be passivated by the addition of the B-type dimer. There are also A-type dimers which are significantly less stable than most, again due to changes of bonding at the edges between facets.
![(colour online) $a)$ Adsorption energy of single dimers added to either a clean facet (), or an existing dimer on the facet to make a dimer pair ($\blacksquare$). The values relative to the adsorption energy of an ad-dimer on the wetting layer are shown. $b)$ Adsorption energy for a single dimer added to an existing dimer pair on the facet to make a 3-dimer structure ($\blacktriangle / \blacktriangledown$). $c)$ Local environment of a pair of ad-dimers (labeled 8, yellow spheres), and two 3-dimer structures, which have an ad-dimer (orange spheres) added upward/downward ($\blacktriangle / \blacktriangledown$) to that pair.[]{data-label="fig3"}](Fig3 "fig:"){width="8cm"}\
We plot the energy for individual dimers in Fig. \[fig3\]$a$. There are several dimers within 0.2eV of the wetting layer energy, all A-type dimers associated with the edge or top of the facet (1, 8, 12, 20; we discount for now the dimers at the base, but will discuss them later). These sites are candidates for the formation of the nucleus for a new facet. Growth of the facet will proceed by addition of extra dimers to the stable nucleus, and we find that pairing one A-type and one B-type dimer gives a stable unit. An example of a stable pair is shown in Fig. \[fig3\]$c$ on the left, notated as $\blacksquare$(8), while the energy of all pairs in Fig. \[fig2\]$a$ is also plotted in Fig. \[fig3\]$a$, using the same numbering as before. Comparing the structures of dimers 8A and 8B in Fig. \[fig2\]$a$ with the dimer pair 8 in Fig. \[fig3\]$c$ shows the stabilisation mechanism: the B-type dimer passivates dangling bonds caused by the adsorption of the A-type dimer (and, equivalently, the A-type dimer reduces strain in the substrate atoms, releasing the strain in the B-type dimer bonds to the substrate). The energy of all pairs, except one, is lower than a corresponding pair on the wetting layer, giving a strong thermodynamic driving force for the growth of the hut. While the total energy of some pairs is significantly lower than others, the location of the nucleation will depend on the *first* dimer to adsorb.
Once a pair of dimers has formed, growth will proceed by adding a third dimer to the nucleus. There are two ways to add another dimer, illustrated in Fig. \[fig3\]$c$: a B-type dimer can be added *above* the dimer pair ($\blacktriangle$); or an A-type dimer can be added *below* the dimer pair ($\blacktriangledown$). Near the edges of the facet, however, there are restrictions: at the base, with the interface to the wetting layer, only B-type dimers can be added above; at the edges, with the interface to other facets, only A-type dimers can be added below. The change in energy for adding a dimer in both ways, at all sites on the facet, is plotted in Fig. \[fig3\]$b$. It is striking that, for almost all sites, adding an A-type dimer below the pair is energetically favourable, while adding a B-type dimer above the pair is unfavourable, revealing a strong thermodynamic driving force for growth of a new facet from the top of the hut downwards. In particular, adding new dimers above a pair at the base of the hut is energetically costly, while adding new dimers below existing pairs near the top of the facet is favourable, and more stable than adsorbing individual dimers on other parts of the facet. The different behaviours of the dimers above and below the dimer pair is largely due to the difference in stability and bonding for A-type and B-type dimers. However, the A-type dimer below the dimer pair also stabilises the dimer pair by completing the local structure of the Uss, while the B-type dimer above the dimer pair is bridging across a large gap caused by rebonding of the substrate atoms in the Uss above the dimer pair. It is, in effect, forming part of two Uss features, both of which are incomplete and hence less stable than wetting layer dimers. While the actual pathway for growth of a new facet of a hut cluster is likely to be highly complex, and to vary from facet to facet, we can say from these calculations that top-to-bottom growth is strongly favoured. The adsorption sites for individual dimers, coupled with the difference in stability of dimers added above and below A-B dimer pairs, drive this energetic favourability.
![(colour online) $a)$ Completion of {105} facet layers from base($\blacktriangle$)/apex($\blacktriangledown$) by adding rows of dimers one by one (consecutive rows are coloured in yellow and orange). $b)$ Energies (per ad-dimers) of various partial facet configurations ($\blacktriangle/\vartriangle$ - facet ending with a filled/half-filled terrace) with respect to the energy of target hut. $c)$ Changes in energy/dimer for two consecutive partial facet configurations ($\blacktriangle/\vartriangle$ - filling/half-filling a terrace).[]{data-label="fig4"}](Fig4 "fig:"){width="8cm"}\
We test a somewhat simplified model of hut growth to explore further the energetics, completing the new facet row-by-row starting from either the base or the top of the hut, as illustrated in Fig. \[fig4\]$a$. We study the energetics in two ways: first, by evaluating the energy per dimer of the partial facet (relative to the energy of a dimer on the wetting layer, as before), plotted in Fig. \[fig4\]$b$; second, by evaluating the *change* in energy per dimer from row to row, plotted in Fig. \[fig4\]$c$. Each row consists entirely of either A-type or B-type dimers, as can be seen in Fig. \[fig2\]$a$, and we consider two rows, one of A-type dimers and another of B-type dimers, to form a terrace.
Considering first the energy per dimer of the facet (Fig. \[fig4\]$b$), the growth from the top and the base show completely different behaviours. The initial nucleation of the base is costly, while it is much smaller at the apex (particularly as this will involve fewer dimers). Adding successive rows from the apex down gives a gradual and almost monotonic decrease in the energy, while the growth from the base oscillates, showing the high energetic cost for upward growth from dimer pairs seen in Fig. \[fig3\]. This adds weight to the top-down growth model for the facet, as it will nucleate easily, and present a consistent, thermodynamically favourable energy surface for dimers on the facet.
This is further underlined when considering the change in energy from row to row (Fig. \[fig4\]$c$). The growth from the base up will face a significant if not insuperable energetic barrier for each complete terrace (shown as open symbols). By contrast, the growth from the apex downwards lowers the energy relative to the wetting layer with each row, whether starting or completing a terrace (with the possibly exception of two points where the energy change is essentially zero). This clearly indicates that the downward growth of the facet is strongly favoured thermodynamically. Of course, we have not provided a complete growth model, and it is not possible to do so. However, this partial model clearly emphasises the key physics in the growth of new facets, and explains the experimentally observed trend for top-down growth. In conclusion, we have used linear scaling DFT calculations of complete Ge hut clusters to explain why their facets grow from the top downwards. Two key factors determine this growth direction. First, the energetics of the underlying reconstruction, which favour the completion of single Uss structures on the surface over partial completion of pairs of Uss structures. Second, the stability of dimers at the boundaries between facets and at the peak of the facet, which give stable nucleation sites for the new layer. These insights require a number of capabilities only possible with linear scaling DFT: a realistic model of the complete hut; detailed structure of the facets and the edges between facets; and quantum mechanical modelling, which is crucial in determining the relative stabilities of different structures and reconstructions on semiconductor surfaces. Given recent developments showing that highly scalable linear scaling molecular dynamics is possible[@Arita2014a], we expect that applications of this technique will further insight into complex, important scientific problems.
We acknowledge Prof M. J. Gillan for helpful discussions and suggestions regarding calculations and interpretations of results. This work is partly supported by KAKENHI projects by MEXT (No. 22104005) and JSPS (No. 26610120 and 26246021), Japan. The support from the Strategic Programs for Innovative Research (SPIRE) and the Computational Materials Science Initiative (CMSI) is also acknowledged. Calculations were performed using K computer at RIKEN Advanced Institute for Computational Science(AICS), Kobe, Japan, and the Numerical Materials Simulator at NIMS, Tsukuba, Japan.
[22]{} 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 discuss the evidence for a period-luminosity (PL) relation and period-luminosity-colour (PLC) at maximum light for Mira variables. We confirm the existence of such relations in the J, H and K bands, and also based on bolometric magnitudes, for oxygen-rich (O) and carbon-rich (C) Miras in the LMC. We demonstrate that in the J and H bands the maximum light PL relations have a significantly smaller dispersion than their counterparts at mean light, while the K band and bolometric PL relations have a dispersion comparable to that at mean light. In the J, H and K bands the fitted PL relations for the O Miras are found to have smaller dispersion than those for the C Miras, at both mean and maximum light, while the converse is true for the relations based on bolometric magnitudes. The reduction in dispersion between mean and maximum light is generally found to be more significant in the J, H and bolometric bands than in the K band – in which the mean PL relations display the smallest dispersion. The inclusion of a non-zero log period term in the regression fits is found to be highly significant in all cases except that of the C Miras in the J band, for which the data are found to be consistent with having constant absolute magnitude. Moreover, if we consider only C Miras with periods in excess of 250 days, then the data are also substantially more consistent with constant absolute magnitude at mean and maximum light in the H, K and bolometric wavebands. This suggests the possibility of employing C Miras as standard candles.
We suggest both a theoretical justification for the existence of Mira PL relations at maximum light and a possible explanation of why these relations should have a smaller dispersion than at mean light. The existence of such maximum light relations offers the possibility of extending the range and improving the accuracy of the Mira distance scale to Galactic globular clusters and to other galaxies.
author:
- |
S. M. Kanbur$^{1}$, M. A. Hendry$^{1,2}$ and D. Clarke$^{1}$\
$^1$Department of Physics and Astronomy, University of Glasgow, Glasgow G12 8QQ\
$^2$Astronomy Centre, University of Sussex, Falmer, Brighton, East Sussex BN1 9QH\
date: 'Accepted ——. Received ——; in original form '
title: 'Period-luminosity and period-luminosity-colour relations for Mira variables at maximum light'
---
Miras – period luminosity – Galactic structure – distance scale
Introduction
============
Miras are long period variable stars lying on the asymptotic giant branch (AGB) of the HR diagram with periods in the range 100 to 700 days. Their masses lie between 0.5$M_{\odot}$ and 3$M_{\odot}$, and their K band mean absolute magnitudes lie in the range $-5 < M_K < -7$ (c.f. Wood 1995). Whitelock (1995) reviews a number of reasons for the astrophysical importance of Mira variables, highlighting in particular their suitability as distance indicators – a fact which makes them useful tracers of galactic structure. The use of Miras as distance estimators relies upon the existence of period-luminosity (PL) and period-luminosity-colour (PLC) relations at mean light, which may be calibrated with nearby stars whose distance is otherwise known and then applied to more remote objects to estimate their distance. In e.g. Feast et al. (1989, hereafter F89) PL and PLC relations were derived for a calibrating sample of about 50 oxygen-rich (O) and carbon-rich (C) Miras in the Large Magellanic Cloud (LMC), using time-averaged mean J, H, K and bolometric magnitudes. F89 found that the O Miras displayed a well-defined relations in the K band, and also based on J, H and bolometric magnitudes, but with a larger dispersion in these latter three cases. For the C Miras F89 confirmed the existence of a PL relation in the K band. These derived relations were then applied to determine distance moduli to a number of galactic globular clusters.
In this paper, using the same calibrating sample of LMC Miras as in F89 we investigate the evidence for PL and PLC relations at maximum light. The primary motivation for this work comes from Kanbur & Hendry (1996, hereafter KH), who derived V band PL and PLC relations at maximum light for a sample of Cepheids in the LMC, previously published in Martin, Warren & Feast (1979). KH outlined specific physical reasons why the use of Cepheid maximum light relations might be preferable to those at mean light, developing the earlier theoretical work of Simon, Kanbur & Mihalas (1993). In a similar manner, in this paper we derive maximum light PL and PLC relations for Mira variables and discuss a possible physical justification for their existence. In particular we consider a physical explanation for the smaller observed dispersion of maximum light relations when compared with the corresponding relations at mean light. For completeness we also consider PL and PLC relations at minimum light and compare them with their counterparts and mean and maximum light.
A number of authors (c.f. Sandage 1958, Madore & Freedman 1991, KH) have discussed the theoretical justification for the existence of a PL and PLC relation for Cepheid variables, deriving the so-called [*pulsation equation*]{}, $$\log P + {1\over 2}\log {\cal{M}} - {3\over 4}\log L + 3 \log T_e =
\log {\cal{Q}}
\label{eq:eq1}$$ where $P$, ${\cal{M}}$, $L$, and $T_e$ are the period, total mass, equilibrium luminosity and effective temperature respectively of the star and ${\cal{Q}}$ is a slowly varying function of stellar parameters. Cepheids occupy an instability strip of finite width in the HR diagram. If a similar situation holds for Miras (Feast 1989, Wood 1990, Shibahashi 1993), then this equation can be used to explain the existence of PL and PLC relations for both Miras and Cepheids since it assumes only the period–mean density theorem and the Stefan–Boltzmann law. Mira and Cepheid PL relations arise from the collapse of equation (1) over the variables $\log {\cal{M}}$ and $\log T_e$. In the case of Miras, however, the equilibrium luminosity is a strong function of the core mass (Shibahashi 1993). Assuming that the equilibrium luminosity is close to the mean luminosity over a pulsational cycle, both the range of core masses and total masses therefore contribute to the scatter in a PL relation at mean light for a Mira of a given period. We discuss the effect of metallicity on the scatter of the PL relation in section 4 below.
If it is the case that $$R_{\rm{eq}} \approx R_{\rm{max}}
\label{eq:eq2}$$ that is, the equilibrium photospheric radius of the star is approximately equal to the photospheric radius at [*maximum*]{} light, then, following essentially the same reasoning as in KH, we can use the period–mean density theorem and the Stefan–Boltzmann law to write $$\log P + {1\over2} \log {\cal{M}} -
{3\over 4}\log L_{\rm{max}} + 3 \log T_{\rm{max}} = \log {\cal {Q}}
\label{eq:eq3}$$ where $L_{\rm{max}}$ and $T_{\rm{max}}$ denote the luminosity and temperature at maximum light. In the case of Cepheids, Cox (1974) provides good evidence that equation (2) is a reasonable assumption. Support for the validity of equation (2) in the case of Mira variables is given in Wood (1995) and references therein. Assuming equation (2) to be valid for Miras, equation (3) can then be used to justify theoretically the existence of PL and PLC relations at [*maximum*]{} light for these stars, as a result of collapsing equation (3) over the variables $\log {\cal{M}}$ and $\log T_{\rm{max}}$.
In the Cepheid case, Simon, Kanbur & Mihalas (1993) showed that at maximum light the range of photospheric temperatures is significantly smaller than the range of effective temperatures at mean light. Motivated by this work KH suggested that Cepheid PL and PLC relations at maximum light could have significantly smaller dispersion than at mean light – a result which was investigated in detail in KH. A similar effect may not be present for Mira variables, but another advantage of the use of maximum light is nonetheless apparent for Miras. Their pulsations, like Cepheids, are envelope phenomena – energy modulation and amplitude limitation occurring in the outer envelope. The maximum luminosity depends on the envelope mass as well as the core mass. In equation (2), on the other hand, the equilibrium luminosity is strongly dependent on the core mass (Shibahashi 1993). Thus the quantities $P$, and ${\cal{M}}$ in equation (1) have dependencies on both the envelope and core mass, whereas $L$ is dependent strongly on the core mass. All the quantities in equation (3) have dependencies on the core and enevlope mass. We conjecture that, even if the range of $T_{\rm max}$ were no different to the range of $T_{\rm eff}$, this situation could lead to Mira PL and PLC relations at maximum light to have smaller dispersion than rheir counterparts at mean light. Further work is needed to examine this proposition.
This paper is organised as follows. Section 2 describes the LMC calibrating data, and the methods used to derive PL and PLC relations and evaluate their statistical significance. In sections 3 and 4 we present our results for PL and PLC relations respectively, which are then discussed further and compared with those of F89 in section 5, highlighting some important consequences for the use of Miras as probes of galactic structure. Finally, in section 6 we present our conclusions and possibilities for further study.
Data
====
The data used in this study were taken from Glass et al (1990), which was also the primary reference for the analysis of F89. These data consisted of multi-epoch observations at a number of wavelengths of a large sample of O and C Miras in the LMC. O Miras are oxygen-rich objects whilst C Miras are carbon-rich; the classification of each star as a C or O Mira can be made from spectral type (if known) or from colour measurements and we adopt the same classifications as those published in F89. In addition we adopt the Mira periods as given in Glass et al (1990); since the average number of epochs of observation for each star in the Glass et al. study was more than eleven, with good phase coverage, it is unlikely that the published periods are subject to any significant uncertainty.
The data for the O and C Miras are summarised in Tables 1 and 2 respectively. These Tables list the star name, its period, taken from Glass et al (1990), and the mean, maximum and minimum magnitudes at J, H, K and bolometric wavebands. All magnitudes in Tables 1 and 2 have been corrected for extinction following F89, assuming $A_J=0.06$, $A_H=0.03$ and $A_K=0.02$. We adopted as the maximum and minimum magnitude simply the maximum and minimum [*observed*]{} value (or interpolated value in the case of bolometric magnitude) reported in Glass et al (1990). Mean magnitudes were calculated as the average of the maximum and minimum observed (interpolated) magnitudes, which was also the definition of mean magnitude adopted for the relations derived in F89. Multi-epoch observations were available in Glass et al. (1990) for 48 of the 49 Miras studied in F89; in the case of the star ‘GR13’ only the mean magnitudes published in F89 were available.
[lrrrrrrrrrrrrr]{} &Period& & & &\
Star&(days)&mean&max&min&mean&max&min&mean&max&min&mean&max&min\
w132 &155&12.66&12.40&12.91&11.85&11.52&12.19&11.65&11.37&11.92&14.48&14.20&14.75\
w151 &172&12.84&12.46&13.17&11.99&11.65&12.33&11.72&11.37&12.07&14.67&14.31&14.98\
w148 &185&12.96&12.60&13.35&12.12&11.74&12.50&11.80&11.38&12.22&14.74&14.41&15.16\
w158 &185&12.90&12.61&13.11&11.97&11.76&12.29&11.74&11.51&11.98&14.75&14.42&14.90\
w19 &189&12.60&12.39&12.82&11.73&11.52&11.93&11.49&11.32&11.66&14.33&14.19&14.62\
w77 &217&12.38&12.19&12.66&11.51&11.35&11.66&11.23&11.09&11.36&14.20&14.00&14.40\
w94 &220&12.44&12.29&12.59&11.55&11.42&11.67&11.27&11.15&11.39&14.23&14.09&14.37\
w74 &227&12.74&12.50&13.00&11.79&11.54&12.04&11.48&11.24&11.71&14.53&14.27&14.76\
w1 &233&12.80&12.34&13.54&11.90&11.50&12.79&11.62&11.12&12.13&14.55&14.16&15.39\
w140 &244&12.38&11.94&12.81&11.39&10.99&11.78&11.17&10.71&11.63&14.17&13.71&14.54\
w48 &279&12.12&11.85&12.39&11.25&10.95&11.53&10.97&10.68&11.26&13.97&13.64&14.19\
517-6551 &117&13.32&13.03&13.60&12.53&12.29&12.76&12.23&12.02&12.44&15.14&14.86&15.42\
512-6559 &141&13.31&12.83&13.79&12.49&12.04&12.94&12.11&11.70&12.51&15.05&14.66&15.61\
526-6754 &160&12.80&12.46&13.14&11.95&11.64&12.31&11.77&11.43&12.10&14.64&14.27&14.94\
528-6531 &195&12.46&12.18&12.73&11.62&11.37&11.95&11.45&11.15&11.76&14.31&13.99&14.54\
507-6639 &208&12.72&12.37&13.06&11.87&11.52&12.22&11.54&11.20&11.89&14.51&14.19&14.88\
533-6807 &247&12.51&12.02&13.00&11.73&11.24&12.20&11.36&10.93&11.79&14.28&13.85&14.84\
524-6543 &312&11.88&11.46&12.31&11.05&10.58&11.50&10.69&10.29&11.09&13.71&13.27&14.15\
505-6657 &311&11.92&11.53&12.31&11.07&10.72&11.41&10.65&10.29&11.01&13.73&13.36&14.11\
w126 &323&12.22&11.67&12.76&11.31&10.76&11.86&10.87&10.42&11.31&13.91&13.46&14.54\
c38 &128&13.08&12.73&13.47&12.37&12.08&12.65&12.10&11.78&12.42&14.87&14.58&15.28\
c11 &202&12.69&12.24&13.14&11.84&11.40&12.27&11.49&11.15&11.82&14.52&14.05&14.95\
c20 &210&12.88&12.37&13.40&12.01&11.49&12.31&11.52&11.08&11.96&14.56&14.17&15.18\
r120 &217&12.50&12.15&12.85&11.61&11.28&11.97&11.36&11.06&11.67&14.27&13.95&14.61\
r141 &255&12.13&11.73&12.53&11.27&10.84&11.65&10.97&10.63&11.31&13.87&13.54&14.36\
r110 &261&12.64&11.95&13.32&11.78&11.10&12.45&11.27&10.72&11.81&14.23&13.77&15.10\
r105 &420&11.58&11.22&11.93&10.64&10.29&10.98&10.27& 9.95&10.59&13.40&13.01&13.66\
537-6607 &284&12.26&11.85&12.68&11.39&10.95&11.81&11.00&10.60&11.39&13.99&13.64&14.49\
gr13 &202&12.68& & &11.83& & &11.57& & &14.50& &\
[lrrrrrrrrrrrrr]{} &Period& & & &\
Star&(days)&mean&max&min&mean&max&min&mean&max&min&mean&max&min\
w220 &286&12.70&12.06&13.33&11.47&11.03&11.92&10.81&10.57&11.04&14.12&13.77&14.48\
w46 &286&12.63&12.38&12.88&11.52&11.32&11.72&10.98&10.78&11.00&14.20&14.05&14.39\
w103 &351&13.64&12.95&14.70&12.00&11.47&12.53&10.76&10.35&11.17&14.14&13.81&14.52\
w30 &400&12.28&12.10&12.47&11.07&10.92&11.23&10.46&10.34&10.57&13.73&13.66&13.93\
530-6437 &157&13.06&12.42&13.70&12.39&11.77&13.00&12.06&11.52&12.60&14.71&14.25&15.57\
515-6617 &211&13.20&12.48&13.91&12.04&11.42&12.66&11.14&10.74&11.53&14.56&14.09&15.01\
528-6520 &231&12.65&12.06&13.24&11.63&11.19&12.06&11.06&10.82&11.29&14.29&13.87&14.70\
529-6759 &274&12.64&12.24&13.09&11.53&11.20&11.86&10.89&10.70&11.09&14.25&13.94&14.49\
515-6451 &284&12.90&12.41&13.43&11.67&11.31&12.04&10.79&10.54&11.03&14.21&13.95&14.49\
514-6605 &305&12.70&12.34&13.06&11.43&11.18&11.69&10.62&10.46&10.78&14.02&13.87&14.24\
502-6711 &308&12.58&11.91&13.25&11.35&10.90&11.81&10.51&10.26&10.76&13.92&13.58&14.22\
534-6531 &312&13.58&12.69&14.46&12.15&11.47&12.84&10.96&10.43&11.48&14.28&13.90&14.85\
529-6739 &319&12.92&12.29&13.46&11.59&11.10&12.07&10.58&10.25&10.91&13.91&13.69&14.37\
541-6631 &328&13.16&12.38&14.73&11.91&11.13&12.72&10.48&10.12&11.27&14.09&13.59&14.55\
515-6438 &365&13.06&12.45&13.98&11.88&11.15&12.56&10.88&10.38&11.38&14.27&13.80&14.91\
537-6740 &418&12.52&12.23&12.80&11.25&11.06&11.45&10.45&10.33&10.57&13.90&13.72&14.02\
c7 &326&12.58&12.02&12.98&11.36&11.01&11.66&10.67&10.50&10.83&13.99&13.75&14.27\
r153 &370&13.10&12.39&13.82&11.58&11.10&12.06&10.50&10.23&10.81&14.03&13.67&14.22\
519-6454 &242&12.79&12.31&13.27&11.67&11.31&12.01&11.07&10.84&11.29&14.33&14.03&14.70\
520-6528 &234&12.46&12.36&12.56&11.57&11.35&11.73&11.26&11.16&11.36&14.17&14.02&14.43\
Using these data we carried out linear regression fits to PL relations of the form $$m = a + b \log P
\label{eq:eq4}$$ $$m_{\rm{max}} = a + b \log P
\label{eq:eq5}$$ and $$m_{\rm{min}} = a + b \log P
\label{eq:eq6}$$ where $m$ denotes apparent J, H, K and bolometric magnitude, corrected for extinction, as appropriate and $a$ and $b$ are constants.
As will be clear from Figures 1 to 6 in Section 4, in most cases the existence of a tightly correlated PL relation at mean and maximum light was immediately evident from a scatterplot of apparent magnitude against log period. Notwithstanding this, we considered it instructive – particularly for the more marginal PL relations – to determine quantitatively the statistical significance of including a log period term in each of our regression fits. In order to do this we applied the same statistical test which was introduced in KH, and which is described in detail in the appendix of that paper, involving the partial multiple correlation coefficient, $\rho$, of the regression (c.f. Graybill 1976). If $\rho$ equals zero then the log period term makes no contribution to a reduction in the dispersion of the fit and is effectively redundant. For each regression we computed the sample value of $\rho$, denoted by $\hat{\rho}$. Under the null hypothesis that the true value of $\rho$ is equal to zero then $\hat{\rho}^2$ has an F distribution (c.f. KH).
We also carried out fits to PLC relations of the form $$m = a + b \log P + c (J-K)
\label{eq:eq7}$$ $$m_{\rm{max}} = a + b \log P + c (J-K)_{\rm{max}}
\label{eq:eq8}$$ and $$m_{\rm{min}} = a + b \log P + (J-K)_{\rm{min}}
\label{eq:eq9}$$ where $(J-K)$ denotes dereddened colour, and also to the corresponding equations for $(J-H)$ colour. We defined maximum and minimum colour as the colour at the phase at which respectively the maximum and minimum magnitude was observed. We applied the same statistical test as for the PL relations to determine the significance of adding the colour term in each PLC relation.
We compared the fitted relations obtained using mean, maximum and minimum magnitudes as defined above with those derived using magnitudes calculated from a first order fourier fit to the light curve of each Mira. In all cases we found no significant difference in the slopes, zero points and dispersions of the fitted relations at either maximum or mean light. The same conclusion regarding the robustness of mean magnitudes was reached in F89, where mean values obtained from averaging the maximum and minimum magnitudes were compared with the average of maximum and minimum intensities and also with the results of fourier fits to both magnitudes and intensities. The robustness of mean and maximum PL relations to the choice of definition for mean and maximum light is in complete accordance with the results of Hendry, Kanbur & Clarke (1997, in prep.), in which we investigate the statistical properties of various different estimators of mean and maximum light – including those adopted here and those derived from fitting low order fourier series – as a function of number of sampled phase points, phase coverage, light curve shape and limiting magnitude.
PL Relation Results
===================
The results of our regression fits to equations 4, 5 and 6 are presented in Tables 3 to 6 and illustrated in Figures 1 to 6. Column 1 in each Table lists the type of regression fit – i.e. to mean, maximum or minimum apparent magnitude. Column 2 indicates the type of Mira sample used: C Miras only, O Miras only or both Mira types (denoted ‘O+C’), and column 3 gives the number, $n$, of Miras in each sample. Note that for the mean relations we used the full sample of 29 O Miras, identical to that used in F89, while for the minimum and maximum light relations we used the sample of 28 O Miras for which phase information was available. Columns 4 to 7 give the fitted values of the zero point, $a$, and slope, $b$, of the relations with their associated standard errors, $\sigma_a$ and $\sigma_b$. Column 8 indicates the dispersion, $\sigma$ (in magnitudes) of the regression fit and column 9 gives the percentage root mean square error, $\Delta$, of the corresponding distance indicator which one would derive from the PL relation, i.e. $\Delta \simeq 46.1 \sigma \%$. Finally, column 10 gives the value of $\hat{\rho}$, the partial multiple correlation coefficient computed for the sample data and column 11 indicates the probability (denoted by ‘Prob’) that $\hat{\rho}^2$ be equal to (or greater than) its computed value under the null hypothesis that the true value of $\rho$ is equal to zero.
Figures 1 and 2 show scatterplots of mean and maximum magnitude respectively against log period for the oxygen-rich Miras. The fitted regression lines in each waveband are also drawn on the plots. Figures 3 and 4 show the corresponding scatterplots for the carbon Miras in our sample and Figures 5 and 6 show the scatterplots for the composite sample of C and O Miras.
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ---------- ---------- -------------- ------------------------
mean C 20 13.33 2.05 -0.19 0.83 0.368 16.9 -0.054 0.820
max C 20 12.42 1.34 -0.04 0.54 0.243 11.2 -0.017 0.942
min C 20 13.18 3.68 0.11 1.49 0.668 30.8 0.02 0.940
mean O 29 19.37 0.67 -2.92 0.28 0.187 8.6 -0.891 $8.43 \times 10^{-11}$
max O 28 19.63 0.57 -3.19 0.24 0.158 7.3 -0.932 $5.46 \times 10^{-13}$
min O 28 19.12 1.05 -2.65 0.45 0.292 13.4 -0.760 $3.13 \times 10^{-6}$
mean O+C 49 15.06 1.02 -0.99 0.43 0.392 18.1 -0.321 0.020
max O+C 48 15.57 0.83 -1.39 0.35 0.317 14.6 -0.508 $2.27 \times 10^{-4}$
min O+C 48 14.03 1.56 -0.37 0.65 0.593 27.3 -0.083 0.577
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ---------- ---------- -------------- ------------------------
mean C 20 15.90 1.56 -1.72 0.63 0.280 12.9 -0.541 $1.38 \times 10^{-2}$
max C 20 14.69 0.89 -1.40 0.36 0.161 7.4 -0.678 $1.02 \times 10^{-3}$
min C 20 16.96 2.48 -1.97 1.01 0.451 20.7 -0.42 $6.55 \times 10^{-2}$
mean O 29 19.06 0.64 -3.15 0.27 0.180 8.3 -0.912 $6.31 \times 10^{-12}$
max O 28 19.48 0.53 -3.48 0.23 0.147 6.8 -0.949 $1.37 \times 10^{-14}$
min O 28 18.71 1.07 -2.84 0.46 0.299 13.8 -0.771 $1.56 \times 10^{-6}$
mean O+C 49 16.75 0.70 -2.12 0.29 0.269 12.4 -0.724 $4.01 \times 10^{-9}$
max O+C 48 16.93 0.57 -2.35 0.24 0.216 9.9 -0.823 $5.18 \times 10^{-13}$
min O+C 48 16.66 1.03 -1.91 0.43 0.394 18.1 -0.546 $5.89 \times 10^{-5}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ---------- ---------- -------------- ------------------------
mean C 20 18.96 0.98 -3.29 0.40 0.176 8.1 -0.890 $1.45 \times 10^{-7}$
max C 20 17.75 0.93 -2.92 0.38 0.169 7.8 -0.877 $3.97 \times 10^{-7}$
min C 20 19.64 1.54 -3.45 0.62 0.279 12.8 -0.794 $2.89 \times 10^{-5}$
mean O 29 19.48 0.45 -3.47 0.19 0.126 5.8 -0.961 $1.29 \times 10^{-16}$
max O 28 19.71 0.45 -3.72 0.19 0.124 5.7 -0.967 $5.28 \times 10^{-17}$
min O 28 19.25 0.77 -3.23 0.33 0.215 9.9 -0.886 $3.48 \times 10^{-10}$
mean O+C 49 19.69 0.39 -3.57 0.16 0.150 6.9 -0.954 $2.86 \times 10^{-26}$
max O+C 48 19.21 0.39 -3.50 0.16 0.148 6.8 -0.953 $1.15 \times 10^{-25}$
min O+C 48 20.00 0.65 -3.57 0.27 0.247 11.4 -0.889 $3.06 \times 10^{-17}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ---------- ---------- -------------- ------------------------
mean C 20 18.73 0.74 -1.85 0.30 0.132 6.1 -0.826 $7.21 \times 10^{-6}$
max C 20 17.34 0.59 -1.42 0.24 0.107 5.0 -0.937 $1.85 \times 10^{-13}$
min C 20 21.25 1.39 -2.73 0.56 0.252 11.6 -0.753 $1.29 \times 10^{-4}$
mean O 29 21.36 0.57 -3.00 0.24 0.160 7.4 -0.920 $1.26 \times 10^{-12}$
max O 28 21.61 0.55 -3.26 0.24 0.154 7.1 -0.937 $1.85 \times 10^{-13}$
min O 28 21.11 1.07 -2.73 0.46 0.297 13.7 0.760 $2.62 \times 10^{-6}$
mean O+C 49 19.86 0.45 -2.34 0.19 0.173 8.0 -0.875 $2.11 \times 10^{-16}$
max O+C 48 19.41 0.50 -2.29 0.21 0.190 8.7 -0.851 $1.85 \times 10^{-14}$
min O+C 48 20.55 0.74 -2.47 0.31 0.280 12.9 -0.76 $2.74 \times 10^{-10}$
It is evident from Tables 3 to 6 and from Figures 2, 4 and 6 that statistically significant PL relations clearly exist at maximum light in almost all of the cases considered – a fact which is supported quantitatively by the values of the partial multiple correlation coefficient, which are generally different from zero at a very high level of significance. The only clear exception to this trend is the case of the J band maximum light PL relation for C Miras, which we discuss further below.
For the O Miras the K band PL relation has the smallest dispersion at maximum light – about 20% smaller than for the relations in the other wavebands. The maximum light relation for the composite sample of both types of Miras also has the smallest dispersion in the K band. Similar behaviour was found in F89 for the mean light relations, which we also confirm here.
It is also clear from Tables 3 to 6 that the dispersion of the minimum light PL relations is considerably larger than that of both the mean light and maximum light relations in all cases. We comment on this in section 5 below.
Comparing all of our results for mean light relations with those of F89, we see that our fitted coefficients, standard errors and dispersions are in excellent agreement in all cases.
We can see from Tables 3 to 6 that the dispersions of the maximum light PL relations are smaller than those of the corresponding mean light relations in every case considered, with the sole exception of the ‘O+C’ relation for bolometric magnitudes. Note also that the standard errors of the fitted regression coefficients are also consistently smaller for the maximum light relations. To assess the statistical significance of this result requires some care, however. We cannot simply apply a standard ratio-of-variance F test (c.f. Graybill 1976) to the data since such a test assumes that the variances are statistically independent. This condition is clearly not satisfied here, as the residuals of our maximum and mean light PL relations are likely to be highly correlated. A failure to account for this correlation would result in underestimating the significance of the reduction in dispersion. We tackle this problem numerically, first computing the correlation coefficient of the mean and maximum light residuals and then – with this correlation coefficient – generating a large number of Monte Carlo simulations to estimate the probability density function of the sample ratio of mean to maximum light dispersion, under the null hypothesis that these dispersions are equal. We then determine the statistical significance of the observed reduction in dispersion by considering the the extent of the tail of our estimated probablility density function in the standard manner.
The results of applying this significance test are given in Table 7. The first and second columns indicate the waveband and type of Mira sample under consideration, the third column indicates the ratio, $R$, of the variance at mean light to the variance at maximum light. The fourth column gives $r$, the sample correlation coefficient of the residuals at mean and maximum light, and the final column indicates the probability (denoted by ‘Prob’) of obtaining as large (or larger) a value of $R$ under the null hypothesis that the true value of the ratio is equal to unity – i.e. the dispersion at mean and maximum light is identical – and the true correlation coefficient of the mean and maximum light residuals is equal to $r$.
waveband sample $R$ $r$ Prob
---------------- -------- ------- ------- ----------------------
J C 2.293 0.827 $1.4 \times 10^{-3}$
H C 3.025 0.748 $6.0 \times 10^{-4}$
K C 1.084 0.701 0.398
$m_{\rm{bol}}$ C 1.494 0.513 0.160
J O 1.401 0.795 $7.8 \times 10^{-2}$
H O 1.500 0.753 $5.8 \times 10^{-2}$
K O 1.032 0.688 0.454
$m_{\rm{bol}}$ O 1.079 0.858 0.348
J O+C 1.529 0.904 $5.0 \times 10^{-4}$
H O+C 1.551 0.818 $4.4 \times 10^{-3}$
K O+C 1.027 0.677 0.451
$m_{\rm{bol}}$ O+C 0.829 0.830 0.875
: Significance of the reduction in dispersion of maximum light PL relations compared with mean light PL relations. Column headings are explained in the text
We can see from Table 7 that the ratio, $R$, is greater than unity in all cases except that of the ‘O+C’ relation for bolometric magnitudes. The reduction in scatter is least significant for the K band Miras – i.e. the tight PL relation already displayed at mean light is not improved as much by the use of maximum light as in the other wavebands – but is still non-negligible. A significant reduction is seen for both the H and J band relations. Although the dispersion at maximum light is slightly larger than at mean light for the ‘O+C’ bolometric relation, the increase in dispersion is not statistically significant. Table 7 illustrates the importance of accounting for the correlation between the mean and maximum light residuals: the reduction in dispersion for the C Miras in the J band is marginally smaller than that in the H band, but is marginally more significant because the J band residuals are more highly correlated.
As mentioned above, the J band PL relation for C Miras at maximum light is seen from Figure 4 to be essentially flat. This is confirmed in Table 3, where we see that the fitted coefficient of log period is only -0.02, consistent with zero, and the partial multiple correlation coefficient is not significantly different from zero. A similarly flat relation is seen in Figure 3, for the J band relation at mean light – as previously reported by F89. The existence of an H, K and bolometric magnitude PL relation, at both mean and maximum light, for the C Miras is somewhat more convincing in Figures 3 and 4: all have significantly non-zero partial multiple correlation coefficients and regression coefficients of log period. If we consider only those C Miras with periods greater than 250 days, however, then the H, K and bolometric relations are considerably flatter – i.e. the longer period Miras are more consistent with having constant absolute magnitude. We comment further on this in the next section.
Finally, F89 found that the slope of the bolometric magnitude PL relation at mean light was shallower for C Miras than for O Miras. Our results confirm this conclusion and indicate that it is also true – and indeed is considerably more pronounced – at maximum light.
PLC Results
===========
F89 also presented evidence for the existence of PLC relations at mean light, for both the carbon and oxygen Miras based on $(J-K)$ colours. By considering the correlation of the PL relation residuals with $(J-K)$, F89 found evidence of a significant mean colour term for the O Miras at all wavelengths, but for the C Miras the colour term was highly significant only for the J band PLC relation, and was found to be marginal for the K band and bolometric relations. F89 also showed that, where significant, the colour term was intrinsic and could not be attributed to differential reddening. In this paper we have derived PLC relations at mean, maximum and minimum light using both $(J-K)$ and $(J-H)$ colours. We present the results of our regression fits to equations 7 to 9 in Tables 8 to 10 below, with the corresponding results for $(J-H)$ colours in Tables 11 to 13. The columns of these Tables are as in Tables 3 to 6, with two additional columns giving the fitted value and standard error of the colour coefficient, $c$. Note that in the case of the PLC relations with $(J-K)$ colours we do not present the K band results since it is straightforward to show that these are trivially related to those at J band: i.e. $a_K = a_J$, $b_K = b_J$ and $c_K = c_J - 1$. Moreover, one may also show that the dispersions of the J and K band PLC relations, and the standard error of the coefficients, are identical. The J and H band PLC relations based on $(J-H)$ colour are similarly related in a trivial way.
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------ ------------ ---------- ---------- -------------- ------------------------
mean C 20 18.37 1.18 -2.97 0.53 0.89 0.11 0.177 8.2 0.884 $5.14 \times 10^{-7}$
max C 20 16.03 1.01 -1.99 0.46 0.68 0.12 0.144 6.6 0.817 $1.96 \times 10^{-5}$
min C 20 20.88 1.68 -4.14 0.74 1.19 0.12 0.268 12.4 0.921 $2.30 \times 10^{-8}$
mean O 29 19.57 0.37 -3.98 0.21 1.91 0.24 0.103 4.8 0.840 $2.29 \times 10^{-8}$
max O 28 19.76 0.43 -4.04 0.27 1.61 0.36 0.120 5.5 0.665 $1.55 \times 10^{-4}$
min O 28 19.37 0.69 -3.77 0.35 1.91 0.32 0.191 8.8 0.767 $3.02 \times 10^{-6}$
mean O+C 49 19.36 0.47 -3.39 0.22 0.93 0.06 0.149 6.9 0.927 $3.35 \times 10^{-21}$
max O+C 48 18.97 0.47 -3.37 0.22 0.94 0.07 0.149 6.8 0.886 $1.28 \times 10^{-16}$
min O+C 48 20.01 0.77 -3.58 0.35 1.00 0.06 0.249 11.5 0.910 $9.04 \times 10^{-19}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------ ------------ ---------- ---------- -------------- ------------------------
mean C 20 19.06 1.31 -3.46 0.59 0.56 0.13 0.197 9.1 0.729 $4.00 \times 10^{-4}$
max C 20 16.50 0.91 -2.38 0.42 0.34 0.10 0.130 6.0 0.619 $4.69 \times 10^{-3}$
min C 20 21.37 1.80 -4.41 0.80 0.68 0.13 0.288 13.3 0.784 $7.25 \times 10^{-5}$
mean O 29 19.24 0.40 -4.11 0.22 1.74 0.26 0.111 5.1 0.796 $4.17 \times 10^{-7}$
max O 28 19.59 0.42 -4.23 0.26 1.42 0.35 0.117 5.4 0.627 $4.65 \times 10^{-4}$
min O 28 18.93 0.83 -3.81 0.42 1.67 0.39 0.231 10.6 0.653 $2.23 \times 10^{-4}$
mean O+C 49 19.27 0.52 -3.53 0.24 0.55 0.06 0.165 7.6 0.794 $1.62 \times 10^{-11}$
max O+C 48 18.95 0.44 -3.53 0.21 0.56 0.07 0.138 6.4 0.773 $1.89 \times 10^{-10}$
min O+C 48 19.76 0.87 -3.58 0.39 0.52 0.08 0.280 12.9 0.711 $2.14 \times 10^{-8}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------ ------------ ---------- ---------- -------------- -----------------------
mean C 20 19.79 0.73 -2.44 0.33 0.19 0.07 0.112 5.2 0.445 $5.65 \times 10^{-2}$
max C 20 17.55 0.77 -1.53 0.36 0.04 0.09 0.111 5.1 0.010 0.671
min C 20 22.80 1.41 -3.59 0.62 0.24 0.10 0.226 10.4 0.493 $3.21 \times 10^{-2}$
mean O 29 21.51 0.39 -3.80 0.22 1.43 0.26 0.111 5.1 0.736 $7.96 \times 10^{-6}$
max O 28 21.74 0.43 -4.08 0.26 1.54 0.36 0.119 5.5 0.651 $2.36 \times 10^{-4}$
min O 28 21.36 0.73 -3.83 0.37 1.88 0.34 0.204 9.4 0.741 $1.00 \times 10^{-5}$
mean O+C 49 21.13 0.43 -3.05 0.20 0.27 0.05 0.138 6.3 0.618 $2.88 \times 10^{-6}$
max O+C 48 20.82 0.48 -3.11 0.23 0.39 0.07 0.151 7.0 0.616 $4.04 \times 10^{-6}$
min O+C 48 21.85 0.80 -3.17 0.36 0.22 0.07 0.257 11.8 0.420 $3.28 \times 10^{-3}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------ ------------ ---------- ---------- -------------- ------------------------
mean C 20 18.35 1.78 -3.17 0.86 1.95 0.42 0.253 11.7 0.744 $2.59 \times 10^{-4}$
max C 20 16.63 0.98 -2.57 0.49 1.86 0.28 0.134 6.2 0.845 $5.21 \times 10^{-6}$
min C 20 21.28 2.14 -4.37 0.96 2.15 0.29 0.337 15.4 0.874 $9.77 \times 10^{-7}$
mean O 29 19.02 0.69 -3.18 0.33 1.13 0.75 0.183 8.4 0.282 0.145
max O 28 19.46 0.54 -3.51 0.28 1.11 0.57 0.150 6.9 0.366 $6.03 \times 10^{-2}$
min O 28 19.05 1.11 -2.68 0.48 0.15 0.74 0.297 13.7 0.041 0.838
mean O+C 49 18.53 0.63 -3.30 0.31 2.06 0.19 0.210 9.7 0.848 $2.65 \times 10^{-14}$
max O+C 48 18.59 0.47 -3.54 0.24 2.23 0.18 0.154 7.1 0.876 $7.05 \times 10^{-16}$
min O+C 48 18.84 0.97 -3.21 0.45 1.84 0.18 0.325 15.0 0.841 $1.47 \times 10^{-13}$
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------- ------------ ---------- ---------- -------------- -----------------------
mean C 20 18.21 1.24 -2.84 0.60 -0.29 0.30 0.176 8.1 -0.233 0.338
max C 20 16.60 1.20 -2.22 0.60 -0.51 0.35 0.164 7.6 -0.335 0.161
min C 20 20.83 1.75 -4.11 0.78 0.32 0.24 0.273 12.6 0.307 0.200
mean O 29 19.29 0.47 -3.61 0.22 0.59 0.51 0.125 5.8 0.221 0.259
max O 28 19.59 0.43 -3.95 0.23 0.81 0.45 0.120 5.5 0.335 $8,69 \times 10^{-2}$
min O 28 19.42 0.81 -3.16 0.35 -0.39 0.54 0.217 10.0 -0.142 0.478
mean O+C 49 19.38 0.45 -3.37 0.22 -0.19 0.13 0.148 6.8 -0.200 0.173
max O+C 48 19.25 0.46 -3.54 0.24 0.03 0.18 0.150 6.9 0.028 0.851
min O+C 48 19.84 0.74 -3.48 0.34 -0.06 0.14 0.249 11.5 -0.069 0.646
fit sample $n$ $a$ $\sigma_a$ $b$ $\sigma_b$ $c$ $\sigma_c$ $\sigma$ $\Delta$ $\hat{\rho}$ Prob
------ -------- ----- ------- ------------ ------- ------------ ------- ------------ ---------- ---------- -------------- -----------------------
mean C 20 19.42 0.92 -2.26 0.45 0.27 0.22 0.130 6.0 0.286 0.236
max C 20 17.81 0.80 -1.70 0.40 0.21 0.23 0.109 5.0 0.211 0.385
min C 20 22.60 1.53 -3.48 0.69 0.36 0.21 0.239 11.0 0.389 $9.98 \times 10^{-2}$
mean O 29 20.99 0.58 -3.28 0.27 1.18 0.63 0.153 7.0 0.346 $7.14 \times 10^{-2}$
max O 28 21.47 0.54 -3.53 0.28 0.91 0.57 0.150 6.9 0.307 0.119
min O 28 21.18 1.13 -2.70 0.49 -0.15 0.76 0.303 14.0 -0.039 0.846
mean O+C 49 20.87 0.44 -3.01 0.22 0.60 0.13 0.145 6.7 0.557 $3.90 \times 10^{-5}$
max O+C 48 20.72 0.46 -3.22 0.23 0.97 0.18 0.148 6,8 0.635 $1.65 \times 10^{-6}$
min O+C 48 21.28 0.81 -2.90 0.38 0.28 0.15 0.273 12.6 0.271 $6.54 \times 10^{-2}$
The PLC results presented here correspond to the ‘Method a’ case presented in F89, i.e. ordinary least squares with the errors in the magnitudes. These solutions do not account for the effect of correlated errors on apparent magnitude and colour excess – an effect treated in detail by e.g. Caldwell & Coulson (1985). Following F89 we conclude that this effect is negligible for these data, since the extinction in the J, H and K bands is very small.
Our results at mean light using $(J-K)$ colours are, as expected, in complete agreement with those of F89. Moreover our conclusions concerning the significance of the colour term are concordant with F89, and indeed are reinforced by considering the value of the sample multiple correlation coefficient, $\hat{\rho}$, and associated P value. Clearly the magnitude of the colour term itself, compared with its standard error, also gives an indication of its statistical significance. On this basis, we can see that there is good evidence for a $(J-K)$ PLC relation at mean, maximum and minimum light in almost all cases considered. The only exceptions are the bolometric PLC relation for the C Miras (for which the P values listed in Table 10 are at least several orders of magnitude larger than in other cases) and the K band PLC relation for the C Miras, which – although not listed in the Tables – can be seen to have a marginal colour term at mean, maximum and minimum light by considering the $c_J$ coefficients and using $c_K = c_J-1$ as noted above. The colour term is highly significant for the J band relations, slightly less significant for the H band relations and least significant (although still clearly present) for the bolometric relations.
For the results using $(J-H)$ colours there is clear evidence of a significant colour term for the J band ‘C’ and ‘O+C’ relations and – in view of the values of $c_J$ – also for the H band ‘C’ and ‘O+C’ relations. The colour term is also significant for the bolometric PLC relation with the ‘O+C’ sample. In all other cases, however, there is no evidence for a significant colour term – i.e. the addition of $(J-H)$ colour does not significantly reduce the dispersion of the PL relation.
It is clear from Tables 8 to 13 that in all cases the dispersion of the PLC relations at minimum light is considerably larger than at mean and maximum light, for both $(J-K)$ and $(J-H)$ colours. Comparing the dispersion at mean and maximum light, however, we see that our results are somewhat more ambiguous than was the case for the PL relations considered in the previous section. For $(J-K)$ colours the dispersion at maximum light is in fact slightly [*larger*]{} than that at mean light for the O Miras with J band, H band and bolometric magnitudes, and also for the ‘O+C’ sample bolometric relation. In the remaining five cases the dispersion at maximum light is smaller than at mean light. What is noteworthy, however, is the particular success of the J and H band maximum light relations for the C Miras: these are the two cases for which the dispersion at mean light is largest, and the reduction in dispersion at maximum light is found to be about $20 - 30 \%$.
For the PLC relations with $(J-H)$ colour, where the colour term was not significant the ratio of the maximum to mean dispersion was very similar to that for the corresponding PL relations – i.e. the dispersion of the maximum light relation was generally found to be comparable to, or slightly smaller than, that at mean light. This is not surprising since in these cases the fitted PLC relation shows no statistically significant difference from the PL relation. For the cases where a statistically significant $(J-H)$ colour term [*was*]{} found, on the other hand, the maximum light relations scored two notable successes: the dispersion of the J band relations for the ‘C’ and ‘O+C’ samples was reduced by almost 50% and 25% respectively. Note that, as for the $(J-K)$ relations, the largest reduction in dispersion at maximum light occurred for the PLC relations with largest dispersion at mean light. Note also that for the C Miras the J band PL dispersion at maximum light is already smaller than the corresponding PLC relation at mean light with $(J-H)$ colour, while the maximum light PLC relation reduces the dispersion by almost another factor of two.
Tables 14 and 15 list the results of applying to the fitted PLC relations the test, introduced in the previous section, to determine the statistical significance of the reduction (or increase!) in scatter between mean and maximum light. The columns are as in Table 7 above. Table 14 gives the results using $(J-K)$ colours while Table 15 is for $(J-H)$ colours.
waveband sample $R$ $r$ Prob
---------------- -------- ------- ------- ----------------------
J C 1.511 0.800 $7.5 \times 10^{-2}$
H C 2.296 0.705 $8.7 \times 10^{-3}$
$m_{\rm{bol}}$ C 1.018 0.709 0.477
J O 0.737 0.597 0.836
H O 0.916 0.541 0.604
$m_{\rm{bol}}$ O 0.854 0.646 0.696
J O+C 1.086 0.667 0.355
H O+C 1.430 0.575 $7.2 \times 10^{-2}$
$m_{\rm{bol}}$ O+C 0.835 0.737 0.752
: Significance of the reduction in dispersion of maximum light PLC relations compared with mean light PLC relations, both using $(J-K)$ colours. Column headings are as explained in Section 3 above
waveband sample $R$ $r$ Prob
---------------- -------- ------- ------- ----------------------
J C 3.565 0.493 $1.5 \times 10^{-3}$
K C 1.152 0.704 0.341
$m_{\rm{bol}}$ C 1.422 0.484 0.192
J O 1.488 0.742 $1.0 \times 10^{-1}$
K O 1.085 0.662 0.389
$m_{\rm{bol}}$ O 1.040 0.811 0.424
J O+C 1.860 0.494 $8.4 \times 10^{-3}$
K O+C 0.974 0.672 0.535
$m_{\rm{bol}}$ O+C 0.960 0.679 0.561
: Significance of the reduction in dispersion of maximum light PLC relations compared with mean light PLC relations, both using $(J-H)$ colours. Column headings are as explained in Section 3 above
The results of Tables 14 and 15 confirm that in the cases where the dispersion at maximum light is greater than at mean light ($\Delta < 1$) the increase in dispersion is never significant at less than the 15% level, while the [*reduction*]{} in dispersion at maximum light is in several other cases significant at the 1% level.
Discussion
==========
The principal result of this paper concerns the existence of Mira PL relations at maximum light and the fact that in all cases these relations display less scatter than the corresponding mean light relations, a reduction in dispersion which is statistically significant in the J and H bands. This result is apparent not only in the values of the dispersion derived for the mean and maximum light relations, but also is suggested by the behaviour of some of the outliers in the scatterplots of magnitude against log period. In Figures 3 and 4 for example, there are a small number of outliers, with log periods of around 2.5, in the J and H band mean light relations which are in much better agreement with the fitted regression line at maximum light. We also find that in several cases for the C Miras the dispersion of the maximum light PLC relation is significantly smaller, by up to 50%, than at mean light.
It is obviously important now to ask what is the most likely source of the reduction in dispersion which we have observed at maximum light. Figures 7 and 8 show plots of dereddened $(J-K)$ colour against log period at mean (a) and maximum (b) light, for the C and O Miras respectively. (Note that ‘maximum $(J-K)_0$’, as indicated on the axes of the plots in Figures 7 and 8, in fact means the dereddened colour at the phase of maximum light in the J band, which need not be the same as the maximum observed value of J-K colour, although the difference is likely to be quite small). Similarly Figures 9 and 10 show plots of dereddened $(J-H)$ colour against log period for the C and O Miras respectively.
Whilst the properties of Mira PLC relations at maximum light will be the focus of future work, we note from these plots that the scatter in $(J-K)_0$ at a given period is very similar at mean and maximum light. This suggests that the range of effective temperatures at given period will not be greatly different at mean and maximum light, as was also claimed by Feast (1995). In the light of our discussion in Section 1, we therefore conjecture that the smaller dispersion of maximum light PL relations compared with mean PL relations is primarily due to the fact that both maximum luminosity and period depend on the total mass, whereas equilibrium luminosity – and hence mean luminosity – depends strongly on core mass. If the above explanation is correct then – insofar as a considerably larger reduction in dispersion is found in the J and H bands for the C Miras compared with the O Miras – one might suppose this to be due to there being a larger difference between the range of core masses and total masses for C Miras compared with that for O Miras. It would be interesting to investigate this possibility further, and carry out a more detailed study of the systematic differences between C and O Miras.
In comparing the values of the regression coefficients obtained in the PL fits at mean and maximum light, the general trend which one observes is as follows. For the C Miras the zero point, $b$, and slope, $a$, are found to be smaller and more shallow (i.e. less negative) respectively for the maximum light relations in all wavebands. Based on the standard errors of the regression coefficients this systematic difference appears to be quite significant, although we have not carried out a specific statistical test of this hypothesis. For the O Miras, on the other hand, precisely the converse is the case: the zero point and slope of the PL relations at maximum light are found to be respectively larger and steeper (i.e. more negative) than at mean light in all wavebands. Aside from investigating whether the use of the maximum light relations derived in this study lead to distance estimates significantly different from those determined using Mira PL relations at mean light (c.f. F89, Whitelock 1995, Feast 1995), it would be interesting to investigate if any systematic difference in the slope and zero point of maximum light relations can be explained in terms of our existing knowledge of the physics of Mira variables. We will address this problem further in a future paper.
In the case of Cepheids it was shown in Simon, Kanbur & Mihalas (1993) that [*both*]{} maximum and minimum light occur as the star is passing through its equilibrium radius. If this were also true for Miras, then equation (3) might also be valid at minimum light, but with $L_{\rm{max}}$ and $T_{\rm{max}}$ replaced by $L_{\rm{min}}$ and $T_{\rm{min}}$. As we commented above, the Mira PL and PLC relations at minimum light were in all cases found to have a larger dispersion than at mean light. In view of our discussion in Section 1, perhaps one reason for this result is that minimum luminosity is dependent on the total mass of the star in a different way to maximum luminosity; in other words it may be the case that equation (3) – in its equivalent form – does indeed hold at minimum light but that when we collapse the equation over the variables $\log {\cal{M}}$ and $\log T_{\rm{min}}$ the resultant PL relation has a larger dispersion than at either maximum or mean light. We intend to investigate further the properties of PL relations at minimum light in future work.
Whilst the mean and maximum light PL relations which we have derived here clearly have practical use in terms of in terms of distance estimation, the relations involving bolometric magnitudes are also very interesting for the purpose of better understanding stellar pulsation and evolution. Indeed, the fact that we have established bolometric PL relations for both C and O Miras directly supports the validity of equation (3).
Finally it is important to comment explicitly on the practical application of the results of this paper – the use of maximum light PL relations for Miras as distance indicators. Aside from the advantage of the small reduction in dispersion which our analysis in this paper has identified, the use of maximum light relations in distance estimation can also be justified on the grounds that one can extend their application to greater distances before the effects of luminosity selection bias become important. In Hendry, Kanbur & Clarke (1997, in prep.) we examine in detail the robustness of PL relations derived for Miras detected close to an apparent magnitude limit, and find that – as one approaches the magnitude limit – the measurement of mean light becomes biased, and is subject to a increasingly large root mean squared error, substantially more quickly than does maximum light. Moreover, for a range of different light curve shapes we find that the identification of maximum light simply with the brightest observed phase point (as was the definition adopted in this paper) remains a robust and reliable estimate of maximum light as one approaches the magnitude limit – provided one has of the order of ten or more sampled phase points – and is certainly considerably more robust than the identification of mean light with the average of the observed magnitudes. This work suggests that maximum light PL relations can easily be constructed without recourse to exhaustive observing programmes and can therefore prove useful in extending the range and reliability of Mira-based distance indicators. It would seem to us, therefore, that a priority for future work is to establish the existence of Mira PL relations at maximum light in different stellar environments, such as the SMC and Galactic globular clusters, and to test the uniformity of such relations.
Wood (1990) has used equations for the position of AGB in the HR diagram together with the period-mean density theorem to obtain a pulsation equation similar to equation (1), but also incorporating a metallicity dependence. This work is discussed in Feast (1995), which suggests that the available evidence indicates little variation in the mean light Mira PL relations at K or bolometric magnitudes in environments with a range of different metallicities. There is no reason to believe why any metallicity dependence of equation (1) would act differentially between mean and maximum light, although of course any possible effect should certainly be checked observationally. In any case, such a metallicity gradient with environment – if present – would have no bearing on our discussion of the relative dispersion of mean and maximum light PL relations in this paper.
Some examples of the recent application of distance indicators based on Mira PL relations at mean light include the following. In F89 Mira distances were determined to Galactic globular clusters, thus providing an absolute calibration of RR Lyrae and horizontal branch stars. On the other hand, Whitelock (1995) and references therein used Mira PL relations at mean light to study the dimensions and kinematics of the disk, halo and bulge of the Galaxy.
Conclusions
===========
In this paper we have demonstrated the existence of PL relations for Miras at maximum light in the J, H and iK bands and for bolometric magnitudes. Our results were based on analysis of a sample of oxygen-rich and carbon-rich Miras in the LMC, as previously studied in F89. In the J, H and K bands the PL relations at maximum light have a smaller dispersion for the oxygen-rich Miras than for the carbon rich Miras, while the converse was found to be true for the PL relation based on bolometric magnitudes. We have shown that for the J and H bands the Mira PL relations at maximum light have a significantly smaller dispersion than their counterparts at mean light. Our results also are suggestive that C Miras with periods in excess of 250 days have constant mean and maximum absolute magnitude.
Based on similar reasoning to that outlined in KH, we present a theoretical justification for the existence of such maximum light relations. The crucial assumption made in this justification is that the photospheric radius at mean light is roughly equal to the photospheric radius at maximum light, for which there exists some evidence. Amongst other factors – including metallicity and temperature – the dispersion at given period in a mean light PL relation is influenced by both the range of core masses and the range of total masses found in Miras. However, at maximum light we suggest that – amongst these same other factors – the dispersion at given period is influenced only by the range of total masses, and that it is this fact which is responsible for the smaller dispersion of maximum light PL relations which we have observed.
In Section 5 we have outlined a number of topics for future work, but in summary it seems clear that the main direction of future work should be the study of larger samples of Miras, in order to investigate the prevalence, uniformity and reliability of maximum light relations in other environments. The relative robustness of maximum light relations when the corresponding mean light relations are pushed close to an apparent magnitude limit makes their application in external galaxies an important and exciting possibility – particularly with the installation of J and H band filters in the NICMOS camera on the newly refurbished Hubble Space Telescope. We are confident that Mira PL relations at maximum light can become a powerful tool for galactic and extragalactic astronomy.
The authors thank Patricia Whitelock for supplying the Mira observations of F89 in a convenient electronic form, and Shaun Hughes, Tom Lloyd Evans, Norman Simon and Dimitri Mihalas for useful discussions. The authors also thank the anonymous referee for useful comments. MAH acknowledges the PPARC, for the award of a Personal Research Fellowship. MAH and SMK ackowledge the use of computer facilities supported by the STARLINK project. DC acknowledges the assistance of Mrs. Margaret Morris in analysing the data.
\*Cox, J.P. 1974, Rep. Prog. Phys., 37, 563
\* Feast, M.W., Glass,I.S, Whitelock, P.A., Catchpole, R.M. 1989, MNRAS, 241, 375 (F89)
\* Feast, M.W. 1996, MNRAS, 278, 11
\* Glass, I.S., Whitelock, P.A., Catchpole, R.M., Feast, M.W., Laney, C.D. 1990, SAAO Circ., 14, 63
\* Graybill, F.A. 1976, Theory and Applications of the Linear Model, Duxbury Press
\* Hendry, M.A., Kanbur, S.M. 1996, in ‘Mapping, Measuring and Modelling the Universe’, ASP Conference Series, eds. P. Coles, V. J. Martinez, M.J. Pons-Borderia, [**94**]{}, 357-361
\* Hughes, S. 1996, private communication
\* Kanbur, S.M., Hendry, M.A. 1996, A&A, 305, 1 (KH)
\* Kennicutt, R.C., Freedman, W.L., Mould, J.R. 1995, AJ, 110, 1476
\* Madore, B.F., Freedman, W.L. 1991, PASP, 103, 933
\* Martin, W.L., Warren, P.R., Feast, M.W. 1979, MNRAS, 188, 139
\* Pierce, M.J., McClure, R.D., Welch, D.L., Racine, R., van den Bergh, S. 1993, in ‘New Perspectives on Stellar Pulsation and Pulsating Variable Stars’, IAU Colloq. 139, eds. J. Nemec, J. Matthews, p. 81
\* Pierce, M.J., Welch, D.L., McClure, R.D., van den Bergh, S., Racine, R., Stetson, P.B. 1994, Nature, 371, 385
\* Saha, A., Sandage, A., Labhardt, L., Tammann, G.A., Macchetto, F.D., Panagia, N. 1996a, ApJ, in press
\* Saha, A., Sandage, A., Labhardt, L., Tammann, G.A., Macchetto, F.D., Panagia, N. 1996b, ApJ, in press
\* Sandage, A. 1958, Astrophys. J., 127, 513
\* Shibahashi, H. 1993, in ‘New Perspectives on Stellar Pulsation and Pulsating Variable Stars’, IAU Colloq. no. 139, eds. J. Nemec J. Matthews, p. 103
\* Simon, N.R., Kanbur, S.M., Mihalas, D. 1993, Astrophys. J., 414, 310
\* Whitelock, P. 1995, in ‘Astrophysical Applications of Stellar Pulsation’, ASP Conference Series no. 83, eds. R. S. Stobie, P. A. Whitelock, p. 165
\*Whitmore, B.C., Sparks, W.B., Lucas, R.A., Macchetto, F.D., Biretta, J.A. 1995, ApJ, 454, L73
\* Wood, P.R., 1990, in ’From Miras to Planetary Nebulae’, eds. Mennessier M.O., Oment A., Editions Frontieres, Gif-sur-Yvette, p. 67
\* Wood,P. 1995, in ‘Astrophysical Applications of Stellar Pulsation’, ASP Conference Series no. 83, eds. R. S. Stobie, P. A. Whitelock, p. 127
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Grégoire Misguich\
Institut de Physique Théorique\
CEA, IPhT, CNRS, URA 2306\
F-91191 Gif-sur-Yvette, France
title: 'Quantum spin liquids[^1]'
---
Introduction: band and Mott insulators
======================================
Depending on the context (experiments, theory, simulations,...), “Quantum spin liquid” is sometimes used with rather different meanings. But let us start with a first simple definition: the ground state of a lattice quantum spin model is said to be a quantum spin liquid (QSL) if it spontaneously breaks [*no*]{} symmetry. According to this first definition, a QSL is realized if the spins fail to develop any kind of long range order at zero temperature ($T=0$) (hence the word “liquid”, as opposed to solids which are ordered and break some symmetries). Of course, this first definition raises a number of questions: Does this define new distinct states of matter ? Do QSL have some interesting properties ? Are there some experimental examples ? To answer these questions, it is useful to go back to the origin of magnetism in insulators.
Generally speaking, there are two kinds of insulators: [*band*]{} insulators, and [*Mott*]{} insulators. The first ones can be qualitatively understood from the limit of non-interacting (or weakly interacting) electrons. Consider for instance a periodic lattice[^2] with an [*even*]{} number $n$ of sites per unit cell, with an average electron density of one electron per site (so-called half filling). The Hamiltonian describing how the electrons hop from sites to sites looks like $H_K=-t \sum_{\langle i,j\rangle,\sigma=\uparrow,\downarrow}\left( c^\dagger_{i\sigma}c_{i\sigma}+H.c\right)$, where only first neighbor hopping is considered for simplicity. $H$ can be diagonalized in Fourier space and gives $n$ dispersing bands. The ground state is just the Fermi sea obtained by filling the lowest energy states. Since the density is one electron per site, the $n/2$ lowest energy bands are completely filled (one up and one down electron for in each single particle state). Assuming that the band $n/2+1$ is separated by a gap $\Delta$ in energy from the $n/2$ lower bands, all the excitations are gapped and, at temperatures smaller than the gap, there is no charge carrier to carry an electric current. This is the well known picture for a band insulator: there are no low energy charge degrees freedom, no magnetic (spin) degrees of freedom, the ground state (Fermi sea) is unique and breaks no symmetry. To get some interesting QSL, we should instead look at Mott insulators. There, the number of sites per unit cell is [*odd*]{} and the non-interacting limit is unable to give the correct insulating behavior (at least one band is partially filled, hence with low energy charge excitations). It is more useful to look at the system in the opposite limit of very large electron-electron repulsion, as with the large $U$ limit of the Hubbard model: $H=H_K+ U\sum_i, c^\dagger_{i\uparrow}c_{i\uparrow}c^\dagger_{i\downarrow}c_{i\downarrow}$. At $U=\infty$ and $t=0$ (still at half filling), the ground state is highly degenerate ($=2^V$, where $V$ is the total number of sites) since any state with one electron per site is a ground state, whatever the spins orientations. To describe how this degeneracy is lifted at weak but finite $t/U$, a second order perturbation has to be computed.[^3] The result is an effective Hamiltonian acting in the subspace spin configurations, and takes the form of a quantum spin-$\frac{1}{2}$ Heisenberg model: $$H=\frac{1}{2}\sum_{ij} J_{ij} \vec S_i \cdot \vec S_j
\label{eq:Heisenberg}$$ where $J_{ij}=t_{ij}^2/U$ involves the hopping amplitude $t_{ij}$ between sites $i$ and $j$ and measures the strength of the antiferromagnetic (AF) interaction between the (electron) spins $\vec S_i$ and $\vec S_j$.[^4]
Although the model of Eq. \[eq:Heisenberg\] is in general a complicated quantum many body problem with very few exact results,[^5] its ground state and low energy properties are qualitatively well understood in many cases. In particular, the ground state can be [*antiferromagnetically ordered*]{} (also called Néel state). Such state can be approached from a semi classical point of view described in Sec. \[sec:SW\]: the spins point well defined directions and form a regular structure. Most of the Mott insulators studied experimentally belong to this family. The simplest example is the nearest neighbor Heisenberg model on bipartite lattices such as the square, cubic or hexagonal lattices. There, on average, all the spins of the sublattice $A$ point in direction $+\vec S_0$ (spontaneous symmetry breaking of the $SU(2)$ rotation symmetry), and all the spins of sublattice $B$ point in direction $-\vec S_0$. The difference with a classical spin configuration is that the magnetization of one sublattice (it is the order parameter for a Néel state) is reduced by the quantum zero-point fluctuations of the spins, even at $T=0$. Such ordered states are not QSL (they might instead be called spin “solids”) since they break the rotation symmetry.
The main question addressed in these notes is the fate of the ground state of Eq. \[eq:Heisenberg\] when the lattice and the interactions $J_{ij}$ are such that the spins [*fail*]{} to develop any such Néel ordered state. A state without any order is not necessarily interesting from a theoretical point of view. For instance, a spin system at very high temperature is completely disordered and does not have any rich structure. As we will see, the situation in Mott insulators at $T=0$ is completely different. A first hint that Mott QSL host some interesting topological properties will be discussed in Sec. \[sec:LSMH\] (Lieb-Shultz-Mattis [@lsm61] Hastings [@hastings04] theorem). A concrete (but qualitative) picture for QSL wave functions is given in Sec. \[sec:RVB\], in terms of short range valence bond configurations and deconfined spinons (magnetic excitations carrying a spin $\frac{1}{2}$). Finally, Sec. \[sec:SB\], presents a formalism which puts some of the ideas above on firmer grounds. It is based on a large-$N$ generalization of the Heisenberg models ($SU(2)\to Sp(N)$) which allow to describe some gapped QSL and to establish a connection by topologically ordered state of matter, such as the ground state of Kiatev’s toric code [@kitaev97].
Some materials without magnetic order at $T=0$
==============================================
There are many magnetic insulators that do order at $T=0$.[^6] For instance, the magnetic, properties of many compounds are described by 1D spin chains of spin ladder Hamiltonians. Thanks to the Mermin-Wagner theorem and the reduced dimensionality, these system cannot develop long range spin-spin correlations, even at $T=0$.[^7] They certainly deserves to be called QSL and represent a very rich field of activity. In these notes we will instead focus on QSL in $D>1$ systems, where our present understanding is less complete.
CaV$_4$O$_9$ is the first Heisenberg system in $D>1$ where the magnetic excitations were experimentally shown to be gapped, in 1995 [@taniguchi95]. This compound can be modeled by an antiferromagnetic spin-$\frac{1}{2}$ Heisenberg model on a [*depleted*]{} square lattice where one site out of five is missing (Fig. \[fig:cav4o9\]). The remaining sites correspond to the locations of the Vanadium ions, which carry the magnetically active electrons (spins). The magnetic interactions $J_{ij}$ turned out to be significant not only between nearest neighbors, but also between second nearest neighbors (the electron hops through oxygen orbitals, which have a complex geometry). Through magnetic susceptibility measurements, it was shown that the ground state is a rotationally invariant spin singlet, thus excluding any Néel ordering. This QSL behavior can be understood by taking a limit where only the strongest $J_{ij}$ are kept, and the other are set to zero. It turns out these strongest couplings are between second-nearest neighbors, and form a set of [*decoupled*]{} four-site plaquettes (of area $\sqrt{2}\times\sqrt{2}$ and surrounding a missing site). Since the ground state of such a four-site Heisenberg cluster is a unique singlet $S=0$ state, separated by a gap from other states, the model is trivially a gapped and without any broken symmetry in this limit. But this is not the kind of QSL we want to focus on here, since it can be adiabatically transformed into a band insulator. Switching off the electron-electron interactions would make the system metallic, but one can proceed in a different way. Starting with realistic values of the $J_{ij}$, the inter plaquette couplings are gradually turned off. Doing so, one can check (numerically for instance) that the spin gap does not close and no (quantum) phase transition in encountered. Then, in this systems of decoupled four-electron cluster, the Hubbard repulsion $U$ can be switched to zero, without causing any phase transition. The final model is evidently a band insulator and smoothly connected to the initial Heisenberg model.
![Depleted square lattice model for the magnetic properties of CaV$_4$O$_9$. The different exchange energies are shown by different types of line. The strongest $J$ correspond to the fat lines forming the large tilted square plaquettes. []{data-label="fig:cav4o9"}](cav4o9){height="4cm"}
Since then, numerous 2D and 3D (Heisenberg) magnetic systems with an [*even*]{} number of spin-$\frac{1}{2}$ per unit cell have been found to be gapped. To our knowledge, their ground state can be qualitatively understood from a limit of weakly coupled clusters in all cases and can therefore be “classified” as band insulators (as CaV$_4$O$_9$ above). Some of them can be very interesting for different reasons,[^8] but their ground states are not fundamentally new states of matter.
In the recent years, experimentalists have also uncovered a number of materials which are well described by 2D Heisenberg models with an [*odd*]{} number of spin-$\frac{1}{2}$ per crystal unit cell,[^9] and which do not develop any Néel order when $T\to0$. Some examples are the Herbertsmithite (ZnCu$_3$(OH)$_6$Cl$_2$)[^10] [@herbert] and Volborthite (Cu$_3$V$_2$O$_7$(OH)$_2$ $\cdot$2H$_2$O)[@volbo] minerals (both with a kagome lattice geometry), triangular based organics materials [@shimizu03; @itou08], or triangular atomic layers of He$^3$ adsorbed onto graphite [@masutomi04] (there the spin is not electronic, but nuclear). It turns out that all these systems seem to have [*gapless*]{} magnetic excitations and a complete theoretical understanding of these system is still lacking. The present theories for gapless QSL are rather elaborate [@gaplessQSL] and many questions remain open (stability, nature of the excitations, correlation exponents, etc.). However, as we will see, [*gapped*]{} QSL are simplest from a theoretical point of view. Intriguingly, to our knowledge, no gapped QSL has been discovered so far in nature, although many spin models do have gapped QSL ground states.
Spin wave theory, zero modes and breakdown of the $1/S$ expansion {#sec:SW}
=================================================================
To understand why an AF Heisenberg spin model can [*fail*]{} to order at zero temperature, is is useful to briefly review the standard approach to Néel phases: the semi-classical $1/S$ spin-wave expansion [@anderson52]. This approach i) starts from a classical spin configuration which minimizes the Heisenberg interaction ii) assumes that the quantum deviations from this ordered direction are [*small*]{} iii) treats this deviations as collection of harmonic oscillators (the leading term in a $1/S$ expansion). In this approximation the Hamiltonian is written using boson creation and annihilation operators, is quadratic, and can be diagonalized by a Bogoliubov transformation. One can then check [*a posteriori*]{} if the spin deviations are indeed small. If it is not the case, we have a strong indication that the magnetic long range order is in fact “destroyed” by the quantum fluctuations, thus opening a route for a QSL ground state.
Holstein-Primakoff representation
---------------------------------
The starting point is the representation of the spin operators using Holstein-Primakoff [@hp40] bosons $$\begin{aligned}
S^z_i &=& S - a^\dagger_i a_i \;,\; S^+=\sqrt{2S-a^\dagger_i a_i} \;\;a_i
\;,\; S^-=a^\dagger_i \sqrt{2S-a^\dagger_i a_i},
\label{eq:HP}\end{aligned}$$ from which on can check that the commutation relations $[S^\alpha_i,S^\beta_i]=i\epsilon^{\alpha\beta\delta}S^\delta_i$ and $\vec S^2_i=S(S+1)$ are satisfied (using $[a_i,a_i^\dagger]=1$).
Let $\left\{\vec z_i\right\}$ be a classical ground state of Eq. \[eq:Heisenberg\], minimizing $E=\frac{1}{2}\sum_{ij}J_{ij} \vec z_i\cdot\vec z_j$ with $\vec z_i^2=1$. These directions can be used as local quantization axes: we use Eq. \[eq:HP\] in a local (orthogonal) frame $(\vec x_i,\vec y_i,\vec z_i=\vec x_i \wedge \vec y_i)$ adapted to the classical ground state. Under the assumption that $\vec S_i$ shows small deviations from the classical vector $S\vec z_i$, the typical number $\langle a^\dagger a\rangle$ of Holstein-Primakoff bosons should be small compared to $S$. We can therefore simplify $S^+$ (and $S^-$) in Eq. \[eq:HP\] by keeping only $\sqrt{2S}$ in the square roots, to obtain [@anderson52] $$\begin{aligned}
\vec S_i \simeq \left((S+\frac{1}{2})-\vec\pi_i^2\right)\vec z_i
+\sqrt{2S}\vec \pi_i
\label{eq:Spi}\end{aligned}$$ where $$\begin{aligned}
\vec\pi_i&=&\frac{1}{2}(a_i+a_i^\dagger)\vec x_i + \frac{1}{2i}(a_i-a_i^\dagger)\vec y_i\\
\vec\pi_i^2&=&a_i^\dagger a_i+\frac{1}{2}\\
{\rm and}\; \vec z_i\cdot \vec \pi_i&=&0.\end{aligned}$$ Replacing Eq. \[eq:Spi\] in the Hamiltonian gives $$\begin{aligned}
H=\frac{1}{2}(S+\frac{1}{2})^2\sum_{ij}J_{ij}\; \vec z_i\cdot\vec z_j
+ S\sum_{ij}J_{ij}\; \vec \pi_i\cdot\vec \pi_j \nonumber\\
-\frac{1}{2}S \sum_{ij}J_{ij}\; \left( \vec\pi_i^2 +\vec\pi_j^2\right)\vec z_i\cdot\vec z_j
+\mathcal{O}(S^0).
\label{eq:Hpi2}\end{aligned}$$ The first term is a constant, proportional to the classical energy $E_0$. The two other terms, proportional to $S$, are quadratic in the boson operators and describe the spin fluctuations as a set of coupled harmonic oscillators.[^11] The positions $q_i=\frac{1}{\sqrt{2}}(a_i+a^\dagger_i)$ and momenta $p_i=\frac{1}{\sqrt{2}i}(a_i-a^\dagger_i)$ operators of these oscillators can be conveniently grouped into a column vector of size $2N$ ($N$ is the total number of spins): $${\bf V}=\left[
\begin{array}{c} q_1\\ \colon \\q_N\\ p_1 \\ \colon \\p_N\end{array}
\right]$$ so that $H$ becomes $$\begin{aligned}
H=(S+\frac{1}{2})^2 E_0+\frac{S}{2} {\bf V}^t \mathcal{M} {\bf V},
\label{eq:HM}\end{aligned}$$ where $\mathcal{M}$ is a $2N\times2N$ matrix given by $$\mathcal{M}=\left[
\begin{array}{cc}
J^{xx}-J^{zz} & J^{xy} \\
(J^{xy})^t & J^{yy}-J^{zz}
\end{array}
\right]$$ and the $N\times N$ matrices $J^{xx}$, $J^{yy}$, $J^{xy}$ and $J^{zz}$ are defined by: $$\begin{aligned}
J^{xx}_{ij}=J_{ij}\;\vec x_i \cdot \vec x_j \;\;,\;\; J^{yy}_{ij}=J_{ij}\;\vec y_i \cdot \vec y_j \;\;,\;\; J^{xy}_{ij}=J_{ij}\;\vec x_i \cdot \vec y_j \\
{\rm and} \;\;\; J^{zz}_{ij}=\delta_{ij} \sum_k\; J_{ik}\vec z_i\cdot \vec z_k.\end{aligned}$$
Bogoliubov transformation
-------------------------
Diagonalizing $H$ amounts to find bosonic creation operators $b_\alpha^\dagger$ and corresponding energies $\omega_\alpha \geq 0$ such that $H=\sum_\alpha \omega_\alpha \left( b_\alpha^\dagger b_\alpha+\frac{1}{2}\right)$ (up to a constant). A necessary condition is that the operator $b_\alpha^\dagger$ and $b_\alpha$ are “eigenoperators” of the commutation with $H$, for the eigenvalues $\omega_\alpha$ and $-\omega_\alpha$ respectively: $\left[H,b_\alpha^\dagger\right]=\omega_\alpha b_\alpha^\dagger$ and $\left[H,b_\alpha\right]=-\omega_\alpha b_\alpha$. We thus seek the eigenvectors of the action of $\left[H,\bullet\right]$ in the space of linear combinations of $q_i$ and $p_j$. The commutators of $H$ (Eq. \[eq:HM\]) with the operators $q$ and $p$ are simple to obtain using $[q_i,q_j]=[p_i,p_j]=0$ and $[q_i,p_j]=i\delta_{ij}$. For an arbitrary linear combinations of the $q_i$ qnd $p_i$ parametrized by the complex numbers $x_1,\cdots,x_{2N}$ the result is $$\begin{aligned}
\left[H,x_1 q_1+x_N q_N + x_{N+1}p_1+\cdots x_{2N}p_N\right] \nonumber\\
= y_1 q_1+y_N q_N + y_{N+1}p_1+\cdots y_{2N}p_N\end{aligned}$$ with the coefficients $y_1,\cdots,y_{2N}$ given by $$\begin{aligned}
\left[\begin{array}{c}y_1 \\ \vdots \\y_{2N}\end{array}\right]=i S \; \mathcal{M} \left[
\begin{array}{ccc}
0 & &{\bf 1} \\
\\
-{\bf 1} & & 0
\end{array}
\right]
\left[\begin{array}{c}x_1 \\ \vdots \\x_{2N}\end{array}\right]\end{aligned}$$ where ${\bf 1}$ is the $N\times N$ identity matrix. So, finding the operators $b_\alpha^\dagger$ (spin-wave creation operators) amounts to find the eigenvectors of the “commutation matrix” $\mathcal{C}=i \mathcal{M} \left[\begin{array}{cc}0&{\bf 1}\\-{\bf 1}&0\end{array}\right]$.
But $\mathcal{C}$ is not symmetric and cannot always be fully diagonalized (contrary to $\mathcal{M}$). It can be shown that if all the eigenvalues of $\mathcal{M}$ were [*strictly*]{} positive, $\mathcal{C}$ could be diagonalized, its eigenvalues would be real and come in pairs $-\omega$,$\omega$.[^12]
However, $\mathcal{M}$ does have some zero eigenvalues. The matrix $\mathcal{M}$ is not specific to the quantum spin problem. The quadratic form describing the classical energy variation for a small perturbation around the chosen classical ground state $\left\{\vec z_i\right\}$, is described by [the same matrix]{} $\mathcal{M}$.[^13] In particular, if the classical ground state admits some zero energy (infinitesimal) spin rotations, $\mathcal{M}$ posses some eigenvector for the eigenvalue 0. Because global rotations should not change the energy, $\mathcal{M}$ has at least two zero eigenvalues. Still, these global rotations do not cause difficulties in diagonalizing the spin-wave Hamiltonian, they just correspond to some $\omega_\alpha=0$ (the associated collective coordinate $Q$ [*and*]{} conjugate momentum $P$ simply do not appear in $H$).
Zero modes on the kagome lattice
--------------------------------
However, some Heisenberg models admit classical zero modes (hence zero eigenvalues in $\mathcal{M}$) which do not correspond to global rotations. As an example, consider the Heisenberg model on the kagome lattice [@chs92] (for another classic example, the $J_1$-$J_2$ model on the square lattice, see [@cd88]). Any classical spin configuration such that the sum $\vec z_i + \vec z_j + \vec z_k$ vanishes on each triangle $(ijk)$ minimizes the classical energy. Among the numerous ways to achieve these conditions, are the [*planar*]{} ground states, where all the spins lie in the same plane. In such a state, the spins take only three possible directions, $\vec a$, $\vec b$ and $\vec c$ at 120 degrees from each other. On the kagome lattice, there is an exponential number of ways to assign these three orientations such that the same letter is never found twice on the same triangle (three-coloring problem, see Fig. \[fig:kagomeABC\]). Now, choose one of these “$abc$” states, and find a closed loop of the type $ababab\cdots$. Because of the three-coloring rule, the spins which are neighbors of this loop all point in the $\vec c$ direction. Now, we can rotate the spins of the loop about the $\vec c$ axis by any angle. This transforms the planar ground state into another (non planar) ground state, without any energy cost. So, for a generic planar ground state, we get as many zero modes (in $\mathcal{M}$) as closed loops with two alternating “colors”. This number typically grows like the number of sites in the system.
![Classical planar ground state on the kagome lattice. The loops where the spins alternate between the $\vec a$ and $\vec b$ directions are marked with dashed lines, they host independent zero modes (by rotation around the $\vec c$ axis).[]{data-label="fig:kagomeABC"}](kagomeABC){height="5cm"}
What are the consequences of such classical zero modes for the quantum problem ? As explained previously, the operators describing the two transverse directions along which the spins can deviate from the $\vec z_i$ axis obey the same commutation rules (at leading order in the $1/S$ expansion) as the position $q$ and momentum $p$ of an harmonic oscillator. In the case of the kagome loop modes discussed above, the energy is zero in one direction (rotation about the $\vec c$ direction), and quadratic in the other direction. Using the associated collective coordinate $P$ and $Q$, we expect the Hamiltonian to be proportional to $H=\frac{1}{2}(P^2+\omega^2 Q^2)$ with $\omega=0$, since there is no classical energy cost for spin deviation in the direction $Q$. The corresponding commutation matrix is $C=i\left[\begin{array}{cc}0&\omega^2\\-1 &0\end{array}\right]$ and cannot be diagonalized when $\omega=0$, as anticipated. In general, each such local zero mode will lead to an irreducible $2\times2$ Jordan block of this kind.[^14] The ground state $|0\rangle$ of the oscillator is simple to obtain and corresponds to a zero point motion of the coordinate $Q$ which [*diverges*]{} when $\omega\to 0$ (no restoring force, like for a free particle) : $\langle 0|Q^2|0\rangle=\frac{1}{2\omega}$.
As long as the the number of such zero modes is [*finite*]{} in the thermodynamic limit (this is the case when the classical ground state has no special degeneracy, beyond those implied by global rotations), the divergences above have a zero measure and do not cause divergences in the number of bosons $\langle 0|a^\dagger_i a_i|0\rangle$,[^15] which measures the strength of the deviations from the classical state. In such a case, the Néel ordered state is stable with respect to quantum fluctuations, at least for large enough $S$.[^16] On the other hand, if the number of such modes grow like $N$, the average number of bosons diverge and the spin-wave expansion breaks down (the initial assumption that $\langle 0|a^\dagger_i a_i|0\rangle$ is finite and small compared to $S$ cannot be satisfied).
At this point, a route to obtain a QSL appears to look for a lattice where the classical model has a sufficient number of “soft” modes, so that the zero point motion of the spins restore the rotation invariance and destroy the long range spin spin correlations. This condition is realized on the kagome lattice, where indeed all numerical studies concluded to the absence of Néel order in this system (at least for $S=\frac{1}{2}$). However, the semi classical spin wave theory described here breaks down. As discussed in the next sections, QSL states in Mott insulators possess some internal topological properties which are missed by the simple picture of a “disordered” state which would just be the quantum analog of a high temperature phase.
Lieb-Schultz-Mattis theorem, and Hastings’s extension to $D>1$: ground state degeneracy in gapped spin liquids {#sec:LSMH}
==============================================================================================================
The Lieb-Schultz-Mattis theorem [@lsm61] was originally derived for spin chains and spin ladders [@affleck88; @oya97] and was recently extended to higher dimensions in an important work by Hastings [@hastings04] (see also [@oshikawa00] for an intuitive topological argument valid in any dimension, and [@ns07] for a mathematically rigorous proof). It applies to spin Hamiltonians which are translation invariant in one direction (say $x$), have a conserved magnetization $S^z_{\rm tot}=\sum_i S^z_i$, and short range interactions. In addition, the model should have periodic boundary conditions in the $x$ direction. Although more general interactions can easily be considered,[^17] we concentrate for simplicity on spin-$S$ Heisenberg models, written as in Eq. \[eq:Heisenberg\] (with $J_{i,j}=J_{i+x,j+x}$ to respect the translation invariance).
Following [@oshikawa00], we define the [*cross section*]{} as all the sites sitting at a given value of $x$. By translation invariance, all cross sections are equivalent and contain $C$ sites (Fig. \[fig:ring\]). In a spin chain, each cross section contains a single site. In an $n-$ leg spin ladder, $C=n$ sites. In a square lattice, $C=L_y$. On a $D$-dimensional lattice with $n$ sites per unit cell, $C=nL^{D-1}$, etc. We note $L_x$ the system length in the $x$ direction, and therefore $CL_x$ is the total number of sites. Finally we define $m^z=\frac{1}{CL_x}\langle0|S^z_{\rm tot}|0\rangle$ as the ground state magnetization per site.
![A lattice model which is translation invariant and periodic in the $x$ direction can be viewed as a ring. The interactions $J_{ij}$, indicated by dashed lines, are invariant in the $x$ direction but otherwise arbitrary. In this example, each cross section has $C=3$ sites. []{data-label="fig:ring"}](ring){height="3cm"}
The theorem says that if $C(S+m^z)$ is [*not an integer*]{}, the ground state is either degenerate, or the spectrum has gapless excitations in the thermodynamic limit. In other words, if $C(S+m^z)\notin\mathbb{Z}$ the system [*cannot have a unique ground state and a finite gap to excited states*]{} in the thermodynamic limit. Although the proof in 1D [@lsm61] and Oshikawa’s topological argument [@oshikawa00] (Sec. \[ssec:oshikawa\]) are relatively simple, the proof appears quite involved for $D>1$, and will not be discussed here.
What is the relation between the LSMH theorem and QSL ? In most AF Heisenberg models on a finite-size lattice, $|0\rangle$ is a singlet and $m^z=0$. If we focus on the case $S=\frac{1}{2}$, the theorem forbids a single ground state and a gap when $C$ is [*odd*]{}. In particular, if the lattice is two dimensional and describes a Mott insulator, the unit cell has an odd number $n$ of sites and any odd $L_y$ can be chosen to get and odd $C=n L_y$ (note that the total number of sites is still even if $L_x$ is even). If we assume that a [*gapped*]{} QSL is realized (for an example which fits in the LSMH conditions, see for instance [@bfg02]), its ground state must be [*degenerate*]{} (with periodic boundary conditions). Usually, ground state degeneracies are the signature of some spontaneous symmetry breaking. However, by definition, a QSL respect all lattice symmetries. The degeneracy imposed by the LSMH theorem cannot be understood from this conventional point of view and is a hint that (gapped) QSL wave function possess some interesting topological properties, which correspond to the notion of “topological order” introduced by Wen [@wen89; @wen91] for spin systems and Wen and Niu [@wn90] in the context of the fractional quantum Hall effect. As we will briefly discuss at the end, this topological degeneracy is deeply related to the exotic nature of the elementary excitations in a QSL.[^18]
Oshikawa’s topological argument {#ssec:oshikawa}
-------------------------------
Oshikawa’s argument is somehow related to Laughlin’s argument [@laughlin81] for the quantization of the transverse conductivity in the quantum Hall effect. First, a “twisted” version of the Hamiltonian is introduced: $$H_\theta=\frac{1}{2}\sum_{ij} J_{ij}\left[
S_i^zS_j^z
+\frac{1}{2}\left(
e^{i\theta(x_i-x_j)/L_x} S_i^+S_j^- + {\rm H.c}
\right)
\right]\label{eq:Htwisted}$$ where $ 0 \leq x_i < L_x$ is the $x$-coordinate of site $i$. It is simple to show that the spectra of $H_0$ and $H_{2\pi}$ are the same, since the unitary operator $$U=\prod_{i} \exp\left(2i\pi \frac{x_i}{L_x}S^z_i\right)$$ maps $H_0$ onto $H_{2\pi}$: $$U H_0 U^{-1} = H_{2\pi}
\label{eq:UH}$$ (the calculation simply uses $e^{i\theta S^z_i}S^+_ie^{-i\theta S^z_i}=S^+_i e^{i\theta}$).
Starting with a spectrum of $H_0$ which is gapped, we further assume that [*the gap of $H_\theta$ remains finite when $\theta$ goes from 0 to $2\pi$*]{}.[^19] On can follow the ground state of $H_{\theta}$, which does not cross any other energy level as $\theta$ is varied. Assuming that the ground state $|0\rangle$ of $H_0$ is unique and using the finite gap hypothesis, it must evolve to the ground state of $H_{2\pi}$, denoted $|2\pi\rangle$. Through Eq. \[eq:UH\], both states are related: $|2\pi\rangle=U^{-1}|0\rangle$. However, the operator $U$ does not always commute with the translation operator $T$ and may change the momentum. The precise relation is $$T U = U T \exp\left(2i\pi \frac{S^z_{\rm tot}}{L_x}\right)\exp\left(2i\pi C S\right) .$$ The first phase factor, also equal to $2\pi C m^z$, comes from the shift by $2\pi/L_x$ of the local rotation angles after a translation. The second phase factor corrects the $2\pi$ jump of the rotation angle when passing from $x=L_x-1$ to $x=0$. This relation implies that the momentum $k_0$ of $|0\rangle$ (defined by $T|0\rangle=e^{ik_0}|0\rangle$) and the momentum $k_{2\pi}$ of $|2\pi\rangle=U^{-1}|0\rangle$ are related by $$k_{0}=k_{2\pi} +2\pi C(S+m^z) \label{eq:k}$$ But $H_{\theta}$ is translation invariant (commutes with $T$) and the momentum of each state (quantized for finite $L_x$) cannot change with $\theta$. So $|0\rangle$ and $|2\pi\rangle$ have the same momentum and $k_0=k_{2\pi}\;[2\pi]$. From Eq. \[eq:k\], we get that $C(S+m^z)$ must be a integer.
![Schematic spectrum of the twisted Hamiltonian (Eq. \[eq:Htwisted\]) as a function the angle $\theta$, in the case where $C(S+m^z)$ is a half integer. []{data-label="fig:twist"}](twist){height="3cm"}
Anderson’s short range resonating valence-bond picture {#sec:RVB}
======================================================
In the $1/S$ expansion, it is assumed that the spins experience small fluctuations about a well defined direction and that spin-spin correlations are long ranged. This is of course incompatible with having a rotationally invariant QSL state. To gain some intuition about what a QSL wave function may look like, it is instructive to start from a completely opposite limit: a spin singlet state with extremely short range correlations. A [*short range valence-bond*]{} (VB) state is such a wave function, it is the direct product of $S=0$ states $|[ij]\rangle=\frac{1}{\sqrt{2}}\left(|\uparrow_i\downarrow_j\rangle-|\downarrow_i\uparrow_j\rangle\right)$ on pairs of sites : $$|{\rm VB}\rangle =|[i_0i_1]\rangle\otimes|[i_2i_3]\rangle\otimes|[i_4i_5]\rangle\otimes\cdots|[i_{N-1}i_N]\rangle$$ where each site of the lattice appears exactly once (Fig. \[fig:triVB\]). Such a VB state is said to be short range if all pairs of sites coupled in a singlet are at a distance $|{\bf r}_{i_{p}} - {\bf r}_{i_{p+1}}|$ smaller than or equal to some fixed length $r_{\rm max}$(much smaller that the lattice size). The simplest case is $r_{\rm max}=1$, where each spin forms a singlet with one of its nearest neighbors.
![A short range valence bond state on the triangular lattice. The singlet pairs are marked with ellipses.[]{data-label="fig:triVB"}](triVB){height="3cm"}
In a VB state, the spin-spin correlations are short ranged: $\langle {\rm VB} | \vec S_i\cdot \vec S_j|{\rm VB}\rangle=0$ if $|{\bf r}_{i_{p}} - {\bf r}_{i_{p+1}}|>r_{\rm max}$. For a nearest neighbor Heisenberg model on a bipartite lattice, one can compare the (expectation value of the) energy of a nearest neighbor VB state, with that of the simple two-sublattice Néel state $|\uparrow\downarrow\uparrow\downarrow\cdots\rangle$. The VB energy is $e_{\rm VB}=-J\frac{3}{8}$ per site and the Néel one is $e_{\rm N}=-J\frac{z}{8}$, where $z$ is the number of nearest neighbors. If the lattice is not bipartite but admits a three-sublattice classical ground states (with spins pointing at 120 degrees from each other), the energy of a classical Néel state is $e_{\rm N}=-J\frac{z}{16}$. From this, we observe for instance that the VB energy is [*lower*]{} than $e_{\rm N}$ on the kagome lattice. More generally, this simple variational comparison shows that a low coordination $z$ and frustrated interactions (which increase the number of sublattices in the classical ground state) tend to favor VB states, and thus possible QSL states.
In fact there are many (frustrated and Heisenberg-like) toy models where some/the nearest neighbor VB states are [*exact*]{} ground states. The most famous example is the Majumdar-Gosh model [@mg69]. Consider the spin-$\frac{1}{2}$ Heisenberg chain with first- ($J_1$) and second- ($J_2$) neighbor couplings. At $J_1=2J_2>0$ we have $$H_{\rm MG}=2\sum_i \vec S_i \cdot \vec S_{i+1}+\sum_i \vec S_i \cdot \vec S_{i+2}$$ and the (two-fold degenerate) ground states are exactly given: $$\begin{aligned}
|a\rangle &=&\cdots\otimes|[01]\rangle\otimes|[23]\rangle\otimes|[46]\rangle\otimes\cdots \\
|b\rangle &=&\cdots\otimes|[12]\rangle\otimes|[34]\rangle\otimes|[56]\rangle\otimes\cdots\end{aligned}$$ The proof can be done three steps. First, the Heisenberg Hamiltonian on three site $H_{ijk}=\vec S_i \cdot \vec S_j + \vec S_j \cdot \vec S_k + \vec S_k \cdot \vec S_i$ is written as $H_{ijk}=\frac{1}{2}(\vec S_i+\vec S_j+\vec S_k)^2-\frac{9}{8}$. In this form, proportional to the square of the total spin, it is clear that the eigenvalues of $H_{ijk}$ are $\frac{1}{2}S(S+1)-\frac{9}{8}$ with $S=\frac{1}{2}$ or $S=\frac{3}{2}$ (the only possible values of $S$ for three spin-$\frac{1}{2}$). So, if the sites $ijk$ are in a $S=\frac{1}{2}$ state, they minimize exactly $H_{ijk}$. Second, one expresses the Majumdar-Gosh Hamiltonian as $$H_{\rm MG}=\sum_i H_{i-1,i,i+1}.\label{eq:HMG}$$ Finally, one remarks that the dimerized states $|a\rangle$ and $|b\rangle$ always have one singlet among the sites $i-1,i,i+1$, which are therefore in a $S=1/2$ state. We conclude that $|a\rangle$ and $|b\rangle$ minimize all the terms in Eq. \[eq:HMG\] and are thus ground states of $H_{\rm MG}$.
The Majumdar-Gosh model is the simplest model of a family of spin models where exact VB ground states can be found.[^20] For instance, the Husimi cactus [@cd94] is a lattice constructed as a tree (no loops) of corner sharing triangles. Its geometry is locally similar to the kagome lattice but it has no closed loop (except of course for the triangles themselves). The argument above (writing the Hamiltonian as a sum of $H_{ijk}$) directly generalizes to this case and shows that any nearest neighbor VB state is a ground state. One can also mention the 2D Shastry-Sutherland Heisenberg model [@ss81], where a particular nearest neighbor VB is the unique ground state, and which has an experimental realization in SrCu$_2$(BO$_3$)$_2$ [@k99].
So far, we do not yet have any gapped [*liquid*]{} state.[^21] To obtain a qualitative idea of how VB states can be the building blocks of a gapped QSL, we will briefly explain the short range resonating valence bond (RVB) picture proposed by Anderson [@anderson73]. If we exclude the toy models discussed above, a VB state is generally not an eigenstate of the Heisenberg model. Starting from a nearest neighbor VB state, the Heisenberg Hamiltonian will induce some dynamics among the VB states. If we take the kagome example, a nearest neighbor VB state inevitably contains some “defect” triangles [*without*]{} any singlet.[^22] While the term $H_{ijk}$ leaves the VB state unchanged if the corresponding triangle has a singlet bond, the three VB touching $i$, $j$ and $k$ will be moved by $H_{ijk}$ if $(ijk$) is a defect triangle. The ground state can be viewed as a linear combination of (many) VB configurations (not necessarily nearest neighbor). Anderson suggested that, with appropriate interactions and lattice geometry, the ground state wave function could be “delocalized” over a large part of the subspace spanned by short range VB states. By forming a linear superposition of a large number of very different VB states, the system may restore all the lattice symmetries (which are broken by an individual VB state) and form a QSL.
A more formal approach to this idea will be discussed in Sec. \[sec:SB\], but this picture can already be used to anticipate the nature of the magnetic excitations in such a short range RVB liquid. To this end, we first consider a 2D model where one ground state is equal to (or dominated by) [*one*]{} particular VB state. Contrary to the Anderson’s RVB liquid, the wave function is localized in the vicinity of one particular VB state. It can be thought as a 2D analog of the Majumdar Gosh chain, where the ground state is a spatially regular arrangement of singlet bonds. Many 2D models are known to realize such VB [*crystals*]{} (VBC) [@ml05], and we refer to Ref. [@glkm07] for a recent example where the exact ground states are known. In a VBC, a finite energy excitation can be created by replacing a singlet bond by a triplet ($S=1$), with an energy cost proportional to $J$. But is it possible to construct two [*separated*]{} spin-$\frac{1}{2}$ excitations in such a system ? As a trial state, one can place two remote spins “up” (two spinons excitations) at sites 0 and $i$. Then, to minimize the energy, the regular VB structure of the ground state should be reconstructed as much as possible. However, due to the spinons, the regular pattern cannot be fully reconstructed between 0 and $i$, and a “string” of misaligned VB is unavoidable. The unpaired spins behave as a topological defect in the crystalline order. So, two remote spinons perturb the ordered VB background, not only in their vicinity, but all the way between them. They lead to an energy cost which is proportional to their separation.[^23] So, isolated spinons are not finite energy excitations in a VBC. The ordered VB background is a medium which confines the spinons in pairs. Since an RVB state should instead be viewed as a [*liquid*]{} (no broken symmetry, no long range order), it is reasonable to expect the spinons to be able to propagate as independent particles. As we will see in the next section, the proper way to address this question of confinement and deconfinement of spinons is to understand the emergence of gauge degrees of freedom in these systems.
Schwinger bosons, large-$\mathcal N$ limit, and $\mathbb{Z}_2$ topological phase {#sec:SB}
================================================================================
Schwinger bosons representation
-------------------------------
The spin wave approach is a large-$S$ approach and is unable to capture highly quantum states which are rotationally symmetric, such as RVB wave functions. From the discussion of Sec. \[sec:RVB\], it is natural to look for a description in terms of [*singlet fields*]{} leaving on bonds, and able to describe the presence or absence of a singlet between two sites. Such variables appear naturally when using the Schwinger boson representation of the spin operators [@aa88; @auerbachbook].
At each site, two types of bosons carrying a spin “up” and “down” are introduced: $a^\dagger_{i\uparrow}$ and $a^\dagger_{i\downarrow}$, and the spin operators are represented as bilinears in the boson creation and annihilation operators $$\begin{aligned}
S^z_i=\frac{1}{2}\left(
a^\dagger_{i\uparrow}a_{i\uparrow}
-a^\dagger_{i\downarrow}a_{i\downarrow}
\right)
\;\;,\;\;
S^+_i=a^\dagger_{i\uparrow}a_{i\downarrow}
\;\;,\;\;
S^-_i=a^\dagger_{i\downarrow}a_{i\uparrow}\end{aligned}$$ With these relations, the commutation relations $[S^\alpha_i,S^\beta_i]=i\epsilon^{\alpha\beta\delta}S^\delta_i$ are automatically verified. The total spin reads $
\vec S_i^2=\frac{n_i}{2}\left(\frac{n_i}{2}+1\right)
$, where $n_i=a^\dagger_{i\uparrow}a_{i\uparrow}
+a^\dagger_{i\downarrow}a_{i\downarrow}$ is the total number of bosons at site $i$. To fix the length of the spins, the following constraint must therefore be imposed on physical states: $$a^\dagger_{i\uparrow}a_{i\uparrow}
+a^\dagger_{i\downarrow}a_{i\downarrow}=2S
\label{eq:cons}$$
With this representation,[^24] the Heisenberg interaction is of degree four in the boson operators and can be written $$\begin{aligned}
\vec S_i \cdot \vec S_j = S^2-\frac{1}{2} (A_{ij})^\dagger A_{ij} \label{eq:SSA2}\\
{\rm with }\;\;A_{ij}=a_{i\uparrow}a_{j\downarrow}-a_{i\downarrow}a_{j\uparrow}.\end{aligned}$$ The bond operators $A_{ij}^\dagger$ behave as a singlet creation operators: $A_{ij}^\dagger$, when applied onto the boson vacuum, creates a spin singlet $|\uparrow_i\downarrow_j\rangle-|\downarrow_i\uparrow_j\rangle$ and, from Eq. \[eq:SSA2\], $A_{ij}^\dagger A_{ij}$ is proportional to the number (0 or 1) of a singlet between sites $i$ and $j$. In addition, $A_{ij}$ is invariant under rotations: redefining the bosons by an $SU(2)$ matrix $P$: $\left[\begin{array}{c}a_\uparrow\\a_\downarrow\end{array} \right]
\to P \left[\begin{array}{c}a_\uparrow\\a_\downarrow\end{array} \right]$ leaves $A_{ij}$ unchanged.[^25]
Mean field approximation
------------------------
Arovas and Auerbach [@aa88] suggested an approximation in which the interaction is decoupled using mean-field expectation values $$A_{ij}^\dagger A_{ij} \longrightarrow A_{ij}^\dagger \langle A_{ij} \rangle
+ \langle A_{ij}^\dagger\rangle A_{ij}
- |\langle A_{ij}^\dagger\rangle|^2
\label{eq:MF}$$ and to replace the constraint (Eq. \[eq:cons\]) by a condition on the [*average*]{} number of boson per site $$\langle a^\dagger_{i\uparrow}a_{i\uparrow}
+a^\dagger_{i\downarrow}a_{i\downarrow} \rangle =2S. \label{eq:n2S}$$ By this replacement, the Hamiltonian becomes quadratic in the boson operator $$\begin{aligned}
H \longrightarrow H_{\rm MF}[Q_{ij}^0,\lambda_j^0]=-\frac{1}{2}\sum_{ij}
\left(
A_{ij}^\dagger Q_{ij}^0+ \bar{ Q_{ij}^0} A_{ij}
\right)\nonumber\\
-\sum_i \lambda_i^0 \left(
a^\dagger_{i\uparrow}a_{i\uparrow}
+a^\dagger_{i\downarrow}a_{i\downarrow} -2S
\right)
+{\rm cst}.\end{aligned}$$ A chemical potential $\lambda_i^0$ has been introduced at each site to tune the boson densities so that they satisfy Eq. \[eq:n2S\]. The mean field Hamiltonian $H_{\rm MF}$ (and thus its ground state $|0\rangle$) depends on the complex parameters $Q_{ij}^0$ (one for each pair of sites $ij$ where $J_{ij}\ne 0$). These parameters have to be adjusted to satisfy the self-consistency conditions on each bond $$Q_{ij}^0 = \frac{1}{2}J_{ij}
\langle 0 | a_{i\uparrow}a_{j\downarrow}-a_{i\downarrow}a_{j\uparrow}|0\rangle.
\label{eq:SC}$$ As in the spin wave approach, the Heisenberg model has been reduced to a quadratic boson model (here with some self consistency conditions). However, the crucial difference is that the present formalism does not impose any preferred spin direction: giving a finite expectation value $A_{ij}^0\ne 0$ to the operator $A_{ij}$ does not break the $SU(2)$ symmetry – which is a necessary condition to describe a QSL.
Generally speaking, two family of solutions can be found at this mean field level. In the first class, favored when $S$ is large, the Schwinger boson [*Bose-condense*]{} in some particular mode. Because they carry a spin index, such condensate state (spontaneously) breaks the $SU(2)$ symmetry. These solutions describe Néel states with long range spin-spin correlations. In such cases, the Schwinger boson mean-field theory is essentially equivalent to the spin wave approach (Sec. \[sec:SW\]).
The second class corresponds to (mean field) QSL states. There, the ground state is rotationally invariant, and the Bogoliubov quasi particles obtained by diagonalizing $H_{\rm MF}$ are gapped. Since the corresponding creation operators, $b_{\uparrow,\alpha}$ and $b_{\downarrow,\alpha}$, are linear combinations of the original bosons, these excitations also carry a spin $\frac{1}{2}$. The most important question is whether the existence of these deconfined (free in the mean field approximation) spinons is an artifact of the mean field approximation, or if they could survive in some Heisenberg spin model. In the first a case, the inclusion of the fluctuations that were neglected would confine the spinons and would deeply change the nature of the ground state. The mean-field picture of a fully symmetric state with non interacting spinons excitation is then qualitatively incorrect. Another possibility is that the spinons remains deconfined, even in presence of fluctuations. In that case, the mean-field approximation is a very useful starting point. We will discuss in Sec. \[ssec:Z2\] a scenario where it is the case. But before, we need to introduce the basic formalism that is needed to describe the fluctuations about the mean field solution, and emergence of gauge degrees of freedom in the system. The central question concerning the long distance and low energy properties of the system will be whether these gauge degrees of freedom confine or not the spinons.
Large $\mathcal N$, saddle point
--------------------------------
To discuss the role of the fluctuations neglected in Eq. \[eq:MF\], it is necessary to formulate the mean field approximation as a saddle point approximation in path integral formulation of the model. It will then be possible to identify the structure of the most important fluctuations about the saddle point. To do so, one duplicates $\mathcal N$ times the two species of bosons ($\uparrow$ and $\downarrow$). In addition to the site and up/down indices $\sigma$, the boson operators now carry an additional “flavor” index $m=1,\cdots,\mathcal{N}$. The Hamiltonian and the constraint are then generalized to $$\begin{aligned}
H&=&-\frac{1}{2\mathcal{N}}\sum_{ij}J_{ij} A_{ij}^\dagger A_{ij} \\
A_{ij}&=&\sum_{m=1}^\mathcal{N} a_{im\uparrow}a_{jm\downarrow}-a_{im\downarrow}a_{jm\uparrow}\end{aligned}$$ and $$\sum_{m=1}^\mathcal{N}
a^\dagger_{im\uparrow}a_{im\uparrow}
+a^\dagger_{im\downarrow}a_{im\downarrow}=2\mathcal{N}S. \label{eq:Ncons}$$ For $\mathcal{N}=1$, this model is Heisenberg model with $SU(2)$ symmetry. For $\mathcal{N}>1$, this model has an enlarged symmetry given by the group $Sp(\mathcal{N})$.[^26] $S$ is a parameter of the model, and is no longer related to a representation of $SU(2)$ if $\mathcal{N}>1$. The bond operator $A_{ij}$ is a sum over all the flavors. For this reason, in the limit where $\mathcal{N}$ is very large, the fluctuations of $A_{ij}$ become negligible compare to its expectation value and the approximation made in Eq. \[eq:MF\] becomes exact.
A formal way to establish this result is to adopt a formulation of model where the partition function $\mathcal{Z}={\rm Tr}\left[e^{-\beta H}\right]$ at temperature $T=\beta^{-1}$ is expressed as a coherent state path integral over complex variables $z_{im\sigma}(\tau)$ (in correspondence with the boson operators $a_{im\sigma}$) which are periodic functions of the imaginary time $\tau\in[0,\beta [$. In this formalism the partition function reads [^27] $$\begin{aligned}
\mathcal{Z}&=&\int \mathcal{D}[z_{im\sigma}(\tau),\lambda_i(\tau)]
\exp\left(-\int_0^\beta L_0\;d\tau\right) \\
L_0&=&\sum_{i\;m\;\sigma} \bar z_{im\sigma} \partial_\tau z_{im\sigma}
-\frac{1}{2\mathcal{N}}\sum_{ij}J_{ij} A_{ij}^\dagger A_{ij} \nonumber \\
&&+i\sum_{i\;m} \lambda_i\left(
\bar z_{im\uparrow}z_{im\uparrow}+\bar z_{im\downarrow}z_{im\downarrow}-2S
\right)
\\
A_{ij}&=&\sum_{m=1}^\mathcal{N} \left(
z_{im\uparrow}z_{jm\downarrow}-z_{im\downarrow}z_{jm\uparrow}
\right),\end{aligned}$$ where a Lagrange multiplier $\lambda$ has been introduced at each lattice site and each time step to enforce the constraint (Eq. \[eq:Ncons\]) exactly (to simplify the notations, the $\tau$ dependence of all fields is implicit).
Now, a Hubbard-Stratonovich transformation is performed : $$\begin{aligned}
\mathcal{Z}&=&\int \mathcal{D}[z_{im\sigma}(\tau),\lambda_i(\tau),Q_{ij}(\tau)]
\exp\left(-\int_0^\beta L_1\;d\tau\right) \\
L_1&=&\sum_{i\;m\;\sigma} \bar z_{im\sigma} \partial_\tau z_{im\sigma}
+\sum_{ij}\left(
\frac{2\mathcal{N}}{J_{ij}} |Q_{ij}|^2 - \bar Q_{ij} A_{ij} - Q_{ij} \bar A_{ij}
\right) \nonumber \\
&&+i\sum_{i\;m} \lambda_i\left(
\bar z_{im\uparrow}z_{im\uparrow}+\bar z_{im\downarrow}z_{im\downarrow}-2S
\right)
\label{eq:L1}\end{aligned}$$ This new formulation involves an additional complex field $Q_{ij}$ on each bond. The equivalence of $L_1$ with the initial Lagrangian $L_0$ can simply be checked by performing the Gaussian integrations over $Q_{ij}(\tau)$ for each bond and each time step: $\int \mathcal{D}[Q_{ij}(\tau)] \exp\left(-\int_0^\beta L_1\;d\tau\right)
= \exp\left(-\int_0^\beta L_0\;d\tau\right)$ (up to a multiplicative constant). At this point, the $\mathcal N$ flavors of particles are no longer coupled to each other, but are coupled to a common bond field $Q_{ij}$. So, for a fixed space-time configuration of $Q$, we have $\mathcal{N}$ independent copies of the same boson system. In addition, the Lagrangian $L_1$ is now quadratic in the $z$ variable. We note $G_{Q,\lambda}^{-1}$ the corresponding quadratic form, a big matrix which has space ($i$), time ($\tau$), spin ($\sigma$) and complex conjugacy ($z$ versus $\bar z$) indices (but no flavor index), and depends on the auxiliary field $Q$ and $\lambda$. $L_1$ is then $$\begin{aligned}
L_1&=&\sum_{ij}
\frac{2\mathcal{N}}{J_{ij}} |Q_{ij}|^2 -2i\mathcal{N}S\sum_{i} \lambda_i \nonumber \\
&&+\sum_m
\left[\bar z_{i\sigma}(\tau) ; z_{i\sigma}(\tau)\right]
G_{Q,\lambda}^{-1}
\left[\begin{array}{c}z_{j\sigma'}(\tau')\\\bar z_{j\sigma'}(\tau')\end{array}\right] \end{aligned}$$ Performing the Gaussian integral over the $z$ fields is now simple, as it gives $({\rm det}[G])^\mathcal{N}$, also equivalent to $e^{\mathcal{N}{\rm Tr}[\log(G)]}$. The partition function is now expressed as a path integral with the fields $Q$ and $\lambda$ only, but with a complicated non-Gaussian weight: $$\begin{aligned}
\mathcal{Z}&=&\int \mathcal{D}[z_{i\sigma}(\tau),\lambda_i(\tau),Q_{ij}(\tau)]
\exp\left(-\mathcal{N} \int_0^\beta L_2\;d\tau\right) \\
L_2&=& +\sum_{ij}
\frac{2}{J_{ij}} |Q_{ij}|^2 -2iS\sum_{i} \lambda_i+{\rm Tr}[\log(G_{Q,\lambda})]\end{aligned}$$ Here, the flavor indices $m$ have disappeared and $\mathcal{N}$ only appears a global multiplicative factor in the action. With this formulation of the $Sp(\mathcal N)$ “spin” model, it is clear that, in the limit $\mathcal{N}\to \infty$ the partition function will be dominated by the configurations $(Q^0,\lambda^0)$ which are [*saddle points*]{} of the action $\mathcal{S}[Q,\lambda]=\int_0^\beta L_2\;d\tau$. In other words [*the fluctuations of $Q_{ij}$ and $\lambda_i$ are frozen when $\mathcal{N}\to\infty$*]{}. Such saddle points are obtained by requiring $$\begin{aligned}
\left.\frac{\partial \mathcal{S}}{\partial \lambda_i(\tau)}\right|_{Q^0,\lambda^0}=0
\;\;,\;\;
\left.\frac{\partial \mathcal{S}}{\partial Q_{ij}(\tau)}\right|_{Q^0,\lambda^0}=0. \\\end{aligned}$$ and in most cases they are found to be time independent $Q_{ij}^0(\tau),\lambda_i^0(\tau)\to Q_{ij}^0,\lambda_i^0$. The equations above can then be shown to be equivalent to the self consistency conditions of Eqs. \[eq:n2S\] and \[eq:SC\], with $Q_{ij}^0=\frac{J_{ij}}{2\mathcal{N}}\sum_m
\langle0|
a_{im\uparrow}a_{jm\downarrow}-a_{im\downarrow}a_{jm\uparrow}
|0\rangle
$.
Fluctuations about a saddle point and gauge invariance
------------------------------------------------------
We are now ready to discuss the fluctuations that are present when $\mathcal{N}$ is finite, where the field $Q_{ij}(\tau)$ is able to fluctuate around its mean field value $Q_{ij}^0$. Treating all the possible fluctuations is certainly very difficult, as it would amount to solve the original spin problem. A possible approach is to compute perturbatively the first $1/\mathcal{N}$ corrections to the mean field results [@aa88]. However, this can miss some important effect (instabilities) which are not perturbative in $1/\mathcal{N}$, and will generally not shed light on the issue of spinon confinement that we are interested in. Instead, as in [@rs89; @rs91], we will examine the qualitative structure of the fluctuation modes which are important for the long distance properties of the system. In particular, we would like to know if some fluctuations could confine the spinons (in which case the mean field picture is incorrect), or if the QSL state is stable at finite $\mathcal{N}$. As we will see, there are some fluctuations modes which are described by a gauge field [@rs89; @rs91] and mediate some (possibly long ranged) interaction between the spinon. The dynamics of this gauge field is therefore crucial to the physics of the spin system. In some cases this gauge field will be in a confining phase, and the $\mathcal{N}=\infty$ limit (where the fluctuations are frozen out) does not represent the physics of the finite $\mathcal{N}$ models [@rs89]. In some other situations, the gauge field has a deconfined phase and a QSL state with elementary spinon excitation is possible [@rs91].
First, it should be noticed that the description of the spin operators with Schwinger bosons is redundant in the sense that an arbitrary local change of phase in the boson operators does not change the physical spin operators. In the path integral formulation, this becomes a full space-time gauge invariance. The Lagrangian $L_1$ (Eq. \[eq:L1\]) is invariant under $$\begin{aligned}
z_{im\sigma}(\tau)&\longrightarrow& e^{i\Lambda_i(\tau)} z_{im\sigma}(\tau) \label{eq:zL}\\
Q_{ij}(\tau)&\longrightarrow& e^{i(\Lambda_i(\tau)+\Lambda_j(\tau))} Q_{ij}(\tau) \label{eq:QL}\\
\lambda_i(\tau)&\longrightarrow& \lambda_i(\tau)-\partial_\tau \Lambda_i(\tau)\end{aligned}$$ where $\Lambda_i(\tau)$ is some arbitrary angle at each site and time step.
However, this local $U(1)$ gauge invariance is broken to a smaller invariance group in the vicinity of a saddle point $(Q^0,\lambda^0)$. This can be illustrated the simpler context of a classical ferromagnetic Heisenberg model. A ground state is magnetized in one particular direction and thus breaks the $O(3)$ symmetry of the Hamiltonian. The theory for the (transverse) spin [*deviations*]{} around this ferromagnetic state has an $O(2)$ symmetry, an not $O(3$). The situation is similar for the fluctuations of the bond field $Q_{ij}$. Although the model has a local $U(1)$ gauge invariance, [*the action describing the fluctuations around $Q^0_{ij}$ have a lower invariance group*]{}. In the ferromagnet example, we look at the rotations under which the ground state is unchanged. Similarly, we look for the gauge transformations which leave $Q^0_{ij}$ unchanged. These transformations form the [*invariant gauge group*]{} (IGG) of the saddle point, a concept introduced by X. G. Wen [@wen02]. A gauge transformation $i\mapsto\Lambda_i$ belongs to the IGG of $Q^0_{ij}$ if it is static and satisfies $$Q_{ij}^0 = Q_{ij}^0 e^{i(\Lambda_i+\Lambda_j)}
\label{eq:IGG}$$ If the lattice made of the bonds where $Q^0_{ij}$ is non zero is [*bipartite*]{}, it is easy to show that $\Lambda_i=\theta$ on sublattice A and $\Lambda_i=-\theta$ on sublattice $B$ satisfies Eq. \[eq:IGG\] for any (global) angle $\theta$. In such a case, the IGG is isomorphic to $U(1)$. On the other hand, if the lattice of the bonds where $Q^0_{ij}\ne0$ is [*not*]{} bipartite, the IGG is isomorphic to $\mathbb{Z}_2$, since $\Lambda_i=\pi$ and $\Lambda_i=0$ are the only two solutions to Eq. \[eq:IGG\] when $Q_{ij}^0\ne0$.
The general result [@wen02] is that, among the fluctuations around the saddle point $Q^0$, some modes are described by a [*gauge field*]{}. with a gauge group given by the IGG. We will illustrate this result in the simple case IGG$=\mathbb{Z}_2$.[^28]
$\mathbb{Z}_2$ gauge field {#ssec:Z2}
--------------------------
If the IGG is $\mathbb Z_2$, the important fluctuations turn out to be fluctuations of the [*sign*]{} of $Q_{ij}$. We therefore parametrize these fluctuations in the following way $$Q_{ij}(\tau) = Q_{ij}^0 \;e^{i\mathcal{A}_{ij}(\tau)}
\;\;,\;\;
\mathcal{A}_{ij}(\tau)\in \{0,\pi\}. \label{eq:A}$$ where the field $\mathcal{A}_{ij}$ will play the role of a “discrete” ($\mathbb Z_2$) vector potential living on the links of the lattice (pairs of sites where $Q_{ij}^0\ne0)$.
Doing the integration over all the other fluctuation modes (amplitude fluctuations the bond field $Q_{ij}$, fluctuations of $\lambda_i$, etc.) in order to obtain an effective action for $\mathcal{A}_{ij}$ and the bosons $z_{i\sigma}$ only[^29] is formally possible, but it is of course a very difficult task in practice. One can instead determine the symmetry constraints, and, in a Landau-Ginzburg type of approach, construct the simplest action compatible with these symmetries.
For this, we consider the (static) local gauge transformation $i\mapsto\Lambda_i$ with the restriction $\Lambda_i\in \{0,\pi\}$. Because $\mathcal{A}_{ij}$ is defined modulo $2\pi$, $-\Lambda_j$ is equivalent to $+\Lambda_j$ and the transformation rules take the usual form (except for the discrete nature of $\mathcal A_{ij}$): $$\begin{aligned}
z_{i\sigma}&\longrightarrow& e^{i \Lambda_i} z_{i\sigma} \\
\mathcal{A}_{ij}&\longrightarrow & \mathcal{A}_{ij} + \Lambda_i-\Lambda_j.\end{aligned}$$ These local transformations form a very large symmetry group (2 to the power of the number of lattice sites) and severely constrain the effective Hamiltonian for these degrees of freedom. Because of this invariance, a term like $\mathcal{A}_{ij}$, $\mathcal{A}_{ij}^2$ or even $\cos(\mathcal{A}_{ij})$ cannot appear as an energy term.[^30] Instead, only the products of $e^{i\mathcal{A}_{ij}}$ on closed loops are gauge invariant. As a circulation of the a vector potential, these loop terms are the analog of the magnetic flux in electromagnetism. Such products can thus appear in an effective description of the fluctuations about the mean field solution. Terms like $\mathcal{E}_{ij}=\partial_\tau \mathcal{A}_{ij}+\lambda_i-\lambda_j$, which are equivalent to the electric field, are also gauge invariant. As for the couplings to the bosons, the coupling to $\mathcal{A}$ allowed by the gauge invariance (an spin-rotations) are of the type $\bar z_{i\sigma}\;e^{i\mathcal{A}_{ij}}\;z_{j\sigma}$.
A simple effective model
------------------------
We can combine the gauge invariant terms above into a simple Hamiltonian which can phenomenologically, when IGG=$\mathbb Z_2$, describe the gauge fluctuations about a saddle point and their effect on the spinons: $$\begin{aligned}
H&=&-K\sum_{\Box} \sigma^z_{ij}\sigma^z_{jk}\sigma^z_{kl}\sigma^z_{li}
-\Gamma\sum_{\langle ij\rangle} \sigma^x_{ij} \nonumber \\
&&-t\sum_{\langle ij\rangle,\sigma=\uparrow,\downarrow}
\left( b^\dagger_{i\sigma}\; \sigma^z_{ij} \;b_{j\sigma} +{\rm H.c}\right)
+\Delta \sum_{i \sigma} b^\dagger_{i\sigma} b_{i\sigma}\nonumber \\
&&+V \sum_i \left[
\left( b^\dagger_{i\uparrow} b_{i\uparrow} + b^\dagger_{i\downarrow} b_{i\downarrow} -\frac{1}{2}\right)^2
-\frac{1}{4}\right]
\label{eq:HK}\end{aligned}$$ The operator $\sigma^z_{ij}$ has eigenvalues $\pm 1$, like a pseudo spin-$\frac{1}{2}$, and corresponds to $e^{i\mathcal{A}_{ij}}$ in the path integral formulation (Eq. \[eq:A\]). $\sigma^x_{ij}$ corresponds to the electric field operator. In the path integral, $\mathcal{A}_{ij}$ and $\mathcal{E}_{ij}$ are conjugated. So $\sigma^x_{ij}$ and $\sigma^z_{ij}$ should not commute on the same bond. The natural choice in our discrete case is $\sigma^x_{ij}\sigma^z_{ij}=-\sigma^z_{ij}\sigma^x_{ij}$. So, $\sigma^x_{ij}$ and $\sigma^z_{ij}$ are the $x$ and $z$ components of the pseudo spin-$\frac{1}{2}$. The bosons represent the Bogoliubov quasi particles (spinon) of the mean field Hamiltonian. The first term ($K$) is a sum over all the elementary plaquettes (square here for simplicity) and corresponds to the magnetic energy of the gauge field. The second term ($\Gamma$) is the electric energy, which generates fluctuations in the magnetic flux. The third one ($t$) describes the spinon hopping and their interaction with the gauge field. The last terms represents the energy cost $\Delta>0$ to create a spinon (related to the spin gap of the spin model) and some (large) penalty $V$ when more than one spinon are on the same site.
This model is of course not directly related to the original spin model but contains the same two important ingredients that have been identified in the large $\mathcal{N}$ limit (spinon coupled to $\mathbb Z_2$ gauge field fluctuations) and can provide as a simplified and phenomenological description to a gapped QSL.
Because of the gauge symmetry, the physical Hilbert space of the model should be constrained to avoid spurious degrees of freedom: two states which differ by a gauge transformation correspond to a single physical state and should not appear twice in the spectrum. In the Hamiltonian formulation of gauge theories, the solution is to construct the operators $U_{i_0}$ which generate the local gauge transformations, and impose that all the physical states should be invariant under these transformations: $U_{i_0}|{\rm phys.}\rangle=|{\rm phys.}\rangle\;\forall i_0$. In the present case, an elementary gauge transformation at site $i_0$ changes the value $\sigma^z_{i_0j}$ for all neighbors $j$ of $i_0$ (noted $j\in +$). In addition, it changes the sign of the boson operators in $i_0$. This transformation is implemented by the following unitary operator $$U_{i_0}=\exp\left[i\pi (
b^\dagger_{i_0\uparrow}b_{i_0\uparrow} + b^\dagger_{i_0\downarrow}b_{i_0\downarrow}
)\right]
\prod_{j\in +} \sigma^x_{i_0j}
\label{eq:Uio}$$ The constraint $U_{i_0}=1$ is the lattice version of the Gauss law, ${\rm div} \vec E=\rho$, in electromagnetism, and the spinons appear to play the role of the “electric” charges.
Readers familiar with lattice gauge theories will have recognized the Hamiltonian formulation of a $\mathbb Z_2$ gauge theory [@kogut79]. However, to show that the ground state of this model realizes a topological phase (when $\Gamma$ is small enough), we will show that it is very close to the toric code model introduced by Kitaev [@kitaev97].
Toric code limit
----------------
One goal of these notes was to show that (gapped) QSL in Mott insulators are topologically ordered states with emerging gauge degrees of freedom. To conclude, we will now take advantage of Kitaev’s lectures of topological states of matter (in this school), and show the close connection between the large $\mathcal{N}$ description of gapped QSL and Kitaev’s toric code [@kitaev97].
We consider the limit of Eq. \[eq:HK\] when $t=0$, $\Gamma=0$ and $V=\infty$. In this limit, the bosons cannot hop any more, and can only be zero or one per site: $n_i=b^\dagger_{i_0\uparrow}b_{i_0\uparrow} + b^\dagger_{i_0\downarrow}b_{i_0\downarrow}\in\{0,1\}$. Using $U_{i}=1$ (Eq. \[eq:Uio\]) we find: $e^{i\pi n_i}=\prod_{j\in +} \sigma^x_{ij}$, so that the boson occupation numbers are expressed in terms of the (lattice divergence of the) electric field operators: $2n_i=1-\prod_{j\in +} \sigma^x_{ij}$. Taking The Hamiltonian can then be written as $$\begin{aligned}
H&=&-K\sum_{\Box} \sigma^z_{ij}\sigma^z_{jk}\sigma^z_{kl}\sigma^z_{li}
-\frac{1}{2}\Delta \sum_{i} \prod_{j\in +} \sigma^x_{ij}
\label{eq:HT}\end{aligned}$$ which is [*exactly*]{} the (solvable) toric code Hamiltonian [@kitaev97].
We can now import some results from the toric code analysis. Although simple derive in the framework of Eq. \[eq:HT\], they are highly non trivial from the point of view of the original spin model. First, the ground state breaks no symmetry and the spinons (here at the sites $i$ with $\prod_{j\in +} \sigma^x_{ij}=-1$) are free particles, they are not confined by the gauge field fluctuations. Secondly, the ground state is degenerate on a cylinder or on a torus (periodic boundary conditions), as required by the LSMH theorem. The ground state are topologically ordered in the sense that no local observable can distinguish the different ground states. Beyond the spinons, the model also have $\mathbb Z_2$-vortex excitations, which correspond to plaquettes with $\sigma^z_{ij}\sigma^z_{jk}\sigma^z_{kl}\sigma^z_{li}=-1$. These gapped excitations are singlet states in the original spin model since the bond field $Q_{ij}$ and its sign fluctuations $\sigma^z_{ij}$ are rotationally invariant.[^31] These excitation have a non trivial mutual statistics with respect to the spinon, and a bound state of a spinon and a vison behaves as a fermion. Finally, the topological properties of the model (fractional excitations, topological degeneracy) are robust to perturbations, and should persist in presence of a small $\Gamma$ and small $t$ (Eq. \[eq:HK\]).
[99]{}
E. H. Lieb, T.D. Schultz, and D.C. Mattis, [[Ann. Phys. (N.Y.) [**16**]{}, 407 (1961)](http://dx.doi.org/10.1016/0003-4916(61)90115-4)]{}.
M. Hastings, [ ]{}.
A. Kitaev, [[Ann. Phys. [**303**]{}, 2 (2003)](http://dx.doi.org/10.1016/S0003-4916(02)00018-0)]{}.
S. Taniguchi [*et al.*]{}, [ ]{}.
H. Kageyama [*et al.*]{}, [ ]{}. K. Takatsu, W. Shiramura and H. Tanaka, [ ]{}; W. Shiramura [*et al.*]{}, [ ]{}.
P. P. Shores [*et al.*]{}, [ ]{}; J. S. Helton [*et al.*]{}, [ ]{}; P. Mendels [*et al.*]{}, [ ]{}; A. Olariu [*et al.*]{}, [ ]{}; T. Imai [*et al.*]{}, [ ]{}.
Z. Hiroi [*et al.*]{}, [ ]{}; F. Bert [*et al.*]{}, [ ]{}
Y. Shimizu [*et al.*]{}, [ ]{}.
T. Itou [*et al.*]{}, [ ]{}
M. Ryuichi, K. Yoshitomo and I. Hidehiko, [ ]{}
W. Rantner and X.-G. Wen, [ ]{}; M. Hermele [*et al.*]{}, [ ]{}; Y. Ran, [*et al.*]{}, [ ]{} J. Alicea [*et al*]{}, [ ]{}. P. W. Anderson, [ ]{}.
T. Holstein and H. Primakoff, [ ]{}.
J. T. Chalker, P. C. Holdsworth, and E. F. Shender, [ ]{}.
P. Chandra and B. Douçot, [ ]{}.
V. I. Arnold, [*Mathematical Methods of Classical Mechanics*]{}, Graduate Texts in Mathematics, Springer-Verlag, 1989.
I. Affleck, [ ]{}.
M. Oshikawa, M. Yamanaka, and I. Affleck, [ ]{}.
M. Oshikawa, [ ]{}.
B. Nachtergaele and R. Sims, [[Com. Math. Phys. [**276**]{}, 437 (2007)](http://dx.doi.org/10.1007/s00220-007-0342-z)]{}.
L. Balents, M. P. A. Fisher and S. M. Girvin, [ ]{}; D. N. Sheng and L. Balents, [ ]{}.
X.-G. Wen, [ ]{}.
X.-G. Wen, [ ]{}.
X.-G. Wen and Q. Niu, [ ]{}.
R. B. Laughlin, [ ]{}.
C. K. Majumdar and D. K. Ghosh, [ ]{}.
D. J. Klein, [[J. Phys. A: Math. Gen. [**15**]{}, 661 (1982)](http://dx.doi.org/10.1088/0305-4470/15/2/032)]{}. K. S. Raman, R. Moessner and S. L. Sondhi, [ ]{}
P. Chandra and B. Douçot, [[J. Phys. A: Math. Gen. [**27**]{}, 1541 (1994)](http://dx.doi.org/10.1088/0305-4470/27/5/019)]{}
B. S. Shastry and B. Sutherland, [[Physica (Amsterdam) [**108B**]{}, 1069 (1981)](http://dx.doi.org/10.1016/0378-4363(81)90838-X)]{}.
P. W. Anderson, [[Mat. Res. Bull. [**8**]{}, 153 (1973)](http://dx.doi.org/10.1016/0025-5408(73)90167-0)]{}.
G. Misguich and C. Lhuillier, in [*Frustrated spin systems*]{} edited by H. T. Diep (World-Scientific, Singapore 2005) (also on [ ]{}).
A. Gellé [*et al.*]{}, [ ]{}.
D. Arovas and A. Auerbach, [ ]{}.
A. Auerbach, [*Interacting electrons and Quantum Magnetism*]{}, Springer-Verlag, 1994.
I. Affleck, [ ]{} I. Affleck and J. B. Marston, [ ]{}; J. B. Marston and I. Affleck, [ ]{}
N. Read and S. Sachdev, [ ]{}.
N. Read and S. Sachdev, [ ]{}. X.-G. Wen, [ ]{}.
J. B. Kogut, [ ]{}.
[^1]: Lectures given at the [*Les Houches*]{} summer school on “Exact Methods in Low-dimensional Statistical Physics and Quantum Computing” (July 2008).
[^2]: We use a tight binding model where the solid is modeled by one state per site, neglecting (or, more precisely, integrated out) filled orbitals or high energy empty states.
[^3]: At first order in $t$, a single electron hopping inevitably leads to a doubly occupied site.
[^4]: In real materials, there are often tens or hundreds of electron per unit cell, several ions and many atomic orbitals. Although the description of the magnetic properties in terms of lattice spin models if often very accurate, the spin-spin interactions is often more complicated than this antiferromagnetic Heisenberg model. It is quite frequent that some interactions violate the $SU(2)$ symmetry of the Heisenberg model, due to spin-orbit couplings in a crystalline environment. In these notes, we focus on models with an $SU(2)$ symmetry.
[^5]: In these notes, we focus here on dimension $D>1$, but much more is known about one-dimensional (1D) spin chains.
[^6]: Some order at a temperature with is very small compared the typical energy scale of the Heisenberg spin-spin interactions. This is often due to perturbations that are not included in the simplest Heisenberg model description.
[^7]: Due to some residual 3D couplings, there can be a finite temperature phase transition to an ordered state at very low temperature.
[^8]: For instance: TlCuCl$_3$ [@tlcucl3] is coupled dimer system with a Bose-Einstein condensation of magnetic excitations in presence of an external magnetic field, and SrCu$_3$(BO$_3$)$_2$ [@k99] has a magnetization curve with quantized plateaus.
[^9]: In such case, the absence of long range order cannot be attributed to some band insulator physics.
[^10]: Although the spin-spin interaction strength is of the order of $J\sim 200$ K, no order has be found down to 50mK.
[^11]: Due to the fact that $\left\{\vec z_i\right\}$ minimizes the classical energy, $\sum_j J_{ij} \vec z_j$ is perpendicular to $\vec z_i$ and thus orthogonal to $\vec \pi_i$, and there is no term [*linear*]{} in $\vec \pi$.
[^12]: Let $P$ be an orthogonal matrix which diagonalizes symmetric $M$: $M=P^{-1}\lambda P$, where $\lambda$ is a diagonal matrix and $P P^t=1$. If the eigenvalues of $M$ are [*strictly*]{} positive, $K=P^{-1}\sqrt{\lambda}$ is invertible and $M=K K^t$. We write $C=iS\;K K^t\;\sigma$, where $\sigma=
\left[
\begin{array}{cc}
0 &{\bf 1} \\
-{\bf 1} & 0
\end{array}
\right]$. Then, $\tilde C=K^{-1} C K =i S\;K^t\;\sigma K$ is Hermitian (since $\sigma$ is real antisymmetric, and K is real). $\tilde C$ can therefore be diagonalized and its spectrum is real. Since $C$ and $\tilde C$ have the same spectrum, $C$ can also be diagonalized and has real eigenvalues. Finally, we use $C^t=-C$. Since $C$ and $C^t$ should have the same spectrum, the eigenvalues of $C$ go in pairs $-\omega,\omega$.
[^13]: The Eqs. \[eq:Spi\] and \[eq:Hpi2\] also hold if $\vec\pi_i$ is a classical spin deviation of length $\vec \pi_i^2\ll1$.
[^14]: The general theory for possible Jordan forms of $\mathcal{C}$ (size and nature of the irreducible blocks) is in fact a result of [*classical mechanics*]{}, found by Williamson and exposed in [@arnold].
[^15]: $\langle 0|a^\dagger_i a_i|0\rangle=\frac{1}{2}\langle 0|p_i^2+q_i^2-1|0\rangle$ can be computed by expressing $q_i$ and $p_i$ in terms of $b^\dagger_\alpha$ and $b_\alpha$, or in terms of the new position and momenta $Q_\alpha=\frac{1}{\sqrt 2}(b_\alpha+b^\dagger_\alpha)$ and $P_\alpha=\frac{1}{\sqrt 2 i}(b_\alpha-b^\dagger_\alpha)$ Concentrating on the term $\langle 0|q_i^2|0\rangle$, $q_i$ is a linear combination of the type $q_i=\sum_{\alpha=1}^N u_\alpha^i Q_\alpha + \sum_{\beta=1}^N v_\beta^i P_\beta$, ($u$ and $v$ are related to the eigenvectors of $\mathcal{C}$). From the fact that $|0\rangle$ is the vacuum of the $b_\alpha$ bosons, we have $\langle 0|P_i P_j |0\rangle=\langle 0|Q_i Q_j |0\rangle$ if $i\ne j$, and $\langle 0|P_i Q_j + Q_jP_i|0\rangle=0$ $\forall i,j$. Then the square of the spin deviation at site $i$ (here the $\vec x_i$ component) is a linear combination of the zero point fluctuations of the normal harmonic oscillators $\langle 0|q_i^2|0\rangle=\sum_\alpha (u^i_\alpha)^2 \langle 0|Q_\alpha^2|0\rangle+\sum_\alpha (v^i_\alpha)^2 \langle 0|P_\alpha^2|0\rangle$. Assuming a regular behavior of the coefficient $(u^i_\alpha)^2$ and $(v^i_\alpha)^2$, $\langle 0|q_i^2|0\rangle$ is typically the sum of terms proportional to $\sim 1/\omega_\alpha$ when the mode frequency $\omega_\alpha$ is small.
[^16]: This does not imply that the order should persists down to $S=\frac{1}{2}$.
[^17]: In particular, the interaction can be anisotropic: $S^z_iS^z_j+\Delta (S^x_iS^x_j+S^y_iS^y_j)$, and an external magnetic field parallel to the $z$ direction can be present.
[^18]: QSL have “spinons” excitations which carry a spin $\frac{1}{2}$ (like an electron) but no electric charge.
[^19]: Hastings argument does not directly use $H_{\theta}$ for a finite $\theta$ and does not rely on this assumption. This assumption is however reasonable by the fact that, under an appropriate choice of gauge (frame), $H=H_0$ and $H_\theta$ only differ for the terms connecting the cross section at $x=L_x-1$ to the cross section at $x=0$ (boundary terms), and are identical in the bulk.
[^20]: There exists a general method for constructing an $SU(2)$ symmetric spin model with short range interactions such that all the nearest neighbor VB states are ground states [@klein82]. Building on this idea, it was possible to construct $SU(2)$ symmetric spin-$\frac{1}{2}$ models (with short ranged interaction) with a gapped QSL ground state [@rms05]. Although complicated, these models are among the very few examples where the ground state is well established to be a short ranged resonating VB liquid.
[^21]: The ground states of $H_{MG}$ spontaneously break the translation symmetry. On the Husimi cactus, the ground state is highly degenerate. The Shastry-Sutherland ground state does not break any symmetry (the ground state is unique), but the lattice has an [*even*]{} number of spins per unit cell and should be considered as a band insulator in our classification.
[^22]: Whatever the nearest neighbor VB state, exactly 1/4 of the triangles have no singlet bonds.
[^23]: The situation is very different in 1D. In the Majumdar-Gosh model, one can get a [*finite energy*]{} state with two remote spinons by introducing a domain wall in the dimerization pattern in 0 and $i$.
[^24]: Fermions can also be used, leading to other very interesting theories for (gapped of gapless) QSL [@a85; @am88; @gaplessQSL].
[^25]: $A_{ij}$ can be written using the $2\times 2$ antisymmetric tensor $\epsilon=\left[\begin{array}{cc}0&-1\\1&0\end{array} \right]$: $A_{ij}=\sum_{\sigma,\sigma'=\uparrow,\downarrow} \epsilon_{\sigma\sigma'} a_{i\sigma} a_{j\sigma'}$. The rotation invariance of $A_{ij}$ follows from the fact that any $P\in SU(2)$ satisfies $P^t \epsilon P=\epsilon$.
[^26]: The simplectic group of $2\mathcal{N}\times 2\mathcal{N}$ matrices $Sp(\mathcal{N})$ is the set of matrices $P$ which satisfies $P^t \mathcal{J} P = \mathcal{J}$, where $\mathcal{J}=\left[
\begin{array}{ccccc}
0 &1 & & & \\
-1 &0 & & & \\
& &\ddots & & \\
& & & 0 & 1 \\
& & & -1 & 0
\end{array}\right]$ generalizes the antisymmetric $\epsilon$ tensor.
[^27]: For an introduction to the path integral formalism in this context of quantum magnetism, see for instance Ref. [@auerbachbook]. We sketch the main steps of the derivation in the case of a single bosonic mode $[a,a^\dagger]=1$. For any complex number $z$, a coherent state $|z\rangle=e^{z a^\dagger}|0\rangle$ is defined. These states satisfy: $a|z\rangle=z|z\rangle$, $\langle z |z'\rangle=e^{\bar z z'}$ and the resolution of the identity $\frac{1}{\pi}\int d^2 z\;|z\rangle\langle z|e^{-|z|^2}={\bf 1}$. On writes the partition function as a product over $N_\tau$ imaginary time steps $\mathcal{Z}={\rm Tr}\left[
e^{-d\tau H}e^{-d\tau H}\cdots
\right]=\lim_{N_{\tau}\to\infty}{\rm Tr}\left[
(1-d\tau H)(1-d\tau H)\cdots
\right]$ with $d\tau=\beta/N_{\tau}$. Then, the identity is inserted at each step: $\mathcal{Z}=\lim_{N_\tau\to\infty}\int\left(\prod_{\tau=1}^{N_{\tau}} d^2 z_\tau\right)
e^{-|z_1|^2}\langle z_1 | 1-d\tau H|z_{N_{\tau}} \rangle e^{-|z_{N_\tau}|^2} \langle z_{N_\tau} |1-d\tau H |z_{N_\tau-1} \rangle
\cdots
e^{-|z_2|^2}\langle z_{2} | 1-d\tau H|z_{1} \rangle
$. Next, we write $e^{-|z_i|^2} \langle z_i |1-d\tau H |z_{i-1} \rangle
\simeq \exp{\left[
-\bar z_i(z_i-z_{i-1})-d\tau H(\bar z_i,z_{i-1})
\right]}$, where the complex number $H(\bar z,z')=\langle z'|H |z \rangle $ is obtained by writing the Hamiltonian in a normal-ordered form an replacing $a^\dagger$ by $\bar z$ and $a$ by $z'$. Taking the continuous time limit $d\tau\to 0$ is formally written as $z_i-z_{i-1}\to\partial_\tau z(\tau)d\tau$ and finally leads to $\mathcal{Z}=\int \mathcal{D}[z] \exp(-\int_0^\beta Ld\tau)$ with the Lagrangian $L=\bar z(\tau) \partial_\tau z(\tau) + H(\bar z(\tau),z(\tau))$.
[^28]: The cases where IGG$=U(1)$ are generically unstable saddle points: the gauge fluctuations lead to spinon confinement, and lattice symmetry breaking (VBC) when $S=\frac{1}{2}$ [@rs89]. This will not be discussed here.
[^29]: From now on, we go back to $\mathcal{N}=1$ a drop the flavor index $m$ for simplicity.
[^30]: In the same way, a a mass term like the square of the vector potential $A_{\mu\nu}^2$ is forbidden by gauge invariance in conventional electromagnetism.
[^31]: Ref. [@bfg02] is an example of spin-$\frac{1}{2}$ model (with $U(1)$ symmetry) where such a $\mathbb Z_2$ QSL is realized and where these vortex excitations, dubbed [*visons*]{}, have been studied. Visons have also bee studied in the context of quantum dimer models, which are simplified models for the the short range VB dynamics in frustrated quantum antiferromagnets (see [@ml05] for a brief introduction).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the effects of decoherence on the entanglement generated by Unruh effect in noninertial frames by using bit flip, phase damping and depolarizing channels. It is shown that decoherence strongly influences the initial state entanglement. The entanglement sudden death can happens irrespective of the acceleration of the noninertial frame under the action of phase flip and phase damping channels. It is investigated that an early sudden death happens for large acceleration under the depolarizing environment. Moreover, the entanglement increases for a highly decohered phase flip channel.PACS: 03.65.Ud; 03.65.Yz; 03.67.Mn;04.70.Dy
Keywords: Entanglement; Decoherence; Noninertial frames.
author:
- |
Salman Khan[^1] and M. K. Khan\
Department of Physics, Quaid-i-Azam University,\
Islamabad 45320, Pakistan
title: Open Quantum Systems in Noninertial Frames
---
Introduction
============
Entanglement is one of the potential sources of quantum theory. It is the key concept and major resource for quantum communication and computation [@springer]. In the last few years, enormous efforts has been made to investigate various aspects of quantum entanglement and its benefits in a number of setups, such as teleportation of unknown states [@Bennett] , quantum key distribution [@Ekert], quantum cryptography [@Bennett2] and quantum computation [@Grover; @Vincenzo]. Recently, the study of quantum entanglement of various fields has been extended to the relativistic setup [@Alsing; @Ling; @Gingrich; @Pan; @Schuller; @Terashima] and interesting results about the behavior of entanglement have been obtained. The study of entanglement in the relativistic framework is important not only from quantum information perspective but also to understand deeply the black hole thermodynamics [@Bombelli; @Callen] and the black hole information paradox [@Hawking; @Terashima2].
The earlier investigations on quantum entanglement in the relativistic framework is mainly focused by considering isolated quantum systems. In fact, no quantum system can be completely isolated from its environment and may results in a non-unitary dynamics of the system. Therefore, it is important to study the effect of environment on the entanglement in an initial state of a quantum system during its evolution. The interaction between an environment and a quantum system leads to the phenomenon of decoherence and it gives rise to an irreversible transfer of information from the system to the environment [@Zurik; @Breuer; @Zurik2].
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- --
![Rindler spacetime diagram: A uniformly accelerated observer Rob (R) moves on a hyperbola with acceleration $a$ in region $I$ and is causally disconnected from region $II$.[]{data-label="Figure1"}](Fig1.eps "fig:")(-350,220)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- --
In this paper we work out the effect of decoherence on the entanglement of Dirac field in a noninertial system. Alsing *et al* [@Alsing] have shown that the entanglement between two modes of a free Dirac field is degraded by the Unruh effect and asymptotically reaches a nonvanishing minimum value in the infinite acceleration. We investigate that how the loss of entanglement through Unruh effect is influenced in the presence of decoherence by using a phase flip, a phase damping and a depolarizing channel in the Kraus operators formalism. The effect of amplitude damping channel on Dirac field in a noninertial system is recently studied by Wang and Jing [@Wang]. We consider two observers, Alice and Rob, that share a maximally entangled initial state of two qubits at a point in flat Minkowski spacetime. Then Rob moves with a uniform acceleration and Alice stays stationary. To achieve our goal, we consider two cases. In one instance we allow only Rob’s qubit to interact with a noisy environment and in the second instance both qubits of the two observers interact with a noisy environment. Let the two modes of Minkowski spacetime that correspond to Alice and Rob are, respectively, given by $|n\rangle _{A}$ and $|n\rangle
_{R}$. Moreover, we assume that the observers are equipped with detectors that are sensitive only to their respective modes and share the following maximally entangled initial state$$|\psi \rangle _{A,R}=\frac{1}{\sqrt{2}}\left( |00\rangle _{A,R}+|11\rangle
_{A,R}\right) , \label{1}$$where the first entry in each ket corresponds to Alice and the second entry corresponds to Rob. From the accelerated Rob’s frame, the Minkowski vacuum state is found to be a two-mode squeezed state [@Alsing],$$|0\rangle _{M}=\cos r|0\rangle _{I}|0\rangle _{II}+\sin r|1\rangle
_{I}|1\rangle _{II}, \label{2}$$where $\cos r=\left( e^{-2\pi \omega c/a}+1\right) ^{-1/2}$. The constant $\omega $, $c$ and $a$, in the exponential stand, respectively, for Dirac particle’s frequency, light’s speed in vacuum and Rob’s acceleration. In Eq. (\[2\]) the subscripts $I$ and $II$ of the kets represent the Rindler modes in region $I$ and $II$, respectively, in the Rindler spacetime diagram (see Fig. (\[Figure1\])). The excited state in Minkowski spacetime is related to Rindler modes as follow [@Alsing]$$|1\rangle _{M}=|1\rangle _{I}|0\rangle _{II}. \label{3}$$
In terms of Minkowski modes for Alice and Rindler modes for Rob, the maximally entangled initial state of Eq. (\[1\]) by using Eqs. (\[2\]) and (\[3\]) becomes$$|\psi \rangle _{A,I,II}=\frac{1}{\sqrt{2}}\left( \cos r|0\rangle
_{A}|0\rangle _{I}|0\rangle _{II}+\sin r|0\rangle _{A}|1\rangle
_{I}|1\rangle _{II}+|1\rangle _{A}|1\rangle _{I}|0\rangle _{II}\right) .
\label{4}$$Since Rob is causally disconnected from region $II$, we must take trace over all the modes in region $II$. This leaves the following mixed density matrix between Alice and Rob, that is,$$\begin{aligned}
\rho _{A,I} &=&\frac{1}{2}[\cos ^{2}r|00\rangle _{A,I}\langle 00|+\cos
r(|00\rangle _{A,I}\langle 11|+|11\rangle _{A,I}\langle 00|) \nonumber \\
&&\sin ^{2}r|01\rangle _{A,I}\langle 01|+|11\rangle _{A,I}\langle 11|].
\label{5}\end{aligned}$$
---------------------------------------------------------------------------------
phase flip channel $E_{o}=\sqrt{1-p}\left(
\begin{array}{cc}
1 & 0 \\
0 & 1\end{array}\right) ,\qquad E_{1}=\sqrt{p}\left(
\begin{array}{cc}
1 & 0 \\
0 & -1\end{array}\right) $
----------------------- ---------------------------------------------------------
phase damping channel $E_{o}=\left(
\begin{array}{cc}
1 & 0 \\
0 & \sqrt{1-p}\end{array}\right) ,\qquad E_{1}=\left(
\begin{array}{cc}
0 & 0 \\
0 & \sqrt{p}\end{array}\right) $
depolarizing channel $\begin{array}{c}
E_{o}=\sqrt{1-p}\left(
\begin{array}{cc}
1 & 0 \\
0 & 1\end{array}\right) ,\qquad E_{1}=\sqrt{p/3}\left(
\begin{array}{cc}
0 & 1 \\
1 & 0\end{array}\right) , \\
E_{2}=\sqrt{p/3}\left(
\begin{array}{cc}
0 & -i \\
i & 0\end{array}\right) ,\qquad E_{3}=\sqrt{p/3}\left(
\begin{array}{cc}
1 & 0 \\
0 & -1\end{array}\right)\end{array}$
---------------------------------------------------------------------------------
\[htb\]
single qubit in a noisy environment
===================================
In this section we consider that only the Rob’s qubit is coupled to a noisy environment. The final density matrix of the system in the Kraus operators representation becomes$$\rho _{f}=\sum_{i}\left( \sigma _{o}\otimes E_{i}\right) \rho _{A,I}\left(
\sigma _{o}\otimes E_{i}^{\dag }\right) , \label{6}$$where $\rho _{A,I}$ is the initial density matrix of the system given by Eq. (\[5\]), $\sigma _{o}$ is the single qubit identity matrix and $E_{i}$ are a single qubit Kraus operators of the channel under consideration. The Kraus operators of the channels we use are given in Table $1$. The spin-flip matrix of the final density matrix of Eq. (\[6\]) is defined as $\tilde{\rho}_{f}=\left( \sigma _{2}\otimes \sigma _{2}\right) \rho _{f}\left(
\sigma _{2}\otimes \sigma _{2}\right) $, where $\sigma _{2}$ is the Pauli matrix. The degree of entanglement in the two qubits mixed state in a noisy environment can be quantified conveniently by concurrence $C$, which is given as [@Wootter; @Coffman]$$C=\max \left\{ 0,\sqrt{\lambda _{1}}-\sqrt{\lambda _{2}}-\sqrt{\lambda _{3}}-\sqrt{\lambda _{4}}\right\} \qquad \lambda _{i}\geq \lambda _{i+1}\geq 0,
\label{7}$$where $\lambda _{i}$ are the eigenvalues of the matrix $\rho _{f}\tilde{\rho}_{f}$. The eigenvalues under the action of phase-flip channel becomes
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
![The concurrence $C$ under the action of phase flip channel is plotted against decoherence parameter $p$ for the case when only Rob’s qubit is coupled to a noisy environment.[]{data-label="Figure2"}](Fig2.eps "fig:")(-350,220)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
$$\begin{aligned}
\lambda _{1}^{\mathrm{PF}} &=&(1-2p+p^{2})\cos ^{2}r, \nonumber \\
\lambda _{2}^{\mathrm{PF}} &=&p^{2}\cos ^{2}r, \nonumber \\
\lambda _{3}^{\mathrm{PF}} &=&\lambda _{4}^{\mathrm{PF}}=0, \label{8}\end{aligned}$$
where the superscript corresponds to phase flip channel. Similarly, the eigenvalues under the action of phase damping and depolarizing channels are, respectively, given by$$\begin{aligned}
\lambda _{1,2}^{\mathrm{PD}} &=&\frac{1}{4}(2-p\pm 2\sqrt{1-p})\cos ^{2}r,
\nonumber \\
\lambda _{3}^{\mathrm{PD}} &=&\lambda _{4}^{\mathrm{PD}}=0, \label{10}\end{aligned}$$$$\begin{aligned}
\lambda _{1}^{\mathrm{DP}} &=&(-1+p)^{2}\cos ^{2}r, \nonumber \\
\lambda _{2}^{\mathrm{DP}} &=&\lambda _{3}^{\mathrm{DP}}=\lambda _{4}^{\mathrm{DP}}=\frac{1}{9}p^{2}\cos ^{2}r, \label{11}\end{aligned}$$where the superscripts and stand for phase damping and depolarizing channels, respectively. In all these equations $p\in \left[
0,1\right] $ is the decoherence parameter. The upper and lower values of $p$ correspond to undecohered and fully decohered case of the channels, respectively. The concurrence under the action of every channel reduces to the result of Ref. [@Alsing] when the decoherence parameter $p=0$.
To see how the concurrence and hence the entanglement is influenced by decoherence parameter $p$ in the presence of Unruh effect, we plot the concurrence for each channel against $p$ for various values of $r$. In Fig. (\[Figure2\]), the concurrence under the action of phase flip channel is plotted against $p$. The figure shows that for smaller values of $p$, the entanglement is strongly acceleration dependent, such that for large values of Rob’s acceleration (the value of $r$) it gets weakened. However, as $p$ increases the dependence of entanglement on acceleration decreases and the increasing value of $p$ causes a rapid loss of entanglement.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
![The concurrence $C$ under the action of phase damping channel is plotted against decoherence parameter $p$ for the case when only Rob’s qubit is coupled to a noisy environment.[]{data-label="Fig3"}](Fig3.eps "fig:")(-350,220)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
The entanglement sudden death happens irrespective of the acceleration of Rob’s frame for a $50\%$ decoherence. Fig. (\[Fig3\]) shows the effect of decoherence on the concurrence under the action of phase damping channel. In this case, the degradation of entanglement due to decoherence is smaller as compare to the the degradation in the case of phase flip. The entanglement vanishes for all values of acceleration only when the channel is fully decohered. The concurrence under the action of the depolarizing channel is exactly equal to the one for phase flip channel. Hence it influences the entanglement in a way exactly similar to the phase flip channel as shown in Fig. ($2$).
Both qubits in a noisy environment
==================================
In this section we consider that both Alice’s and Rob’s qubits are influenced simultaneously by a noisy environment. The final density matrix in this case can be written in the Kraus operators formalism as follows$$\rho _{f}=\sum_{k}E_{k}\rho _{A,I}E_{k}^{\dag }, \label{12}$$where $\rho _{A,I}$ is given by Eq. (\[5\]) and $E_{k}$ are the Kraus operators for a two qubit system, satisfying the completeness relation $\sum_{k}E_{k}E_{k}=I$ and are constructed from a single qubit Kraus operators of a channel by taking tensor product of all the possible combinations in the following way$$E_{k}=\sum_{i,j}E_{i}\otimes E_{j}, \label{13}$$where $E_{i,j}$ are the single qubit Kraus operators of a channel given in Table $1$. We consider that both Alice’s and Bob’s qubits are influenced by the same environment, that is, the decoherence parameter $p$ for both qubits is same. Proceeding in a similar way like the case of single qubit coupled to the environment, the eigenvalues of the matrix $\rho _{f}\tilde{\rho}_{f}$ under the action of phase flip channel become
$$\begin{aligned}
\lambda _{1}^{\mathrm{PF}} &=&(1+2(-1+p)p)^{2}\cos ^{2}r, \nonumber \\
\lambda _{2}^{\mathrm{PF}} &=&4(-1+p)^{2}p^{2}\cos ^{2}r, \nonumber \\
\lambda _{3}^{\mathrm{PF}} &=&\lambda _{4}^{\mathrm{PF}}=0, \label{14}\end{aligned}$$
Likewise the eigenvalues for phase damping and depolarizing channels, respectively, becomes
$$\begin{aligned}
\lambda _{1}^{\mathrm{PD}} &=&\frac{1}{4}(-2+p)^{2}\cos ^{2}r, \nonumber \\
\lambda _{2}^{\mathrm{PD}} &=&\frac{1}{4}p^{2}\cos ^{2}r, \nonumber \\
\lambda _{3}^{\mathrm{PD}} &=&\lambda _{4}^{\mathrm{PD}}=0, \label{16}\end{aligned}$$
$$\begin{aligned}
\lambda _{1,3}^{\mathrm{DP}} &=&\frac{1}{1296}[324+p(-3+2p)(387+152p(-3+2p))
\nonumber \\
&&+4(3-4p)^{2}(9+5p(-3+2p))\cos 2r \nonumber \\
&&+(3-4p)^{2}p(-3+2p)\cos 4r\pm 4(3-4p)^{2}\cos r \nonumber \\
&&\times \{3(54+p(-3+2p)(33+8p(-3+2p))) \nonumber \\
&&+(3-4p)^{2}(2(9-6p+4p^{2})\cos 2r+p(-3+2p)\cos 4r)\}^{1/2}], \nonumber \\
\lambda _{2}^{\mathrm{DP}} &=&\lambda _{4}^{\mathrm{DP}}=\frac{1}{648}p(-3+2p)(-9+4p \nonumber \\
&&+(-3+4p)\cos 2r)(3+4p+(-3+4p)\cos 2r), \label{17}\end{aligned}$$
The $"\pm "$ sign in Eq. (\[17\]), correspond to the eigenvalues $\lambda
_{1}$, and $\lambda _{3}$ respectively. It is necessary to point out here that the concurrence under the action of each channel reduces to the result of Ref. [@Alsing] when we set the decoherence parameter $p=0$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- --
![The concurrence $C$ under the action of phase flip channel is plotted against decoherence parameter $p$ for the case when both qubits are coupled to a noisy environment.[]{data-label="Fig4"}](Fig4.eps "fig:")(-350,220)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- --
To see how the entanglement behaves when both the qubits are coupled to the noisy environment, we plot the concurrence against the decoherence parameter $p$ for different values of $r$ under the action of each channel separately. Fig. (\[Fig4\]) shows the dependence of concurrence on decoherence parameter $p$ under the action of phase flip channel. The dependence of entanglement on acceleration of Rob’s frame is obvious in the region of lower values of $p$. However, this dependence diminishes as $p$ increases and a rapid decrease in the degree of entanglement develops. At a $50\%$ decoherence level, the entanglement sudden death occurs irrespective of Rob’s acceleration. It’s interesting to see that beyond this point onward, the entanglement regrows as $p$ increases. The dependence of entanglement on acceleration of the Rob’s frame reemerges and the entanglement reaches to the corresponding undecohered maximum value for a fully decohered case. The concurrence varies as a parabolic function of decoherence parameter $p$ with its vertex at $p=0.5$.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
![The concurrence $C$ under the action of phase damping channel is plotted against decoherence parameter $p$ for the case when both qubits are coupled to a noisy environment.[]{data-label="Fig5"}](Fig5.eps "fig:")(-350,220)
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
The dependence of entanglement on $p$ under the action of phase damping channel is shown in Fig. (\[Fig5\]). In this case the entanglement decreases linearly as $p$ increases and the dependence on acceleration diminishes. Whatever the acceleration of Rob’s frame may be, the entanglement sudden death occurs when the channel is fully decohered. The influence of depolarizing channel on the entanglement is shown in Fig. ([Fig6]{}). Unlike the other two channels, the depolarizing channel does not diminish the effect of acceleration on the entanglement as the $p$ increases. However a rapid decrease in entanglement appears which leads to entanglement sudden death at different values of decoherence parameter for different acceleration of Rob’s frame. The larger the acceleration the earlier the entanglement sudden death occurs.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
![The concurrence $C$ under the action of depolarizing channel is plotted against decoherence parameter $p$ for the case when both qubits are coupled to a noisy environment.[]{data-label="Fig6"}](Fig6.eps "fig:")(-350,220)
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- --
If we compare the single qubit and the both qubits decohering situations, it becomes obvious that the entanglement loss is rapid when both the qubits are coupled to the noisy environment. For example, in the case of bit flip channel the concurrence behaves as a linear function of $p$ for single qubit decohering case whereas in the case of both qubits decohering case it varies as a parabolic function. Nevertheless, the sudden death happens at the same value of $p$, irrespective of the acceleration, for both cases under the action of bit flip and phase damping channels. For depolarizing channel, however, this is not true.
Conclusion
==========
In conclusion, we have investigated that the entanglement in Dirac fields is strongly dependent on coupling with a noisy environment. This result is contrary to the case of an isolated system in which the entanglement of Dirac fields survives even in the limit of infinite acceleration of Rob’s frame. In the presence of decoherence, the entanglement rapidly decreases and entanglement sudden death occurs even for zero acceleration. Under the action of phase flip channel, the entanglement can regrow when both qubits are coupled to a noisy environment in the limit of large values of decoherence parameter. The entanglement disappears, irrespective of the acceleration, under the action of phase damping channel only when the channel is fully decohered both for single qubit and the two qubits decohering cases. However, under the action of depolarizing channel an early sudden death occurs for larger acceleration when both qubits are coupled to the environment. In summary, the entanglement generated by Unruh effect in noninertial frame is strongly influenced by decoherence.
[99]{} *The Physics of Quantum Information*, D. Bouwmeester, A. Ekert, A. Zeilinger (Springer-Verlag, Berlin, 2000)
C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, and W. K. Wootters, Phys. Rev. Lett. 70, 1895 (1993).
A. Ekert, Phys. Rev. Lett. 67, 661 (1991).
C.H. Bennett, G. Brassard, N.D. Mermin, Phys. Rev. Lett. 68, 557 (1992).
L.K. Grover, Phys. Rev. Lett. 79, 325 (1997).
D.P. DiVincenzo, Science 270, 255 (1995)
P.M. Alsing, I. Fuentes-Schuller, R. B.Mann, and T. E. Tessier, Phys. Rev. A 74, 032326 (2006).
Yi Ling et al, J. Phys. A: Math. Theor. 40, 9025 (2007).
R. M. Gingrich and C. Adami Phys. Rev. Lett. 89, 270402 (2002).
Q. Pan and J. Jing, Phys. Rev. A 77, 024302 (2008).
I. Fuentes-Schuller and R. B. Mann, Phys. Rev. Lett. 95, 120404 (2005).
H. Terashima and M. Ueda, Int. J. Quantum Inf. 1, 93 (2003).
L. Bombelli, R. K. Koul, J. Lee, and R. Sorkin, Phys. Rev. D 34, 373 (1986).
C. Callen and F. Wilzcek, Phys. Lett. B 333, 55 (1994).
S. W. Hawking, Commun. Math. Phys. 43, 199 (1975); Phys. Rev. D 14, 2460 (1976).
H. Terashima, Phys. Rev. D 61, 104016 (2000).
Zurek W H et al. Phys. Today 44, 36 (1991).
H. P. Breuer and F. Petruccione, *The Theory of Open Quantum Systems* (Oxford University Press, Oxford), 2002; H. Carmichael, *An Open Systems Approach to Quantum Optics* (Springer, Berlin, 1993).
W. H. Zurek, Rev. Mod. Phys. 75, 715 (2003).
J. Wang and J. Jing, arxiv: 1005.2865v4 (2010).
W. K. Wootters, Phys. Rev. Lett. 80, 2245 (1998).
V. Coffman, J. Kundu, and W. K. Wootters, Phys. Rev. A 61, 052306 (2000).
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study randomly initialized residual networks using mean field theory and the theory of difference equations. Classical feedforward neural networks, such as those with tanh activations, exhibit exponential behavior on the average when propagating inputs forward or gradients backward. The exponential forward dynamics causes rapid collapsing of the input space geometry, while the exponential backward dynamics causes drastic vanishing or exploding gradients. We show, in contrast, that by adding skip connections, the network will, depending on the nonlinearity, adopt subexponential forward and backward dynamics, and in many cases in fact polynomial. The exponents of these polynomials are obtained through analytic methods and proved and verified empirically to be correct. In terms of the “edge of chaos” hypothesis, these subexponential and polynomial laws allow residual networks to “hover over the boundary between stability and chaos,” thus preserving the geometry of the input space and the gradient information flow. In our experiments, for each activation function we study here, we initialize residual networks with different hyperparameters and train them on MNIST. Remarkably, our [*initialization time*]{} theory can accurately predict [*test time*]{} performance of these networks, by tracking either the expected amount of gradient explosion or the expected squared distance between the images of two input vectors. Importantly, we show, theoretically as well as empirically, that common initializations such as the Xavier or the He schemes are not optimal for residual networks, because [*the optimal initialization variances depend on the depth*]{}. Finally, we have made mathematical contributions by deriving several new identities for the kernels of powers of ReLU functions by relating them to the zeroth Bessel function of the second kind.'
author:
- |
Greg Yang[^1]\
[Microsoft Research AI]{}\
`[email protected]`\
Samuel S. Schoenholz\
[Google Brain]{}\
`[email protected]`\
bibliography:
- 'neural\_dynamics.bib'
title: 'Mean Field Residual Networks: On the Edge of Chaos'
---
Introduction
============
Previous works [@poole_exponential_2016; @daniely_toward_2016; @schoenholz_deep_2017] have shown that randomly initialized neural networks exhibit a spectrum of behavior with depth, from stable to chaotic, which depends on the variance of the initializations: the cosine distance of two input vectors converges exponentially fast with depth to a fixed point in \[0, 1\]; if this fixed point is 1, then the behavior is stable; if this fixed point is 0, then the behavior is chaotic. It has been argued in many prior works [@bertschinger_real-time_2004; @poole_exponential_2016] that effective computation can only be supported by a dynamical behavior that is on [**the edge of chaos**]{}. Too much stability prevents the neural network from telling apart two different inputs. While some chaotic behavior can increase the expressivity of a network, too much chaos makes the neural network think two similar inputs are very different. At the same time, the same initialization variances also control how far gradient information can be propagated through the network; the networks with chaotic forward dynamics will tend to suffer from exploding gradients, while networks with stable forward dynamics will tend to suffer from vanishing gradients.
These works have focused on vanilla (fully connected) feedforward networks. Here we consider residual networks [@he_deep_2016; @he_identity_2016] (with fully-connected layers and without batchnorm), which are a family of recently proposed neural network architectures that has achieved state-of-the-art performance on image recognition tasks, beating all other approaches by a large margin. The main innovation of this family of architectures is the addition of a passthrough (identity) connection from the previous layer to the next, such that the usual nonlinearity computes the “residual” between the next-layer activation and the previous-layer activation.
In this work, we seek to characterize randomly initialized residual networks. One of our main results is that random residual networks for many nonlinearities such as $\tanh$ [**live on the edge of chaos**]{}, in that the cosine distance of two input vectors will converge to a fixed point at a polynomial rate, rather than an exponential rate, as with vanilla tanh networks. Thus a typical residual network will slowly cross the stable-chaotic boundary with depth, hovering around this boundary for many layers. In addition, for most of the nonlinearities considered here, the mean field estimate of the gradient grows subexponentially with depth. In fact, for $\alpha$-ReLU, the $\alpha$th-power of ReLU, for $\alpha < 1$, the gradient grows only polynomially. These theoretical results provide some theoretical justification for why residual networks work so well in practice. In our experiments, we are also able to predict surprisingly well the relative performances of [*trained*]{} residual networks based only on their initialization hyperparameters, in a variety of settings. In particular, we find that the quality of initialization for tanh resnets is determined by [*trainability*]{} (how much gradient explosion on average) while that for ($\alpha$-)ReLU resnets is determined by expressivity (how far can two different input vectors be pulled apart) (see \[sec:experiments\]). To the best of our knowledge, this is the first time that a quantity other than gradient explosion/vanishing has been found to control the quality of initialization. We establish theoretically and empirically that the best initialization variances for residual networks depend on the depth of the network (contrary to the feedforward case [@schoenholz_deep_2017]), so that **common initialization schemes like Xavier [@glorot_understanding_2010] or He [@he_delving_2015] cannot be optimal**. In fact, even the rationale of He initialization is incorrect for ReLU residual networks because it tries to control gradient dynamics rather than expressivity. However we want to emphasize that we study a simplified model of residual networks in this work, with no batchnorm or convolutional layers, so that these results are not necessarily indicative of the MSRA residual network used in practice [@he_deep_2016].
In the body of this paper, we give account of general intuition and/or proof strategy when appropriate for our theoretical results, but we relegate all formal statements and proofs to the appendix.
Background {#sec:background}
==========
Consider a vanilla feedforward neural network of $L$ layers, with each layer $l$ having $\p N l$ neurons; here layer 0 is the input layer. For the ease of presentation we assume all hidden layer widths are the same $\p N l = N$ for all $l > 0$. Let $\p x 0 = (\p x 0_1, \ldots, \p x 0_{\p N 0})$ be the input vector to the network, and let $\p x l$ for $l > 0$ be the activation of layer $l$. Then a neural network is given by the equations $$\begin{aligned}
\p x l _i &= \phi(\p h l _i), & \p h l _i &= \sum_{j=1}^N \p w l _{ij} \p x {l-1}_j + \p b l_i
\end{aligned}$$ where
$\p h l$ is the pre-activation at layer $l$,
$\p w l$ is the weight matrix,
$\p b l$ is the bias vector, and
$\phi$ is a nonlinearity, for example $\tanh$ or ReLU, which is applied coordinatewise to its input.
To lighten up notation, we suppress the explicit layer numbers $l$ and write $$\begin{aligned}
x_i &= \phi(h_i), & h_i &= \sum_j w_{ij} {\underline}x_j + b_i
\end{aligned}$$ where $\bullet$ implicitly denotes $\p \bullet l$, and ${\underline}\bullet$ denotes $\p \bullet {l-1}$ (and analogously, ${\overline}\bullet$ denotes $\p \bullet {l+1}$).
A series of papers [@poole_exponential_2016; @raghu_expressive_2016; @schoenholz_deep_2017] investigated the “average behavior” of random neural networks sampled via $\p w l_{ij} \sim {\mathcal{N}}(0, \sigma_w^2/N), \p b l_i \sim {\mathcal{N}}(0, \sigma_b^2)$, for fixed parameters $\sigma_w$ and $\sigma_b$, independent of $l$. Consider the expectation of $\f 1 N \sum_{i=1}^N x^2_i$, the normalized squared length of $x$, over the sampling of $w$ and $b$. @poole_exponential_2016 showed that this quantity converges to a fixed point exponentially fast for sigmoid nonlinearities. Now suppose we propagate two different vectors $\p x 0$ and $(\p x 0)'$ through the network. @poole_exponential_2016 also showed that the expectation of the normalized dot product $\f 1 N \sum_{i=1}^N x_i x'_i$ converges exponentially fast to a fixed point. The ratio between the normalized squared length and the normalized dot product is the cosine distance between $x$ and $x'$. Thus these two exponential convergence results show that the cosine distance converges exponentially fast to a fixed point as well. Intuitively, this means that a vanilla feedforward network “forgets” the geometry of the input space “very quickly,” after only a few layers.
In addition, @schoenholz_deep_2017, under certain independence assumptions, showed that the expected normalized squared norm of the gradient also vanishes or explodes in an exponential fashion with depth, with the ”half-life” controlled by $\sigma_w$ and $\sigma_b$. They verified that this theoretical ”half-life” correlates in practice with the maximal number of layers that are admissible to good performance.
At the same time, @daniely_toward_2016 published work of similar nature, but phrased in the language of reproducing kernel Hilbert spaces, and provided high probability estimates that are meaningful for the case when the width $N$ is finite and the depth is logarithmic in $N$. However, they essentially fixed the variance parameters $\sigma_\bullet$, and furthermore, their framework (for example the notion of a “skeleton”) does not immediately generalize to the residual network case.
In this work, we show that residual networks have very different dynamics from vanilla feedforward networks. In most cases, the cosine distance convergence rate and the gradient growth rate are subexponential in a residual network, and in most cases, these rates may be polynomial.
Preliminaries
=============
Residual networks were first introduced by [@he_deep_2016] and later refined by [@he_identity_2016], and they are now commonplace among deployed neural systems. The key innovation there is the addition of a shortcut connection from the previous layer to the next. We define the following idealized architectures for ease of analysis. Note that we only consider fully-connected affine layers instead of convolutional layers. A [**reduced residual network (RRN)**]{} has the recurrence $$\begin{aligned}
x_i &= \phi(h_i) + {\underline}x, & h_i &= \sum_j w_{ij} {\underline}x_j + b_i.
\end{aligned}$$ A [**(full) residual network (FRN)**]{} in addition has an affine connection given by weights $v$ and biases $a$ from the nonlinearity $\phi(h)$ to the next layer: $$\begin{aligned}
x_i &= \sum_j v_{ij} \phi( h_j) + {\underline}x_i + a_i, & h_i &= \sum_j w_{ij} {\underline}x_j + b_i
\end{aligned}$$
We are interested in the “average behavior” of these network when the weights and biases, $\p w l_{ij}, \p b l_i, \p v l_{ij}$, and $\p a l_i$ are sampled i.i.d. from Gaussian distributions resp. with standard deviations $\sigma_w, \sigma_b, \sigma_v,$ and $\sigma_a$, independent from $l$. Here we take the variance of $\p w l_{ij}$ to be $\sigma_w^2/N$ so that the variance of each $h_i$ is $\sigma_w^2$, assuming each ${\underline}x_j$ is fixed (similarity for $\p v l_{ij}$). Such an initialization scheme is standard in practice.
We make several key “physical assumptions” to make theoretical computations tractable:
\[ass:symAct\] (a) We assume $\la (\p h l _i)^2 \ra = \la (\p h l _j)^2 \ra$ and $\la (\p x 0 _i)^2 \ra = \la (\p x 0 _j)^2 \ra$ for any $i, j, l$. (b) We also assume that the gradient $\pd E/\pd \p x l _i$ with respect to the loss function $E$ satisfies $\la (\pd E/\pd \p x l _i)^2 \ra = \la (\pd E/ \pd \p x l _j)^2 \ra$ for any $i, j, l$.
One can see that \[ass:symAct\](a) is satisfied if the input $\p x 0 \in \{\pm 1\}^N$ and \[ass:symAct\](b) is satisfied if \[ass:gradInd\] below is true and the gradient at the last layer $\pd E/\pd x L \in \{\pm 1 \}^N$. But in general it is justified both empirically and theoretically as an approximation, because $(\p h l _i)^2 - (\p h l _j)^2$ stays about constant with $l$, but $(\p h l _i)^2$ and $(\p h l _j)^2$ grow rather quickly at the same pace with $l$ (as will be seen later in calculations), so that their additive difference becomes negligible; similarly for $(\p x l _i)^2$ and $(\pd E/\pd \p h l _i)^2$.
\[ass:gradInd\] (a) We assume the we use a different set of weights for backpropagation than those used to compute the network outputs, but sampled i.i.d. from the same distributions. (b) For any loss function $E$, we assume that the gradient at layer $l$, $\pd E/\pd \p x l _i$, is independent from all activations $\p h {l} _j$ and $\p x {l-1} _j$ from the previous layer.
\[ass:gradInd\](a) was first made in [@schoenholz_deep_2017] for computing the mean field theory of gradients for feedforward tanh networks. This is similar to the practice of feedback alignment [@lillicrap_random_2016]. Even though we are the first to explicitly formulate \[ass:gradInd\](b), in fact it was already applied implicitly in the gradient calculations of [@schoenholz_deep_2017]. Note that a priori \[ass:gradInd\](b) is not true, as $\pd E/\pd \p x l _i$ depends on $\dot \phi(\p h {l+1} _k)$ for every $k$, which depend on $\p h {l} _j$ for each $j$, and which depends on $\p x {l-1} _k$ for every $k$. Nevertheless, in practice both subassumptions hold very well.
Now we define the central quantities studied in this paper. Inevitably, our paper involves a large amount of notation that may be confusing for the first-time reader. We have included a glossary of symbols (\[tab:glossary\]) to ameliorate notation confusion.
\[defn:length\] Fix an input $\p x 0$. Define the [**length quantities**]{} $\p {\mathbf{q}}l := \la (\p h l_1)^2 \ra$ and $\p {\mathbf{p}}l := \la (\p x l_1)^2 \ra$ for $l > 0$ and $\p {\mathbf{p}}0 = \|\p x 0\|^2/N$. Here the expectations $\la \bullet \ra$ are taken over all random initialization of weights and biases for all layers $l$, as $N \to \infty$ (large width limit).
Note that in our definition, the index $1$ does not matter by \[ass:symAct\].
\[defn:corr\] Fix two inputs $\p x 0$ and $\p x 0{}'$. We write $\bullet'$ to denote a quantity $\bullet$ with respect to the input ${\p x 0}'$. Then define [**the correlation quantities**]{} $\p {\boldsymbol{\oldgamma}}l:= \la \p h l_1 \p h l_1{}' \ra$ and $\p {\boldsymbol{\oldlambda}}l:= \la \p x l_1 \p x l_1{}'\ra$ for $l > 0$ and $\p {\boldsymbol{\oldgamma}}0 = \p x 0 \cdot \p x 0 {}' / N$, where the expectations $\la \bullet \ra$ are taken over all random initialization of weights and biases for all layers $l$, as $N \to \infty$ (large width limit). Again, here the index $1$ does not matter by \[ass:symAct\]. By [**metric expressivity**]{}, we mean $\p {\mathbf{s}}l := \f 1 {2N} \la \|\p x l - \p x l {}'\|^2\ra = \f 1 {2N} (\la \|\p x l\|^2\ra + \la \|\p x l {}'\|^2 \ra - 2 \la \p x l \cdot \p x l {}'\ra) = \f 1 2 (\p {\mathbf{p}}l + \p {\mathbf{p}}l {}') - \p {\boldsymbol{\oldgamma}}l$. Additionally, define [**the cosine distance quantities**]{} $\p {\mathbf{e}}l := \p {\boldsymbol{\oldgamma}}l / \sqrt{\p {\mathbf{p}}l \p {\mathbf{p}}l {}'}$ and $\p {\mathbf{c}}l := \p {\boldsymbol{\oldlambda}}l / \sqrt{\p {\mathbf{q}}l \p {\mathbf{q}}l{}'}$, and we will also call $\p {\mathbf{e}}l$ [**angular expressivity**]{}.
In this paper, for the ease of presentation, we assume $\p {\mathbf{p}}0 = \p {\mathbf{p}}0 {}'$. Then, as we will see, $\p {\mathbf{p}}l = \p {\mathbf{p}}l{}', \p {\mathbf{q}}l = \p {\mathbf{q}}l{}'$ for all $l$, and as a result, $\p {\mathbf{e}}l = \p {\boldsymbol{\oldgamma}}l / \p {\mathbf{p}}l$ and $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = (1 - \p {\mathbf{e}}l) \p {\mathbf{p}}l$.
\[defn:grad\] Fix an input $\p x 0$ and a gradient vector $(\pd E/ \pd{\p x L_i})_i$ of some loss function $E$ with respect to the last layer $\p x L$. Then define [**the gradient quantities**]{} $\p {{\boldsymbol{\oldchi}}}l:= \la (\pd E/\pd \p x l _1)^2 \ra, \p {\boldsymbol{\oldchi}}l _\bullet := \la (\pd E/\pd \p \bullet l _1)^2 \ra$ for $\bullet = a, b$, and $\p {\boldsymbol{\oldchi}}l _\bullet := \la (\pd E/\pd \p \bullet l _{11})^2 \ra$ for $\bullet = w, v$. Here the expectations are taken with \[ass:gradInd\] in mind, over both random initialization of forward and backward weights and biases, as $N \to \infty$ (large width limit). Again, the index $1$ or $11$ does not matter by \[ass:symAct\].
#### Asymptotic notations.
The expressions $f = O(g) \iff g = \Omega(f)$ have their typical meanings, and $f = \Theta(g)$ iff $f = O(g), g = O(f)$. We take $f(x) = \tilde O(g(x)) \iff g(x) = \tilde \Omega(f(x))$ to mean $f(x) = O(g\log^k x)$ for some $k \in \Z$ (this is slightly different from the standard usage of $\tilde O$), and $f = \tilde\Theta(g) \iff f = \tilde O(g) \And g = \tilde O(f).$ We introduce a new notation: $f = {{\check\Theta}}(g)$ if $f(x) = O(g(x) \cdot x^{\epsilon})$ and $f(x) = \Omega(g(x) \cdot x^{-{\epsilon}})$, as $x \to \infty$, for any ${\epsilon}> 0$. All asymptotic notations are sign-less, i.e. can indicate either positive or negative quantities, unless stated otherwise.
Overview
========
The primary reason we may say anything about the average behavior of any of the above quantities is the central limit theorem: every time the activations of the previous layer pass through an affine layer whose weights are sampled i.i.d., the output is a sum of a large number of random variables, and thus follows approximately Gaussian distributions. The mean and variance of these distributions can be computed by keeping track of the mean and variances of the activations in the previous layer.
In what follows, we use this technique to derive recurrence equations governing ${\mathbf{p}}, {\mathbf{q}}, {\boldsymbol{\oldgamma}}, {\boldsymbol{\oldlambda}}, {{\boldsymbol{\oldchi}}}$ for different architectures and different activation functions. We use these equations to investigate the dynamics of ${\mathbf{e}}$ and ${\mathbf{s}}$, the key quantities in the forward pass, and the dynamics of ${{\boldsymbol{\oldchi}}}$, the key quantity in the backward pass.
The cosine distance ${\mathbf{e}}$ in some sense measures the angular geometry of two vectors. If ${\mathbf{e}}= 1$, then the vectors are parallel; if ${\mathbf{e}}= 0$, then they are orthogonal. Just as in [@poole_exponential_2016] and [@schoenholz_deep_2017], we will show that in all of the architectures and activations we consider in this paper, $\p {\mathbf{e}}l$ converges to a fixed point ${\mathbf{e}}^*$ as $l \to \infty$ [^2]. Thus, on the average, as vectors propagate through network, the geometry of the original input space, for example, linear separability, is “forgotten” by residual networks as well as by vanilla networks. But we will prove and verify experimentally that, while @poole_exponential_2016 and [@schoenholz_deep_2017] showed that the convergence rate to ${\mathbf{e}}^*$ is exponential in a vanilla network, the convergence rate is rather only polynomial in residual networks, for tanh and $\alpha$-ReLU (\[defn:alphaReLU\]) nonlinearities; see \[thm:dalethRecReduced\], \[thm:eDynamicsFullResTanh\], \[thm:ReLUSquaredConvergence\], and \[thm:alphaReLUeConvergence\]. This slow convergence preserves geometric information in the input space, and allows a typical residual network to “hover over the edge of chaos”: Even when the cosine distance $\p{\mathbf{e}}l$ converges to 0, corresponding to “chaos”, (resp. 1, corresponding to “stability”), for the number of layers usually seen in practice, $\p {\mathbf{e}}l$ will reside well away from 0 (resp. 1).
Similarly, the quantity ${\mathbf{s}}$ measures the metric geometry of two vectors. The evolution of $\p {\mathbf{s}}l$ with $l$ tells us the ability of the average network to separate two input points in terms of Euclidean distance. Again, for tanh and $\alpha$-ReLU ($\alpha < 1$) nonlinearities, ${\mathbf{s}}$ varies only polynomially with $l$.
On the other hand, $\p {{\boldsymbol{\oldchi}}}l$ measures the size of gradient at layer $l$, and through it we track the dynamics of gradient backpropagation, be it explosion or vanishing. In contrast to vanilla tanh networks, which can experience both of these two phenomenon depending on the initialization variances, typical residual networks cannot have vanishing gradient, in the sense of vanishing $\p {{\boldsymbol{\oldchi}}}l$ as $l \to 1$; see \[thm:dalethRecReduced\] and \[thm:dalethRecFull\]. Furthermore, while vanilla tanh networks exhibit exponentially vanishing or exploding gradients, all of the activation/architecture pairings considered here, except the full residual network with ReLU, have subexponential gradient dynamics. While tanh residual networks (reduced or full) has $\p {{\boldsymbol{\oldchi}}}0 \approx \exp(\Theta(\sqrt l)) \p {{\boldsymbol{\oldchi}}}l$ (\[thm:dalethExpSqrtTanhFullRes\]), $\alpha$-ReLU residual networks for $\alpha < 1$ have $\p {{\boldsymbol{\oldchi}}}0 \approx {\mathsf{poly}}(l) \p {{\boldsymbol{\oldchi}}}l$ (\[thm:dalethDynamicsAlphaReLU\]). Instead of $\pd E/\pd x_i$, we may also consider the size of gradients of actual trainable parameters. For tanh and $\alpha$-ReLU with $\alpha < 1$, they are still subexponential and polynomial (\[thm:alphaReLUAllGradients\]). On the other hand, while $\p {{\boldsymbol{\oldchi}}}0 = \exp(\Theta(l))\p {{\boldsymbol{\oldchi}}}l$ for a ReLU resnet, its weight gradients have size independent of layer, within $O(1)$ (\[thm:alphaReLUAllGradients\])! This is the only instance in this paper of gradient norm being completely preserved across layers.
The above overviews the theoretical portion of this paper. Through experiments, we discover that we can very accurately predict whether one random initialization leads to better performance than another on the test set, after training, by leveraging this theory we build. Residual networks of different nonlinearities have different [*controlling quantities*]{}: for resnets with tanh, the optimal initialization is obtained by controlling the gradient explosion $\p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$; whereas for ReLU and $\alpha$-ReLU, the optimal initialization is obtained by maximizing ${\mathbf{s}}$ without running into numerical issues (with floating point computation). See \[sec:experiments\] for details.
Over the course of our investigation of $\alpha$-ReLU, we derived several new identities involving the associated kernel functions, first defined in [@cho_kernel_2009], which relate them to the zeroth Bessel functions (\[lemma:JalphaBessel,lemma:LAlphaRec,lemma:JalphaRec,lemma:JalphaGrad\]).
Theoretical Results
===================
In what follows in the main text, we assume $\sigma_\bullet > 0$ for all $\bullet = w, v, b, a$; in the appendix, the formal statement of each main theorem will contain results for other cases. We are interested in the two major categories of nonlinearities used today: tanh-like and rectified units. We make the following formal definitions as a foundation for further consideration.
We say a function $\phi$ is [**tanh-like**]{} if $\phi$ is antisymmetric ($\phi(-x) = -\phi(x)$), $|\phi(x)| \le 1$ for all $x$, $\phi(x) \ge 0, \forall x \ge 0$, and $\phi(x)$ monotonically increases to 1 as $x \to \infty$.
\[defn:alphaReLU\] Define the $\alpha$-ReLU $\psi_\alpha(x) = x^\alpha$ if $x > 0$ and 0 otherwise. [^3]
Antisymmetric/RRN Any/FRN
-------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------- --
Theorems \[thm:p\_q\_linear\], \[thm:lambda\_gamma\_recurrence\], \[thm:dalethRecReduced\] Theorems \[thm:fullResPQRec\], \[thm:full\_res\_l\_g\_recurr\], \[thm:dalethRecFull\]
: Main Recurrences[]{data-label="tab:recurrences"}
By applying the central limit theorem as described in the last section, we derive a set of recurrences for different activation/architecture pairs, shown in \[tab:recurrences\] (see appendix for proofs). They leverage certain integral transforms [^4] as in the following
\[defn:integralTransform\] Define the transforms ${\mathrm{V}}$ and ${\mathrm{W}}$ by ${\mathrm{V}}\phi(q) := \EV[\phi(z)^2: z \sim {\mathcal{N}}(0, q)]$ and ${\mathrm{W}}\phi(\rho, \nu) := \EV[\phi(z)\phi(z'): (z, z') \sim {\mathcal{N}}(0, \begin{pmatrix}\rho & \nu \\ \nu & \rho \end{pmatrix})]$.
These recurrences are able to track the corresponding quantities in practice very well. For example, \[fig:theory\_tracks\_pratice\] compares theory vs experiments for the tanh/FRN pair. The agreement is very good for tanh/RRN (not shown, but similar to the case of tanh/FRN with $\sigma_v = 1$ and $\sigma_a = 0$) and $\alpha$-ReLU/FRN as well (see \[fig:alphaReLUTheoryVsEmpirics\]).
As mentioned in previous sections, we seek to characterize the long term/high depth behavior of all of the quantities defined in \[sec:background\]. To do so, we solve for the asymptotics of the recurrences in \[tab:recurrences\], where $\phi$ is instantiated with tanh or $\alpha$-ReLU. Our main dynamics results are summarized in \[tab:dynamics\].
Tanh/RRN Tanh/FRN ReLU/FRN $\alpha$-ReLU/FRN, $\alpha<1$
------------------------------------- ------------------------------------------------------- ---------------------------------------------------- ---------------------------------------------------- --------------------------------------------------------------------------------------- --
$\p {\mathbf{p}}l$ $\Theta(l)$, \[thm:p\_q\_linear\] $\Theta(l)$, \[thm:pIsLinearTanh\] $\exp(\Theta(l))$, \[thm:pDynamicAlphaReLU\] $\Theta(l^{1/(1-\alpha)})$, \[thm:pDynamicAlphaReLU\]
$\p {\mathbf{s}}l$ $\Theta(l)$, \[thm:edynamics\] $\Theta(l)$, \[thm:eDynamicsFullResTanh\] $\exp(\Theta(l))$, \[thm:ReLUSquaredConvergence\] $\Theta(l^{1/(1-\alpha)})$, \[thm:alphaReLUeConvergence\]
$\p {\mathbf{e}}l - {\mathbf{e}}^*$ ${{\check\Theta}}(l^{\f2 \pi -1})$, \[thm:edynamics\] ${\mathsf{poly}}(l)$, \[thm:eDynamicsFullResTanh\] $\Theta(l^{-2})$, \[thm:ReLUSquaredConvergence\] ${\mathsf{poly}}(l)$, \[thm:alphaReLUeConvergence\]
$\p {{\boldsymbol{\oldchi}}}l$ $\exp(\Theta(\sqrt l))$, \[thm:dalethExpSqrtTanh\] $\exp(\Theta(\sqrt l))$, \[thm:dalethRecFull\] $\exp(\Theta(l))$, \[thm:dalethDynamicsAlphaReLU\] $\Theta(l^{\f{\alpha^2}{(1-\alpha)(2 \alpha - 1)}})$, \[thm:dalethDynamicsAlphaReLU\]
: Summary of Main Dynamics Results. Note that while $\p{{\boldsymbol{\oldchi}}}l$ is exponential for ReLU/FRN, the gradients with respect to weight parameters have norms (${\boldsymbol{\oldchi}}_w$ and ${\boldsymbol{\oldchi}}_v$) constant in $l$ (\[thm:alphaReLUAllGradients\]). Also, the $\p {{\boldsymbol{\oldchi}}}l$ entry for $\alpha$-ReLU is for $\alpha \in (3/4, 1)$ only[]{data-label="tab:dynamics"}
Tanh
----
![Our equations predict the relevant quantities very well in practice. These plots make the comparison between prediction and measurements for the full resnet with tanh activation, with $\sigma_v^2 = 1.5$, $\sigma_a^2 = .5$, $\sigma_w^2 = 1.69$, $\sigma_b^2 = .49$. Left-to-right: [**(a)**]{} $\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ against layer $l$ for 200 layers. [**(b)**]{} $\p {\mathbf{e}}l = \p {\boldsymbol{\oldgamma}}l /\p {\mathbf{p}}l$ against $l$ for 200 layers. Both (a) and (b) trace out curves for different initial conditions. [**(c)**]{} Different gradient quantities against $l$ for 50 layers. From left to right the layer number $l$ decreases, following the direction of backpropagation. Notice that the gradient increases in norm as $l \to 1$. All three figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:theory_tracks_pratice"}](graphics/tanh_full_res_theory_vs_exp-chi.pdf){height=".17\textheight"}
#### Forward dynamics.
When $\phi = \tanh$, $\p {\mathbf{p}}l$ and $\p {\mathbf{q}}l$ increase as $\Theta(l)$ in either RRN or FRN (\[thm:p\_q\_linear\]), as one might expect by observing that ${\mathrm{V}}\tanh( {\mathbf{q}}) \to 1$ as ${\mathbf{q}}\to \infty$ so that, for example in the RRN case, the recurrence ${\mathbf{p}}= {\mathrm{V}}\tanh( {\mathbf{q}}) + {\underline}{\mathbf{p}}$ becomes ${\mathbf{p}}= 1 + {\underline}{\mathbf{p}}$. This is confirmed graphically by the black lines of the leftmost chart of \[fig:theory\_tracks\_pratice\]. We carefully verify that this intuition is correct in its proof in the appendix, and find that in fact $\p {\mathbf{p}}l \sim l$ in the RRN case and $\p {\mathbf{p}}l \sim (\sigma_v^2 + \sigma_a^2)l$ in the FRN case.
What about $\p {\boldsymbol{\oldgamma}}l$? The middle chart of \[fig:theory\_tracks\_pratice\] shows that over time, $\p {\mathbf{e}}l = \p {\boldsymbol{\oldgamma}}l / \p {\mathbf{p}}l$ contracts toward the center of the interval $[0, 1]$, but from the looks of it, it is not clear whether there is a stable fixed point ${\mathbf{e}}^*$ of ${\mathbf{e}}$ or not. We prove that, in fact, [**all trajectories of ${\mathbf{e}}$ not starting at 1 do converge to a single fixed point, but only at a polynomial rate**]{}, in both the RRN and FRN cases (\[thm:p\_q\_linear\] and \[thm:full\_res\_l\_g\_recurr\]); we can even explicitly compute the fixed point and the rate of convergence: For FRN, there is a [**unique stable fixed point**]{} ${\mathbf{e}}^* < 1$ determined by the equation $${\mathbf{e}}^* = \f 1 {\sigma_v^2 + \sigma_a^2}[\sigma_v^2 \f 2 \pi \arcsin\lp {\mathbf{e}}^* \rp + \sigma_a^2],$$ and $|{\mathbf{e}}^* - \p {\mathbf{e}}l|$ decreases like $l^{-\delta^*}$, where $$\delta^* := 1 - \f 2 \pi \f 1 {\sqrt{1 - ({\mathbf{e}}^*)^2}} \f{\sigma_v^2 }{\sigma_v^2 + \sigma_a^2}.$$ Since ${\mathbf{e}}^* < 1$, ${\mathbf{s}}= (1 - {\mathbf{e}}) {\mathbf{p}}= \Theta({\mathbf{p}}) = \Theta(l).$ The case of RRN can be viewed as a special case of the above, setting $\sigma_v^2 = 1$ and $\sigma_a^2 = 0$, which yields ${\mathbf{e}}^* = 0$ and $\delta^* = 1 - \f 2 \pi$. We observe that both ${\mathbf{e}}^*$ and $\delta^*$ only depend on the ratio $\rho := \sigma_a/\sigma_v$, so in \[fig:edelta\_plot\] we graph these two quantities as a function of $\rho$. ${\mathbf{e}}^*$ and $\delta^*$ both increase with $\rho$ and asymptotically approach 1 and $\nicefrac 1 2$ respectively from below. When $\rho = \sigma_a = 0$, ${\mathbf{e}}^* = 0$ and $\delta^* = 1 - \f 2 \pi$. Thus the rate of convergence at its [**slowest**]{} for tanh/FRN is $\delta^* = 1 - \f 2 \pi \approx 0.36338$, where asymptotically the network tends toward a [**chaotic regime**]{} ${\mathbf{e}}^* = 0$, corresponding to a large weight variance and a small bias variance; it at its [**fastest**]{} is $\delta^* = \nicefrac 1 2$, where asymptotically the network tends toward a [**stable regime**]{} ${\mathbf{e}}^* = 1$, corresponding to a large bias variance and small weight variance. We verify $\delta^*$ by comparing $\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}$ to $l^{-\delta^* - 1}$ in log-log scale. If $\p {\mathbf{e}}l = \Theta(l^{-\delta^*})$, then $\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1} = \Theta(l^{-\delta^* - 1})$ and should obtain the same slope as $l^{-\delta^* - 1}$ as $l \to \infty$. The middle figure of \[fig:edelta\_plot\] ascertains that this is indeed the case, starting around layer number 400.
![Left-to-right: [**(a)**]{} Plots of ${\mathbf{e}}^*$ and $\delta^*$ against $\sigma_a/\sigma_v$. [**(b)**]{} In log-log scale: the dashed line is $l^{-\delta^* - 1}$, and the colored lines are $\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}$ for different initial conditions $\p {\mathbf{e}}0$. That they become parallel at about $l = 400$ on verifies that $\p {\mathbf{e}}l = \Theta(l^{-\delta^*})$. [^5] [**(c)**]{} In log-log scale: The dashed line is $\mathcal A \sqrt l$ ($\mathcal A$ given in \[thm:dalethExpSqrtTanhFullRes\]), and the colored lines are $\log(\p \bullet 1/\p \bullet l)$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$. That they all converge together starting around $l=1000$ indicates that the approximation in \[thm:dalethExpSqrtTanhFullRes\] is very good for large $l$.[]{data-label="fig:edelta_plot"}](graphics/edelta_plot.pdf "fig:"){height=".16\textheight" width=".3\textwidth"} ![Left-to-right: [**(a)**]{} Plots of ${\mathbf{e}}^*$ and $\delta^*$ against $\sigma_a/\sigma_v$. [**(b)**]{} In log-log scale: the dashed line is $l^{-\delta^* - 1}$, and the colored lines are $\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}$ for different initial conditions $\p {\mathbf{e}}0$. That they become parallel at about $l = 400$ on verifies that $\p {\mathbf{e}}l = \Theta(l^{-\delta^*})$. [^6] [**(c)**]{} In log-log scale: The dashed line is $\mathcal A \sqrt l$ ($\mathcal A$ given in \[thm:dalethExpSqrtTanhFullRes\]), and the colored lines are $\log(\p \bullet 1/\p \bullet l)$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$. That they all converge together starting around $l=1000$ indicates that the approximation in \[thm:dalethExpSqrtTanhFullRes\] is very good for large $l$.[]{data-label="fig:edelta_plot"}](graphics/verify_deltastar_n_grad-chi.pdf "fig:"){height=".16\textheight"}
#### Backward dynamics.
Finally, we show that the gradient is approximated by $$\begin{aligned}
\p {{\boldsymbol{\oldchi}}}{m} &= \exp(\mathcal A(\sqrt{l} - \sqrt{m}) + O(\log l - \log m))\p {{\boldsymbol{\oldchi}}}{l} \label{eqn:tanhGradEst}\tag{$\star$}\end{aligned}$$ where $\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \sigma_w$ in the RRN case and $
\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \f{\sigma_v^2 \sigma_w}{\sqrt{\sigma_v^2 + \sigma_a^2}}$ in the FRN case (\[thm:dalethExpSqrtTanh\] and \[thm:dalethExpSqrtTanhFullRes\]). The rightmost plot of \[fig:edelta\_plot\] verifies that indeed, for large $l \ge 1000$, this is a very good approximation. This demonstrates that the mean field assumption of independent backpropagation weights is very practical and convenient even for residual networks.
Note that in the FRN case, the constant $\mathcal A$ can be decomposed into $\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \cdot \sigma_v \cdot \sigma_w \cdot (1 + \sigma_a^2/\sigma_v^2)^{-1/2}$. Consider the ratio $\rho := \sigma_a/\sigma_v$. If $\rho \gg 1$, then ${\mathbf{e}}^* \approx 1$ (\[fig:jjj\_vs\_id\_main\]), meaning that the typical network essentially computes a constant function, and thus unexpressive; at the same time, large $\rho$ makes $\mathcal A$ small, and thus ameliorating the gradient explosion problem, making the network more trainable. On the other hand, if $\rho \ll 1$, then ${\mathbf{e}}^* \approx 0$ (\[fig:jjj\_vs\_id\_main\]), the typical network can tease out the finest differences between any two input vectors, and a final linear layer on top of such a network should be able to express a wide variety of functions [@poole_exponential_2016]; at the same time, small $\rho$ increases $\mathcal A$, worsening the gradient explosion problem, making the network less trainable. This is the same expressivity-trainability tradeoff discussed in [@schoenholz_deep_2017].
$\alpha$-ReLU
-------------
#### Forward dynamics.
As with the tanh case, to deduce the asymptotic behavior of random $\alpha$-ReLU resnets, we need to understand the transforms ${\mathrm{V}}\psi_\alpha$ and ${\mathrm{W}}\psi_\alpha$. Fortunately, ${\mathrm{V}}\psi_\alpha$ has a closed form, and ${\mathrm{W}}\psi_\alpha$ has been studied before [@cho_kernel_2009]. In particular, if $\alpha > -\f 1 2$, then ${\mathrm{V}}\psi_\alpha( {\mathbf{q}}) = {\mathsf{c}}_\alpha {\mathbf{q}}^{\alpha}$, where ${\mathsf{c}}_\alpha$ is a constant with a closed form given by \[lemma:VtPsiAlpha\]. In addition, by [@cho_kernel_2009], we know that ${\mathrm{W}}\psi_\alpha( {\mathbf{q}}, {\mathbf{c}}{\mathbf{q}}) = {\mathrm{V}}\psi_\alpha( {\mathbf{q}}) {\mathbb{J}}_\alpha({\mathbf{c}})$ for ${\mathbb{J}}_\alpha$ given in \[sec:AlphaReluForwardProofs\]. \[fig:jjj\_vs\_id\_main\] shows a comparison of ${\mathbb{J}}_\alpha$ for different $\alpha$s along with the identity function.
Substituting in ${\mathsf{c}}_\alpha {\mathbf{q}}^\alpha$ for ${\mathrm{V}}\psi_\alpha$, we get a difference equation ${\mathbf{p}}- {\underline}{\mathbf{p}}= \sigma_v^2 {\mathsf{c}}_\alpha (\sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2)^\alpha + \sigma_a^2$ governing the evolution of ${\mathbf{p}}$. This should be reminiscent of the differential equation $\dot P(l) = C P(l)^\alpha$, which has solution $\propto l^{1/(1-\alpha)}$ for $\alpha < 1$, and $\propto \exp(Cl)$ when $\alpha = 1$. And indeed, the solutions $\p {\mathbf{p}}l$ to these difference equations behave asymptotically exactly like so (\[thm:pDynamicAlphaReLU\]). Thus [**ReLU behaves very explosively compared to $\alpha$-ReLU with $\alpha<1$**]{}. In fact, in simulations, for $\sigma_w^2 = 1.69$ and $\sigma_v^2 = 1.5$, the ReLU resnets overflows into `inf`s after around 100 layers, while there’s no problem from any other kind of networks we consider.
Regardless, [**$\alpha$-ReLU for all $\alpha$ massages $\p {\mathbf{e}}l$ toward a fixed point ${\mathbf{e}}^*$ that depends on $\alpha$**]{}. [ When $\phi = \psi_1$, the standard ReLU, $\p {\mathbf{e}}l$ converges to 1 asymptotically as $C l^{-2}$ for an explicit constant $C$ depending on $\sigma_v$ and $\sigma_w$ only (\[thm:ReLUSquaredConvergence\]), so that ${\mathbf{s}}= (1 - {\mathbf{e}}){\mathbf{p}}= \Theta(l^{-2}\exp(\Theta(l))) = \exp(\Theta(l)).$ When $\phi = \psi_\alpha$ for $\alpha < 1$, then $\p {\mathbf{e}}l$ converges to the nonunit fixed point ${\mathbf{e}}^*$ of ${\mathbb{J}}_\alpha$ at a rate of ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1-\dot {\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$ is independent of the variances (\[thm:alphaReLUeConvergence\]), so that ${\mathbf{s}}= \Theta({\mathbf{p}})$.]{} These rates are verified in \[fig:alphaReLUVerifyExponents\].
#### Backward dynamics.
Finally, we have also characterized the rate of gradient growth for any $\alpha \in (\f 3 4, 1]$. [^7] [**In the case of $\alpha = 1$, the dynamics of ${{\boldsymbol{\oldchi}}}$ is exponential**]{}, the same as that of ${\mathbf{p}}$, $\p{{\boldsymbol{\oldchi}}}{l-m} = \p {{\boldsymbol{\oldchi}}}{l} B^m$ where $B =\f 1 2 \sigma_v^2 \sigma_w^2 + 1$. [**For $\alpha \in (\f 3 4, 1)$, the dynamics is polynomial**]{}, but with different exponent in general from that of the forward pass: $\p{{\boldsymbol{\oldchi}}}{l-m} = \Theta(1) \p {{\boldsymbol{\oldchi}}}{l} (l/(l-m))^R$ for $R = \f{\alpha^2}{(1-\alpha)(2 \alpha - 1)}$, where the constants in $\Theta(1)$ do not depend on $l$ or $m$. This exponent $R$ is minimized on $\alpha \in [\f 3 4, 1)$ at $\alpha = \nicefrac 3 4$, where $R = \nicefrac 9 2$ (but on $\alpha \in (\f 1 2, 1)$ it is minimized at $\alpha = \nicefrac 2 3$, where $R = 4$); see \[fig:backprop\_exponent\_alpha-relu\]. These exponents are verified empirically in \[fig:alphaReLUVerifyExponents\].
Looking only at ${{\boldsymbol{\oldchi}}}$ and the gradients against the biases, it seems that ReLU suffers from a dramatic case of exploding gradients. But in fact, because ${{\boldsymbol{\oldchi}}}$ gains a factor of $B$ moving backwards while ${\mathbf{p}}$ loses a factor of $B$, the gradient norm ${\boldsymbol{\oldchi}}_w^{(l-m)}$ (and similarly for ${\boldsymbol{\oldchi}}_v^{(l-m)}$) is independent of how far, $m$, the gradient has been propagated (\[thm:alphaReLUAllGradients\]) — this is certainly the best gradient preservation among all of the models considered in this paper. Thus strangely, random ReLU FRN exhibits both the best (constant for $v$ and $w$) and the worse (exponential for $a$ and $b$) gradient dynamics. This begs the question, then, is this a better deal than other $\alpha$-ReLU for which for any learnable parameter we have at most a polynomial blowup with depth in its gradient? Our experiments (discussed below) show that $\alpha$-ReLU is useful to the extent that smaller $\alpha$ avoids numerical issues with exponentiating forward and backward dynamics, but the best performance is given by the largest $\alpha$ that avoids them (\[fig:tanhHeatmaps\](c, d)); in fact, the metric expressivity ${\mathbf{s}}$, determines performance, not gradient explosion (see $\alpha$-ReLU experiments).
Experimental Results {#sec:experiments}
====================
![From left to right, top to bottom: **(a)** and **(b)**: $\sigma_w^2$, $L$, and test set accuracy of a grid of tanh reduced (left) and full (right) resnets trained on MNIST. Color indicates performance, with ligher colors indicating higher accuracy on test set. Other than the values on the axes, we have fixed $\sigma_b^2 = \sigma_a^2 = \f 1 2$ and $\sigma_v^2 = 1$. The white dotted lines are given by $\sigma_w^2 L = C$, where $C = 170$ on the left and $C = 145$ on the right. We see that both dotted lines accurately predict the largest optimal $\sigma_w$ for each depth $L$. **(c)** Varying the ratio $\sigma_a^2/\sigma_v^2$ while fixing $\sigma_v/\sqrt{1 + \sigma_a^2/\sigma_v^2}$, and thus fixing $\mathcal A$, the leading constant of $\log \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$. **(d)** in log-log scale: Heatmap gives the test accuracies of ReLU FRN for varying $\sigma_w^2$ and $L$. Curves give level sets for the log ratios $\log \p {\mathbf{s}}L / \p {\mathbf{s}}0 \approx \log \p {\mathbf{p}}L / \p {\mathbf{p}}0 \approx \log \p{{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L = L \log(1 + \sigma_v^2 \sigma_w^2/2)$. **(e)** Red heatmap shows the test accuracies of a grid of $\alpha$-ReLU FRN with varying $\alpha$ and $L$ as shown, but with all $\sigma_\bullet$s fixed. The white dashed curve gives a typical contour line of $L^R = \text{const}$, where $R = \f {\alpha^2}{(1-\alpha)(2\alpha-1)}.$ The yellow-to-blue curves form a set of level curves for $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = \text{const}$, with yellow curves corresponding to higher levels. []{data-label="fig:tanhHeatmaps"}](graphics/tanh_simple_resnet_trainability "fig:"){height="\height"} ![From left to right, top to bottom: **(a)** and **(b)**: $\sigma_w^2$, $L$, and test set accuracy of a grid of tanh reduced (left) and full (right) resnets trained on MNIST. Color indicates performance, with ligher colors indicating higher accuracy on test set. Other than the values on the axes, we have fixed $\sigma_b^2 = \sigma_a^2 = \f 1 2$ and $\sigma_v^2 = 1$. The white dotted lines are given by $\sigma_w^2 L = C$, where $C = 170$ on the left and $C = 145$ on the right. We see that both dotted lines accurately predict the largest optimal $\sigma_w$ for each depth $L$. **(c)** Varying the ratio $\sigma_a^2/\sigma_v^2$ while fixing $\sigma_v/\sqrt{1 + \sigma_a^2/\sigma_v^2}$, and thus fixing $\mathcal A$, the leading constant of $\log \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$. **(d)** in log-log scale: Heatmap gives the test accuracies of ReLU FRN for varying $\sigma_w^2$ and $L$. Curves give level sets for the log ratios $\log \p {\mathbf{s}}L / \p {\mathbf{s}}0 \approx \log \p {\mathbf{p}}L / \p {\mathbf{p}}0 \approx \log \p{{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L = L \log(1 + \sigma_v^2 \sigma_w^2/2)$. **(e)** Red heatmap shows the test accuracies of a grid of $\alpha$-ReLU FRN with varying $\alpha$ and $L$ as shown, but with all $\sigma_\bullet$s fixed. The white dashed curve gives a typical contour line of $L^R = \text{const}$, where $R = \f {\alpha^2}{(1-\alpha)(2\alpha-1)}.$ The yellow-to-blue curves form a set of level curves for $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = \text{const}$, with yellow curves corresponding to higher levels. []{data-label="fig:tanhHeatmaps"}](graphics/tanh_resnet_trainability "fig:"){height="\height"} ![From left to right, top to bottom: **(a)** and **(b)**: $\sigma_w^2$, $L$, and test set accuracy of a grid of tanh reduced (left) and full (right) resnets trained on MNIST. Color indicates performance, with ligher colors indicating higher accuracy on test set. Other than the values on the axes, we have fixed $\sigma_b^2 = \sigma_a^2 = \f 1 2$ and $\sigma_v^2 = 1$. The white dotted lines are given by $\sigma_w^2 L = C$, where $C = 170$ on the left and $C = 145$ on the right. We see that both dotted lines accurately predict the largest optimal $\sigma_w$ for each depth $L$. **(c)** Varying the ratio $\sigma_a^2/\sigma_v^2$ while fixing $\sigma_v/\sqrt{1 + \sigma_a^2/\sigma_v^2}$, and thus fixing $\mathcal A$, the leading constant of $\log \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$. **(d)** in log-log scale: Heatmap gives the test accuracies of ReLU FRN for varying $\sigma_w^2$ and $L$. Curves give level sets for the log ratios $\log \p {\mathbf{s}}L / \p {\mathbf{s}}0 \approx \log \p {\mathbf{p}}L / \p {\mathbf{p}}0 \approx \log \p{{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L = L \log(1 + \sigma_v^2 \sigma_w^2/2)$. **(e)** Red heatmap shows the test accuracies of a grid of $\alpha$-ReLU FRN with varying $\alpha$ and $L$ as shown, but with all $\sigma_\bullet$s fixed. The white dashed curve gives a typical contour line of $L^R = \text{const}$, where $R = \f {\alpha^2}{(1-\alpha)(2\alpha-1)}.$ The yellow-to-blue curves form a set of level curves for $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = \text{const}$, with yellow curves corresponding to higher levels. []{data-label="fig:tanhHeatmaps"}](graphics/tanh_resnet_sa_sb "fig:"){height="\height"}\
![From left to right, top to bottom: **(a)** and **(b)**: $\sigma_w^2$, $L$, and test set accuracy of a grid of tanh reduced (left) and full (right) resnets trained on MNIST. Color indicates performance, with ligher colors indicating higher accuracy on test set. Other than the values on the axes, we have fixed $\sigma_b^2 = \sigma_a^2 = \f 1 2$ and $\sigma_v^2 = 1$. The white dotted lines are given by $\sigma_w^2 L = C$, where $C = 170$ on the left and $C = 145$ on the right. We see that both dotted lines accurately predict the largest optimal $\sigma_w$ for each depth $L$. **(c)** Varying the ratio $\sigma_a^2/\sigma_v^2$ while fixing $\sigma_v/\sqrt{1 + \sigma_a^2/\sigma_v^2}$, and thus fixing $\mathcal A$, the leading constant of $\log \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$. **(d)** in log-log scale: Heatmap gives the test accuracies of ReLU FRN for varying $\sigma_w^2$ and $L$. Curves give level sets for the log ratios $\log \p {\mathbf{s}}L / \p {\mathbf{s}}0 \approx \log \p {\mathbf{p}}L / \p {\mathbf{p}}0 \approx \log \p{{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L = L \log(1 + \sigma_v^2 \sigma_w^2/2)$. **(e)** Red heatmap shows the test accuracies of a grid of $\alpha$-ReLU FRN with varying $\alpha$ and $L$ as shown, but with all $\sigma_\bullet$s fixed. The white dashed curve gives a typical contour line of $L^R = \text{const}$, where $R = \f {\alpha^2}{(1-\alpha)(2\alpha-1)}.$ The yellow-to-blue curves form a set of level curves for $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = \text{const}$, with yellow curves corresponding to higher levels. []{data-label="fig:tanhHeatmaps"}](graphics/reluHeatmapContour "fig:"){height="\height"} ![From left to right, top to bottom: **(a)** and **(b)**: $\sigma_w^2$, $L$, and test set accuracy of a grid of tanh reduced (left) and full (right) resnets trained on MNIST. Color indicates performance, with ligher colors indicating higher accuracy on test set. Other than the values on the axes, we have fixed $\sigma_b^2 = \sigma_a^2 = \f 1 2$ and $\sigma_v^2 = 1$. The white dotted lines are given by $\sigma_w^2 L = C$, where $C = 170$ on the left and $C = 145$ on the right. We see that both dotted lines accurately predict the largest optimal $\sigma_w$ for each depth $L$. **(c)** Varying the ratio $\sigma_a^2/\sigma_v^2$ while fixing $\sigma_v/\sqrt{1 + \sigma_a^2/\sigma_v^2}$, and thus fixing $\mathcal A$, the leading constant of $\log \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$. **(d)** in log-log scale: Heatmap gives the test accuracies of ReLU FRN for varying $\sigma_w^2$ and $L$. Curves give level sets for the log ratios $\log \p {\mathbf{s}}L / \p {\mathbf{s}}0 \approx \log \p {\mathbf{p}}L / \p {\mathbf{p}}0 \approx \log \p{{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L = L \log(1 + \sigma_v^2 \sigma_w^2/2)$. **(e)** Red heatmap shows the test accuracies of a grid of $\alpha$-ReLU FRN with varying $\alpha$ and $L$ as shown, but with all $\sigma_\bullet$s fixed. The white dashed curve gives a typical contour line of $L^R = \text{const}$, where $R = \f {\alpha^2}{(1-\alpha)(2\alpha-1)}.$ The yellow-to-blue curves form a set of level curves for $\p {\mathbf{s}}l = \p {\mathbf{p}}l - \p {\boldsymbol{\oldgamma}}l = \text{const}$, with yellow curves corresponding to higher levels. []{data-label="fig:tanhHeatmaps"}](graphics/alphaReluGrid_GammaGapOverlay "fig:"){height="\height"}
Our experiments show a dichotomy of what matters in initialization: for tanh resnets, quality of an initialization is determined by how much gradient explosion there is (measured by $\p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$); for ($\alpha$-)ReLU resnets, it is determined by how expressive the random network is (measured by the metric expressivity $\p {\mathbf{s}}L$). We hypothesize this is because in tanh resnets, the gradient dynamics is much more explosive than the expressivity dynamics ($\exp(\Theta(\sqrt l))$ vs $\Theta(l)$), whereas for ReLU it’s somewhat the opposite (${\boldsymbol{\oldchi}}_w, {\boldsymbol{\oldchi}}_v = \Theta(1)$ vs ${\mathbf{s}}= \exp(\Theta(l))$).
#### Tanh, vary $\sigma_w$.
We train a grid of reduced and full tanh resnets on MNIST, varying the variance $\sigma_w^2$ and the number of layers (for FRN we fix $\sigma_v = 1$). The results are indicated in \[fig:tanhHeatmaps\](a, b). We see that in either model, deeper resnets favor much smaller $\sigma_w$ than shallower ones. The white dotted lines in \[fig:tanhHeatmaps\](a, b) confirm our theory: according to \[eqn:tanhGradEst\], for the same gradient ratio $R = \p {{\boldsymbol{\oldchi}}}0 / \p {{\boldsymbol{\oldchi}}}L$, we want $\log R \approx \sigma_w \sqrt L$. Indeed, the white dotted lines in \[fig:tanhHeatmaps\](a, b) trace out such a level curve and it remarkably pinpoints the largest $\sigma_w$ that gives the optimal test set accuracy for each depth $L$. Why isn’t the best initialization given by $R = 1 \iff \sigma_w = 0$? We believe that when $L$ and/or $\sigma_w$ is small, gradient dynamics no longer dominates the initialization quality because it has “less room to explode,” and expressivity issues start to dampen the test time performance.
#### Tanh, vary $\sigma_a^2/\sigma_v^2$.
As suggested in the analysis of \[eqn:tanhGradEst\], the ratio $\rho^2 = \sigma_a^2/\sigma_v^2$ determines the fixed point ${\mathbf{e}}^*$ and its convergence rate by itself while also contributes to the rate of gradient explosion in tanh FRN. We seek to isolate its effect on forward dynamics by varying $\sigma_v$ with $\rho$ such that $\sigma_v/\sqrt{1 + \rho^2}$ is kept constant, so that the leading term of the log gradient ratio is kept approximately equal for each $L$ and $\rho$. \[fig:tanhHeatmaps\](c) shows the test accuracies of a grid of tanh FRN initialized with such an ensemble of $\sigma_\bullet$s. What stands out the most is that performance is maximized essentially around a fixed value of $L$ regardless of $\rho$, which shows that indeed gradient dynamics determines the initialization quality in tanh resnets. There is also a minor increase in performance with increasing $\rho$ regardless of $L$; this is counterintuitive as increasing $\rho$ means “decreasing expressivity.” It is currently not clear what accounts for this effect.
#### ReLU, vary $\sigma_w$
We train a grid of ReLU FRN on MNIST, varying $\sigma_w^2 \in [0, 1.5]$ while fixing $\sigma_v^2 = 1, \sigma_a^2 = \sigma_b^2 = \f 1 2$. The resulting test set accuracies are shown in \[fig:tanhHeatmaps\](d). The dark upper region signifies failure of training caused by numerical issues with exploding activation and gradient norms: This corresponds to the region where $\p {\mathbf{p}}L$, which is a measure of the mean magnitude of an neuronal activation in layer $L$, becomes too big. We see that the best test accuracies are given by depths just below where these numerical issues occur. However, if we were to predict that the optimal init is the one minimizing $\p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}L \ge 1$, then we would be wrong — in fact it is exactly the opposite. In this case, the dynamics of $\p {\mathbf{s}}l, \p {\mathbf{p}}l$, and $\p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}l$ are approximately the same (all $\exp(\Theta(l))$ with the same hidden constants), and optimal performance corresponds to the highest $\p {\mathbf{s}}L$, $\p {\mathbf{p}}L$, and $\p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}L$ without running into `inf`s.
#### $\alpha$-ReLU, vary $\alpha$.
We similarly trained a grid of $\alpha$-ReLU FRN on MNIST, varying only $\alpha$ and the depth, fixing all $\sigma_\bullet$. \[fig:tanhHeatmaps\](e) shows their test accuracies. We see similar behavior to ReLU, where when the net is too deep, numerical issues doom the training (black upper right corner), but the best performance is given by $L$ just below where this problem occurs. In this case, if we were to predict optimality based on minimizing gradient explosion, we would be again wrong, and furthermore, the contour plot of $\p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}L$ (white dashed line) now gives no information at all on the test set accuracy. In contrast, the contours for $\p {\mathbf{s}}l$ succeeds remarkably well at this prediction (yellow/green lines).[^8] By interpolation, this suggests that indeed in the ReLU case, it is expressivity, not trainability, which determines performance at test time.
In all of our experiments, we did not find ${\mathbf{e}}$ dynamics to be predictive of neural network performance.
Conclusion
==========
In this paper, we have extended the mean field formalism developed by [@poole_exponential_2016; @raghu_expressive_2016; @schoenholz_deep_2017] to residual networks, a class of models closer to practice than classical feedforward neural networks as were investigated earlier. We proved and verified that in both the forward and backward passes, most of the residual networks discussed here do not collapse their input space geometry or the gradient information exponentially. We found our theory incredibly predictive of test time performance despite saying nothing about the dynamics of training. In addition, we overwhelmingly find, through theory and experiments, that an optimal initialization scheme must take into account the depth of the residual network. The reason that Xavier [@glorot_understanding_2010] or He [@he_delving_2015] scheme are not the best for residual networks is in fact not that their statistical assumptions are fragile — theirs are similar to our mean field theoretic assumptions, and they hold up in experiments for large width — but rather that their structural assumptions on the network break very badly on residual nets.
#### Open Problems.
Our work thus have shown that optimality of initialization schemes can be very unstable with respect to architecture. We hope this work will form a foundation toward a mathematically grounded initialization scheme for state-of-the-art architectures like the original He et al. residual network. To do so, there are still two major components left to study out of the following three:
Residual/skip connection
Batchnorm
Convolutional layers.
Recurrent architectures and attention mechanisms are also still mostly unexplored in terms of mean field theory. Furthermore, many theoretical questions still yet to be resolved; the most important with regard to mean field theory is: why can we make \[ass:symAct,ass:gradInd\] and still be able to make accurate predictions? We hope to make progress on these problems in the future and encourage readers to take part in this effort.
Acknowledgments {#acknowledgments .unnumbered}
===============
Thanks to Jeffrey Ling for early exploration experiments and help with the initial draft. Thanks to Felix Wong for offering his wisdom and experience working in statistical physics.
Additional Figures
==================
In figures appearing in the appendix, $\olddaleth$ means ${\boldsymbol{\oldchi}}$ (due to legacy reasons).
[m[.1]{}m[0.6]{}]{} $\alpha=1$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.9$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/9relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.8$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/8relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.7$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/7relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.6$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/6relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.55$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/55relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}\
$\alpha=.51$ & ![Empirical vs theoretical dynamics for $\p {\mathbf{p}}l, \p {\mathbf{e}}l$, and different gradient quantities for $\alpha$-ReLU, with format similar to \[fig:theory\_tracks\_pratice\]. We refer to each figure on each row from left to right as (a), (b), and (c). Note that in the $\alpha=1$ case, figure (a) ($\p {\mathbf{p}}l$ and $\p {\boldsymbol{\oldgamma}}l$ for different initial values) has log scale y-axis and (a) and (b) have x-axis ranging from 1 to 50, while for other $\alpha$, (a) has normal y-axis and (a) and (b) have x-axis ranging from 1 to 200. We do so because the norm of the activation vector in a typical ReLU resnet blows up into `NaN` at around layer 90, while this is not a problem for $\alpha < 1$. Our theoretical predictions track the average of empirical values closely for forward quantities $\p {\mathbf{p}}l, \p {\boldsymbol{\oldgamma}}l,$ and $\p {\mathbf{e}}\l$ for all $\alpha$, but variance is extremely large for $\p {\mathbf{e}}l$ at $\alpha = 1$; it also predicts the average gradient norm accurately for $\alpha = 1$ to $\alpha = .7$ (despite the fact that we should not expect so for $\alpha \le .75$ due to exploding variance (\[thm:dalethInfVarAlphaReLU\])), although variance is large for $\alpha = 1$ at earlier layers (i.e. later layers w.r.t backpropagation). However it [*consistently and significantly overestimates*]{} the average gradient norm for $\alpha = .6$ to $\alpha = .5$, where the variance is so large that one standard deviation below the mean results in negative values. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. All figures exhibit smooth curves, which are theoretical estimates, and irregular curves with shades around them, which indicate empirical means and standard deviations (both of which taken in regular scale, not log scale). For each $\alpha$, figures (a) and (b) are made with 20 runs of resnets of width 1000. (c) is made with 25 runs of resnets of width 250.[]{data-label="fig:alphaReLUTheoryVsEmpirics"}](graphics/51relu_full_res_theory_vs_exp "fig:"){height=".08\textheight"}
[m[.1]{}m[0.4]{}]{} $\alpha=.9$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/9relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
$\alpha=.8$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/8relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
$\alpha=.7$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/7relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
$\alpha=.6$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/6relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
$\alpha=.55$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/55relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
$\alpha=.51$ & ![We verify the exponents of the forward and backward dynamics for $\alpha$-ReLU FRN. For each row, the figures are labeled (a) and (b) from left to right. The format is the same as in \[fig:jjj\_vs\_id\_main\]. All figures are in log-log scale. [**(a)**]{} We exhibit our theoretical dynamics of the cosine distance $\p {\mathbf{e}}l$ based on the recurrences \[thm:fullResPQRec\] and \[thm:full\_res\_l\_g\_recurr\] for different initial conditions $\p {\mathbf{e}}0$. We draw $|\p {\mathbf{e}}l - \p {\mathbf{e}}{l-1}|$ for each of these dynamics in colored solid lines. We predict that each dynamic is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1 - \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$, and the dashed line gives $l^{-\mu-1}$ (\[thm:alphaReLUeConvergence\]), shifted vertically to better compare the slope in log scale (i.e. the exponent of the polynomial dynamics). (See footnote \[footnote:plotDelta\] for why we plot the dynamics this way). We see that the our asymptotic prediction is very accurate for the sequence of $\p {\mathbf{e}}l$ that starts with $\p {\mathbf{e}}0 = 0.99$, the closest to ${\mathbf{e}}^*$ for each $\alpha$, while other lines only slowly converge to the same exponent (which is the slope in the log-log plot). This is to be expected based on the proof of \[thm:alphaReLUeConvergence\]. For $\alpha = .9$, the $\p {\mathbf{e}}0 = .99$ line upticks at around $10^3$ and then turn into `NaN`s due to numerical instability. [**(b)**]{} Colored lines are $\p \bullet 0 / \p \bullet l$ for $\bullet = {{\boldsymbol{\oldchi}}}, {\boldsymbol{\oldchi}}_b, {\boldsymbol{\oldchi}}_w$ (we are not taking logs in addition to plotting in log-log scale like in \[fig:jjj\_vs\_id\]). The dashed lines are our asymptotic predictions for the dynamics with corresponding colors, based on \[thm:alphaReLUAllGradients\], again shifted appropriately to easily compare slope visually. We see that for every alpha our asymptotic predictions are highly accurate. For both (a) and (b), we did not show $\alpha = 1$ case as ReLU FRN runs into numerical issues quickly (i.e. with even for 100 layers) because of exponential explosions in $\p {\mathbf{p}}l$ and $\p {{\boldsymbol{\oldchi}}}l$ as predicted by \[thm:pDynamicAlphaReLU,thm:dalethDynamicsAlphaReLU\], so we cannot expect to empirically verify the precise predicted asymptotics. All plots are made with parameters $\sigma_v^2 = 1.5, \sigma_a^2 = .5, \sigma_w^2 = 1.69, \sigma_b^2 = .49$; only $\alpha$ is varied. []{data-label="fig:alphaReLUVerifyExponents"}](graphics/51relu_verify_deltastar_n_grad "fig:"){height=".08\textheight"}\
Symbol Meaning Ref
----------------------------------------- ---------------------------------------------------------------------------------------------------- ----------------------------------
$\sigma_\bullet$ standard deviation of trainable parameter $\bullet$
$\p x l$ activation vector/input vector
$\p h l$ hidden vector
$N$ width (same across all layers)
$\p {\mathbf{p}}l$ m.n. squared length of activation vector $\p x l$ \[defn:length\]
$\p {\mathbf{q}}l$ m.n. squared length of hidden vector $\p h l$ \[defn:length\]
$\p {\boldsymbol{\oldgamma}}l$ m.n. dot product $\p x l \cdot \p x l{}'$ \[defn:corr\]
$\p {\boldsymbol{\oldlambda}}l$ m.n. dot product $\p h l \cdot \p h l{}'$ \[defn:corr\]
$\p {\mathbf{s}}l$ m.n. squared distance $\|\p x l - \p x l {}'\|^2$ \[defn:corr\]
$\p {\mathbf{e}}l$ cosine distance $\p {\boldsymbol{\oldgamma}}l / \sqrt{\p {\mathbf{p}}l \p {\mathbf{p}}l {}'}$ \[defn:corr\]
${\mathbf{e}}^*$ limit value of $\p {\mathbf{e}}l$ as $l \to \infty$
$\p {\mathbf{c}}l$ cosine distance $\p {\boldsymbol{\oldlambda}}l / \sqrt{\p {\mathbf{q}}l \p {\mathbf{q}}l {}'}$ \[defn:corr\]
$\p {\boldsymbol{\oldchi}}l$ m.n. gradient squared norm w.r.t. $\p x l$ \[defn:grad\]
$\p {{\boldsymbol{\oldchi}}_\bullet} l$ m.n. gradient squared norm w.r.t. trainable parameter $\bullet$ \[defn:grad\]
$\phi$ variable nonlinearity $\R \to \R$
$\psi_\alpha$ $\alpha$-ReLU \[defn:alphaReLU\]
${\mathrm{V}}$ variance integral transform \[defn:integralTransform\]
${\mathrm{W}}$ covariance integral transform \[defn:integralTransform\]
$\delta^*$ $\p {\mathbf{e}}l$ converges like $\Theta(l^{-\delta^*})$ in tanh FRN \[thm:eDynamicsFullResTanh\]
$\mathcal A$ leading coeff of $\log \p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}L$ in tanh FRN \[thm:dalethExpSqrtTanhFullRes\]
$R$ $\log \p {\boldsymbol{\oldchi}}0 / \p {\boldsymbol{\oldchi}}L \sim R \log L$ for $(\alpha<1)$-ReLU \[thm:dalethDynamicsAlphaReLU\]
${\mathbb{J}}_\alpha$ kernel function of $\alpha$-ReLU \[lemma:basicJalpha\]
: Glossary of Symbols. “Mean normalized” is abbreviated “m.n.”[]{data-label="tab:glossary"}
A Listing of Main Theorems
==========================
Tanh
----
### Reduced Residual Network
[lemma]{}[pqrecurrence]{}\[lemma:p\_q\_recurrence\] Suppose $\phi$ is antisymmetric. Then in an RRN, ${\mathbf{p}}$ and ${\mathbf{q}}$ satisfy the recurrence $$\begin{aligned}
{\mathbf{q}}&= \sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2\\
{\mathbf{p}}&= {\mathrm{V}}\phi( {\mathbf{q}}) + {\underline}{\mathbf{p}}.\end{aligned}$$
[thm]{}[pqlinear]{}\[thm:p\_q\_linear\] Suppose $\phi$ is tanh-like. Assume RRN architecture.
- If $\sigma_w = 0$, then $\p {\mathbf{p}}l = l {\mathrm{V}}\phi( \sigma_b^2) + \p {\mathbf{p}}0$ and $\p {\mathbf{q}}l = \sigma_b^2$.
- If $\sigma_w > 0$, $\lim_{l \to \infty} \p {\mathbf{p}}l/ l = 1$ and $\lim_{l \to \infty} \p {\mathbf{q}}l /(\sigma_w^2 l) = 1$. If $\phi = \tanh$, then we can obtain more terms of the asymptotic expansions: $$\begin{aligned}
\p {\mathbf{p}}l &= l - 2 C \sigma_w^{-1} l^{1/2} - C^2 \sigma_w^{-2} \log l + O(1)\\
\p {\mathbf{q}}l &= \sigma_w^2 l - 2 C \sigma_w l^{1/2} - C^2 \log l + O(1)
\end{aligned}$$ as $l \to \infty$, where $C = \sqrt{2 / \pi}$.
[thm]{}[lambdagammarecurrence]{}\[thm:lambda\_gamma\_recurrence\] Suppose $\phi$ is antisymmetric. Then in an RRN, ${\boldsymbol{\oldlambda}}$ and ${\boldsymbol{\oldgamma}}$ satisfy the recurrence $$\begin{aligned}
{\boldsymbol{\oldlambda}}&= \sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2\\
{\boldsymbol{\oldgamma}}&= {\mathrm{W}}\phi({\mathbf{q}}, {\boldsymbol{\oldlambda}}) + {\underline}{\boldsymbol{\oldgamma}}.\end{aligned}$$
[thm]{}[edynamics]{}\[thm:edynamics\] Suppose $\phi$ is a tanh-like nonlinearity in an RRN. Assume $\p {\mathbf{e}}0 < 1$.
- If $\sigma_w = 0$, then $\p {\boldsymbol{\oldgamma}}l = l {\mathrm{W}}\phi( \sigma_b^2, \sigma_b^2) + \p {\boldsymbol{\oldgamma}}0 = l {\mathrm{V}}\phi( \sigma_b^2) + \p {\boldsymbol{\oldgamma}}0$ and $\p {\boldsymbol{\oldlambda}}l = \sigma_b^2$, so that $\p {\mathbf{e}}l \to 1$ and $1 - \p {\mathbf{e}}l = \Theta(l^{-1})$. As a result, $\p {\mathbf{s}}l = \p {\mathbf{p}}l (1 - \p {\mathbf{e}}l) = \Theta(1).$
- If $\sigma_w > 0$, then $\p {\boldsymbol{\oldgamma}}l = {{\check\Theta}}(l^{\f 2\pi})$, and $\p {\mathbf{e}}l \to 0$ like ${{\check\Theta}}(l^{\f 2 \pi - 1})$. Thus $\p {\mathbf{s}}l = \Theta(\p {\mathbf{p}}l) = \Theta(l).$
[thm]{}[dalethRecReduced]{}\[thm:dalethRecReduced\] For any nonlinearity $\phi$ in an RRN, under assumptions \[ass:symAct\] and \[ass:gradInd\], whenever $\dot \phi^2(\zeta)$ has finite variance for Gaussian variable $\zeta$, $$\begin{aligned}
{\underline}{{\boldsymbol{\oldchi}}}&= (\sigma_w^2 {\mathrm{V}}\dot \phi( {\mathbf{q}}) + 1){{\boldsymbol{\oldchi}}},&
{\boldsymbol{\oldchi}}_b &= {{\boldsymbol{\oldchi}}}{\mathrm{V}}\dot \phi( {\mathbf{q}}),&
{\boldsymbol{\oldchi}}_w &= {{\boldsymbol{\oldchi}}}{\mathrm{V}}\dot \phi( {\mathbf{q}}) {\underline}{\mathbf{p}}. \end{aligned}$$
[thm]{}[dalethExpSqrtTanh]{}\[thm:dalethExpSqrtTanh\] For $\phi = \tanh$ in an RRN,
- If $\sigma_w = 0$, $\p {{\boldsymbol{\oldchi}}}m = \p {{\boldsymbol{\oldchi}}}l$ for all $l, m$.
- If $\sigma_w > 0$, $$\begin{aligned}
\log(\p {{\boldsymbol{\oldchi}}}{m}/\p {{\boldsymbol{\oldchi}}}l) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B (\log l - \log m) + O(1)
\end{aligned}$$ where $\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \sigma_w$ and $\mathcal B = \f 4 {3\pi} - \sigma_w^2 \f 4 {9\pi}$.
[thm]{}[dalethExpSqrtTanhAllGrad]{}\[thm:dalethExpSqrtTanhAllGrad\] Suppose $\phi = \tanh$. Then in an RRN
- If $\sigma_w = 0$, $\p {\boldsymbol{\oldchi}}l _b = \p {{\boldsymbol{\oldchi}}}L {\mathrm{V}}\dot \phi( \sigma_b^2)$ and $\p {\boldsymbol{\oldchi}}l _w = \p {{\boldsymbol{\oldchi}}}L {\mathrm{V}}\dot \phi( \sigma_b^2) ((l-1) {\mathrm{V}}\phi( \sigma_b^2) + \p {\mathbf{p}}0),$ where $L$ is the last layer.
- If $\sigma_w > 0$, $$\begin{aligned}
\log(\p{\boldsymbol{\oldchi}}m _b / \p {\boldsymbol{\oldchi}}l _b) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B_b (\log l - \log m) + O(1)\\
\log(\p{\boldsymbol{\oldchi}}m _w / \p {\boldsymbol{\oldchi}}l _w) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B_w (\log l - \log m) + O(1)
\end{aligned}$$ where $\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \sigma_w$ (same as $\mathcal A$ in \[thm:dalethExpSqrtTanh\]) and $\mathcal B_b = \mathcal B + \f 1 2, \mathcal B_w = \mathcal B - \f 1 2$, with $\mathcal B = \f 4 {3\pi} - \sigma_w^2 \f 4 {9\pi}$ (same as $\mathcal B$ in \[thm:dalethExpSqrtTanh\]).
### Full Residual Network
[thm]{}[fullResPQRec]{} \[thm:fullResPQRec\] For any nonlinearity $\phi$ in an FRN, $$\begin{aligned}
{\mathbf{q}}&= \sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2\\
{\mathbf{p}}&= \sigma_v^2 {\mathrm{V}}\phi( {\mathbf{q}}) + \sigma_a^2 + {\underline}{\mathbf{p}}\end{aligned}$$
[thm]{}[pIsLinearTanh]{}\[thm:pIsLinearTanh\] Suppose $\phi$ is tanh-like. Assume the FRN architecture.
- If $\sigma_w = 0$, then $\p {\mathbf{p}}l = (\sigma_v^2 {\mathrm{V}}\phi( \sigma_b^2) + \sigma_a^2)l +\p {\mathbf{p}}0$, and $\p {\mathbf{q}}l = \sigma_b^2$.
- If $\sigma_w > 0$, then $\p {\mathbf{p}}l = b_0 l + b_1 l^{1/2} + b_2 \log l + O(1)$, where $$\begin{aligned}
b_0 &= \sigma_v^2 + \sigma_a^2\\
b_1 &= \f{-2C \sigma_v^2 \sigma_w^{-1}}{\sqrt{\sigma_v^2 + \sigma_a^2}}\\
b_2 &= \f{-C^2 \sigma_v^4 \sigma_w^{-2}}{(\sigma_v^2 + \sigma_a^2)^2}
\end{aligned}$$ and $C = \sqrt{\f 2 \pi}$. Additionally, $\p {\mathbf{q}}l = \sigma_w^2 b_0 l + \sigma_w^2 b_1 l^{1/2} + \sigma_w^2 b_2 \log l + O(1)$.
![Empirical verification of \[thm:pIsLinearTanh\].](graphics/tanh_p_asymptotic_expansion.pdf){height=".16\textheight"}
[thm]{}[LGRecFullRes]{}\[thm:full\_res\_l\_g\_recurr\] For any nonlinearity $\phi$, in an FRN $$\begin{aligned}
{\boldsymbol{\oldlambda}}&= \sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2\\
{\boldsymbol{\oldgamma}}&= \sigma_v^2 {\mathrm{W}}\phi({\mathbf{q}}, {\boldsymbol{\oldlambda}}) + \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}}\end{aligned}$$
[thm]{}[eDynamicsFullResTanh]{} \[thm:eDynamicsFullResTanh\] Assume $\phi = \tanh$ in an FRN. Suppose $\p {\mathbf{e}}0 < 1$.
- If $\sigma_w = 0$, then $\p {\boldsymbol{\oldlambda}}l = \sigma_b^2$ and $\p {\boldsymbol{\oldgamma}}l = l (\sigma_v^2 {\mathrm{W}}\phi( \sigma_b^2, \sigma_b^2) + \sigma_a^2) + \p {\boldsymbol{\oldgamma}}0 = l (\sigma_v^2 {\mathrm{V}}\phi( \sigma_b^2) + \sigma_a^2) + \p {\boldsymbol{\oldgamma}}0$. Thus $\p {\mathbf{e}}l \to 1$ and $1 - \p {\mathbf{e}}l = \Theta(l^{-1})$. As a result, $\p {\mathbf{s}}l = \p {\mathbf{p}}l (1 - \p {\mathbf{e}}l) = \Theta(1).$
- If $\sigma_w > 0$, then $\p {\mathbf{e}}l$ converges to the unique fixed point ${\mathbf{e}}^* \not = 1$ determined by the equation $${\mathbf{e}}^* = \f 1 {\sigma_v^2 + \sigma_a^2}[\sigma_v^2 \f 2 \pi \arcsin\lp {\mathbf{e}}^* \rp + \sigma_a^2].$$ Furthermore, $\p {\mathbf{e}}l$ converges to ${\mathbf{e}}^*$ polynomially: $|\p {\mathbf{e}}l - {\mathbf{e}}^*|$ is ${{\check\Theta}}(l^{-\delta^*})$, where $$\delta^* := 1 - \f 2 \pi \f 1 {\sqrt{1 - ({\mathbf{e}}^*)^2}} \f{\sigma_v^2 }{\sigma_v^2 + \sigma_a^2} \in [\f 2 \pi - 1, \f 1 2)$$ Since ${\mathbf{e}}^* < 1$, $\p {\mathbf{s}}l = \Theta(\p {\mathbf{p}}l) = \Theta(l).$
[thm]{}[dalethRecFull]{} \[thm:dalethRecFull\] For any nonlinearity $\phi$ in an FRN, under assumptions \[ass:symAct\] and \[ass:gradInd\], whenever $\dot \phi(\zeta)^2$ has finite variance for Gaussian variable $\zeta$, $$\begin{aligned}
{\underline}{{\boldsymbol{\oldchi}}}&= (\sigma_v^2\sigma_w^2 {\mathrm{V}}\dot \phi( {\mathbf{q}}) + 1){{\boldsymbol{\oldchi}}},&
{\boldsymbol{\oldchi}}_b &= \sigma_v^2{{\boldsymbol{\oldchi}}}{\mathrm{V}}\dot \phi( {\mathbf{q}}),\\
{\boldsymbol{\oldchi}}_w &= \sigma_v^2{{\boldsymbol{\oldchi}}}{\mathrm{V}}\dot \phi( {\mathbf{q}}) {\underline}{\mathbf{p}},&
{\boldsymbol{\oldchi}}_v &= {{\boldsymbol{\oldchi}}}{\mathrm{V}}\phi( {\mathbf{q}}),&
{\boldsymbol{\oldchi}}_a &= {{\boldsymbol{\oldchi}}}\end{aligned}$$
[thm]{}[dalethExpSqrtTanhFullRes]{}\[thm:dalethExpSqrtTanhFullRes\] Assume $\phi = \tanh$ in an FRN.
- If $\sigma_w = 0$, $\p {{\boldsymbol{\oldchi}}}m = \p {{\boldsymbol{\oldchi}}}l$ for all $l, m$.
- If $\sigma_w > 0$, then for $l \ge m \ge 0,$ $$\log(\p {{\boldsymbol{\oldchi}}}{m} / \p {{\boldsymbol{\oldchi}}}l) = \mathcal A (\sqrt l - \sqrt m) + \mathcal B (\log l - \log m) + O(1)$$ where $$\begin{aligned}
\mathcal A &= \f 4 3 \sqrt{\f 2 \pi} \f{\sigma_v^2 \sigma_w}{\sqrt{\sigma_v^2 + \sigma_a^2}}\\
\mathcal B &= \f 4 {9\pi}\f{ \sigma_v^4 }{\sigma_v^2 + \sigma_a^2}\lp \f 3 {\sigma_v^2 + \sigma_a^2} - \sigma_w^2\rp
\end{aligned}$$
\[fig:tanhasymptoticsgrid\] shows empirical verification of the asymptotic expansion of ${{\boldsymbol{\oldchi}}}$ for various values of $\sigma_\bullet$s.
![Empirical verification of the asymptotic expansion of ${{\boldsymbol{\oldchi}}}$ for various values of $\sigma_\bullet$s. Note that we have chosen all small values for $\sigma_\bullet$s. For larger values, the constant term in \[thm:dalethExpSqrtTanhFullRes\] begins to dominate (primarily because of the expansion $\log(1+x) = x + \Theta(x^2)$ has large $\Theta$ term when $x$ is large), and ${{\boldsymbol{\oldchi}}}$ behaves more like $\exp(\Theta(l))$ up to depth 1000.[]{data-label="fig:tanhasymptoticsgrid"}](graphics/tanhAsymptoticsGrid){height=".3\textheight"}
[thm]{}[dalethExpSqrtTanhFullResAllGrad]{}\[thm:dalethExpSqrtTanhFullResAllGrad\] Suppose $\phi = \tanh$ in an FRN.
- If $\sigma_w = 0$, then $$\begin{aligned}
\p {\boldsymbol{\oldchi}}l _b &= \sigma_v^2 \p {{\boldsymbol{\oldchi}}}L {\mathrm{V}}\dot \phi( \sigma_b^2)\\
\p {\boldsymbol{\oldchi}}l _w &= \sigma_v^2 \p {{\boldsymbol{\oldchi}}}L {\mathrm{V}}\dot \phi( \sigma_b^2) ( (\sigma_v^2 {\mathrm{V}}\phi( \sigma_b^2) + \sigma_a^2)(l-1) +\p {\mathbf{p}}0)\\
\p {\boldsymbol{\oldchi}}l _v &= \p {{\boldsymbol{\oldchi}}}L {\mathrm{V}}\phi( \sigma_b^2)\\
\p {\boldsymbol{\oldchi}}l _a &= \p {{\boldsymbol{\oldchi}}}L.
\end{aligned}$$
- If $\sigma_w > 0$, then for $l \ge m \ge 0,$
$$\begin{aligned}
\log(\p{\boldsymbol{\oldchi}}m _b / \p {\boldsymbol{\oldchi}}l _b) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B_b (\log l - \log m) + O(1)\\
\log(\p{\boldsymbol{\oldchi}}m _w / \p {\boldsymbol{\oldchi}}l _w) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B_w (\log l - \log m) + O(1)\\
\log(\p{\boldsymbol{\oldchi}}m _a / \p {\boldsymbol{\oldchi}}l _a) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B (\log l - \log m) + O(1)\\
\log(\p{\boldsymbol{\oldchi}}m _v / \p {\boldsymbol{\oldchi}}l _v) &= \mathcal A (\sqrt l - \sqrt m) + \mathcal B (\log l - \log m) + O(1)
\end{aligned}$$
where $\mathcal A = \f 4 3 \sqrt{\f 2 \pi} \f{\sigma_v^2 \sigma_w}{\sqrt{\sigma_v^2 + \sigma_a^2}}$ and $\mathcal B = \f 4 {9\pi}\f{ \sigma_v^4 }{\sigma_v^2 + \sigma_a^2}\lp \f 3 {\sigma_v^2 + \sigma_a^2} - \sigma_w^2\rp$ are as in \[thm:dalethExpSqrtTanhFullRes\] and $\mathcal B_b = \mathcal B + \f 1 2$ and $\mathcal B_w = \mathcal B - \f 1 2$.
$\alpha$-ReLU
-------------
[lemma]{}[VtPsiAlpha]{}\[lemma:VtPsiAlpha\] If $\alpha > -\f 1 2$, then ${\mathrm{V}}\psi_\alpha( q) = {\mathsf{c}}_\alpha q^{\alpha}$, where ${\mathsf{c}}_\alpha = \f 1 {\sqrt \pi} 2^{\alpha - 1} \Gamma \left(\alpha+ \f 1 2\right)$.
Note that if $\alpha \le - \f 1 2$, then ${\mathrm{V}}\psi_\alpha( q)$ is not defined (its defining integral does not converge).
### Full Residual Network
By \[thm:fullResPQRec\] and \[lemma:VtPsiAlpha\], we have the length recurrences $$\begin{aligned}
{\mathbf{q}}&= \sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2\\
{\mathbf{p}}&= \sigma_v^2 {\mathsf{c}}_\alpha {\mathbf{q}}^\alpha + \sigma_a^2 + {\underline}{\mathbf{p}}\end{aligned}$$
[thm]{}[pDynamicAlphaReLU]{}\[thm:pDynamicAlphaReLU\] Suppose we have the nonlinearity $\phi = \psi_\alpha$. The in an FRN: If $\alpha = 1$, then $\p {\mathbf{p}}l = \Theta((1 + \sigma_v^2 \sigma_w^2/2)^l)$, with the hidden constant depending on the initial condition. If $0 < \alpha < 1$, then $\p {\mathbf{p}}l = \Theta(l^{\f 1 {1- \alpha}})$. More precisely, $\lim_{l \to \infty} {\mathbf{p}}/ l^{\f 1 {1-\alpha}} = [\sigma_v^2 \sigma_w^{2\alpha} {\mathsf{c}}_\alpha (1 - \alpha)]^{\f 1 {1-\alpha}}$.
\[fig:reluverifypasymptotics\] empirically verifies the asymptotics for $\alpha=1$ for various $\sigma_v$ and $\sigma_w$.
![Verification of the exponential asymptotics of $\p {\mathbf{p}}l$ when $\alpha=1$. The lines of each color correspond to different $(\sigma_w, \sigma_v)$ pairs, which are given in the legend. The solid lines are given by the recurrences \[thm:fullResPQRec\], and the dashed lines are given by our asymptotics $(1+\sigma_v^2\sigma_w^2/2)^l$ (\[thm:pDynamicAlphaReLU\]). Note that the y-axis is in log-scale.[]{data-label="fig:reluverifypasymptotics"}](graphics/relu_verify_p_asymptotics){height=".2\textheight"}
Similarly, by \[thm:full\_res\_l\_g\_recurr\], if ${\mathbf{q}}= {\mathbf{q}}'$, then $$\begin{aligned}
{\boldsymbol{\oldlambda}}&= \sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2\\
{\boldsymbol{\oldgamma}}&= \sigma_v^2 {\mathbf{q}}^\alpha {\mathrm{W}}\psi_\alpha( 1, {\mathbf{c}}) + \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}}\end{aligned}$$
[thm]{}[ReLUSquaredConvergence]{}\[thm:ReLUSquaredConvergence\] Suppose $\phi = \psi_1$. Then in an FRN, $\p {\mathbf{e}}l \to 1$ and $1 - \p {\mathbf{e}}l \sim [\f 1 4 \sigma_v^2 \sigma_w^2 \inv B U l]^{-2}$ for $B = 1 + \sigma_v^2 \sigma_w^2/2$ and $U = \f {2\sqrt 2}{3\pi}$. As a result, $\p {\mathbf{s}}l = (1 - \p {\mathbf{e}}\l) \p {\mathbf{p}}l = \Theta(l^{-2}\exp(\Theta(l))) = \exp(\Theta(l)).$
[thm]{}[alphaReLUeConvergence]{} \[thm:alphaReLUeConvergence\] Suppose $\phi = \psi_\alpha$ for $0 < \alpha < 1$ in an FRN. Then ${\mathbf{e}}$ converges to the unique nonunit fixed point ${\mathbf{e}}^*$ of ${\mathbb{J}}_\alpha$, and $|{\mathbf{e}}^* - \p {\mathbf{e}}l|$ is ${{\check\Theta}}(l^{-\mu})$, where $\mu = (1-\dot{\mathbb{J}}_\alpha({\mathbf{e}}^*))/(1-\alpha)$. Additionally, $\p {\mathbf{s}}l = \Theta(\p {\mathbf{p}}l) = \Theta(l^{1/(1-\alpha)}).$
\[fig:6reluverifyestar\] verifies empirically that ${\mathbf{e}}^*$ is indeed the fixed point of $\p {\mathbf{e}}l$. \[fig:alphaReLUVerifyExponents\] verifies empirically the convergence rate $l^{-\mu}$. \[fig:MuPlots\] plots $\dot {\mathbb{J}}_\alpha({\mathbf{e}}^*)$ and $\mu$ versus $\alpha$. It certainly looks like $\mu = \f 1 2 (1 - \alpha)$, but we have no proof for it. Based on this conjecture, we see there is a “discontinuity” of $\mu$ at $\alpha = 1$: $\mu \to 0$ as $\alpha \to 1$, but for $\alpha = 1$, the actual convergence dynamics has exponent $-2$ by \[thm:ReLUSquaredConvergence\].
![Verification of fixed point ${\mathbf{e}}^*$ in \[thm:alphaReLUeConvergence\] for $\alpha = .6$. Different colors correspond to different initial conditions $\p {\mathbf{e}}0$, and the dashed line gives the fixed point.[]{data-label="fig:6reluverifyestar"}](graphics/6relu_verify_estar){height=".2\textheight"}
![[**(a)**]{} A plot of $\dot {\mathbb{J}}_\alpha({\mathbf{e}}^*)$ versus $\alpha$. [**(b)**]{} A plot of the exponent $\mu$ of the dynamics of $|{\mathbf{e}}^* - \p {\mathbf{e}}l|$ (see \[thm:alphaReLUeConvergence\]) []{data-label="fig:MuPlots"}](graphics/JdotPlot "fig:"){height=".2\textheight"} ![[**(a)**]{} A plot of $\dot {\mathbb{J}}_\alpha({\mathbf{e}}^*)$ versus $\alpha$. [**(b)**]{} A plot of the exponent $\mu$ of the dynamics of $|{\mathbf{e}}^* - \p {\mathbf{e}}l|$ (see \[thm:alphaReLUeConvergence\]) []{data-label="fig:MuPlots"}](graphics/MuPlot "fig:"){height=".2\textheight"}
Because of the following theorem, we cannot expect the equations of \[thm:dalethRecFull\] to hold for $\alpha \le \f 3 4$.
[thm]{}[dalethInfVarAlphaReLU]{} \[thm:dalethInfVarAlphaReLU\] Suppose we have the nonlinearity $\psi_\alpha$ in an FRN. $\Var(\dot \psi_\alpha(\zeta)^2)$ diverges for any Gaussian variable $\zeta$ with mean 0 if $\alpha \le \f 3 4$ but is finite if $\alpha > \f 3 4$.
[thm]{}[dalethDynamicsAlphaReLU]{} \[thm:dalethDynamicsAlphaReLU\] Suppose we have the nonlinearity $\psi_\alpha$ in an FRN. If $\alpha = 1$, then $\p{{\boldsymbol{\oldchi}}}{l-m} = \p {{\boldsymbol{\oldchi}}}{l} \left(\f 1 2 \sigma_v^2 \sigma_w^2 + 1\right)^m$. If $\alpha \in (\f 3 4, 1)$, then $\p{{\boldsymbol{\oldchi}}}{l-m} = \Theta(1) \p {{\boldsymbol{\oldchi}}}{l} (l/(l-m))^R$ for $R = \f{\alpha^2}{(1-\alpha)(2 \alpha - 1)}$, where the constants in $\Theta(1)$ do not depend on $l$ or $m$.
This exponent $\f{\alpha^2}{(1 - \alpha)(2\alpha - 1)}$ is minimized at $\alpha = \f 3 4$ on $\alpha \in (3/4, 1)$, where the value is $\f 9 2$ (and at $\alpha = \f 2 3$ on $\alpha \in (1/2, 1)$, where the value achieved is 4) (\[fig:backprop\_exponent\_alpha-relu\](a)).
As a corollary,
[thm]{}[alphaReLUAllGradients]{}\[thm:alphaReLUAllGradients\] If $\phi = \psi_1$ in an FRN, then for $l \ge m \ge 0,$ $$\begin{aligned}
\p {\boldsymbol{\oldchi}}{l-m} _b &= \Theta(1) \p {{\boldsymbol{\oldchi}}}l B^m,&
\p {\boldsymbol{\oldchi}}{l-m} _w &= \Theta(1) \p {{\boldsymbol{\oldchi}}}l B^l,\\
\p {\boldsymbol{\oldchi}}{l-m} _v&= \Theta(1) \p {{\boldsymbol{\oldchi}}}l B^l,&
\p {\boldsymbol{\oldchi}}{l-m} _a &= \Theta(1) \p {{\boldsymbol{\oldchi}}}{l} B^m.\end{aligned}$$ where $B = 1 + \sigma_v^2\sigma_w^2/2$.
If $\phi = \psi_\alpha$ in an FRN, for $\alpha < 1$, then for $l \ge m \ge 0,$ $$\begin{aligned}
\p {\boldsymbol{\oldchi}}{l-m} _b &= \Theta(1) \p {{\boldsymbol{\oldchi}}}l l^R (l-m)^{-R-1},&
\p {\boldsymbol{\oldchi}}{l-m} _w &= \Theta(1) \p {{\boldsymbol{\oldchi}}}l l^R (l-m)^{\f \alpha {1-\alpha} - R},\\
\p {\boldsymbol{\oldchi}}{l-m} _v &= \Theta(1) \p {{\boldsymbol{\oldchi}}}l l^R (l-m)^{\f \alpha {1-\alpha} - R},&
\p {\boldsymbol{\oldchi}}{l-m} _a &= \Theta(1) \p {{\boldsymbol{\oldchi}}}{l} (l/(l-m))^R.\end{aligned}$$
\[fig:alphaReLUVerifyExponents\] verifies the backward asymptotic dynamics empirically for different $\alpha < 1$. \[fig:backprop\_exponent\_alpha-relu\](b) graphs the exponent $\f \alpha {1-\alpha} - R$ in terms of $\alpha$. We see that on $[0.5, 1]$, the maximum of this exponent is at $\alpha = 1$.
![**(a)** The exponent of the polynomial gradient dynamics with respect to $\alpha$-ReLU versus $\alpha$. **(b)** The exponent of the dynamics of ${\boldsymbol{\oldchi}}_v$ and ${\boldsymbol{\oldchi}}_w$.[]{data-label="fig:backprop_exponent_alpha-relu"}](graphics/backprop_exponent_alpha-relu.pdf "fig:"){width=".4\textwidth"} ![**(a)** The exponent of the polynomial gradient dynamics with respect to $\alpha$-ReLU versus $\alpha$. **(b)** The exponent of the dynamics of ${\boldsymbol{\oldchi}}_v$ and ${\boldsymbol{\oldchi}}_w$.[]{data-label="fig:backprop_exponent_alpha-relu"}](graphics/alphaReLU_chi_w_chi_v_exponent.pdf "fig:"){width=".4\textwidth"}
Proofs
======
A brief note about notation: We use $\sim$ to denote both how a random variable is sampled (ex: $x \sim {\mathcal{N}}(0, 1)$ for a Gaussian $x$) and how a function behaves asymptotically, i.e. $f(x) \sim g(x)$ as $x \to a$ iff $\lim_{x \to a} f(x)/g(x) = 1$. Context should be enough to differentiate between these two cases. We in addition use $\simeq$ to denote asymptotic expansion. For example, if $\{\alpha_i\}_{i \ge 0}$ is a sequence of strictly decreasing reals and $\{\beta_i\}_{i \ge 0}$ is a sequence of nonzero reals, then $$f(x) \simeq \sum_{i \ge 0} \beta_i (x - \xi)^{\alpha_i}$$ means that as $x \to \xi$, $f(x) - \sum_{i = 0}^N \beta_i (x - \xi)^{\alpha_i} = \Theta((x- \xi)^{\alpha_{N+1}})$.
Preliminary Lemmas
------------------
\[lemma:cExpansion\] We have $$\f{\sigma_w^2 {\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\mathbf{p}}+ \sigma_b^2} = {\mathbf{e}}(1 + O( {\boldsymbol{\oldgamma}}^{-1})).$$ regardless of whether $\p {\mathbf{e}}l = \p {\boldsymbol{\oldgamma}}l /\p {\mathbf{p}}l$ converges.
But suppose $\p {\mathbf{e}}l = \p {\boldsymbol{\oldgamma}}l /\p {\mathbf{p}}l \to {\mathbf{e}}^*$. If ${\mathbf{e}}^* < 1$, then $$\f{\sigma_w^2 {\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\mathbf{p}}+ \sigma_b^2} = {\mathbf{e}}(1 + \Theta( {\boldsymbol{\oldgamma}}^{-1})).$$ If ${\mathbf{e}}^* = 1$, then $$\f{\sigma_w^2 {\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\mathbf{p}}+ \sigma_b^2} = {\mathbf{e}}(1 + \Theta( {\epsilon}{\mathbf{p}}^{-1})),$$ where ${\epsilon}= 1 - {\mathbf{e}}$.
Write $M = \sigma_b^2/\sigma_w^2$. $$\begin{aligned}
\f{\sigma_w^2 {\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\mathbf{p}}+ \sigma_b^2} &= {\mathbf{e}}(1 + \f{1 + M {\boldsymbol{\oldgamma}}^{-1}}{1 + M {\mathbf{p}}^{-1}})\\
&= {\mathbf{e}}(1 + M(\inv {\boldsymbol{\oldgamma}}- \inv {\mathbf{p}}) + O(\inv {\mathbf{p}}(\inv {\boldsymbol{\oldgamma}}- \inv {\mathbf{p}}))).\end{aligned}$$ In any situation, $\inv {\boldsymbol{\oldgamma}}- \inv {\mathbf{p}}= O(\inv {\boldsymbol{\oldgamma}})$ because ${\boldsymbol{\oldgamma}}\le {\mathbf{p}}$, so this gives the first statement. If ${\mathbf{e}}^*$ exists and ${\mathbf{e}}^* < 1$, then $\inv {\boldsymbol{\oldgamma}}- \inv {\mathbf{p}}= \Theta(\inv {\boldsymbol{\oldgamma}})$, which yields the second statement. If ${\mathbf{e}}^*$ exists and ${\mathbf{e}}^* = 1$, then $\inv {\boldsymbol{\oldgamma}}- \inv {\mathbf{p}}= \inv {\mathbf{p}}((1-{\epsilon})^{-1} - 1) = \inv {\mathbf{p}}({\epsilon}+ O({\epsilon}^2)) = \Theta({\epsilon}\inv {\mathbf{p}})$.
For any function $f$ that is $(k+1)$-times differentiable in a neighborhood of $0$, we have the asymptotic expansion $$f(z) = \sum_{n = 0}^k \f{d^n f}{dz^n}(0) \f{z^n}{n!} + O(z^{k+1}), \text{as } z \to 0.$$ Since $$\begin{aligned}
\left.\f{d^n}{d(1/q)^n}q^{1/2}{\mathrm{V}}\phi( q)\right\rvert_{q \to \infty} &= \f {(-1)^n} {2^n\sqrt{2\pi}}\int_{-\infty}^\infty \phi^2(z) z^{2n} \dd z\end{aligned}$$ whenever the RHS is integrable, we have
\[lemma:Vt\_asymptotic\] Suppose $\phi^2(z) z^{2n}$ is integrable over $z \in \R$ for all $0 \le n \le N + 1$. Then ${\mathrm{V}}\phi( q) = q^{-1/2} (\sum_{n = 0}^N C_n q^{-n} + O(q^{-N-1}))$ as $q \to \infty$, where $$C_n := \f {(-1)^n} {2^n n! \sqrt{2\pi}}\int_{-\infty}^\infty \phi^2(z) z^{2n} \dd z.$$
Note that ${\operatorname{sech}}^d(z) = \Theta(e^{-d|z|})$ for $z \to \infty$ as long as $d > 0$, so that $C_n$ from the above result converges when $\phi = {\operatorname{sech}}^d$. Therefore
\[lemma:Vt\_sech\_asymptotics\] Let $d > 0$. We have ${\mathrm{V}}{\operatorname{sech}}^d( q) \simeq q^{-1/2} \sum_{n \ge 0} C_n q^{-n}$, where $$C_n := \f {(-1)^n} {2^n n! \sqrt{2\pi}}\int_{-\infty}^\infty {\operatorname{sech}}^{2d}(z) z^{2n} \dd z.$$
As corollaries, we obtain the following asymptotics.
\[lemma:Vt\_dot\_tanh\_asymptotics\] ${\mathrm{V}}\dot \tanh( q) = \f 2 3 \sqrt{\f 2 \pi} q^{-1/2} + \Theta(q^{-3/2})$ as $q \to \infty$.
Use \[lemma:Vt\_sech\_asymptotics\] along with the fact that $\dot \tanh(z) = {\operatorname{sech}}^2(z)$ and $\int {\operatorname{sech}}^4 z \dd z = \f 2 3 \tanh z + \f 1 2 {\operatorname{sech}}^2 z \tanh z$.
\[lemma:vtanhSqrtConvergence\] $1 - {\mathrm{V}}\tanh( q) = \sqrt{\f 2 \pi} q^{-1/2} + \Theta(q^{-3/2})$ as $q \to \infty$.
Use \[lemma:Vt\_sech\_asymptotics\] along with the fact that $1 - \tanh^2(z) = {\operatorname{sech}}^2(z)$ and $\int {\operatorname{sech}}^2 z \dd z = \tanh z$.
\[lemma:sech\_lower\_bound\] ${\operatorname{sech}}^2(t) \ge \exp(-t^2)$ for all $t$, with equality iff $t = 0$.
The lower bound is equivalent to $$\begin{aligned}
2 & \ge e^{t-t^2/2} + e^{-t -t^2/2}\end{aligned}$$ The RHS has derivative $(1 - t)e^{t - t^2/2} - (1 + t)e^{-t-t^2/2}$. This is 0 iff $$\begin{aligned}
\f{1 - t}{1 + t} = e^{-2t}\end{aligned}$$ which has a solution 0 and in general can only have solution $t\in (-1, 1)$ (by considering the sign of the LHS). Since each side is analytic in $t \in (-1, 1)$, we expand $$\begin{aligned}
\log \f{1 - t}{1 + t} &= \log e^{-2t}\\
\log(1-t) - \log(1+t) &= -2t\\
(-t-t^2-\cdots) - (t-t^2+\cdots) &= -2t\\
-2t-2t^3-\cdots &= -2t\end{aligned}$$ which shows that the only solution is $t = 0$. A simple plot shows that $t=0$ is a maximum, where the bound in question achieves equality.
\[lemma:V\_dot\_tanh\_lower\_bound\] Suppose $\phi = \tanh$. Then ${\mathrm{V}}\dot \phi( q) \ge \f 1 {\sqrt{4q+1}}$.
As a sanity check, \[lemma:Vt\_dot\_tanh\_asymptotics\] shows that ${\mathrm{V}}\dot \phi( q) \sim C_0 q^{1/2}$ where $C_0 \approx .5319$, which is above the .5 in this lemma.
By \[lemma:sech\_lower\_bound\], $$\begin{aligned}
{\mathrm{V}}\dot\phi( q) &= \int \dd \mu(z) \dot \phi^2(\sqrt q z)\\
&\ge \f 1 {\sqrt{2\pi}} \int \dd z \exp(-z^2/2 - 2qz^2)\\
&= \f 1 {\sqrt{2\pi}} \int \dd z \exp(-(4q + 1)z^2/2)\\
&= \f 1 {\sqrt{4q + 1}}.
\end{aligned}$$
\[fig:V\_dot\_tanh\_lower\_bound\] demonstrates \[lemma:V\_dot\_tanh\_lower\_bound\].
![Illustration of \[lemma:V\_dot\_tanh\_lower\_bound\]: ${\mathrm{V}}\dot \phi( q)$ vs $\f 1 {\sqrt{4q+1}}$ for $\phi = \tanh$. This bound is very tight, and for most purposes, $\f 1 {\sqrt{4q+1}}$ can be taken as a good approximation of ${\mathrm{V}}\dot \phi( q)$.[]{data-label="fig:V_dot_tanh_lower_bound"}](graphics/V_dot_tanh_lower_bound.pdf){width=".4\textwidth"}
\[lemma:power\_sum\_asymptotics\] Let $d \in \R$ and $1 < M < N$ with $N - M \in \Z^{\ge 0}$. Set $\Sigma(M, N, d) := \sum_{a=M}^N a^d$. If we fix $M$ and let $N \to \infty$, $$\Sigma(M, N, d) = \begin{cases}
\Theta(1) & \text{if $d < -1$}\\
\log N + O(1) & \text{if $d = -1$}\\
\f{N^{d+1}}{d+1} + O(1) & \text{if $-1 < d < 0$}\\
N - M + 1 & \text{if $d =0$}\\
\f 1 {d+1}N^{d+1} + \f 1 2 N^d + O(N^{\max(0, d-1)}) & \text{if $d > 0$}
\end{cases}$$
Consider the integrals $A = \int_M^{N+1} a^d \dd a$ and $B = \int^N_{M-1} a^d \dd a $. They evaluate to $A = \f 1 {d+1}((N+1)^{d+1} - M^{d+1})$ and $B = \f 1 {d+1}(N^{d+1} - (M-1)^{d+1})$ when $d \not = -1$ and to $A = \log(N+1) - \log M$ and $B = \log N - \log(M-1)$ when $d = -1$. When $d \le 0$, we have $A \le B$ and $\Sigma(M, N, d) \in [A, B]$; when $d > 0$, $B \le A$ and $\Sigma(M, N, d) \in [B, A].$ Thus, as $N \to \infty$ with $M$ fixed, when $d < -1$, $\Sigma(M, N, d) = \Theta(1)$; when $d = -1$, $\Sigma(M, N, -1) = \log N + O(1)$; and when $d > -1$, we have $\Sigma(M, N, d) = \f{N^{d+1}}{d+1} + O(N^{d})$.
Now for $a > 0$ and $d > -1$ and $d \not = 0, 1$, $$\begin{aligned}
\int_a^{a+1} z^d - a^d \dd z &= \f 1 {d+1} ( (a+1)^{d+1} - a^d )\\
&= (a^d + \f d 2 a^{d-1} + \cdots ) - a^d\\
&= \f d 2 a^{d-1} + \Theta(a^{d-2}).
\end{aligned}$$ where the hidden constants in $\Theta$ depend only on $d$ (and in fact this term vanishes if $d = 1$). Thus $$\begin{aligned}
\Sigma(M, N, d) &= \int_M^{N+1} z^d \dd z - \sum_{a=M}^N [ \f d 2 a^{d-1} + \Theta(a^{d-2})]\\
&= \f 1 {d+1} ((N+1)^{d+1} - M^{d+1}) - \f d 2 \Sigma(M, N, d-1) + \Theta(\Sigma(M, N, d-2))
\end{aligned}$$ If $-1 < d < 0$, then $\Sigma(M, N, d - 1) = \Theta(1)$, so that $\Sigma(M, N, d) = \f{(N+1)^{d+1}}{d+1} + O(1) = \f{N^{d+1}}{d+1} + O(1).$ If $d > 0$ and $d \not = 1$, then $\Sigma(M, N, d - 1) = \f{N^d}d$, so that $$\begin{aligned}
\Sigma(M, N, d) &= \f 1 {d+1} N^{d+1} + N^{d} + \Theta(N^{\max(0, d-1)}) - \f 1 2 N^d + \Theta(\Sigma(M, N, d-2)) \\
&= \f 1 {d+1}N^{d+1} + \f 1 2 N^d + O(N^{\max(0, d-1)}).
\end{aligned}$$
We can obtain more terms in the expansion for higher $d$ via the Euler-Maclaurin formula, but this suffices for our purposes.
Dynamics Zoo
------------
This section deduces the asymptotic behaviors of some sequences governed by recurrence equations. For the most part, the leading term of their asymptotic expansions is as one would expect from the corresponding differential equation. However, in some cases we need subleading terms for later results. They require slightly more nuanced reasoning. First we present a technical lemma.
\[lemma:simpleDynamics\] Let $F: \R \times \N \to \R$ be a function such that for a subset $U \sbe \R$, and for all $z, z' \in U, z \ge z' \implies F(z, n) \ge F(z', n)$ for every $n$. Suppose sequences $\p a l, \p b l, \p {c}l$ satisfy
- $\p a {l+1} = F(\p a l, l)$ for all $l$;
- $\p b {l+1} \le F(\p b l, l)$ for all $l$ above a constant $K_b$.
- $\p {c}{l+1} \ge F(\p {c}l, l)$ for all $l$ above a constant $K_c$.
and furthermore, $\p a l, \p b l, \p {c}l$ all fall into $U$ for $l$ above a constant $K_U$. If for some $m \ge \max(K_b, K_U)$, $\p b m \le \p a m$, then $\p b l \le \p a l, \forall l \ge m$. Similarly, if for some $n \ge \max(K_c, K_U)$, $\p {c}n \ge \p a n$, then $\p {c}l \ge \p a l, \forall l \ge n$.
For the first claim: $\p b m \le \p a m \implies \p b {m+1} \le F(\p b m, m) \le F(\p a m, m) = \p a {m+1}$. Here the last inequality used the monotonicity of $F$. Induction gives the desired result.
It’s similar for the second claim, where the inductive step is $\p {c}m \ge \p a m \implies \p {c}{m+1} \ge F(\p {c}m, m) \ge F(\p a m, m) = \p a {m+1}$.
\[lemma:alphaDeltaRec\] Suppose $\p {\epsilon}l$ satisfies the recurrence $$\p {\epsilon}l = \p {\epsilon}{l-1} (1 + \f \delta {l^\beta}).$$ for some nonzero constant $\delta \in \R$ independent of $l$.
- If $\beta > 1$, then $\p {\epsilon}l = \Theta(1)$.
- If $\beta = 1$, then $\p {\epsilon}l = \Theta(l^{\delta})$.
- If $0 < \beta < 1$, then $\p {\epsilon}l = \exp(\f{\delta}{1 - \beta}l^{1-\beta} + \tilde\Theta(l^{\psi_1(1-2\beta)}))$, where $\psi_1(x) = \max(0, x)$ is the ReLU function.
We have $$\begin{aligned}
\log \p {\epsilon}l &= \log \p {\epsilon}{l-1} + \log(1 + \delta/l^\beta)\\
&= \log \p {\epsilon}{l-1} + \delta/l^\beta + \Theta(\delta^2/l^{2\beta})
\end{aligned}$$ for large $l$. If $\beta > 1$, then $\sum_l l^{-\beta}$ converges, and $$\begin{aligned}
\log \p {\epsilon}l &= \log \p {\epsilon}0 - \Theta(1)\\
\p {\epsilon}l &= \Theta(1).
\end{aligned}$$ If $\beta = 1$, then $$\begin{aligned}
\log \p {\epsilon}l &= \log \p {\epsilon}0 + \delta \log l + \Theta(1)\\
\p {\epsilon}l &= \Theta(l^{\delta}).
\end{aligned}$$ If $\beta < 1$, then $$\begin{aligned}
\log \p {\epsilon}l &= \log \p {\epsilon}0 + \f\delta{1 - \beta} l^{1 - \beta} + \tilde\Theta(l^{1 - 2\beta})\\
\p {\epsilon}l &= \exp( \f\delta{1 - \beta} l^{1 - \beta} + \tilde\Theta(l^{\psi_1(1 - 2\beta)})).
\end{aligned}$$
\[lemma:alphaDeltaDynamics\] Suppose $\p {\epsilon}l = C l^{-\alpha} + \p {\epsilon}{l-1} (1 + \delta/l^{\beta})$ for $\alpha \in \R$, $C\not=0$, and $\delta \not = 0$. Then
- If $\beta > 1$, then
- $\p {\epsilon}l = \Theta(l^{1 - \alpha})$ if $\alpha \in (0, 1)$;
- $\p {\epsilon}l = \Theta(\log l)$ if $\alpha = 1$;
- $\p {\epsilon}l = \Theta(1)$ if $\alpha > 1$.
- If $\beta = 1$, then
- $\p {\epsilon}l = \Theta(l^{\max(\delta, 1-\alpha)})$ if $1-\delta \not = \alpha$.
- $\p {\epsilon}l = \Theta(l^{\delta} \log l)$ if $1-\delta = \alpha$.
Furthermore, for $\beta = -\delta = 1$, $\p {\epsilon}l \sim l^{-1}$ if $\alpha > 2$, $\p {\epsilon}l \sim l^{1-\alpha}$ if $\alpha < 2$, and $\p {\epsilon}l \sim l^{\delta} \log l$ if $\alpha = 2$.
We can unwind the recurrence to get $$\begin{aligned}
\p {\epsilon}l &= \sum_{m=1}^l m^{-\alpha} \prod_{n=m+1}^l (1 + \f \delta {n^\beta}) + \p {\epsilon}0 \prod_{n=1}^l (1 + \f \delta {n^\beta})
\end{aligned}$$
Suppose $\beta > 1$. By \[lemma:alphaDeltaRec\], we get $$\begin{aligned}
\p {\epsilon}l &= \Theta(1)\sum_{m=1}^l m^{-\alpha} + \p {\epsilon}0 \Theta(1)\\
&= \begin{cases}
\Theta(l^{1 - \alpha}) & \text{if $\alpha \in (0, 1)$}\\
\Theta(\log l) & \text{if $\alpha = 1$}\\
\Theta(1) & \text{if $\alpha > 1$.}
\end{cases}
\end{aligned}$$
Now suppose $\beta = 1$. By \[lemma:alphaDeltaRec\], we get $$\begin{aligned}
\p {\epsilon}l &= \sum_{m=1}^l m^{-\alpha} \Theta(m^{-\delta}l^\delta) + \p {\epsilon}0 \Theta(l^{\delta})
\end{aligned}$$ where the constants hidden inside the $\Theta$ are the same in every term of the sum. If $\alpha > 1 - \delta$, then $m^{-\delta - \alpha} = o(m^{-1})$, so that $\sum_{m=1}^l m^{-\delta - \alpha} = \Theta(1)$, and $$\begin{aligned}
\p {\epsilon}l &= \Theta(l^{\delta}) + \p {\epsilon}0 \Theta(l^{\delta})\\
&= \Theta(l^{\delta}).
\end{aligned}$$ On the other hand, if $\alpha < 1 - \delta$, then $\sum_{m=1}^l m^{-\delta - \alpha} = \Theta(l^{1 - \delta - \alpha})$. So $$\begin{aligned}
\p {\epsilon}l &= \Theta(l^{1 - \alpha}) + \p {\epsilon}0 \Theta(l^{\delta})\\
&= \Theta(l^{1 - \alpha}).
\end{aligned}$$ If $\alpha = 1 - \delta$, then $\sum_{m=1}^l m^{-\delta - \alpha} = \Theta(\log l)$. So $$\begin{aligned}
\p {\epsilon}l &= \Theta(l^{\delta}\log l) + \p {\epsilon}0 \Theta(l^{\delta})\\
&= \Theta(l^{\delta}\log l).
\end{aligned}$$
Finally, if $\beta \in (0, 1)$, then $$\begin{aligned}
\p {\epsilon}l &= e^{\f\delta{1-\beta}l^{1-\beta} + \Theta(l^{1-2\beta})}\sum_{m=1}^l m^{-\alpha} e^{\f{-\delta}{1-\beta} m^{1-\beta} + \Theta(m^{1-2\beta})} + e^{\f\delta{1-\beta}l^{1-\beta} + \Theta(l^{1-2\beta})}
\end{aligned}$$ The case of $\delta = -1$ telescopes, so that the upper and lower constants hidden in $\Theta$ can both be taken to be 1.
\[lemma:invlogDynamics\] Suppose for some $\beta > 0$, a sequence $\p {\epsilon}l$ satisfies $$\p {\epsilon}l = \p {\epsilon}{l-1} (1 - \mu (\p {\epsilon}{l-1})^\beta/l),\quad \p {\epsilon}0 \in (0, \f 1 \mu).$$ Then $\p {\epsilon}l \sim (\beta\mu \log l)^{-1/\beta}$.
Consider the differential equation $$\dot x_\mu = - \mu x^{\beta+1}_\mu/t$$ for constant $\mu$ has solution $x_\mu = [\beta(\mu \log t + C)]^{-1/\beta}$ for some constant $C$ determined by initial condition. Note that $$-\mu x_\mu(t)^{\beta+1}/t \le x_\mu(t+1) - x_\mu(t) \le - \mu x_\mu(t+1)^{\beta+1}/(t+1) = - (1 - o(t^{-1}))\mu x_\mu (t)^{\beta+1}/t.$$ For any small enough $\alpha > 0$, we apply \[lemma:simpleDynamics\] with $F({\epsilon}, l) = {\epsilon}- \mu {\epsilon}^{\beta+1}/l$ (which is monotonic in ${\epsilon}$ for small enough ${\epsilon}$), $\p {c}l = x_\mu(l)$, and $\p b l = x_{\mu-\alpha}(l)$ to obtain $$x_{\mu - \alpha}(l) \le \p {\epsilon}l \le x_\mu(l)$$ for large enough $l$ and appropriately chosen initial conditions. This shows that $\p {\epsilon}l = \Theta(\log l ^{-1/\beta})$ Taking $\alpha \to 0$, we also obtain the leading coefficient $\p {\epsilon}l \sim [\beta\mu \log l]^{-1/\beta}$.
\[lemma:polyDynamics\] Suppose a sequence $\p u l$ is governed by the equation $$\p u l - \p u {l-1} = A(\p u {l-1} + B)^\alpha,$$ where $\alpha \in [0, 1)$ and $A > 0$. Then $\p u l = K_1 l^{{\f 1 {1-\alpha}}}- K_2 l^{{\f \alpha {1-\alpha}}}\log l + o(l^{{\f \alpha {1-\alpha}}}\log l)$, where $K_1 = [A(1-\alpha)]^{{\f 1 {1-\alpha}}}$ and $K_2 = \f 1 2 A^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha$.
**Leading term.** The differential equation $$\dot x_{A, B} = A(x_{A, B} + B)^\alpha$$ has solution $x_{A,B}(l) = [A(1-\alpha)(l + S)]^{\f 1 {1-\alpha}} - B$ for some constant $S$. Since $\dot x_{A, B}$ is monotonic, we have (writing $x = x_{A, B}$ for brevity) $$A(x_{A, B}(l) + B)^\alpha = \dot x_{A, B}(l) \le x_{A, B}(l+1) - x_{A, B}(l) \le \dot x_{A, B}(l+1) \le (A + o(1))(x_{A, B}(l) + B)^\alpha$$ for large enough $l$. We apply \[lemma:simpleDynamics\] with $F(x, l) = x + A(x + B)^\alpha$ (which is monotonic in $x$ for large $x$), $\p {c}l = x_{A, B}(l)$, and $\p b l = x_{A-{\epsilon}, B}(l)$ to obtain $$x_{A-{\epsilon},B}(l) \le \p u l \le x_{A, B}(l)$$ for large enough $l$ and appropriate initial conditions. Therefore $\lim \p u l / l^{\f 1 {1-\alpha}} \in [[(A - {\epsilon}) ( 1- \alpha)]^{\f 1 {1 - \alpha}}, [A ( 1- \alpha)]^{\f 1 {1 - \alpha}}].$ Taking ${\epsilon}\to 0$ gives the leading term.
**Subleading term.** Now let $\p v l := \p u l - \aleph l^{\f 1 {1-a}}$, where $\aleph = [A(1-\alpha)]^{\f 1 {1-\alpha}}.$ Then we have the recurrence $$\begin{aligned}
\p v {l+1} + \aleph (l + 1)^{{{\f 1 {1-\alpha}}}} - \p v l - \aleph l^{{{\f 1 {1-\alpha}}}} &= A(\p v l + \aleph l^{{{\f 1 {1-\alpha}}}} + B)^\alpha\\
\p v {l+1} - \p v l + \aleph({{\f 1 {1-\alpha}}}l^{{{\f \alpha {1-\alpha}}}} &+ \f 1 2 ({{\f 1 {1-\alpha}}}) ({{\f \alpha {1-\alpha}}}) l^{{{\f \alpha {1-\alpha}}}-1} + \Theta(l^{{{\f \alpha {1-\alpha}}}- 2})) \\&= A[\aleph^\alpha l^{{{\f \alpha {1-\alpha}}}} + \alpha (\p v l+B) \aleph^{\alpha - 1} \inv l + \Theta((\p v l+B) l^{-1-{{\f 1 {1-\alpha}}}})]\\
\p v {l+1} - \p v l &= {{\f \alpha {1-\alpha}}}\p v l \inv l - \f 1 2 \aleph({{\f 1 {1-\alpha}}})({{\f \alpha {1-\alpha}}}) l^{{{\f \alpha {1-\alpha}}}- 1} + g(l)
\end{aligned}$$ for some $g(l) = O(l^{{{\f \alpha {1-\alpha}}}- 2} + \inv l)$ and where, to get the last equation, we have used $A \alpha^\alpha = {{\f 1 {1-\alpha}}}\aleph$ to cancel the $l^{{{\f \alpha {1-\alpha}}}}$ term and simplified $\alpha A \aleph^{\alpha-1} = {{\f \alpha {1-\alpha}}}$.
For any $J > 0$, the differential equation $\dot v_J(l) = {{\f \alpha {1-\alpha}}}v_J(l) \inv l - J l^{{{\f \alpha {1-\alpha}}}- 1}$ has solution $v_J(l) = C[l(1-\alpha)]^{{{\f \alpha {1-\alpha}}}} - J l^{{{\f \alpha {1-\alpha}}}} \log l$. Note that the functions $F_J(z, n) = z + {{\f \alpha {1-\alpha}}}z \inv n - J n^{{{\f \alpha {1-\alpha}}}- 1}$ and $G_J(z, n) = F_J(z, n) + g(n)$ is monotonic in $z$ (for positive $n$). For large $l$, we also have $\dot v_J(l)$ and $F_J(v_J(l), l) = v_J(l) + \dot v_J(l)$ decreasing in $l$. Thus for any ${\epsilon}> 0$ and $l$ large enough $$G_{J+{\epsilon}}(v_J(l), l) \le F_{J+{\epsilon}/2}(v_J(l), l) \le v_J(l) + \dot v_J(l+1) \le v_J(l+1) \le F_J(v_J(l), l) \le G_{J-{\epsilon}}(v_J(l), l).$$ Now apply \[lemma:simpleDynamics\] with $F = G_{K}$, $\p a l = \p v l, \p {c}l = v_{K-{\epsilon}}, \p b l = v_{K+{\epsilon}}$ where $K := \f 1 2 \aleph({{\f 1 {1-\alpha}}})({{\f \alpha {1-\alpha}}}) = \f 1 2 A^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha$, with appropriately chosen initial conditions. This yields $\lim_{l\to \infty} \p v l / (l^{{{\f \alpha {1-\alpha}}}} \log l) \in [-K - {\epsilon}, -K + {\epsilon}]$ for every ${\epsilon}> 0$, and there it must be equal to $K$. We have thus obtained the asymptotic expansion $$\p u l = [A(1-\alpha) l]^{{\f 1 {1-\alpha}}}- \f 1 2 A^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha l^{{\f \alpha {1-\alpha}}}\log l + o(l^{{\f \alpha {1-\alpha}}}\log l).$$
\[lemma:polyDynamicsPosAlpha\] Suppose a sequence $\p u l$ is governed by the equation $$\p u l - \p u {l-1} = -A(\p u {l-1} + B)^\alpha,$$ where $\alpha > 1$ and $A > 0$. Then $\p u l \sim [A(\alpha-1)l]^{\f 1 {1-\alpha}}$.
Similar to \[lemma:polyDynamics\].
\[lemma:polyDynamicsConstant\] Suppose a sequence $\p u l$ is governed by the equation $$\p u l - \p u {l-1} = A(\p u {l-1} + B)^\alpha + C,$$ where $\alpha \in (0, 1)$. Then $\p u l = K_1 l^{{\f 1 {1-\alpha}}}+ R(l)$, where the remainder $R(l)$ is $$R(l) \sim \begin{cases}
-K_2 l^{{{\f \alpha {1-\alpha}}}} \log l &\text{if $\alpha > \f 1 2$}\\
(C-K_2) l \log l &\text{if $\alpha = \f 1 2$ and $K_2 \not=C$}\\
\f{C(1-\alpha)}{1-2\alpha} l & \text{if $ \alpha < \f 1 2$}
\end{cases}$$ where $K_1 = [A(1-\alpha)]^{\f 1 {1-\alpha}}, K_2 = \f 1 2 A^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha$ as in \[lemma:polyDynamics\].
$u$ is bounded below by the dynamics $\p v l - \p v {l-1} = A(\p v {l-1} + B)^\alpha$ and bounded above by the dynamics $\p w l - \p w {l-1} = (A + o(1))(\p w {l-1} + B)^\alpha$. By \[lemma:polyDynamics\], both $v$ and $w$ are asymptotic to $\p u l \sim [A(1-\alpha)l]^{\f 1 {1-\alpha}}$, which gives the result.
Now define $\p v l = \p u l - [A(1-\alpha)l]^{{\f 1 {1-\alpha}}}$, and similar to the proof of \[lemma:polyDynamics\], we find $$\p v {l+1} - \p v l = {{\f \alpha {1-\alpha}}}\p v l \inv l - K l^{{{\f \alpha {1-\alpha}}}- 1} + C + g(l)$$ where $K = \f 1 2 A^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha$ and $g(l) = O(l^{{{\f \alpha {1-\alpha}}}- 2} + \inv l)$. If ${{\f \alpha {1-\alpha}}}> 1 \iff \alpha > \f 1 2$, then $C + g(l) = o(l^{{{\f \alpha {1-\alpha}}}- 1})$ and we can proceed as in the proof of \[lemma:polyDynamics\] to find $\p v l \sim K l^{{{\f \alpha {1-\alpha}}}} \log l.$ If ${{\f \alpha {1-\alpha}}}= 1 \iff \alpha = 1$ and $K \not = C$, then $\p v {l+1} - \p v l = {{\f \alpha {1-\alpha}}}\p v l l^{-1} - (K-C) l^{{{\f \alpha {1-\alpha}}}- 1} + g(l)$, so that the technique used in \[lemma:polyDynamics\] would obtain $\p v l \sim (K-C) l^{{{\f \alpha {1-\alpha}}}} \log l = (K-C) l \log l.$ If ${{\f \alpha {1-\alpha}}}< 1 \iff \alpha < \f 1 2$, then $\p v {l+1} - \p v l = {{\f \alpha {1-\alpha}}}\p v l l^{-1} + C + o(1)$, then by using the differential equation $\dot v_J(l) = {{\f \alpha {1-\alpha}}}v_J(l) l^{-1} + J$ to approximate the difference equation solution and applying \[lemma:simpleDynamics\] as in the proof of \[lemma:polyDynamics\], we obtain $\p v l(l) \sim \f{C(1-\alpha)}{1-2\alpha} l$.
Forward Dynamical Equations
---------------------------
Here we derive the recurrences governing the forward length and correlation quantities ${\mathbf{p}}, {\mathbf{q}}, {\boldsymbol{\oldlambda}}, {\boldsymbol{\oldgamma}}.$ We start with reduced residual networks.
We have $$\begin{aligned}
{\mathbf{q}}&= \la h_j^2 \ra = \la \sum_i (w_{ji}{\underline}x_i + b_j)^2 \ra\\
&= \la b_j^2 \ra + \sum_i \la w_{ji}^2 {\underline}x_i^2\ra + 2\sum_i \la w_{ji} {\underline}x_i b_j \ra + 2 \sum_{j \not= l} \la w_{ji} w_{li} x_i^2\ra
\end{aligned}$$ But $w_{ji}, w_{li}, {\underline}x,$ and $b_j$ form an independency, so the last two sums are 0, and the terms in the first sum split multiplicatively. Therefore $$\begin{aligned}
{\mathbf{q}}&= \sigma_b^2 + \sum_i \la w_{ji}^2 \ra \la {\underline}x_i^2 \ra\\
&= \sigma_b^2 + N \cdot \f {\sigma_w^2}{N} {\underline}{\mathbf{p}}\\
&= \sigma_b^2 + \sigma_w^2 {\underline}{\mathbf{p}}.
\end{aligned}$$
For the recurrence of ${\mathbf{p}}$, we have $$\begin{aligned}
{\mathbf{p}}&= \la x_i^2 \ra = \la (\phi(h_i) + {\underline}x_i)^2 \ra \\
&= \la \phi(h_i)^2\ra + \la {\underline}x_i^2 \ra + 2 \la\phi(h_i) {\underline}x_i \ra
\end{aligned}$$ As $N \to \infty$, the coefficient $w_{ii}$ of ${\underline}x_i$ in $h_i$ has vanishing covariance, so $h_i$ and ${\underline}x_i$ become independent. Therefore $\la \phi(h_i) {\underline}x_i \ra = \la \phi(h_i) \ra \la {\underline}x_i \ra$. Because $h_i$ is the sum of a large number of independent random variables, by CLT, $h_i$ is a Gaussian with mean $\sum_i \la w_{ji}\ra \la {\underline}x_i\ra + \la b_j\ra = 0$ since $\la w_{ji} \ra = \la b_j \ra = 0$. Our antisymmetry assumption on $\phi$ then implies $\la \phi(h_i) \ra = 0$. Therefore, $$\begin{aligned}
{\mathbf{p}}&= \la \phi(h_i)^2\ra + \la {\underline}x_i^2 \ra\\
&= {\mathrm{V}}\phi( {\mathbf{q}}) + {\underline}{\mathbf{p}}\end{aligned}$$ as desired.
Similar to \[lemma:p\_q\_recurrence\].
Now, for the full residual networks, the proofs are similar, but we no longer need to assume that $\phi$ is antisymmetric because of the randomization via the extra sets of weights.
$$\begin{aligned}
{\mathbf{q}}&= \la h_j^2 \ra
= \la ( w_j^i {\underline}x_i + b_j)^2\ra
= \la ( w_j^i {\underline}x_i)^2 \ra + \la b_j^2\ra\\
&= \sigma^2_w \la {\underline}x_i^2 \ra + \sigma_b^2\\
&= \sigma^2_w {\underline}{\mathbf{p}}+ \sigma^2_b\\
{\mathbf{p}}&= \la x_i^2 \ra
= \la (v^j_i\phi( h_j) +{\underline}x_i + a_i)^2 \ra\\
&= \sigma_v^2\la \phi( h_i)^2 \ra + \la {\underline}x_i^2 \ra + \sigma_a^2\\
&=\sigma_v^2 {\mathrm{V}}\phi( {\mathbf{q}}) + \sigma_a^2 + {\underline}{\mathbf{p}}\\
\end{aligned}$$
where in the third equality for ${\mathbf{p}}$, we are now using the independence of $v_i^j$ from all other variables to cancel out the terms, whereas before we had to rely on $\phi$ being antisymmetric.
Similar to \[thm:fullResPQRec\].
Backward Dynamical Equations
----------------------------
Here we derive the recurrences governing the gradient quantities ${{\boldsymbol{\oldchi}}}$ and ${\boldsymbol{\oldchi}}_{\bullet}$ for different $\bullet$, all under the gradient independence assumption. Write $\p {\beta}l_i = \pdf E {\p x l_i}$ for a cost function $E$.
For a reduced residual network, we have the following derivative computation: $$\begin{aligned}
\pdf {x_i} {{\underline}x_j} &= \delta_{ji} + \dot \phi(h_i) \pdf {h_i}{{\underline}x_j},&
\pdf {x_i}{h_j} &= \delta_{ji} \dot \phi(h_j),&
\pdf {h_i}{{\underline}x_j} &= w_{ij},&
\pdf {h_i}{w_{ij}} &= {\underline}x_j,&
\pdf {h_i}{b_j} &= \delta_{ij}.
\end{aligned}$$
Then $$\begin{aligned}
{\underline}{\beta}_j &= {\beta}_j + \sum_i {\beta}_i \dot \phi(h_i) \pdf {h_i}{{\underline}x_j}\\
&= {\beta}_j + \sum_i {\beta}_i\dot \phi(h_i) w_{ij}\\
\la {\underline}{\beta}_j^2 \ra &= \la [{\beta}_j + \sum_i {\beta}_i\dot \phi(h_i) w_{ij}]^2 \ra\\
&= \la {\beta}_j^2 \ra + \sum_i \la {\beta}_i^2 \dot \phi^2(h_i) (w_{ij})^2\ra \\
&\phantom{={}}+ 2\sum_{i < k} \la {\beta}_i {\beta}_k\dot\phi(h_i)w_{ij}\dot\phi(h_k)w_{kj}\ra + 2 \sum_i\la {\beta}_j {\beta}_i \dot\phi(h_i)w_{ij} \ra
\end{aligned}$$ The last two terms of the above vanish as $w_{ij}$ is independent from $w_{kj}$, $h_i, h_k$ and ${\beta}_i, {\beta}_j, {\beta}_k$ by \[ass:gradInd\], and $\la w_{ij} \ra = 0$.
Therefore, applying \[ass:symAct\], $$\begin{aligned}
\la {\underline}{\beta}_j^2 \ra &= \sigma_w^2\la {\beta}_j^2\ra\la\dot \phi^2(h_i)\ra + \la {\beta}_j^2\ra\\
&= (\sigma_w^2 {\mathrm{V}}\dot \phi( {\mathbf{q}}) + 1)\la {\beta}_j^2 \ra
\end{aligned}$$
We similarly have $$\begin{aligned}
\pdf E {b_j} &= \sum_i \pdf E {x_i} \pdf {x_i}{h_j}
= {\beta}_j \dot \phi(h_j), & \text{since $\pdf {x_i}{h_j} = \delta_{ji} \dot \phi(h_j)$}\\
\la \lp\pdf E {b_j}\rp^2 \ra &= \la {\beta}_j^2 \dot \phi(h_j)^2 \ra
= \la {\beta}_j^2 \ra {\mathrm{V}}\dot \phi( {\mathbf{q}}), & \text{by \cref{ass:gradInd}(b)};\\
\pdf E {w_{ji}} &= \sum_i \pdf E {x_i} \pdf {x_i}{h_j}\pdf {h_j}{w_{ji}}
= {\beta}_j \dot \phi(h_j){\underline}x_i, & \text{since $\pdf {x_i}{h_j} = \delta_{ji} \dot \phi(h_j)$}\\
\la \lp\pdf E {w_{ji}}\rp^2 \ra &= \la {\beta}_j^2 \dot \phi^2(h_j) {\underline}x_i^2 \ra
= \la {\beta}_j^2 \ra {\mathrm{V}}\dot \phi( {\mathbf{q}}) {\underline}{\mathbf{p}}, & \text{by \cref{ass:gradInd}(b)}
\end{aligned}$$ In the last equation we have also used the fact that as $N \to \infty$, $h_j$ and $x_i$ become independent (they are jointly Gaussian and their correlation $\la w_{ji}^2 \ra$ goes to 0 with $N$).
For the full residual network, we have the following derivative computations: $$\begin{aligned}
\pdf {x_i} {{\underline}x_j} &= \delta_{ji} + \sum_k v_{ik}\dot \phi(h_k) \pdf {h_k}{{\underline}x_j},&
\pdf {x_i}{h_j} &= v_{ij} \dot \phi(h_j),&
\pdf {h_i}{{\underline}x_j} &= w_{ij},&
\pdf {h_i}{w_{ij}} &= {\underline}x_j,&
\pdf {h_i}{b_i} &= 1,\\
&&
\pdf {x_i}{v_{ik}} &= \phi(h_k),&
\pdf {x_i}{a_i} &= 1.
\end{aligned}$$
Again let ${\beta}_j = \pdf E {x_j}$. Then $$\begin{aligned}
{\underline}{\beta}_j &= \sum_i {\beta}_i (\delta_{ji} + \sum_k v_{ik}\dot \phi(h_k) \pdf {h_k}{{\underline}x_j})\\
&= \sum_i {\beta}_i (\delta_{ji} + \sum_k v_{ik}\dot \phi(h_k) w_{kj})
\end{aligned}$$
Thus, $$\begin{aligned}
\la {\underline}{\beta}_j^2 \ra &= \la [\sum_i {\beta}_i (\delta_{ji} + \sum_k v_{ik}\dot \phi(h_k) w_{kj})]^2 \ra\\
&= \la {\beta}_j^2 \ra + \sum_{i, k} \la v_{ik}^2 \ra \la w_{kj}^2\ra {\mathrm{V}}\dot \phi( {\mathbf{q}}) \la {\beta}_i^2 \ra\\
&= \la {\beta}_j^2 \ra (1 + \sigma_v^2 \sigma_w^2 {\mathrm{V}}\dot \phi( {\mathbf{q}}))
\end{aligned}$$ where in the second equality we applied the independence argument as in the proof of \[thm:dalethRecReduced\], leveraging \[ass:gradInd\], and in the third equality we used \[ass:symAct\] to get $\la {\beta}_i^2\ra = \la {\beta}_j^2 \ra$.
The other computations are similar to the proof of \[thm:dalethRecFull\].
Tanh: Reduced Residual Network
------------------------------
### Forward Dynamics
The case with $\sigma_w = 0$ is trivial. We assume $\sigma_w > 0$ from here on.
**${\mathbf{p}}$ and ${\mathbf{q}}$ are asymptotically linear with $l$.** We first show that, for any $\omega < 1$, $$l + \p {\mathbf{p}}0 \ge \p {\mathbf{p}}l \ge \omega l$$ and $$\sigma_w^2 (l + \p {\mathbf{p}}0) + \sigma_b^2 \ge \p {\mathbf{q}}l \ge \sigma_w^2 \omega (l-1) + \sigma_b^2,$$ so that $\p {\mathbf{p}}l \sim l$ and $\p {\mathbf{q}}l \sim \sigma_w^2 l$.
The upper bounds are trivial, given ${\mathrm{V}}\phi( {\mathbf{q}}) \le 1$ for any ${\mathbf{q}}$. We show the lower bounds for any $\omega < 1$.
For any ${\epsilon}> 0$, define $\aleph_{\epsilon}$ by $\phi^2(\aleph_{\epsilon}) = \exp(-{\epsilon})$. Then $$\begin{aligned}
{\mathrm{V}}\phi( {\mathbf{q}}) &\ge \exp(-{\epsilon}) \Pr[z \not \in [-\aleph_{\epsilon}, \aleph_{\epsilon}]: z \sim {\mathcal{N}}(0, {\mathbf{q}})]\\
&\ge \exp(-{\epsilon})\lp 1 - \f{2 \aleph_{\epsilon}}{\sqrt{2\pi {\mathbf{q}}}} \rp\end{aligned}$$ where the second inequality follows from an overestimate of the $\Pr[z \in [-\aleph_{\epsilon}, \aleph_{\epsilon}]]$ via the mode of ${\mathcal{N}}(0, {\mathbf{q}})$.
For any ${\mathbf{q}}\ge \p {\mathbf{q}}0$, ${\mathrm{V}}\phi( {\mathbf{q}})$ is then lower bounded by $$\phi^2\lp \sqrt{\p {\mathbf{q}}0} \rp
\lp1 - \f{2\sqrt{\p {\mathbf{q}}0}}{\sqrt{2 \pi \p {\mathbf{q}}0}}\rp
= \phi^2\lp\sqrt{\p {\mathbf{q}}0}\rp
\lp1 - \sqrt{\f 2 \pi}\rp > 0.$$ Thus $\p {\mathbf{p}}l$ and $\p {\mathbf{q}}l$ are unbounded with $l$.
Furthermore, as ${\mathbf{q}}\to \infty$, the lower bound $\exp(-{\epsilon})\lp 1 - \f{2 \aleph_{\epsilon}}{\sqrt{2\pi {\mathbf{q}}}} \rp$ goes to $\exp(-{\epsilon})$, for any ${\epsilon}$. Therefore, for any $\omega < 1, \p {\mathbf{p}}l \ge \omega l$ and $\p {\mathbf{q}}l \ge \sigma_w^2 \omega (l-1) + \sigma_b^2$.
**Asymptotic expansion.** Now we repeat the following to get each successive asymptotic term of $\p {\mathbf{p}}l$ and $\p {\mathbf{q}}l$: We plug in the current asymptotic form of $\p {\mathbf{q}}l$ into ${\mathrm{V}}\tanh( {\mathbf{q}}) = 1 - C {\mathbf{q}}^{-1/2} + \Theta({\mathbf{q}}^{-3/2})$ (\[lemma:vtanhSqrtConvergence\]), where $C = \sqrt{2/\pi}$. Next we take the sum $\p {\mathbf{q}}l = \sum_{r=1}^l {\mathrm{V}}\tanh( \p {\mathbf{q}}r)$, which yields one more term in the asymptotic expansion of ${\mathbf{p}}$ than the last round. We then repeat until we get only constant terms.
The following exhibits a trace of this procedure, where in the summation step for $\p {\mathbf{q}}l$, we implicitly apply $$\begin{aligned}
{\mathbf{q}}&= \sigma_w^2 l + o(l) = \sigma_w^2 l (1 + o(1))\\
{\mathbf{q}}^{-1/2} &= \sigma_w^{-1} l^{-1/2}(1 + o(1)) = \sigma_w^{-1} l^{-1/2} + o(l^{-1/2})\\
{\mathbf{p}}&= \sum_{r=1}^l 1 - C (\p {\mathbf{q}}r)^{-1/2} + \Theta((\p {\mathbf{q}}r)^{-3/2}) \\
&= \sum_{r=1}^l 1 - C (\sigma_w^{-1} r^{-1/2} + o(r^{-1/2})) + \Theta(r^{-3/2}) \\
&= l - 2C \sigma_w^{-1} l^{1/2} + o(l^{1/2})\\
{\mathbf{q}}&= \sigma_w^2 l - 2 C \sigma_w l^{1/2} + o(l^{1/2}) = \sigma_w^2 l (1 - 2 C \sigma_2^{-1} l^{-1/2} + o(l^{-1/2}))\\
{\mathbf{q}}^{-1/2} &= \sigma_w^{-1} l^{-1/2} ( 1 + C \sigma_w^{-1} l^{-1/2} + o(l^{-1/2}))= \sigma_w^{-1} l^{-1/2} + C \sigma_w^{-2} l^{-1} + o(l^{-1})\\
{\mathbf{p}}&= \sum_{r=1}^l 1 - C(\sigma_w^{-1} l^{-1/2} + C \sigma_w^{-2} l^{-1} + o(l^{-1})) + \Theta(l^{-3/2})\\
&= l - 2 C \sigma_w^{-1} l^{1/2} - C^2 \sigma_w^{-2} \log l + o(\log l)\\
{\mathbf{q}}&= \sigma_w^2 l (1 - 2 C \sigma_w^{-1} l^{-1/2} - C^2 \sigma_w^{-2} \f{\log l}{l} + o(\f{\log l}{l}))\\
{\mathbf{q}}^{-1/2} &=\sigma_w^{-1} l^{-1/2}(1 + C \sigma_w^{-1} l ^{-1/2} + \f 1 2 C^2 \sigma_w^{-2} \f{\log l}{l} + o (\f{\log l}{l}))\\
{\mathbf{p}}&= \sum_{r=1}^l 1 - C(\sigma_w^{-1} r^{-1/2} + C \sigma_w^{-2} r ^{-1} + \f 1 2 C^2 \sigma_w^{-3} \f{\log r}{r^{3/2}} + o (\f{\log r}{r^{3/2}}))) + \Theta(r^{-3/2})\\
&= l - 2 C \sigma_w^{-1} l^{1/2} - C^2 \sigma_w^{-2} \log l + O(1)\end{aligned}$$ which is what we want.
\[lemma:WtPhiAntisymmetric\]
Let $\phi$ is antisymmetric. Then for $\tau \in [0, \pi/2]$, $$\begin{aligned}
{\mathrm{W}}\phi( {q}, {q}\cos \tau) &= \lim_{t \to \tau}\f 1 {\pi \sin t} \int_{w' \ge |w|} \dd w \dd w' \Upsilon(w, w';\tau) \phi(\f {\sqrt{{q}}} {\sqrt{2}}(w + w')) \phi(\f {\sqrt {q}} {\sqrt{2}}(w' - w))\\
&= \f 1 \pi \int_{0}^\infty r\dd r e^{-r^2/2} \int_{0}^{\pi}\dd \theta \Sigma(\sqrt {q}r, \theta; \tau)\\
&= \f 1 \pi \int_0^\infty s \dd s \inv {q}e^{-s^2 \inv {q}/2} \int_0^\pi \dd \theta \Sigma(s, \theta; \tau)\\
&= \f 1 \pi \int_0^\pi \dd \theta \int_0^\infty \dd s e^{-s^2 \inv {q}/2} \pdf{}{s} \Sigma(s, \theta; \tau)
\end{aligned}$$ where $\Upsilon(w, w';\tau) :=e^{-\f 1 2(\f{w^2}{1-c} + \f{(w')^2}{1+c})} - e^{-\f 1 2(\f{(w')^2}{1-c} + \f{w^2}{1+c})}$ with $c = \cos \tau$, and $\Sigma(s, \theta; \tau) := \phi(s \sin \theta) \phi(s \sin(\theta - \tau)).$
Of course, in the above lemma, the limit in the first equation is only necessary when $\tau = 0$ or $\tau = \pi/2$.
Let $c := \cos \tau$ and $$\Gamma := {\mathrm{W}}\phi( {q}, cq) = \f 1 {2\pi {q}\sqrt{1 - c^2}} \int \dd \mathbf z \exp(-\mathbf z^T \Sigma^{-1} \mathbf z /2) \phi(z)\phi(z'),$$ where $\Sigma = \begin{pmatrix}{q}& cq \\ cq & {q}\end{pmatrix}$.
Our proof will have two portions: Symmetrization of the $\Gamma$ integral and trigonometric change of variables for evaluation.
**Symmetrization.** $\Sigma$ is diagonalized by $\Omega = \f 1 {\sqrt{2q}}\begin{pmatrix}-1 & 1 \\ 1 & 1 \end{pmatrix}$, $$\Sigma = \Omega^T \Diag(1 - c, 1 + c) \Omega.$$ By a change of variable $\mathbf w = \Omega \mathbf z$, so that $\dd \mathbf w = \inv {q}\dd \mathbf z$, we have $$\begin{aligned}
\Gamma &= \f 1 {2\pi \sqrt{1 - c^2}}\int \dd \mathbf w \exp(-\mathbf w^T \Diag(1 - c, 1 + c)^{-1}\mathbf w/2) \phi(\f {\sqrt {q}} {\sqrt{2}}(w' - w)) \phi(\f {\sqrt {q}} {\sqrt{2}}(w + w'))\\
&= \f 1 {2\pi \sqrt{1 - c^2}} \int \dd w \dd w' e^{-\f 1 2(\f{w^2}{1-c} + \f{(w')^2}{1+c})} \phi(\f {\sqrt{{q}}} {\sqrt{2}}(w' - w)) \phi(\f {\sqrt {q}} {\sqrt{2}}(w + w'))
\end{aligned}$$
By a change of variable swapping $w$ with $w'$, we get
$$\begin{aligned}
\Gamma &= -\f 1 {2\pi \sqrt{1 - c^2}} \int \dd w \dd w' e^{-\f 1 2(\f{(w')^2}{1-c} + \f{w^2}{1+c})} \phi(\f {\sqrt{{q}}} {\sqrt{2}}(w + w')) \phi(\f {\sqrt {q}} {\sqrt{2}}(w' - w))
\end{aligned}$$
Thus $$\begin{aligned}
2 \Gamma &= \f 1 {2\pi \sqrt{1 - c^2}} \int \dd w \dd w' \Upsilon(w, w';\tau) \phi(\f {\sqrt{{q}}} {\sqrt{2}}(w + w')) \phi(\f {\sqrt {q}} {\sqrt{2}}(w' - w))
\end{aligned}$$ where $$\Upsilon(w, w';\tau) = e^{-\f 1 2(\f{w^2}{1-c} + \f{(w')^2}{1+c})} - e^{-\f 1 2(\f{(w')^2}{1-c} + \f{w^2}{1+c})}.$$
![The integrand of $\Gamma$ after symmetrization. Here $c = .2$ and ${q}= 100$ and $\phi = \tanh$.[]{data-label="fig:symmetrization_of_Gamma"}](graphics/symmetrized_integrand.pdf){height=".2\textheight"}
Note that, by the antisymmetry of $\phi$, the integrand $K := \Upsilon(w, w';\tau)\phi(\ldots)\phi(\ldots)$ above has the symmetries $K(w, w') = K(w', w) = K(w, -w')$, and is everywhere nonnegative. \[fig:symmetrization\_of\_Gamma\] displays a contour plot of $K$ for typical values of ${q}$ and $c$. So $$\begin{aligned}
\Gamma &= \f 1 {\pi \sqrt{1 - c^2}} \int_{w' \ge |w|} \dd w \dd w' K(w, w').
\end{aligned}$$ This gives the first equation in the lemma.
**Polar Coordinates.** Let $\f w {\sqrt{1 - c}} = r \cos \theta, \f {w'} {\sqrt{1 + c}} = r \sin \theta$, so that $$\begin{aligned}
w &= r \cos \theta \sqrt{1-c} = \sqrt 2 r \cos \theta \sin \f \tau 2\\
w' &= r \sin \theta \sqrt{1+c} = \sqrt 2 r \sin \theta \cos \f \tau 2\\
\dd w \dd w' &= \sqrt{1-c^2}r \dd r \dd \theta = (\sin^2 \tau) r \dd r \dd \theta.
\end{aligned}$$ Then $$\begin{aligned}
\mathbf A &:= \int_{w' \ge |w|} e^{-(\f{w^2}{1-c} + \f{(w')^2}{1 +c})/2}\phi(\sqrt{{q}/2}(w+w'))\phi(\sqrt{{q}/2}(w'-w)) \dd w \dd w'\\
&= \sin^2 \tau \int_{0}^\infty r\dd r e^{-r^2/2} \int_{\tau/2}^{\pi - \tau/2}\dd \theta \phi(\sqrt {q}r \sin(\theta + \tau/2))\phi(\sqrt {q}r \sin(\theta - \tau/2)).
\end{aligned}$$ Similarly, let $\f w {\sqrt{1 + c}} = r \cos \theta, \f {w'} {\sqrt{1 - c}} = r \sin \theta$, so that $$\begin{aligned}
w &= r \cos \theta \sqrt{1+c} = \sqrt 2 r \cos \theta \cos \f \tau 2\\
w' &= r \sin \theta \sqrt{1-c} = \sqrt 2 r \sin \theta \sin \f \tau 2\\
\dd w \dd w' &= \sqrt{1-c^2}r \dd r \dd \theta = (\sin^2 \tau) r \dd r \dd \theta,
\end{aligned}$$ and $$\begin{aligned}
\mathbf B &= \int_{w' \ge |w|} e^{-(\f{w^2}{1+c} + \f{(w')^2}{1 - c})/2}\phi(\sqrt{{q}/2}(w+w'))\phi(\sqrt{{q}/2}(w'-w)) \dd w \dd w'\\
&= -\sin^2 \tau \int_{0}^\infty r\dd r e^{-r^2/2} \int_{\pi/2 - \tau/2}^{\pi/2 + \tau/2}\dd \theta \phi(\sqrt {q}r \cos(\theta + \tau/2))\phi(\sqrt {q}r \cos(\theta - \tau/2))\\
&= -\sin^2 \tau \int_{0}^\infty r\dd r e^{-r^2/2} \int_{- \tau/2}^{\tau/2}\dd \theta \phi(\sqrt {q}r \sin(\theta + \tau/2))\phi(\sqrt {q}r \sin(\theta - \tau/2)).
\end{aligned}$$ Thus $$\begin{aligned}
\Gamma &= \f 1 {\pi \sqrt{1-c^2}} (\mathbf A - \mathbf B)\\
&= \f 1 \pi \int_{0}^\infty r\dd r e^{-r^2/2} \int_{- \tau/2}^{\pi - \tau/2}\dd \theta \phi(\sqrt {q}r \sin(\theta + \tau/2))\phi(\sqrt {q}r \sin(\theta - \tau/2))\\
&= \f 1 \pi \int_{0}^\infty r\dd r e^{-r^2/2} \int_{0}^{\pi}\dd \theta \phi(\sqrt {q}r \sin(\theta))\phi(\sqrt {q}r \sin(\theta - \tau)).
\end{aligned}$$ This gives the second equation in the lemma, and a change of variables $s = \sqrt {q}r$ gives the third.
For the fourth equality, we start from the third equality, and apply integration by parts: $$\begin{aligned}
&\phantom{{}={}}\f 1 \pi \int_0^\infty s \dd s \inv {q}e^{-s^2 \inv {q}/2} \int_0^\pi \dd \theta \Sigma(s, \theta; \tau)\\
&= \f 1 \pi \int_0^\pi \dd \theta \int_0^\infty \dd s s \inv {q}e^{-s^2 \inv {q}/2} \Sigma(s, \theta; \tau)\\
&= \f 1 \pi \int_0^\pi \dd \theta \biggl(-e^{-s^2 \inv {q}/2} \Sigma(s, \theta; \tau)\biggr\rvert_{s=0}^\infty + \int_0^\infty \dd s e^{-s^2 \inv {q}/2} \pdf{}{s} \Sigma(s, \theta; \tau) \biggr)\\
&= \f 1 \pi \int_0^\pi \dd \theta \int_0^\infty \dd s e^{-s^2 \inv {q}/2} \pdf{}{s} \Sigma(s, \theta; \tau).
\end{aligned}$$ where the last equality follows because $\Sigma(0, \theta; \tau) = 0$ and $e^{-s^2\inv {q}/ 2} \to 0$ as $s \to \infty$.
In the following lemmas, the “2” is not important, and can be any arbitrary finite or infinite value.
Suppose a function $f: (0, 2) \to \R$ is $C^k$ on $(0,2)$. If $\lim_{x \downarrow 0} \p f i (x)$ exists and is finite for every $i \in [0, k]$, then $f$ can be extended to $[0, 2)$ such that one sided $i$th derivatives exist at 0 for all $i \in [0, k]$.
Consider $\overline{\p f i}(0) := \p f i (1) - \int_0^1 \p f {i+1}(x) \dd x$ for $i \in [0, k-1]$, which naturally is also equal to $\p f i ({\epsilon}) - \int_0^{\epsilon}\p f {i+1}(x)\dd x$ for any ${\epsilon}> 0$. Certainly $\p f i (x) \to \overline{\p f i}(0)$ as $x \to 0$ if this limit exists — and by assumption it does, for $0 \le i \le k -1$. Therefore, we can define the extension of $\p f i$ to $x = 0$ to be $\p f i (0) := \overline{\p f i}(0)$. But we need to check that for $i \in [0, k-1]$. $$\begin{aligned}
\lim_{{\epsilon}\to 0} \f 1 {\epsilon}(\p f i({\epsilon}) - \p f i (0)) = \p f {i+1}(0)
\end{aligned}$$ so that all one sided $i$th derivatives exist. But $$\begin{aligned}
\f 1 {\epsilon}(\p f i({\epsilon}) - \p f i (0)) &= \f 1 {\epsilon}\int_0^{\epsilon}\p f {i+1}(x)\dd x\\
&=\p f {i+1}(0) + \int_0^1 (\p f {i+1}(x) - \p f i (0)) \I(x \in [0, {\epsilon}]) \dd x
\end{aligned}$$ Since $\lim_{x \downarrow 0} \p f {i+1}(x) = \p f {i+1} (0)$, $\p f {i+1}(x) - \p f {i+1}(0)$ is bounded for small $x$, and by dominated convergence, $\int_0^1 (\p f {i+1}(x) - \p f i (0)) \I(x \in [0, {\epsilon}]) \dd x \to \int_0^1 0 \dd x = 0$ as ${\epsilon}\to 0$. Thus $$\begin{aligned}
\lim_{{\epsilon}\to 0} \f 1 {\epsilon}(\p f i({\epsilon}) - \p f i (0)) &= \p f {i+1}(0)
\end{aligned}$$ as desired.
If $f: [0, 2) \to \R$ is $C^k$ on $(0, 2)$ and has one sided derivatives at 0 up to order $k$, then $$\begin{aligned}
f({\epsilon}) &= f(0) + {\epsilon}\p f 1 (0) + \cdots + \f{{\epsilon}^{i-1}}{(i-1)!} \p f {i-1} (0) + O({\epsilon}^i)
\end{aligned}$$ for any $i \le k$.
We have $$\begin{aligned}
f({\epsilon}) &= f(0) + \int_0^{\epsilon}\p f 1 (x) \dd x\\
&= f(0) + {\epsilon}\p f 1(0) + \int_0^{\epsilon}\p f 1(x) - \p f 1 (0) \dd x\\
&= f(0) + {\epsilon}\p f 1(0) + \int_0^{\epsilon}\int_0^{x_0} \p f 2 (x_2) \dd x_2 \dd x_1\\
&= f(0) + {\epsilon}\p f 1(0) + \f{{\epsilon}^2}{2} \p f 2(0) + \int_0^{\epsilon}\int_0^{x_1} \p f 2 (x_2) - \p f 2 (0) \dd x_2 \dd x_1\\
&\ \ \mathbin{\vdots}\ \\
f({\epsilon}) &= f(0) + {\epsilon}\p f 1 (0) + \cdots + \f{{\epsilon}^{i-1}}{(i-1)!} \p f {i-1} (0) + \int_0^{\epsilon}\dd x_1 \int_0^{x_1} \dd x_2 \cdots \int_0^{x_{i-1}} \dd x_{i} \p f {i}(x_{i})
\end{aligned}$$ for any $i\le k$. It suffices then to bound the size of the integral. Since $\p f i (x) \to \p f i (0)$ as $x \downarrow 0$ by assumption, $|\p f i (x_i)|$ is bounded by some constant $C$ on the integration region $\mathbb{A} := \{(x_1, \ldots, x_i): {\epsilon}\ge x_1 \ge \cdots \ge x_i\}$ for small enough ${\epsilon}$. Therefore, $$\begin{aligned}
&\phantom{{}={}}\int_0^{\epsilon}\dd x_1 \int_0^{x_1} \dd x_2 \cdots \int_0^{x_{i-1}} \dd x_{i} \p f {i}(x_{i})\\
&= \int \p f {i}(x_i) \I(\vec x \in \mathbb A) \dd \vec x\\
&\le C |\mathbb A|\\
&= \Theta({\epsilon}^i).
\end{aligned}$$
As a corollary,
\[lemma:smoothExtensionBoundary\] If $f: (0, 2) \to \R$ is smooth on $(0, 2)$ and $\lim_{x \to 0}\p f i (x)$ exists and is finite for all $i$, then $f$ can be extended to $[0, 2)$ and be [*one-sided smooth*]{} at 0, and $$\begin{aligned}
f({\epsilon}) &= f(0) + {\epsilon}\p f 1 (0) + \cdots + \f{{\epsilon}^{i-1}}{(i-1)!} \p f {i-1} (0) + O({\epsilon}^i)
\end{aligned}$$ for any $i$.
\[lemma:WtExtension\]
Let $\phi = \tanh$. For any fixed $c$, ${\mathrm{W}}\phi( {q}, cq)$ is smooth (infinitely differentiable) on ${q}\in (0, \infty)$. As a function of $Q := \inv {q}$, it can be extended smoothly to the point $Q = 0$, so that $$\begin{aligned}
{\mathrm{W}}\phi( {q}, cq ) &= \lim_{{q}' \to \infty} {\mathrm{W}}\phi( {q}', cq') + \inv {q}\lim_{{q}' \to \infty} \pd {\mathrm{W}}\phi( {q}', cq')/\pd ({q}')^{-1} + \cdots \\
&\qquad+ \f{{q}^{-i+1}}{(i-1)!} \lim_{{q}' \to \infty} \pd^{i-1} {\mathrm{W}}\phi( {q}', cq')/\pd ({q}')^{-i+1} + O({q}^{-i})
\end{aligned}$$ for any $i \ge 0$. Furthermore, for $c$ bounded away from $1$, the constants hidden $O$ can be taken independent of $c$.
**Smoothness on $(0, \infty)$.** By the third equation of \[lemma:WtPhiAntisymmetric\], for $Q \in (0, \infty) \iff {q}\in (0, \infty)$, $$\begin{aligned}
&\phantom{{}={}}\f 1 \pi \int_0^\infty s \dd s \left|\pdf{^n}{Q^n}\lp Q e^{-s^2 Q/2}\rp\right| \int_0^\pi \dd \theta |\phi(s \sin \theta)\phi(s \sin(\theta - \tau))|\\
&\le \int_0^\infty s \dd s \left|\pdf{^n}{Q^n}\lp Q e^{-s^2 Q/2}\rp\right| < \infty,
\end{aligned}$$ so by Leibniz’s integral rule and a simple induction, all derivatives of ${\mathrm{W}}\phi( {q}, cq)$ against $Q$ exists for any $Q \in (0, \infty)$.
**Extension to $Q = 0$.** By \[lemma:smoothExtensionBoundary\], it suffices to show that the limit of $
\pdf{^k{\mathrm{W}}\phi( {q}, cq)}{Q^k}$ exists and is finite as $Q \to 0$, for all $k$. Let $\tau = \arccos c$. By the fourth equation of \[lemma:WtPhiAntisymmetric\], we have explicitly $$\begin{aligned}
\pdf{^k{\mathrm{W}}\phi( {q}, cq)}{Q^k}
&= \f 1 \pi \int_0^\pi \dd \theta \int_0^\infty \dd s (\nicefrac{-s^2}{2})^k e^{-s^2 Q /2} \pdf{}{s} \Sigma(s, \theta; \tau)\\
&= \f {(-2)^{-k}} \pi \int_0^\pi \dd \theta \int_0^\infty \dd s\ s^{2k} e^{-s^2 Q /2} \pdf{}{s} \Sigma(s, \theta; \tau)
\end{aligned}$$ for any $Q \in (0, \infty)$. Note that for $\phi = \tanh$, $\dot \phi = {\operatorname{sech}}^2$, $$\begin{aligned}
\pdf{}{s} \Sigma(s, \theta; \tau)
&= \sin \theta \dot \phi(s \sin \theta)\phi(s \sin (\theta - \tau)) + \sin(\theta-\tau)\phi(s \sin \theta)\dot \phi(s \sin (\theta - \tau)).\\
\end{aligned}$$ We split the integral of $\pdf{^k {\mathrm{W}}\phi}{Q^k}$ as follows: $$\begin{aligned}
\pdf{^k{\mathrm{W}}\phi( {q}, cq)}{Q^k}
&= \f {(-2)^{-k}} \pi \int_0^\pi \dd \theta \int_0^\infty \dd s\ s^{2k} e^{-s^2 Q /2} \sin \theta \dot \phi(s \sin \theta)\phi(s \sin (\theta - \tau))\\
&\qquad + \f {(-2)^{-k}} \pi \int_0^\pi \dd \theta \int_0^\infty \dd s\ s^{2k} e^{-s^2 Q /2} \sin(\theta-\tau)\phi(s \sin \theta)\dot \phi(s \sin (\theta - \tau))
\end{aligned}$$ We show that for each piece, the limit as $Q \to 0$ exists and is finite, for any $k$. This will prove the smooth extendability of ${\mathrm{W}}\phi$ to $Q = 0$. We will do this for the first piece; the second is similar.
For $Q > 0$, the integrand is absolutely integrable, so we may switch the integrals. $$\begin{aligned}
&\phantom{{}={}} \int_0^\pi \dd \theta \int_0^\infty \dd s\ s^{2k} e^{-s^2 Q /2} \sin \theta \dot \phi(s \sin \theta)\phi(s \sin (\theta - \tau))\\
&= \int_0^\infty \dd s s^{2k} e^{-s^2 Q/2}\int_0^\pi \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))
\end{aligned}$$ We now try to bound the inner integral by an exponentially decreasing term $e^{-s \mu}$ for some $\mu$; clearly, by monotone convergence on the outer integral as $Q \to 0$, this would show the limit of the integral exists and is finite.
Because $\phi$ is odd and $\dot \phi$ is even, the inner integrand is negative on $\theta \in [0, \tau)$ and positive on $\theta \in (\tau, \pi]$. We will break up the inner integral as follows, for some fixed ${\epsilon}> 0$ satisfying $\tau - {\epsilon}> 0$ independent of $s$ (recall $\tau \in (0, \pi/2]$). $$\begin{aligned}
&\phantom{{}={}} \int_0^\pi \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \left(\int_0^{\epsilon}+ \int_{\pi - {\epsilon}}^\pi \right) \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau)) + \int_{\epsilon}^{\pi - {\epsilon}} \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))
\end{aligned}$$ Now because $\dot \phi(z) = {\operatorname{sech}}^2(z) \le 2 e^{-z}$, and $\sin \theta \ge \sin {\epsilon}$ on $\theta \in [{\epsilon}, \pi - {\epsilon}]$, $$\begin{aligned}
&\phantom{{}={}}\left|\int_{\epsilon}^{\pi - {\epsilon}} \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\right|\\
&\le 2\int_{\epsilon}^{\pi - {\epsilon}} \dd \theta \exp(-s \sin{\epsilon}) \\
&= 2(\pi - 2{\epsilon}) \exp(-s \sin{\epsilon}).
\end{aligned}$$
For the other part: $$\begin{aligned}
&\phantom{{}={}}\int_{\pi - {\epsilon}}^\pi \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \int_{\epsilon}^{0} \sin(\pi - \theta) \dot \phi(s \sin \pi - \theta) \phi(s \sin (\pi - \theta - \tau)) \dd(\pi - \theta)\\
&= \int_0^{\epsilon}\dd \theta \sin\theta \dot \phi(s \sin \theta) \phi(s \sin \theta + \tau)
\end{aligned}$$ so that $$\begin{aligned}
&\phantom{{}={}}\left(\int_0^{\epsilon}+ \int_{\pi - {\epsilon}}^\pi \right) \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \int_0^{\epsilon}\dd \theta \sin \theta \dot \phi(s \sin\theta) [\phi(s \sin (\tau + \theta)) - \phi(s \sin(\tau - \theta))]
\end{aligned}$$ But by intermediate value theorem, $\phi(s \sin (\tau + \theta)) - \phi(s \sin(\tau - \theta)) = 2 \theta \pd \phi(s \sin(\tau + \theta))/\pd \theta |_{\theta = \psi} = 2 \theta \dot \phi(s \sin (\tau + \psi))s \cos(\tau + \psi)$ for some $\psi \in [- \theta, \theta]$. By the assumption on ${\epsilon}$, $\phi(s \sin (\tau + \theta)) - \phi(s \sin(\tau - \theta)) \le 2{\epsilon}\dot \phi(s \sin(\tau - {\epsilon}))s \cos(\tau - {\epsilon}).$ Then $$\begin{aligned}
&\phantom{{}={}} \int_0^{\epsilon}\dd \theta \sin \theta \dot \phi(s \sin\theta) [\phi(s \sin (\tau + \theta)) - \phi(s \sin(\tau - \theta))]\\
&\le \int_0^{\epsilon}\dd \theta \sin \theta \dot \phi(s \sin\theta) 2 {\epsilon}\dot \phi(s \sin(\tau - {\epsilon}))s \cos(\tau - {\epsilon})\\
&\le 2 {\epsilon}\dot \phi(s \sin(\tau - {\epsilon}))s \cos(\tau - {\epsilon}) O(1)
\end{aligned}$$ Because $\tau -{\epsilon}> 0$ by assumption on ${\epsilon}$, and because $\dot \phi(z) = \exp(-\Theta_+(z))$, this quantity is $\exp(-\Theta_+(z))$, as desired (here $\Theta_+$ denotes a positive quantity).
Thus $$\begin{aligned}
&\phantom{{}={}} \int_0^\pi \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \left(\int_0^{\epsilon}+ \int_{\pi - {\epsilon}}^\pi \right) \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau)) + \int_{\epsilon}^{\pi - {\epsilon}} \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \exp(-\Theta_+(s))
\end{aligned}$$ and similarly for the other piece of $\pdf{^k {\mathrm{W}}\phi}{Q^k}$, so that $$\begin{aligned}
&\phantom{{}={}} \int_0^\infty \dd s s^{2k} e^{-s^2 Q/2}\int_0^\pi \dd \theta \sin \theta \dot \phi(s \sin\theta) \phi(s \sin (\theta - \tau))\\
&= \int_0^\infty \dd s s^{2k} e^{-s^2\f Q 2 - \Theta_+(z)}\\
&\to \int_0^\infty \dd s s^{2k} e^{-\Theta_+(z)}
\end{aligned}$$ is finite as $Q \to 0$, by monotone convergence.
**Independence of constant hidden in $O(({q}')^{-i})$.** The constant hidden is a function of the ${\epsilon}$ chosen above, which depend on $\tau$, but only to the extent that it must satisfy $\tau - {\epsilon}> 0$. As long as we are interested in a set $\mathcal C$ of $c$ that is bounded away from 1, the corresponding set of $\tau$ is bounded away from $0$, so ${\epsilon}$ can be taken to be some number smaller than all of the corresponding $\tau$.
\[lemma:Wt\_le\_arcsin\]
Suppose $\phi$ is tanh-like. Then for $c \in [0, 1]$, $${\mathrm{W}}\phi( {q}, cq) \le \f 2 \pi \arcsin(c),$$ and weakly increases to this upper bound as ${q}\to \infty$. Furthermore,
- If $c = 0$ or 1, then equality holds regardless of ${q}$.
- If $c \in (0, 1)$ is held constant, $\f 2 \pi \arcsin(c) - {\mathrm{W}}\phi( {q}, cq) = \Theta(\inv {q})$, where the hidden constants in $\Theta$ depend on $c$. But the constants can be made independent of $c$ if $c \in [{\epsilon}, 1-{\epsilon}]$ for some ${\epsilon}> 0$.
The cases of $c = 0$ or 1 are obvious by the definition of ${\mathrm{W}}$. So from here on we assume $c \in (0, 1)$.
Let $\tau := \arccos c$. By the first equation of \[lemma:WtPhiAntisymmetric\] and the assumption that $\phi$ is tanh-like, it is immediate that ${\mathrm{W}}\phi({q},cq)$ is nondecreasing in ${q}$. By dominated convergence, using the second equation of \[lemma:WtPhiAntisymmetric\], we get $$\begin{aligned}
\lim_{{q}\to \infty} {\mathrm{W}}\phi({q},cq) &= \f 1 \pi \int_{0}^\infty r\dd r e^{-r^2/2} (\pi - 2 \tau)\\
&= \f {\pi - 2 \tau}{\pi}\\
&= \f 2 \pi \arcsin c.
\end{aligned}$$
Then the convergence rate is $O(\inv {q})$ by \[lemma:WtExtension\] and Taylor’s theorem. Thus to show the convergence rate is $\Theta(\inv {q})$, it suffices to show that $\mathbf D := \pdf{{\mathrm{W}}\phi( {q}, cq)}{Q} < 0$. But this is apparent from the first equation of \[lemma:WtPhiAntisymmetric\]: For $\tau \in (0, \pi/2)$, $$\begin{aligned}
\mathbf D &= \f 1 {\pi \sin \tau} \int_{w' \ge |w|} \dd w \dd w' \Upsilon(w, w';\tau)(-\f 1 {2\sqrt 2} Q^{-3/2})\\
&\qquad\qquad\qquad\quad \times[\dot\phi(\sqrt{{q}/2}(w + w'))\phi(\sqrt{{q}/2}(w'-w))\\
&\qquad\qquad\qquad\qquad + \phi(\sqrt{{q}/2}(w + w'))\dot\phi(\sqrt{{q}/2}(w'-w))]\\
& < 0
\end{aligned}$$ since $\Upsilon$ is positive on the integration domain, and $\dot \phi$ and $\phi$ are both positive for positive arguments, by the assumption of $\phi$ being tanh-like.
**Independence of the constants in $\Theta(\inv {q})$ from $c$ when $c \in [{\epsilon}, 1-{\epsilon}]$.** By \[lemma:WtExtension\], the upper constant can be made independent from $c$. Since $\mathbf D$ is monotonically decreasing in $c$ (or monotonically increasing in $\tau$) and $|\mathbf D|$ is monotonically increasing in $c$ (or monotonically decreasing in $\tau$), we have $|\mathbf D| > |\mathbf D|\bigg]_{c = {\epsilon}}$, which can be taken to be the lower constant in $\Theta(\inv {q})$.
\[fig:verifywtanhasymptotics\] verifies empirically that the subleading term in ${\mathrm{W}}\tanh( q, cq)$ is linear in $\inv q$, for constant $c$.
![We verify empirically that the subleading term in ${\mathrm{W}}\tanh( q, cq)$ is linear in $\inv q$, for constant $c$. Indeed, observe that the curve of of ${\mathrm{W}}\tanh$ intersects the y-axis at an angle.[]{data-label="fig:verifywtanhasymptotics"}](graphics/verify_Wtanh_asymptotics){height=".15\textheight"}
We have by \[lemma:Wt\_le\_arcsin\], $$\begin{aligned}
{\boldsymbol{\oldgamma}}&= \f 2 \pi \arcsin({\boldsymbol{\oldlambda}}/{q}) - \Theta({q}^{-1}) + {\underline}{\boldsymbol{\oldgamma}}.\end{aligned}$$ Since ${q}= \sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2$ by \[thm:p\_q\_linear\], and ${\boldsymbol{\oldlambda}}= \sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2$ by \[thm:lambda\_gamma\_recurrence\], $$\begin{aligned}
{\boldsymbol{\oldgamma}}&= \f 2 \pi \arcsin\lp\f{\sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2}\rp - \Theta({q}^{-1}) + {\underline}{\boldsymbol{\oldgamma}}.\end{aligned}$$ We claim that $\p {\boldsymbol{\oldgamma}}l \to \infty$ as $l \to \infty$. Otherwise, there is some $C$ such that $\p {\boldsymbol{\oldgamma}}l \le C$ for all $l$. For large enough $l$, $\p {\mathbf{p}}l \ge \omega l$ for any $\omega < 1$ and $\arcsin \lp \f{C}{\sigma_w^2 \p {\mathbf{p}}{l-1} + \sigma_b^2} \rp = \Theta(1/l)$ by linearization of $\arcsin$. Thus $\p {\boldsymbol{\oldgamma}}l = \Theta(\log l)$, but this contradicts our assumption that ${\boldsymbol{\oldgamma}}$ is bounded. This proves our claim.
Therefore, for large enough $l$, $$\f{\sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2} = {\underline}{\boldsymbol{\oldgamma}}/{\underline}{\mathbf{p}}+ \Theta(l^{-1}).$$
\[fig:arcsin-vs-x\] shows $\f2\pi\arcsin x$ vs $x$. One sees that 1 is an unstable fixed point; if ${\mathbf{e}}< 1 - \epsilon$, then $\f 2 \pi \arcsin {\mathbf{e}}< 1 - \epsilon - \delta$ for some $\delta$. Thus ${c}$ drops monotonically until some threshold under which the linearization of $\arcsin$, $\arcsin x = x + \Theta(x^3)$, is applicable. So for large enough $l$, $$\begin{aligned}
{\boldsymbol{\oldgamma}}- {\underline}{\boldsymbol{\oldgamma}}&= \f 2 \pi \arcsin ({\underline}{\boldsymbol{\oldgamma}}/ {\underline}{\mathbf{p}}+ \Theta(\inv l)) - \Theta(\inv l)\\
&= \f 2 \pi {\underline}{\boldsymbol{\oldgamma}}/ {\underline}{\mathbf{p}}+ O(\inv l)\end{aligned}$$ As $\p {\mathbf{p}}l \sim l$ by \[thm:p\_q\_linear\], this difference equation has solution ${\boldsymbol{\oldgamma}}= \Omega(l^{\f 2 \pi -{\epsilon}}), O(l^{\f 2 \pi + {\epsilon}})$ for any ${\epsilon}$ by using the dynamics of \[lemma:alphaDeltaDynamics\] to upper and lower bound this difference equation.
### Backward Dynamics
The $\sigma_w = 0$ case is obvious. We will assume $\sigma_w > 0$ from here on.
Let $\p {\mathbf{p}}l = b_0 l + b_1 l^{1/2} + b_2 \log l + O(1)$. Then for $D = \f 2 3 \sqrt{\f 2 \pi}$, we have (implicitly applying \[lemma:Vt\_dot\_tanh\_asymptotics\] and \[lemma:power\_sum\_asymptotics\]), $$\begin{aligned}
{q}^{-1/2} &= \sigma_w^{-1} b_0^{-1/2} l^{-1/2}(1 - b_1 b_0^{-1} \inv 2 l^{-1/2} - b_2 b_0^{-1} 2^{-1} l^{-1} \log l + O(l^{-1}))\\
{\mathrm{V}}\dot \phi( {q}) & = D {q}^{-1/2} + \Theta({q}^{-3/2})\\
&= D\sigma_w^{-1} b_0^{-1/2} l^{-1/2}(1 - b_1 b_0^{-1} \inv 2 l^{-1/2} - b_2 b_0^{-1} 2^{-1} l^{-1} \log l + O(l^{-1}))\\
\log(B {\mathrm{V}}\dot \phi( {q}) + 1) &= B D \sigma_w^{-1} b_0^{-1/2} l^{-1/2}\\
&\phantom{={}}- (BD \sigma_w^{-1} b_0^{-3/2} b_1 2^{-1} +B^2 D^2 \sigma_w^{-2} b_0^{-1}2^{-1})l^{-1} + \Theta(l^{-3/2}\log l)\\
\sum_{r=1}^l \log(B {\mathrm{V}}\dot \phi( \p {q}r) + 1) &= 2B D \sigma_w^{-1} b_0^{-1/2} l^{1/2}
\\
&\phantom{={}} -(BD \sigma_w^{-1} b_0^{-3/2} b_1 2^{-1} +B^2 D^2 \sigma_w^{-2} b_0^{-1}2^{-1})\log l + O(1)\\
\end{aligned}$$ In our case, we have $b_0 = 1, b_1 = -2C \sigma_w^{-1}, b_2 = C^2 \sigma_w^{-2}, B = \sigma_w^2, C = \sqrt{\f 2 \pi}$, which gives $$\sum_{r=1}^l \log(B {\mathrm{V}}\dot \phi( \p {q}r) + 1) = \f 4 3 \sqrt{\f 2 \pi} \sigma_w l^{1/2} + (\f 4 {3\pi} - \sigma_w^2 \f 4 {9\pi}) \log l + O(1).$$ so that $$\begin{aligned}
\p {{\boldsymbol{\oldchi}}}{m}/\p {{\boldsymbol{\oldchi}}}l &= \exp\left[\f 4 3 \sqrt{\f 2 \pi} \sigma_w (\sqrt l - \sqrt m) + (\f 4 {3\pi} - \sigma_w^2 \f 4 {9\pi}) (\log l - \log m) + O(1)\right]\\
\end{aligned}$$
The $\sigma_w = 0$ case is obvious. We will assume $\sigma_w > 0$ from here on.
As in the proof of \[thm:dalethExpSqrtTanh\], $${\mathrm{V}}\dot \phi( {q}) = D\sigma_w^{-1} b_0^{-1/2} l^{-1/2} + \Theta(\inv l)$$ where $D = \f 2 3 \sqrt{\f 2 \pi}$. Thus by \[thm:dalethRecReduced\], $$\log(\p {{\boldsymbol{\oldchi}}}{m}/\p {{\boldsymbol{\oldchi}}}l) = \f 4 3 \sqrt{\f 2 \pi} \sigma_w (\sqrt l - \sqrt m) + (\f 4 {3\pi} - \sigma_w^2 \f 4 {9\pi}) (\log l - \log m) + O(1)$$ $$\log(\p {\boldsymbol{\oldchi}}m_b/\p {\boldsymbol{\oldchi}}l _b) =\f 4 3 \sqrt{\f 2 \pi} \sigma_w (\sqrt l - \sqrt m) + (\f 4 {3\pi} - \f 1 2 - \sigma_w^2 \f 4 {9\pi}) (\log l - \log m) + O(1)$$ Similarly, since ${\mathbf{p}}= l + \Theta(\sqrt l)$ by \[thm:p\_q\_linear\], we have $$\log(\p {\boldsymbol{\oldchi}}m_w/\p {\boldsymbol{\oldchi}}l _w) = \f 4 3 \sqrt{\f 2 \pi} \sigma_w (\sqrt l - \sqrt m) + (\f 4 {3\pi} + \f 1 2 - \sigma_w^2 \f 4 {9\pi}) (\log l - \log m) + O(1)$$
Tanh: Full Residual Network
---------------------------
### Forward Dynamics
The $\sigma_w = 0$ case is obvious. We will assume $\sigma_w > 0$ from here on.
As in \[thm:p\_q\_linear\], ${\mathbf{p}}$ will have expansion ${\mathbf{p}}= b_0 l + b_1 l^{1/2} + b_2 \log l + O(1).$ Then, for $C = \sqrt{\f 2 \pi}$, $$\begin{aligned}
{q}^{-1/2} &= \sigma_w^{-1} b_0^{-1/2} l^{-1/2}(1 - b_1 b_0^{-1} \inv 2 l^{-1/2} - b_2 b_0^{-1} 2^{-1} l^{-1} \log l + O(l^{-1}))\\
\sum_{r=1}^l {\mathrm{V}}\phi( \p {q}r) &= \sum_{r=1}^l 1 - C(\p {q}r)^{-1/2} + \Theta((\p {q}r)^{-3/2})\\
&= l - 2 C \sigma_w^{-1} b_0^{-1/2} l^{1/2} + C \sigma_w^{-1} b_1 b_0^{-3/2} 2^{-1} \log l + O(1)\\
\p {\mathbf{p}}l &= \sigma_v^2 \sum_{r=1}^l + \sigma_a^2 l\\
&= (\sigma_v^2 + \sigma_a^2) l - 2C \sigma_v^2 \sigma_w^{-1} b_0^{-1/2} l^{1/2} + C \sigma_v^2 \sigma_w^{-1} b_1 b_0^{-3/2} 2^{-1} \log l + O(1)
\end{aligned}$$ which yields $$\begin{aligned}
b_0 &= \sigma_v^2 + \sigma_a^2\\
b_1 &= -2C \sigma_v^2 \sigma_w^{-1} b_0^{-1/2} = \f{-2C \sigma_v^2 \sigma_w^{-1}}{\sqrt{\sigma_v^2 + \sigma_a^2}}\\
b_2 &= \f{-C^2 \sigma_v^4 \sigma_w^{-2}}{(\sigma_v^2 + \sigma_a^2)^2}
\end{aligned}$$
Suppose $\phi$ is tanh-like. Then $${\boldsymbol{\oldgamma}}\le \sigma_v^2 \f 2 \pi \arcsin\lp { {\boldsymbol{\oldlambda}}}/{{q}}\rp + \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}},$$ and $$\sigma_v^2 \f 2 \pi \arcsin\lp { {\boldsymbol{\oldlambda}}}/{{q}}\rp + \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}}- {\boldsymbol{\oldgamma}}= \Theta({q}^{-1}).$$
Similar to the proof of \[lemma:Wt\_le\_arcsin\].
\[lemma:timeDependentConvergence\] Let $u^* \in [0, 1)$. Let $f_t: [0, 1) \to [0, 1]$ be a continuous function for each $t \in \N$, to each of which we associate two numbers $0 \le a_t \le u^* \le b_t \le 1$. Suppose for each $t$, $f_t(u) > u$ for all $u \in [0, a_t)$ and $f_t(u) < u$ for all $u \in (b_t, 1)$. Assume that for each $u$, $f_t(u) - u \to 0$ as $t \to \infty$ uniformly over $u$. If $a_t \nearrow u^*$ and $b_t \searrow u^*$, then for any $u_0 \in [0, 1)$, the dynamics $u_t = f_t(u_{t-1})$ has a limiting point. Furthermore, either $u_t \to u^*$ or $u_t$ eventually converges monotonically (decreasing or increasing) to a limit point.
Fix a $u_0 \in [0, 1)$. If $u_t \to u^*$ then we are done. Otherwise, suppose there is a neighborhood $[u^* - {\epsilon}, u^* + {\epsilon}]$ such that for an infinite sequence $t_1, t_2, \ldots$, $u_{t_i} \not \in [u^* - {\epsilon}, u^* + {\epsilon}]$. WLOG assume $u_{t_i} < u^* - {\epsilon}$ for all $i$ and $(t_i)_i$ is the sequence of all $t$s that satisfy this inequality.
If $(t_i)_i$ contains $\{s: s \ge N\}$ for some $N$, then for some $M > N$, for every $t > M$, $a_t > u^* - {\epsilon}> u_t$. By assumption, $u_t$ is monotonic for all $t > M$ but is bounded above. Thus $u_t$ has a fixed point $\hat u \le u^* - {\epsilon}$ as desired.
Now assume there are infinite $i$s such that $t_i - 1 \not= t_{i-1}$ (i.e. $t_i - 1$ is not part of the sequence $(t_i)_i$). We will show that this case is contradictory. Take $T$ large enough such that $a_t > u^* - {\epsilon}/2$ and $|f_t(u) - u| < {\epsilon}/4$ for all $u$ and for all $t \ge T$ ($T$ exists by premise). Let $j$ be the smallest index such that $t_j > T$ and $t_j - 1 \not = t_{j-1}$. By the definition of $j$, $u_{t_j - 1} \ge u^* - {\epsilon}$. If $u_{t_j - 1} \ge u^* - {\epsilon}/2$, then by definition of $T$, $u^* - {\epsilon}> u_{t_j} = f_{t_j}(u_{t_j - 1}) > u_{t_j-1} - {\epsilon}/4 > u^* - 3{\epsilon}/4 > u^* - {\epsilon}$, a contradiction. If $u^* - {\epsilon}\le u_{t_j - 1} \le u^* - {\epsilon}/2$, then by the definition of $T$, $u_{t_j-1} \le a_{t_j-1}$ so that $u_{t_j } = f_{t_j}(u_{t_j - 1}) > u_{t_j - 1} \ge u^* - {\epsilon},$ a contradiction.
The “furthermore” claim is clear from our proof above.
The $\sigma_w = 0$ case is obvious. We will assume $\sigma_w > 0$ from here on.
If $\sigma_a = 0$, then ${\mathbf{e}}^*$ as defined above is 0, and ${\mathbf{e}}= \f {\boldsymbol{\oldgamma}}{\mathbf{p}}$ decreases as $\Theta(l^{\f 2 \pi - 1})$ to 0, by the same reason as before.
So from now on suppose $\sigma_a > 0$. We apply \[lemma:timeDependentConvergence\] first to show that ${\mathbf{e}}$ converges. We have $$\begin{aligned}
\sigma_v^2 {\mathrm{W}}\phi( {q}, cq) + \sigma_a^2 &= {\mathbf{e}}{\mathbf{p}}- {\underline}{\mathbf{e}}{\underline}{\mathbf{p}}\\
&= {\mathbf{e}}{\mathbf{p}}- {\underline}{\mathbf{e}}{\mathbf{p}}+ {\underline}{\mathbf{e}}{\mathbf{p}}- {\underline}{\mathbf{e}}{\underline}{\mathbf{p}}\\
&= ({\mathbf{e}}- {\underline}{\mathbf{e}}) {\mathbf{p}}+ {\underline}{\mathbf{e}}({\mathbf{p}}- {\underline}{\mathbf{p}})\\
&= ({\mathbf{p}}- {\underline}{\mathbf{p}})[({\mathbf{e}}- {\underline}{\mathbf{e}})\f {\mathbf{p}}{{\mathbf{p}}- {\underline}{\mathbf{p}}} + {\underline}{\mathbf{e}}]\\
\f{\sigma_v^2 {\mathrm{W}}\phi( {q}, cq) + \sigma_a^2}{\sigma_v^2 {\mathrm{V}}\phi( {q}) + \sigma_a^2} &= ({\mathbf{e}}- {\underline}{\mathbf{e}}) \f {\mathbf{p}}{{\mathbf{p}}- {\underline}{\mathbf{p}}} + {\underline}{\mathbf{e}}\\
\f {{\mathbf{p}}- {\underline}{\mathbf{p}}}{\mathbf{p}}\left[\f{\sigma_v^2 {\mathrm{W}}\phi + \sigma_a^2}{\sigma_v^2 {\mathrm{V}}\phi + \sigma_a^2} - {\underline}{\mathbf{e}}\right] &= {\mathbf{e}}- {\underline}{\mathbf{e}}\end{aligned}$$ If we define $f_l(u) := \f{\p {\mathbf{p}}l - \p {\mathbf{p}}{l-1}}{\p {\mathbf{p}}l}\left[\f{\sigma_v^2 {\mathrm{W}}\phi( \p {q}l, \p {c}l \p {q}l) + \sigma_a^2}{\sigma_v^2 {\mathrm{V}}\phi( \p {q}l) + \sigma_a^2} - u\right] + u$ (the LHS of the above), then $f_l(u) - u = O(l^{-1})$ uniformly for all $u$ because $\p {\mathbf{p}}l = \Theta(l)$, $\p {\mathbf{p}}l - \p {\mathbf{p}}{l-1} = \Theta(1)$, and the part in the bracket is $O(1)$, with constants all (able to be taken) independent of $u$. We divide $[0, 1)$ into the following intervals $I_1 = [1, 1/2), I_2 = [1/2, 3/4), I_3 = [3/4, 7/8), \ldots$. For each $I_k$, it is clear that the trajectories of $\p {\mathbf{e}}l = f_l(\p {\mathbf{e}}{l-1})$ with $\p {\mathbf{e}}0 \in I_k$ will fall into some interval $J_k$ bounded away from 1 for all $l \ge L$, for large enough $L$ (dependent on $k$). Then we can apply \[lemma:cExpansion,lemma:vtanhSqrtConvergence,lemma:Wt\_le\_arcsin\] to get $f_l(u) = \f{\p {\mathbf{p}}l - \p {\mathbf{p}}{l-1}}{\p {\mathbf{p}}l}\left[\f{\sigma_v^2\f 2 \pi \arcsin(u) + \sigma_a^2}{\sigma_v^2 + \sigma_a^2} - u + o(1)\right] + u$ where the constants in $o(1)$ is uniform for all $\p {\mathbf{e}}0 \in I_k$. For $u < {\mathbf{e}}^*$ (as defined in the theorem statement), $\f{\sigma_v^2\f 2 \pi \arcsin(u) + \sigma_a^2}{\sigma_v^2 + \sigma_a^2} > u$ and for $u > {\mathbf{e}}^*$, $\f{\sigma_v^2\f 2 \pi \arcsin(u) + \sigma_a^2}{\sigma_v^2 + \sigma_a^2} < u$ (see \[fig:arcsin-vs-x\]). Thus as $l \to \infty$, the $o(1)$ term gets smaller and smaller, and this monotonicity holds for $f_l(u) - u = \left[\f{\sigma_v^2\f 2 \pi \arcsin(u) + \sigma_a^2}{\sigma_v^2 + \sigma_a^2} - u + o(1)\right] > 0$ (resp. $<0$) on larger and larger intervals $[0, a_l] \cap J_k$ (resp. $[b_l, 1) \cap J_k$). This proves all the preconditions for \[lemma:timeDependentConvergence\], which yields that $I_k$ converges to a limit point. As this argument is independent of $k$, we have that for all $\p {\mathbf{e}}0 \in [0, 1)$, $\p {\mathbf{e}}l$ converges.
Now we solve for the limit point.
Suppose ${\mathbf{e}}$ has limit point ${\mathbf{e}}^\dagger$ (possibly different from ${\mathbf{e}}^*$ described in the theorem); if we express $\p {\boldsymbol{\oldgamma}}l = ({\mathbf{e}}^\dagger + \p {\epsilon}l ) \p {\mathbf{p}}l$, then $$\begin{aligned}
\sigma_v^2 {\mathrm{W}}\phi( {q}, cq) + \sigma_a^2 &= {\boldsymbol{\oldgamma}}- {\underline}{\boldsymbol{\oldgamma}}\\
&= ({\mathbf{e}}^\dagger + {\epsilon}){\mathbf{p}}- ({\mathbf{e}}^\dagger + {\underline}{\epsilon}) {\underline}{\mathbf{p}}\\
&= {\mathbf{e}}^\dagger({\mathbf{p}}- {\underline}{\mathbf{p}}) + {\epsilon}{\mathbf{p}}- {\underline}{\epsilon}{\underline}{\mathbf{p}}\\
\f{\sigma_v^2 {\mathrm{W}}\phi( {q}, cq) + \sigma_a^2}{\sigma_v^2 {\mathrm{V}}\phi( {q}) + \sigma_a^2} &= {\mathbf{e}}^\dagger + {\epsilon}+ ({\epsilon}- {\underline}{\epsilon})\f{{\underline}{\mathbf{p}}}{{\mathbf{p}}- {\underline}{\mathbf{p}}}\end{aligned}$$ As $l \to \infty$, ${c}\sim {\mathbf{e}}\to {\mathbf{e}}^\dagger$, and ${\mathrm{W}}\phi( {q}, {\mathbf{e}}^\dagger {q}) \to \f 2 \pi \arcsin({\mathbf{e}}^\dagger)$, and ${\mathrm{V}}\phi( {q}) \to 1$. Additionally, ${\underline}{\mathbf{p}}/({\mathbf{p}}- {\underline}{\mathbf{p}}) = \Theta(l)$ and ${\epsilon}= o(1)$ so that ${\epsilon}- {\underline}{\epsilon}= o(l^{-1})$. Then we have, taking limits $l \to \infty$, $$\begin{aligned}
\f{\sigma_v^2 \f 2 \pi \arcsin({\mathbf{e}}^\dagger)+\sigma_a^2}{\sigma_v^2 + \sigma_a^2} &= {\mathbf{e}}^\dagger.\end{aligned}$$ Since $f_l$ (as defined above) repels points away from 1, the only solution for ${\mathbf{e}}^\dagger$ when $\p {\mathbf{e}}0 < 1$ is ${\mathbf{e}}^\dagger = {\mathbf{e}}^*$ as specified in the theorem statement.
We defer the proof of the convergence rate to ${\mathbf{e}}^*$ to \[thm:TanhFullResConvergenceRate\].
![Graph of $y({\mathbf{e}}) = \f 1 {\sigma_v^2 + \sigma_a^2}[\sigma_v^2\f 2 \pi \arcsin({\mathbf{e}}) + \sigma_a^2]$ for various $\sigma_v$ and $\sigma_a$.[]{data-label="fig:arcsin-vs-x"}](graphics/arcsin_vs_x.pdf){height=".15\textheight"}
\[lemma:fixed\_point\_ineq\] Let ${\mathbf{e}}^*$ be the stable fixed point determined by $\sigma_a$ and $\sigma_v$. Then as long as $\sigma_v > 0$, $$\begin{aligned}
\f 2 \pi \f 1 {\sqrt{1 - ({\mathbf{e}}^*)^2}} \f{\sigma_v^2 }{\sigma_v^2 + \sigma_a^2} \in (\f 1 2, \f 2 \pi]\end{aligned}$$
Write $\rho := \f{\sigma_a^2}{\sigma_v^2}$. By definition of ${{{\mathbf{e}}^*}}$, we get $$\begin{aligned}
{{{\mathbf{e}}^*}}&= (1 - \rho) \f 2 \pi \arcsin {{{\mathbf{e}}^*}}+ \rho\\
\rho = &= \f{{{{\mathbf{e}}^*}}- \f 2 \pi \arcsin {{{\mathbf{e}}^*}}}{1 - {{{\mathbf{e}}^*}}}
\end{aligned}$$ Substituting $\rho$ into the expression in question, it follows that we want to show $$\begin{aligned}
\f 2 \pi (1 - {{{\mathbf{e}}^*}}^2)^{-1/2} (1 + \rho)^{-1}= \f 2 \pi (1 - {{{\mathbf{e}}^*}}^2)^{-1/2} \left(\f{1 - \f 2 \pi \arcsin {{{\mathbf{e}}^*}}}{1 - {{{\mathbf{e}}^*}}}\right)^{-1}\in (\f 1 2, \f 2 \pi]
\end{aligned}$$ for ${{{\mathbf{e}}^*}}\in [0, 1)$ (the endpoint at 1 is not included since $\sigma_v > 0$. But this is $$\begin{aligned}
&\phantom{={}} \f 2 \pi (1 - {{{\mathbf{e}}^*}})^{1/2} (1 + {{{\mathbf{e}}^*}})^{-1/2} (1 - \f 2 \pi \arcsin {{{\mathbf{e}}^*}})^{-1}.
\end{aligned}$$ Set $g({{{\mathbf{e}}^*}})$ to be this expression. We could proceed by finding critical points, but a simple plot \[fig:deltastarBound\] shows that $g$ is decreasing on $[0, 1)$, with extremal values at the end points: $$g({{{\mathbf{e}}^*}}) \in [\lim_{{{{\mathbf{e}}^*}}\to 1} g({{{\mathbf{e}}^*}}), g(0)), \quad \text{for }{{{\mathbf{e}}^*}}\in [0, 1).$$ Obviously $g(0) = \f 2 \pi$. For the limit, we note that $\arcsin {{{\mathbf{e}}^*}}$ has an asymptotic expansion $\f \pi 2 - \sqrt 2 (1 - e)^{1/2} + \Theta((1-e)^{3/2})$ at 1, so that $(1 - {{{\mathbf{e}}^*}})^{1/2}(1 - \f 2 \pi \arcsin {{{\mathbf{e}}^*}})^{-1} \to \dfrac{\pi}{2 \sqrt 2}$, and $g({{{\mathbf{e}}^*}}) \to \dfrac{1}{2}$ as ${{{\mathbf{e}}^*}}\to 1$.
![Plot of $g({{{\mathbf{e}}^*}})$ in the proof of \[lemma:fixed\_point\_ineq\][]{data-label="fig:deltastarBound"}](graphics/deltastarBound.pdf){width=".4\textwidth"}
\[thm:TanhFullResConvergenceRate\] If $\p {\mathbf{e}}0 < 1$, then $|\p {\mathbf{e}}l - {\mathbf{e}}^*|$ is $\Omega(l^{-\delta^* - \varepsilon})$ and $O(l^{-\delta^* + \varepsilon})$ for any $\varepsilon > 0$, where $$\delta^* := 1 - \f 2 \pi \f 1 {\sqrt{1 - ({\mathbf{e}}^*)^2}} \f{\sigma_v^2 }{\sigma_v^2 + \sigma_a^2} \in [1-\f 2 \pi, \f 1 2),$$ where the bounds on the right follow from \[lemma:fixed\_point\_ineq\].
Define $\omega(q, c) = \f 2 \pi \arcsin(c) - {\mathrm{W}}\tanh( q, c q)$. By \[lemma:Wt\_le\_arcsin\], for large enough $l$, ${c}$ is close to ${\mathbf{e}}^*$ bounded away from 0 or 1, so that $\omega({q}, {c}) = \Theta(\inv {q})$ with the constant hidden in $\Theta$ independent of ${c}$. Additionally, by \[lemma:vtanhSqrtConvergence\], $1 - {\mathrm{V}}\tanh( {q}) = \Theta({q}^{-1/2})$. Therefore, $$\begin{aligned}
({\mathbf{e}}^* + \epsilon){\mathbf{p}}&= \sigma_v^2 (\f 2 \pi \arcsin({\mathbf{e}}^* + {\underline}\epsilon) - \omega({q}, {c}))+ \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}}\\
&= \sigma_v^2 \f 2 \pi [\arcsin({\mathbf{e}}^*) + \f{{\underline}\epsilon}{\sqrt{1 - ({\mathbf{e}}^*)^2}} + \Theta({\underline}{\epsilon}^2)] - \Theta(\inv l) + \sigma_a^2 + {\underline}{\boldsymbol{\oldgamma}}\\
&= {\mathbf{e}}^*(\sigma_v^2 + \sigma_a^2) + ({\mathbf{e}}^* + {\underline}{\epsilon}) {\underline}{\mathbf{p}}+ \sigma_v^2 \f 2 \pi\f{{\underline}\epsilon}{\sqrt{1 - ({\mathbf{e}}^*)^2}} + \Theta({\underline}{\epsilon}^2) - \Theta(\inv l)\\
{\mathbf{e}}^*({\mathbf{p}}- {\underline}{\mathbf{p}}- \sigma_v^2 - \sigma_a^2) &= {\underline}{\epsilon}{\underline}{\mathbf{p}}- {\epsilon}{\mathbf{p}}+ \sigma_v^2 \f 2 \pi\f{{\underline}\epsilon}{\sqrt{1 - ({\mathbf{e}}^*)^2}} + \Theta({\underline}{\epsilon}^2)- \Theta(\inv l)\\
{\mathbf{e}}^*\sigma_v^2({\mathrm{V}}\phi( {q}) - 1) &= {\underline}{\epsilon}{\underline}{\mathbf{p}}- {\epsilon}{\mathbf{p}}+ \sigma_v^2 \f 2 \pi\f{{\underline}\epsilon}{\sqrt{1 - ({\mathbf{e}}^*)^2}} + \Theta({\underline}{\epsilon}^2) - \Theta(\inv l)\\
{\epsilon}&= \f 1 {\mathbf{p}}({\mathbf{e}}^*\sigma_v^2(1 - {\mathrm{V}}\phi( {q})) + \Theta({\underline}{\epsilon}^2) - \Theta(\inv l) + {\underline}{\epsilon}({\underline}{\mathbf{p}}+ \sigma_v^2 \f 2 \pi\f{1}{\sqrt{1 - ({\mathbf{e}}^*)^2}}))\\
&= \Theta(l^{-3/2}) + {\underline}{\epsilon}(1 - \p \delta l / l)\\
$$ where $$\begin{aligned}
\p \delta l &= \f l {\mathbf{p}}(\sigma_v^2 {\mathrm{V}}\phi( {q}) + \sigma_a^2 -\sigma_v^2 \f 2 \pi\f{1}{\sqrt{1 - ({\mathbf{e}}^*)^2}}) + \Theta({\underline}{\epsilon}/l)\\
&= (1 + \Theta(l^{-1/2}))(\sigma_v^2 (1 - \Theta(l^{-1/2})) + \sigma_a^2 -\sigma_v^2 \f 2 \pi\f{1}{\sqrt{1 - ({\mathbf{e}}^*)^2}})/(\sigma_v^2 + \sigma_a^2) + \Theta({\underline}{\epsilon}/ l)\\
&= \delta^* + O(l^{-1/2}),\end{aligned}$$ where $\delta^* := 1 - \f 2 \pi \f 1 {\sqrt{1 - ({\mathbf{e}}^*)^2}} \f{\sigma_v^2 }{\sigma_v^2 + \sigma_a^2}$, which is positive by \[lemma:fixed\_point\_ineq\]. By taking the $\delta$ of \[lemma:alphaDeltaDynamics\] to be $\delta^* + \varepsilon$ or $\delta^* - \varepsilon$ respectively for lower and upper bounding the dynamics of $\p {\epsilon}l$, the solution $\p {\epsilon}l$ is $\Omega(l^{-\delta^* - \varepsilon})$ and $O(l^{-\delta^* + \varepsilon})$ for any $\varepsilon > 0$ since $\f 1 2 > \delta^*$.
### Backward Dynamics
The $\sigma_w = 0$ case is obvious. We will assume $\sigma_w > 0$ from here on.
As in the proof of \[thm:dalethExpSqrtTanh\], $$\begin{aligned}
\log(\p {{\boldsymbol{\oldchi}}}{m} / \p {{\boldsymbol{\oldchi}}}l) &= 2B D \sigma_w^{-1} b_0^{-1/2} (\sqrt l - \sqrt m)
\\
&\phantom{={}} -(BD \sigma_w^{-1} b_0^{-3/2} b_1 2^{-1} +B^2 D^2 \sigma_w^{-2} b_0^{-1}2^{-1})(\log l - \log m) + O(1)\end{aligned}$$ where $B = \sigma_v^2 \sigma_w^2, D = \f 2 3 \sqrt{\f 2 \pi},$ $$\begin{aligned}
b_0 &= \sigma_v^2 + \sigma_a^2\\
b_1 &= \f{-2C \sigma_v^2 \sigma_w^{-1}}{\sqrt{\sigma_v^2 + \sigma_a^2}}\\
b_2 &= \f{-C^2 \sigma_v^4 \sigma_w^{-2}}{(\sigma_v^2 + \sigma_a^2)^2}.\end{aligned}$$ with $C = \sqrt{\f 2 \pi}$. This simplifies to the desired form.
Similar to \[thm:dalethExpSqrtTanhAllGrad\].
$\alpha$-ReLU: Full Residual Network
------------------------------------
The following can be checked readily
Since $\dot \psi_\alpha = \alpha \psi_{\alpha - 1}$, we have as a corollary,
\[lemma:Vt\_dotpsi\_alpha\] If $\alpha > \f 1 2$, then ${\mathrm{V}}\dot \psi_\alpha( q) = \alpha^2 {\mathsf{c}}_{\alpha - 1} q^{\alpha-1}$.
As a special case, when $\alpha = 1$, ${\mathsf{c}}_\alpha = \f 1 2$.
The following is a trivial computation, but useful for many simplifications.
${\mathsf{c}}_{\alpha+1}/{\mathsf{c}}_\alpha = 2\alpha+1$.
### Forward Dynamics {#sec:AlphaReluForwardProofs}
\[thm:pDynamic1ReLU\] Suppose we have the nonlinearity $\phi = \psi_1$. Then $\p {\mathbf{p}}l = \Theta((1 + \sigma_v^2 \sigma_w^2/2)^l)$, with the hidden constant depending on the initial condition.
We have $$\begin{aligned}
{\mathbf{p}}&= \f 1 2\sigma_v^2(\sigma_w^2 {\underline}{\mathbf{p}}+ \sigma_b^2) + \sigma_a^2 + {\underline}{\mathbf{p}}\\
&= (\f 1 2\sigma_v^2\sigma_w^2 + 1) {\underline}{\mathbf{p}}+ \f 1 2 (\sigma_v^2 \sigma_b^2 + \sigma_a^2).\end{aligned}$$ By the standard method of characteristic equation, we get that $$\p {\mathbf{p}}l = A + C B^l$$ where $A = - \f{\sigma_a^2 + \sigma_b^2 \sigma_v^2}{\sigma_v^2 \sigma_w^2}$, $B = 1 + \f{\sigma_v^2 \sigma_w^2}{2}$, and $C$ is a coefficient determined by initial conditions.
\[thm:pDynamicLT1ReLU\] Suppose $\alpha < 1$. We have the following asymptotic expansion $$\p {\mathbf{p}}l = K_1 l^{{\f 1 {1-\alpha}}}+ R(l)$$ where the remainder term $$R(l) \sim \begin{cases}
-K_2 l^{{{\f \alpha {1-\alpha}}}} \log l &\text{if $\alpha > \f 1 2$}\\
(C-K_2) l \log l &\text{if $\alpha = \f 1 2$ and $K_2 \not=C$}\\
\f{C(1-\alpha)}{1-2\alpha} l & \text{if $ \alpha < \f 1 2$}
\end{cases}$$ where $K_1 = [\sigma_v^2 \sigma_w^{2\alpha} {\mathsf{c}}_\alpha (1 - \alpha)]^{\f 1 {1-\alpha}}, K_2 = \f 1 2 [\sigma_v^2 {\mathsf{c}}_\alpha \sigma_w^{2\alpha}]^{{{\f 1 {1-\alpha}}}} (1-\alpha)^{{{\f \alpha {1-\alpha}}}- 1} \alpha$ and $C = \sigma_a^2$.
\[fig:55reluverifyleadingcoeffp\] verifies the leading coefficient and the exponent of the leading term.
The difference equation governing the evolution of ${\mathbf{p}}$ is $${\mathbf{p}}- {\underline}{\mathbf{p}}= A({\underline}{\mathbf{p}}+ B)^\alpha + C$$ where $A =\sigma_v^2 {\mathsf{c}}_\alpha \sigma_w^{2\alpha}$, $B = \sigma_b^2/\sigma_w^2$, and $C = \sigma_a^2$. Then \[lemma:polyDynamicsConstant\] yields the result.
![Verification of leading term of \[thm:pDynamic1ReLU\] for $\alpha = 0.55$.[]{data-label="fig:55reluverifyleadingcoeffp"}](graphics/55relu_verify_leading_coeff_p){height=".2\textheight"}
\[thm:pDynamicLT1ReLU\] combined with \[thm:pDynamic1ReLU\] gives the following result.
By [@cho_kernel_2009], we know that ${\mathrm{W}}\psi_\alpha( q, q c) = {\mathrm{V}}\psi_\alpha( q) {\mathbb{J}}_\alpha(c)$, where ${\mathbb{J}}_\alpha(c) = J_\alpha(\arccos c)$ and $$J_\alpha(\theta) := \f 1 {2\pi {\mathsf{c}}_\alpha} (\sin \theta)^{2\alpha + 1} \Gamma(\alpha + 1)\int_0^{\pi / 2} \f{\dd \eta \cos^\alpha \eta}{(1 - \cos \theta \cos \eta)^{1 + \alpha}}.\tag{$\triangle$}\label{eqn:JalphaIntegralFormula}$$
Note that ${\mathbb{J}}_\alpha(c) \in (-\infty, \infty)$ for $\alpha \in (-1, \infty)$ and any $c \in (0, 1)$, even though ${\mathrm{V}}\psi_\alpha$ is only defined for $\alpha > -1/2$.
![(a) ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. (b) shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. (c) Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$.[]{data-label="fig:jjj_vs_id"}](graphics/jjjVsId2.pdf "fig:"){width=".3\textwidth"} ![(a) ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. (b) shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. (c) Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$.[]{data-label="fig:jjj_vs_id"}](graphics/jjj_vs_id_log.pdf "fig:"){width=".3\textwidth"} ![(a) ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. (b) shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. (c) Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$.[]{data-label="fig:jjj_vs_id"}](graphics/jjj_vs_id_close.pdf "fig:"){width=".3\textwidth"}
\[fig:jjj\_vs\_id\] shows a comparison of ${\mathbb{J}}_\alpha$ for different $\alpha$s along with the identity function. By [@daniely_toward_2016 Lemma 11], ${\mathbb{J}}_\alpha$ is an increasing and convex function as long as $\psi_\alpha^2$ is Gaussian-integrable, which is precisely when $\alpha > -1/2$. We can compute ${\mathbb{J}}_\alpha(1) = {\mathrm{W}}\psi_\alpha(q, q)/{\mathrm{V}}\psi_\alpha(q) = 1$, and ${\mathbb{J}}_\alpha(0) = {\mathrm{W}}\psi_\alpha(q, 0)/{\mathrm{V}}\psi_\alpha(q) = {\mathrm{V}}\psi_{\alpha/2}(q)^2/{\mathrm{V}}\psi_\alpha(q) = {\mathsf{c}}_{\alpha/2}^2/{\mathsf{c}}_\alpha = \f 1{2\sqrt \pi} \f{\Gamma(\f \alpha 2 + \f 1 2)^2}{\Gamma(\alpha + \f 1 2)}$. We record these observations as a lemma.
\[lemma:basicJalpha\] ${\mathbb{J}}_\alpha(c)$ is an increasing and convex function for each $\alpha > -1/2$ on $c \in [0, 1]$. ${\mathbb{J}}_\alpha(1) = 1$ and ${\mathbb{J}}_\alpha(0) = \f 1{2\sqrt \pi} \f{\Gamma(\f \alpha 2 + \f 1 2)^2}{\Gamma(\alpha + \f 1 2)}$.
![${\mathbb{J}}_1$ vs identity[]{data-label="fig:jjj1_vs_id"}](graphics/jjj1_vs_id.pdf){width=".3\textwidth"}
For $\alpha = 1$, @cho_kernel_2009 computed $${\mathbb{J}}_1(c) = \f 1 \pi (\sqrt{1 - c^2} + (\pi - \arccos(c))c).$$ \[fig:jjj1\_vs\_id\] shows a plot of ${\mathbb{J}}_1$ vs identity. It has derivative $\dot{\mathbb{J}}_1(c) = 1 - \f 1 \pi \arccos c$, which shows that $\dot{\mathbb{J}}_1(c) < 1$ with equality iff $c = 1$, and consequently ${\mathbb{J}}_1(c) \ge c$ with equality iff $c = 1$. At the same time, $\dot {\mathbb{J}}_1(c) \ge 0$ with equality iff $c = -1$, so ${\mathbb{J}}_1$ is increasing on $[-1, 1]$. It has an asymptotic expansion ${\mathbb{J}}_1(1 - \varepsilon) = 1 - \varepsilon + \f {2\sqrt 2}{3\pi} {\epsilon}^{3/2} + \Theta({\epsilon}^{5/2})$ at 1.
The zeroth Bessel function of the second kind is defined by ${{\mathcal{K}_0}}(z) = \int_1^\infty e^{-z x} (x^2 - 1)^{-1/2} \dd x$. It is one of the fundamental solutions to the homogeneous differential equation $x^2 \dot y + x \dot y - x^2 y = 0$. The following lemma shows that $J_\alpha$ can be expressed in terms of ${{\mathcal{K}_0}}$.
\[lemma:JalphaBessel\] For any $\alpha > -1$, $J_\alpha(\theta) = \f 1 {2\pi {\mathsf{c}}_\alpha}\sin^{2\alpha+1} \theta \int_0^\infty \dd x {{\mathcal{K}_0}}(x) e^{x\cos \theta} x^\alpha$
@cho_kernel_2009 gave the expression $$2\pi {\mathsf{c}}_\alpha J_\alpha(\theta) = \csc \theta \int_0^\infty \dd u \int_0^\infty \dd v e^{-(u^2+v^2 - 2uv \cos \theta)/2\sin^2\theta} u^\alpha v^\alpha.$$ Note that the integrand is symmetric in $u$ and $v$.
Thus, if ${{\mathsf V}}= \{(u, v): u, v \ge 0 \And v \ge u\}$, then $$2\pi {\mathsf{c}}_\alpha J_\alpha(\theta) = 2\csc \theta \int_{{\mathsf V}}\dd u \dd v e^{-(u^2+v^2 - 2uv \cos \theta)/2\sin^2\theta} u^\alpha v^\alpha.$$
Now make the change of variables from ${{\mathsf V}}$ to $\{({\mathbbm{p}}, {\mathbbm{q}}): {\mathbbm{q}}\ge 2 \sqrt{{\mathbbm{p}}}\}$: $$\begin{aligned}
{\mathbbm{p}}&= uv & \dd {\mathbbm{p}}&= v \dd u + u \dd v\\
{\mathbbm{q}}&= u + v& \dd {\mathbbm{q}}&= \dd u + \dd v\\
\dd {\mathbbm{p}}\dd {\mathbbm{q}}&= (v - u) \dd u \dd v & \dd u \dd v &= ({\mathbbm{q}}^2 - 4 {\mathbbm{p}})^{-1/2} \dd {\mathbbm{p}}\dd {\mathbbm{q}}\end{aligned}$$ so that we have $$2 \pi {\mathsf{c}}_\alpha J_\alpha(\theta) = 2\csc \theta \int_0^\infty \dd {\mathbbm{p}}e^{{\mathbbm{p}}(1 + \cos \theta)\csc^2\theta} {\mathbbm{p}}^\alpha \int_{2\sqrt{{\mathbbm{p}}}}^\infty \dd {\mathbbm{q}}e^{-{\mathbbm{q}}^2\csc^2 \theta}({\mathbbm{q}}^2 - 4 {\mathbbm{p}})^{-1/2}.$$ The inner integral in ${\mathbbm{q}}$ can be expressed in terms of ${{\mathcal{K}_0}}$ by a change of variable $x = {\mathbbm{q}}^2/2\sqrt {\mathbbm{p}}$: $$\begin{aligned}
2 \pi {\mathsf{c}}_\alpha J_\alpha(\theta) &= 2\csc \theta \int_0^\infty \dd {\mathbbm{p}}e^{{\mathbbm{p}}(1 + \cos \theta)\csc^2\theta} {\mathbbm{p}}^\alpha \f 1 2 e^{-{\mathbbm{p}}\csc^2 \theta} {{\mathcal{K}_0}}({\mathbbm{p}}\csc^2 \theta)\\
&= \csc\theta \int_0^\infty \dd {\mathbbm{p}}{{\mathcal{K}_0}}({\mathbbm{p}}\csc^2\theta) e^{{\mathbbm{p}}\cos \theta \csc^2\theta} {\mathbbm{p}}^\alpha\\
&= \sin^{2\alpha + 1} \theta \int_0^\infty \dd x {{\mathcal{K}_0}}(x) e^{x \cos \theta} x^\alpha
\end{aligned}$$
Define $L_\alpha(\theta) = 2 \pi {\mathsf{c}}_\alpha J_\alpha(\theta)\csc^{2\alpha+1}\theta = \int_0^\infty \dd x {{\mathcal{K}_0}}(x) e^{x \cos \theta} x^\alpha$.
\[lemma:LAlphaRec\] If $\alpha > 1$, then $$L_\alpha(\theta) = \csc^2 \theta [(2\alpha-1) \cos \theta L_{\alpha-1}(\theta) + (\alpha-1)^2 L_{\alpha - 2}(\theta)].$$
We will prove this claim for $\theta < 1$, and by continuity this also proves the case $\theta = 1$. As remarked above, ${{\mathcal{K}_0}}(z) = \ddot {{\mathcal{K}_0}}(z) + \inv z \dot {{\mathcal{K}_0}}(z).$ Thus $$\begin{aligned}
L_\alpha(\theta) &= \int_0^\infty \dd x (\ddot {{\mathcal{K}_0}}(x) + \inv x \dot {{\mathcal{K}_0}}(x)) e^{x \cos\theta} x^\alpha\\
&=\dot {{\mathcal{K}_0}}e^{x\cos\theta} x^\alpha \rvert_0^\infty + {{\mathcal{K}_0}}e^{x\cos \theta} x^{\alpha - 1} \rvert_0^\infty\\
&\phantom{={}} - \int \dd x[\cos \theta e^{x \cos \theta} x^\alpha + \alpha e^{x \cos \theta} x^{\alpha - 1}] \dot {{\mathcal{K}_0}}\\
&\phantom{={}} - \int \dd x[\cos \theta e^{x \cos \theta} x^{\alpha - 1} + (\alpha-1) e^{x\cos\theta} x^{\alpha - 2}] {{\mathcal{K}_0}}\end{aligned}$$ Asymptotically, ${{\mathcal{K}_0}}(z) \sim \sqrt{\f{\pi}{2z}} e^{-z}$ as $z \to \infty$ and ${{\mathcal{K}_0}}(z) \sim -\ln(z)$ as $z \searrow 0$, and $\dot{{\mathcal{K}_0}}(z) \sim -\sqrt{\f{\pi}{2z}} e^{-z}$ as $z \to \infty$ and $\dot{{\mathcal{K}_0}}(z) \sim -\inv z$ as $z \searrow 0$. Thus, as $\alpha > 1$, $$\begin{aligned}
\dot {{\mathcal{K}_0}}e^{x\cos\theta} x^\alpha \rvert_0^\infty = -\lim_{x \to \infty} \sqrt{\pi/2} e^{-x(1 - \cos \theta)}x^{\alpha - 1} + \lim_{x\searrow 0} e^{x\cos\theta} x^{\alpha-1} = 0\\
{{\mathcal{K}_0}}e^{x\cos\theta} x^{\alpha-1} \rvert_0^\infty = -\lim_{x \to \infty} \sqrt{\pi/2} e^{-x(1 - \cos \theta)}x^{\alpha - 2} + \lim_{x\searrow 0} e^{x\cos\theta} x^{\alpha-1} \ln x = 0
\end{aligned}$$ So $$\begin{aligned}
L_\alpha(\theta) &= -\cos \theta L_{\alpha-1}(\theta) - (\alpha-1)L_{\alpha-2}(\theta) - \int \dd x[\cos \theta e^{x \cos \theta} x^\alpha + \alpha e^{x \cos \theta} x^{\alpha - 1}] \dot {{\mathcal{K}_0}}\end{aligned}$$ Via another integration by parts, the integral on the right is $$\begin{aligned}
&\phantom{={}} \cos \theta e^{x \cos \theta} x^\alpha {{\mathcal{K}_0}}\rvert_0^\infty + \alpha e^{x\cos \theta} x^{\alpha-1} {{\mathcal{K}_0}}\rvert_0^\infty\\
&\phantom{={}} - \int \dd x[\cos^2 \theta e^{x \cos \theta} x^\alpha + 2 \alpha \cos \theta e^{x \cos \theta} x^{\alpha - 1} + \alpha (\alpha - 1) e^{x \cos \theta} x^{\alpha - 2}] {{\mathcal{K}_0}}\\
&= - [\cos^2 \theta L_\alpha(\theta) + 2 \alpha \cos \theta L_{\alpha-1}(\theta) + \alpha(\alpha-1) L_{\alpha-2}(\theta)]
\end{aligned}$$ where the evaluation terms vanish just like before. Altogether, we have $$\begin{aligned}
L_\alpha(\theta) &= \cos^2 \theta L_\alpha(\theta) + (2\alpha-1)\cos\theta L_{\alpha-1}(\theta) + (\alpha-1)^2 L_{\alpha-2}(\theta)\\
&= \csc^2 \theta[(2\alpha-1) \cos \theta L_{\alpha-1}(\theta) + (\alpha-1)^2 L_{\alpha-2}(\theta)]
\end{aligned}$$
As a corollary we get
\[lemma:JalphaRec\] Suppose $\alpha > 1$. Then $$\begin{aligned}
J_\alpha(\theta) &= \cos \theta J_{\alpha-1}(\theta) + (\alpha-1)^2 (2\alpha-1)^{-1}(2\alpha-3)^{-1} \sin^2 \theta J_{\alpha-2}(\theta)\\
{\mathbb{J}}_\alpha(c) &= c {\mathbb{J}}_{\alpha-1}(c) + (\alpha-1)^2 (2\alpha-1)^{-1}(2\alpha-3)^{-1}(1-c^2) {\mathbb{J}}_{\alpha-2}(c)
\end{aligned}$$
The derivative of $J_\alpha(\theta)$ turns out to be quite simple.
\[lemma:JalphaGrad\] Suppose $\alpha > 0$. Then $$\begin{aligned}
\dot J_\alpha(\theta) &= -\alpha^2(2 \alpha-1)^{-1} J_{\alpha-1}(\theta) \sin \theta\\
\dot {\mathbb{J}}_\alpha(c) &= \alpha^2 (2 \alpha-1)^{-1} {\mathbb{J}}_{\alpha-1}(c)
\end{aligned}$$
We will prove the first formula. The second follows from chain rule. By \[lemma:JalphaBessel\], $$\begin{aligned}
J_\alpha(\theta) &= \f 1 {2\pi {\mathsf{c}}_\alpha}\sin^{2\alpha+1} \theta \int \dd x {{\mathcal{K}_0}}(x) e^{x\cos \theta} x^\alpha\\
\dot J_\alpha(\theta) &= \f 1 {2\pi {\mathsf{c}}_\alpha}[(2\alpha+1)\sin^{2\alpha} \theta \cos\theta \int \dd x {{\mathcal{K}_0}}(x) e^{x\cos \theta} x^\alpha\\
&\phantom{={}} -\sin^{2\alpha+2}\theta \int \dd x {{\mathcal{K}_0}}(x) e^{x\cos \theta} x^{\alpha+1}]\\
&= (2\alpha+1) \cot \theta J_\alpha(\theta) - \f{c_{\alpha+1}}{c_\alpha}\csc\theta J_{\alpha+1}(\theta)\\
&= (2\alpha+1) \csc \theta [\cos \theta J_\alpha(\theta) - J_{\alpha+1}(\theta)].\\
\end{aligned}$$ As $\alpha + 1 > 1$, by \[lemma:JalphaRec\], this is $$\begin{aligned}
&\phantom{={}}-(2\alpha+1)\csc \theta [(\alpha-1)^2 (2\alpha+1)^{-1}(2\alpha-1)^{-1} \sin^2 \theta J_{\alpha-1}(\theta)]\\
&= -(\alpha-1)^2(2\alpha-1)^{-1} \sin \theta J_{\alpha-1}(\theta).
\end{aligned}$$
Thus $\dot {\mathbb{J}}_\alpha(1) = \alpha^2(2\alpha-1)^{-1} {\mathbb{J}}_{\alpha-1}(1) = \alpha^2(2\alpha-1)^{-1}$ for any $\alpha > 0$ by \[lemma:basicJalpha\]. For $1/2 < \alpha \le 1$, $\dot {\mathbb{J}}_\alpha(1) \ge 1$ with equality iff $\alpha = 1$, and for $\alpha = 1/2$, $\dot {\mathbb{J}}_\alpha(1) =\infty > 1$ by continuity of $\dot {\mathbb{J}}_\alpha(c)$ in $\alpha$. Because for $\alpha > -1/2$, ${\mathbb{J}}_\alpha$ is increasing and convex on $[0, 1]$ and ${\mathbb{J}}_\alpha(0) > 0$ by \[lemma:basicJalpha\], ${\mathbb{J}}_\alpha$ intersects identity at a unique point away from 1 when $\alpha \in [1/2, 1)$. We record this as a theorem.
\[thm:stableFixedPointsJJ\] For $\alpha \in [1/2, 1)$, ${\mathbb{J}}_\alpha(c) = c$ has two solutions: an unstable solution at 1 (“unstable” meaning $\dot{\mathbb{J}}_\alpha(1) > 1$) and a stable solution in ${\mathbf{e}}^* \in (0, 1)$ (“stable” meaning $\dot {\mathbb{J}}_\alpha({\mathbf{e}}^*) < 1$).
![Left-to-right: **(a)** ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function (black, dashed line). ${\mathbb{J}}_1$ is highlighted in red. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. **(b)** shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. **(c)** Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$. (b) and (c) demonstrate the existence of stable fixed points away from 1 for ${\mathbb{J}}_\alpha, \alpha \in (1/2, 1)$, which is confirmed rigorously by \[thm:stableFixedPointsJJ\]. []{data-label="fig:jjj_vs_id_main"}](graphics/jjjVsId2.pdf "fig:"){height=".1\textheight"} ![Left-to-right: **(a)** ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function (black, dashed line). ${\mathbb{J}}_1$ is highlighted in red. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. **(b)** shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. **(c)** Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$. (b) and (c) demonstrate the existence of stable fixed points away from 1 for ${\mathbb{J}}_\alpha, \alpha \in (1/2, 1)$, which is confirmed rigorously by \[thm:stableFixedPointsJJ\]. []{data-label="fig:jjj_vs_id_main"}](graphics/jjj_vs_id_log.pdf "fig:"){height=".1\textheight"} ![Left-to-right: **(a)** ${\mathbb{J}}_\alpha$ for different $\alpha$s and the identity function (black, dashed line). ${\mathbb{J}}_1$ is highlighted in red. From this plot, it looks like ${\mathbb{J}}_\alpha(c) \ge c$ and $\dot{\mathbb{J}}_\alpha(c) \le 1$ for all $\alpha \in (\f 1 2, 1]$ with equality iff $c = 1$, but this is misleading. **(b)** shows $|{\mathbb{J}}_\alpha(c) - c|$ in log scale. Where the curves dip below the x-axis indicate points where ${\mathbb{J}}_\alpha(c) = c$. We see that in fact every ${\mathbb{J}}_\alpha$ has a solution ${\mathbb{J}}_\alpha(c) = c$ for a $c < 1$, when $\alpha < 1$. **(c)** Furthermore, at each such $c$, $\dot{\mathbb{J}}_\alpha < 1$. (b) and (c) demonstrate the existence of stable fixed points away from 1 for ${\mathbb{J}}_\alpha, \alpha \in (1/2, 1)$, which is confirmed rigorously by \[thm:stableFixedPointsJJ\]. []{data-label="fig:jjj_vs_id_main"}](graphics/jjj_vs_id_close.pdf "fig:"){height=".1\textheight"}
This result confirms that pictures presented in \[fig:jjj\_vs\_id\_main\]b,c are qualitatively correct, that there are indeed stable fixed points of ${\mathbb{J}}_\alpha$ away from 1.
If ${\underline}{\mathbf{e}}< 1$, then $$\begin{aligned}
{c}= \f{\sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\underline}{\mathbbm{p}}+ \sigma_b^2} &\ge {\underline}{\mathbf{e}}\\
{\mathbb{J}}_1({c}) &\ge {\mathbb{J}}_1({\underline}{\mathbf{e}})\\
{\mathbf{e}}= \f{\sigma_v^2 {\mathsf{c}}_\alpha {\mathbbm{q}}^\alpha{\mathbb{J}}_1({c}) + \sigma_b^2}{\sigma_v^2 {\mathsf{c}}_\alpha {\mathbbm{q}}^\alpha + \sigma_b^2} &\ge {\mathbb{J}}_1({\underline}{\mathbf{e}}) \end{aligned}$$ but ${\mathbf{e}}\ge {\mathbb{J}}_1({\underline}{\mathbf{e}}) > {\underline}{\mathbf{e}}$ as noted above. Thus by monotone convergence ${\mathbf{e}}$ converges, and ${\mathbf{e}}^* = 1$ is the only possible fixed point.
By \[lemma:cExpansion\], ${c}= {\underline}{\mathbf{e}}(1 + \Theta({\underline}{\epsilon}{\underline}{\mathbbm{p}}^{-1})) = 1 - {\underline}{\epsilon}+ \Theta({\underline}{\epsilon}\inv {\mathbbm{p}}) = 1 - u{\underline}{\epsilon}$ where $u := 1 - \Theta({\underline}{\mathbbm{p}}^{-1})$. Using the asymptotic expansion ${\mathbb{J}}_1(1 - {\epsilon}) = 1 - {\epsilon}+ U {\epsilon}^{3/2} + \Theta({\epsilon}^{5/2})$, we have $$\begin{aligned}
(1 - {\epsilon}){\mathbbm{p}}&= \sigma_v^2 \f {\mathbbm{q}}2 {\mathbb{J}}_1(1 - u{\underline}{\epsilon}) + \sigma_a^2 + (1 - {\underline}{\epsilon}) {\underline}{\mathbbm{p}}\\
-{\epsilon}{\mathbbm{p}}&= \sigma_v^2 \f {\mathbbm{q}}2 ({\mathbb{J}}_1(1 - u{\underline}{\epsilon}) - 1) - {\underline}{\epsilon}{\underline}{\mathbbm{p}}\\
&= \sigma_v^2 \f {\mathbbm{q}}2 [- u{\underline}{\epsilon}+ U u^{3/2}{\underline}{\epsilon}^{3/2} + \Theta(u^{5/2}{\underline}{\epsilon}^{5/2})] - {\underline}{\epsilon}{\underline}{\mathbbm{p}}\\
{\epsilon}&= {\underline}{\epsilon}\f 1 {\mathbbm{p}}[{\underline}{\mathbbm{p}}+ \sigma_v^2 \f {\mathbbm{q}}2 (u - U u^{3/2} {\underline}{\epsilon}^{1/2} + \Theta(u^{5/2}{\underline}{\epsilon}^{3/2}))]\\
&= {\underline}{\epsilon}\f 1 {\mathbbm{p}}[{\mathbbm{p}}- \sigma_a^2 + \sigma_v^2 \f {\mathbbm{q}}2 (\Theta(\inv {{\underline}{\mathbbm{p}}}) - U u^{3/2} {\underline}{\epsilon}^{1/2} + \Theta(u^{5/2}{\underline}{\epsilon}^{3/2}))]\\
&= {\underline}{\epsilon}[1 + \f{-\sigma_a^2 + \sigma_v^2 \f {\mathbbm{q}}2 (\Theta(\inv {{\underline}{\mathbbm{p}}}) - U u^{3/2} {\underline}{\epsilon}^{1/2} + \Theta(u^{5/2}{\underline}{\epsilon}^{3/2}))}{{\mathbbm{p}}}]\\
&= {\underline}{\epsilon}[1 + \f{-\sigma_a^2\inv {\mathbbm{q}}+ \f 1 2 \sigma_v^2 (\Theta(\inv {{\underline}{\mathbbm{p}}}) - U u^{3/2} {\underline}{\epsilon}^{1/2} + \Theta(u^{5/2}{\underline}{\epsilon}^{3/2}))}{{\mathbbm{p}}\inv {\mathbbm{q}}}]
$$ Let the content of the bracket on the RHS be $\aleph$. We have ${\mathbbm{p}}\inv {\mathbbm{q}}= (1+ o(1))B/\sigma_w^2$. If ${\epsilon}= O({\underline}{\mathbbm{p}}^{-1})$, then $\aleph = 1 - O(\inv {\mathbbm{p}})$, but because ${\mathbbm{p}}$ is exponentially decreasing, this means ${\epsilon}= \Theta(1)$ and does not converge to 0 — this is a contradiction. Therefore, ${\underline}{\epsilon}= \omega({\underline}{\mathbbm{p}}^{-1})$, and $$\begin{aligned}
{\epsilon}&= {\underline}{\epsilon}[1 - \f 1 2 B^{-1}\sigma_v^2 \sigma_w^2 U{\underline}{\epsilon}^{1/2}(1 + o(1))]\\
{\epsilon}- {\underline}{\epsilon}&= - \f 1 2 B^{-1}\sigma_v^2 \sigma_w^2 U{\underline}{\epsilon}^{3/2}(1 + o(1))\end{aligned}$$ Using \[lemma:polyDynamicsPosAlpha\] to upper and lower bound our dynamics, we get that $\p {\epsilon}l \sim [\f 1 4 \sigma_v^2 \sigma_w^2 \inv B U l]^{-2}$.
\[lemma:separableDynamics\] Let $\phi$ be any nonlinearity. Suppose ${\mathrm{W}}\phi( r, rd) = {\mathrm{V}}\phi( r){\mathbb{K}}(d)$ for some twice differentiable function ${\mathbb{K}}(d)$ independent of ${\mathbbm{q}}$, where ${\mathbb{K}}(1) = 1$ naturally. Suppose further that
- ${\mathbb{K}}(d) = d$ has a solution $d = {\mathbf{e}}^* > 0$ where $\dot{\mathbb{K}}({\mathbf{e}}^*) = \delta < 1$;
- ${\mathbb{K}}(d) > d$ for all $d < {\mathbf{e}}^*$ and ${\mathbb{K}}(d) < d$ for all $1 > d > {\mathbf{e}}^*$; and
- ${\mathbb{K}}$ is nondecreasing.
Let $\p {\epsilon}l := \p {\mathbf{e}}l - {\mathbf{e}}^*$ and suppose $\p {\mathbf{e}}0 < 1$. If $\p {\boldsymbol{\oldgamma}}l \to \infty$ and ${\mathrm{V}}\phi( \p {\mathbbm{q}}l) \to \infty$, then $\p {\epsilon}l \to 0$ and satisfies $${\epsilon}= {\underline}{\epsilon}\left(1 - \f{\sigma_a^2 + (1 - \delta + O({\underline}{\epsilon}))\sigma_v^2 {\mathrm{V}}\phi( {\mathbbm{q}})}{{\mathbbm{p}}}\right) + {\mathrm{V}}\phi( {\mathbbm{q}})\Theta(\inv {\boldsymbol{\oldgamma}}\inv {\mathbbm{p}}).$$
First we note that because ${\mathbf{e}}^*$ is the only stable fixed point of the dynamics $x \mapsto {\mathbb{K}}(x)$, with the basin of attraction $[0, 1)$, we can show $\p {\mathbf{e}}l \to {\mathbf{e}}^*$ as in the proof of \[thm:eDynamicsFullResTanh\] (using \[lemma:timeDependentConvergence\]).
Write $\p V l := {\mathrm{V}}\phi( \p {\mathbbm{q}}l)$. We first show that $\p {\mathbf{e}}l \to {\mathbf{e}}^*$. When $l$ is large, $$\begin{aligned}
{c}= \f{\sigma_w^2 {\underline}{\boldsymbol{\oldgamma}}+ \sigma_b^2}{\sigma_w^2 {\underline}{\mathbbm{p}}+ \sigma_b^2} & = {\underline}{\mathbf{e}}( 1 + O(\inv {\boldsymbol{\oldgamma}}))\\
{\mathbf{e}}= \f{\sigma_v^2V{\mathbb{K}}({c}) + \sigma_a^2}{\sigma_v^2 V + \sigma_a^2} &= {\mathbb{K}}({c})(1 + O(\inv V {\mathbb{K}}({c})^{-1})).\end{aligned}$$ If $\p {\boldsymbol{\oldgamma}}l$ is bounded for all $l$, then ${\mathbf{e}}\to 0$ because $\p {\mathbbm{p}}l \to \infty$. Since ${\mathbb{K}}({c}) > 0$ for ${c}\in [0, 1]$ and $\p V l\to \infty$, we have that in the limit $l\to \infty$, $\lim_{l \to \infty} {\mathbf{e}}= 0 = {\mathbb{K}}(\lim_{l \to \infty}{\mathbf{e}}) = {\mathbb{K}}(0)$ (by the continuity of ${\mathbb{K}}$), which is impossible by our assumptions. Thus $\p {\boldsymbol{\oldgamma}}l \to \infty$, and we have $\lim_{l \to \infty} {\mathbf{e}}= {\mathbb{K}}(\lim_{l \to \infty} {\mathbf{e}})$. By our assumptions, ${\mathbf{e}}^*$ is the only stable fixed point of ${\mathbb{K}}$ with basin of attraction $[0, 1)$, so this shows that ${\mathbf{e}}\to {\mathbf{e}}^*$ as desired.
Now we derive the equation in question. Note that ${c}= {\underline}{\mathbf{e}}(1 + \Theta(\inv {\boldsymbol{\oldgamma}}))$ because ${\mathbf{e}}^* < 1$. We use the Taylor expansion ${\mathbb{K}}({\mathbf{e}}^* + {\epsilon}) = {\mathbf{e}}^* + \delta {\epsilon}+ O({\epsilon}^2)$.
$$\begin{aligned}
({\mathbf{e}}^* + {\epsilon}){\mathbbm{p}}&= {\sigma_v^2}V {\mathbb{K}}\lp({\mathbf{e}}^* + {\underline}{\epsilon})(1 + \Theta(\inv {\boldsymbol{\oldgamma}}))\rp + {\sigma_a^2}+ ({\mathbf{e}}^* + {\underline}{\epsilon}) {\underline}{\mathbbm{p}}\\
&= {\sigma_v^2}V({\mathbf{e}}^* + \delta({\underline}{\epsilon}+ \Theta(\inv {\boldsymbol{\oldgamma}})) + O({\underline}{\epsilon}^2)) + {\sigma_a^2}+ ({\mathbf{e}}^* + {\underline}{\epsilon}) {\underline}{\mathbbm{p}}\\
{\epsilon}{\mathbbm{p}}&= {\sigma_v^2}V(\delta({\underline}{\epsilon}+ \Theta(\inv {\boldsymbol{\oldgamma}})) + O({\underline}{\epsilon}^2)) + {\underline}{\epsilon}{\underline}{\mathbbm{p}}\\
{\epsilon}&= {\underline}{\epsilon}(1 - \f{{\sigma_a^2}+ (1 - \delta + O({\underline}{\epsilon})){\sigma_v^2}V}{\mathbbm{p}}) + \Theta(V \inv{\boldsymbol{\oldgamma}}\inv {\mathbbm{p}})\end{aligned}$$
We apply \[lemma:separableDynamics\]. We first check the conditions of the lemma, with ${\mathbb{K}}= {\mathbb{J}}_\alpha$. The following conditions were already verified.
- ${\mathbb{J}}_\alpha$ has a fixed point ${\mathbf{e}}^*$ less than but very close to 1, where its slope is $\upsilon := \dot{\mathbb{J}}_\alpha({\mathbf{e}}^*) < 1$. (\[thm:stableFixedPointsJJ\])
- ${\mathbb{J}}_\alpha(d) > d$ for all $d < {\mathbf{e}}^*$ and ${\mathbb{J}}_\alpha(d) < d$ for all $d > {\mathbf{e}}^*$. (By the convexity shown in \[lemma:basicJalpha\])
- ${\mathbb{J}}_\alpha$ is nondecreasing (\[lemma:basicJalpha\]). Furthermore, from its integral formula (\[eqn:JalphaIntegralFormula\]), we see easily that ${\mathbb{J}}_\alpha$ is smooth at ${\mathbf{e}}^* < 1$.
We also proved the following
- $\p {\mathbbm{p}}l \sim [\sigma_v^2 \sigma_w^{2\alpha} {\mathsf{c}}_\alpha (1 - \alpha)]^{\f 1 {1-\alpha}}l^{\f 1 {1-\alpha}}$ (\[thm:pDynamicLT1ReLU\]) and $\p {\boldsymbol{\oldgamma}}l$ is asymptotically a constant fraction of $\p {\mathbbm{p}}l$ (\[lemma:separableDynamics\]), so both go to $\infty$.
- ${\mathrm{V}}\psi_\alpha( {\mathbbm{q}}) = {\mathsf{c}}_\alpha {\mathbbm{q}}^\alpha = {\mathsf{c}}_\alpha (\sigma_w^2 {\mathbbm{p}}+ \sigma_b^2)^\alpha = \Theta(l^{\alpha/(1-\alpha)})$, so goes to $\infty$. (\[lemma:VtPsiAlpha\])
Thus, for $\upsilon = \dot {\mathbb{J}}({\mathbf{e}}^*)$, $$\begin{aligned}
\f{\sigma_a^2 + (1 - \upsilon + O({\underline}{\epsilon}))\sigma_v^2 {\mathrm{V}}\phi( {\mathbbm{q}})}{{\mathbbm{p}}} &\sim \f{(1 - \upsilon)\sigma_v^2\sigma_w^{2\alpha}{\mathsf{c}}_\alpha}{{\mathbbm{p}}^{1-\alpha}}\\
&= \inv l (1-\upsilon)/(1 - \alpha).\end{aligned}$$
Now, ${\mathrm{V}}\phi( {\mathbbm{q}})\inv {\boldsymbol{\oldgamma}}\inv {\mathbbm{p}}= \Theta(l^{-\f1{1-\alpha} - 1})$. By using the dynamics of \[lemma:alphaDeltaDynamics\] to upper and lower bound our dynamics, we have $\p {\epsilon}l = \Omega(l^{-\mu - {\epsilon}}), O(l^{-\mu + {\epsilon}})$ for any ${\epsilon}> 0$, where $\mu = \min((1-\upsilon)/(1-\alpha), 1/(1-\alpha)) = (1-\upsilon)/(1-\alpha).$
### Backward Dynamics
\[lemma:infVarAlphaLe75\] Suppose random variable $X \sim {\mathcal{N}}(0, \sigma^2)$, and $Y = \psi_{-\beta}(X)$ for some $\beta > 0 $, where $\psi_\alpha$ is $\alpha$-ReLU. Then for $\xi > 0$, $Y$ has density $$\Pr[Y \in [\xi, \xi+\dd \xi]] = \f 1 {\beta\sqrt{2\pi \sigma^2}} \xi^{-\f 1 \beta - 1} e^{-\xi^{-2/\beta}/2\sigma^2}.$$ At $\xi = 0$, $Y$ has density given by a Dirac delta of mass $\f 1 2$.
Furthermore, $Y$ has finite second moment iff $\beta < \f 1 2$.
We have $$\begin{aligned}
\Pr[Y \in [\xi, \infty)] &= \Pr[X \in [0, \xi^{-1/\beta}]]\\
&= \f 1 {\sqrt{2\pi \sigma^2}}\int_0^{\xi^{-1/\beta}} e^{-x^2/2\sigma^2}\dd x.
\end{aligned}$$ Differentiating the RHS against $\xi$ using Leibniz’s rule, we get $$\begin{aligned}
d\Pr[Y \in [\xi, \infty)]/d\xi &= \f 1 {\sqrt{2\pi \sigma^2}} e^{-\xi^{-2/\beta}/2\sigma^2} \f d {d\xi}\xi^{-1/\beta}\\
&= \f {-1} {\beta\sqrt{2\pi \sigma^2}} \xi^{-\f 1 \beta - 1} e^{-\xi^{-2/\beta}/2\sigma^2}.
\end{aligned}$$ Negating both sides gives the density $f_Y$ of $Y$ for $\xi > 0$. For $\xi = 0$, observe that $\lim_{\xi \to 0} f_Y(\xi) = 0$ because, while $\xi^{-\f 1 \beta -1}$ blows up polynomially, $e^{-\xi^{-2/\beta}/2\sigma^2}$ blows up exponentially. Thus the contribution of $Y$’s mass at $Y = 0$ from $X > 0$ is 0. On the other hand, all $X < 0$ gets mapped to $Y = 0$, so $f_Y(0) = \f 1 2 \delta_0$, where $\delta_0$ is the Dirac delta.
For the second assertion, observe that $$\begin{aligned}
f_Y(\xi) \sim \f 1 {\beta \sqrt{2\pi \sigma^2}} \xi^{-\f 1 \beta - 1} & \text{as ${\xi \to \infty}$}.
\end{aligned}$$ Thus, $\xi^2 f_Y(\xi)$ is integrable iff $2 -\f 1 \beta - 1 < -1 \iff \beta < \f 1 2$.
Note that $\dot \psi_\alpha \propto \psi_{\alpha - 1}$, so it suffices to show that $\Var(\psi_{\alpha-1}(\zeta)^2) = \Var(\psi_{2\alpha - 2}(\zeta))$ is infinite for $\zeta \sim {\mathcal{N}}(0, \sigma^2)$. By \[lemma:infVarAlphaLe75\] with $\beta = 2 - 2\alpha$, $\psi_{2\alpha - 2}(\zeta)$ has finite variance iff $\beta < \f 1 2 \iff \alpha > \f 3 4$.
If $\alpha = 1$, then $${\underline}{{\boldsymbol{\oldchi}}}= {{\boldsymbol{\oldchi}}}(1 + \f 1 2\sigma_v^2 \sigma_w^2).$$ So $\p {{\boldsymbol{\oldchi}}}{l-m} / \p {{\boldsymbol{\oldchi}}}l = \Theta(1) B^m$ for $B = 1 + \f 1 2\sigma_v^2 \sigma_w^2$.
If $\f 1 2 < \alpha < 1$, then ${\underline}{{\boldsymbol{\oldchi}}}/{{\boldsymbol{\oldchi}}}- 1$ is $$\begin{aligned}
&\phantom{=} \sigma_v^2 \sigma_w^2 {\mathrm{V}}\dot \phi( {\mathbbm{q}})\\
&= \sigma_v^2 \sigma_w^2 \alpha^2 {\mathsf{c}}_{\alpha - 1} {\mathbbm{q}}^{\alpha - 1}\\
&= \sigma_v^2 \sigma_w^2 \alpha^2 {\mathsf{c}}_{\alpha - 1} (\sigma_w^2 {\mathbbm{p}})^{\alpha - 1} + \Theta({\mathbbm{p}}^{\alpha - 2})\\
&= \sigma_v^2 \sigma_w^{2\alpha} \alpha^2 {\mathsf{c}}_{\alpha - 1}(K_1 l^{{\f 1 {1-\alpha}}}- K_2 l^{{\f \alpha {1-\alpha}}}\log l + o(l^{{\f \alpha {1-\alpha}}}\log l))^{\alpha - 1} + \Theta(l^{\f{\alpha - 2}{1 - \alpha}}) & \text{by \cref{thm:pDynamicLT1ReLU}}\\
&= \sigma_v^2 \sigma_w^{2\alpha} \alpha^2 {\mathsf{c}}_{\alpha - 1} [K_1^{\alpha - 1} \inv l + \Theta(l^{-2} \log l)] + O(l^{-3})\\
&= \sigma_v^2 \sigma_w^{2\alpha} \alpha^2 {\mathsf{c}}_{\alpha - 1} K_1^{\alpha - 1} \inv l + \Theta(l^{-2} \log l)\\
&= R \inv l + \Theta(l^{-2} \log l)\end{aligned}$$ where $R = \sigma_v^2 \sigma_w^{2\alpha} \alpha^2 {\mathsf{c}}_{\alpha - 1} K_1^{\alpha - 1} = \f{\alpha^2}{(1-\alpha)(2 \alpha - 1)} $ and $K_1 = [\sigma_v^2 \sigma_w^{2\alpha} {\mathsf{c}}_\alpha (1 - \alpha)]^{\f 1 {1-\alpha}}$. So $$\begin{aligned}
{\underline}{{\boldsymbol{\oldchi}}}&= {{\boldsymbol{\oldchi}}}\exp(R \inv l + \Theta(l^{-2} \log l))\\
\p {{\boldsymbol{\oldchi}}}{l -m} &= \Theta(1)\p {{\boldsymbol{\oldchi}}}{l} \lp\f l {l-m} \rp^R\end{aligned}$$ as desired.
The proof is similar to that of \[thm:dalethExpSqrtTanhAllGrad\].
[^1]: Work done while at Harvard University
[^2]: Under simplified conditions, @daniely_toward_2016 showed that there exists a fixed point for any “well-behaved” activation function in a feedforward net. However, this result does not apply to architectures with residual connections.
[^3]: Note that in practice, to avoid the diverging gradient $\dot \psi_\alpha(x) \to \infty$ as $x \to 0$, we can use a tempered version $\Psi_\alpha(x)$ of $\alpha$-ReLU, defined by $\Psi_\alpha(x) = (x + {\epsilon})^\alpha - {\epsilon}^\alpha$ on $x > 0$ and 0 otherwise, for some small ${\epsilon}> 0$. The conclusions of this paper on $\psi_\alpha$ should hold similarly for $\Psi_\alpha$ as well.
[^4]: @daniely_toward_2016 called the version of ${\mathrm{W}}\phi$ with fixed $\rho = 1$ the “dual function” of $\phi$.
[^5]: [A more natural visualization is to graph $\p {\mathbf{e}}l - {\mathbf{e}}^*$ versus $l^{-\delta^*}$, but because of floating point precision, $\p {\mathbf{e}}l - {\mathbf{e}}^*$ doesn’t converge to 0, but a small number close to 0, so that the log-log plot wouldn’t look like what is expected.\[footnote:plotDelta\]]{}
[^6]: [A more natural visualization is to graph $\p {\mathbf{e}}l - {\mathbf{e}}^*$ versus $l^{-\delta^*}$, but because of floating point precision, $\p {\mathbf{e}}l - {\mathbf{e}}^*$ doesn’t converge to 0, but a small number close to 0, so that the log-log plot wouldn’t look like what is expected.\[footnote:plotDelta\]]{}
[^7]: Our derivations actually apply to all $\alpha \in (\f 1 2, 1]$, where at $\alpha = \f 1 2$, the expected norm of the gradient diverges within our mean field formalism. However, at $\alpha \le \f 3 4$, the variance of the gradient already diverges (\[thm:dalethInfVarAlphaReLU\]), so we cannot expect the empirical values to agree with our theoretical predictions. But in fact, empirically our theoretical predictions seem to form an upper bound on the gradient norms (see \[fig:alphaReLUTheoryVsEmpirics\]).
[^8]: the contour for $\p {\mathbf{p}}l$ is similar, but its slopes are slightly off from the heatmap contours.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A substantial population of high redshift early-type galaxies is detected in very deep UBVRIJHK images towards the HDF-South. Four elliptical profile galaxies are identified in the redshift range 1$<$z$<$2, with very red SEDs, implying ages of $\gtrsim\hspace{-.15cm}2$ Gyrs for standard passive evolution. We also find later type IR-luminous galaxies at similarly high redshift, (10 objects with z$>$1, H$<$25), with weak UV emission implying single burst ages of $\gtrsim\hspace{-.15cm}1$ Gyr. The number and luminosity-densities of these galaxies are comparable with the local E/SO-Sbc populations for $\Omega_m$$>$0.2, suggesting that the major fraction of luminous Hubble-sequence galaxies have evolved little since $z\sim2$. A highly complete photometric redshift distribution is constructed to H=25 (69 galaxies) showing a broad spread of redshift, peaking at z$\sim$1.5, in reasonable agreement with some analyses of the HDF. Four ‘dropout’ galaxies are detected at $z\approx 3.8$, which are compact in the IR, $\sim$0.5 kpc/h at rest 3500Å. No example of a blue IR luminous elliptical is found, restricting the star-formation epoch of ellipticals to z$\geq$10 for a standard IMF and modest extinction.'
author:
- 'Narciso Benítez, Tom Broadhurst, Rychard Bouwens'
- Joseph Silk
- Piero Rosati
title: 'Detection of Evolved High-Redshift Galaxies in Deep NICMOS/VLT Images'
---
Date: Thu, 11 Feb 1999 14:59:55 -0800 (PST) From: tom broadhurst <[email protected]> To: [email protected]
&
Introduction
============
The detection of early type galaxies at z$>$1 requires deep IR imaging to overcome the observed large restframe optical K-correction and correspondingly minimal inferred passive spectral evolution for this class of galaxy (e.g. Stanford, Eisenhardt & Dickinson 1998). The highest redshift example of an early-type galaxy is a luminous object reported by Dunlop et al. (1996) at z=1.55. Its restframe UV spectrum is dominated by late-type F-stars, equivalent to an age of $\approx 3$Gyrs, for standard stellar synthesis models (Spinrad et al. 1997). Following ordinary ellipticals to $z>1$ requires much deeper IR imaging now possible from space. The NICMOS observation in the Hubble Deep Field South (Fruchter et al 1999) represents a three magnitude increase in depth over the deepest ground based IR data, and with usefully high spatial resolution. This, in combination with deep optical images from the VLT allows easy identification of red $L{^*}$ ellipticals to $z\sim2.5$, where significant spectral and density evolution is widely anticipated.
In section $\S1$ we briefly outline our spectral and profile fitting procedures and simulations as applied to the NICMOS and VLT observations. In $\S2$ we discuss detections of early-type and other IR bright galaxies at high–redshift. In $\S3$ we compare their statistical properties with local populations of Hubble sequence galaxies of similar spectral type, for different evolutionary models. In $\S4$ we briefly discuss a the absence of bright blue precursor ellipticals in the IR.
Photometry, Redshift Estimation and Profile Fitting
===================================================
Three sets of images are utilized, the HDF-S NICMOS observations in JHK (Fruchter et al. 1999), the higher resolution unfiltered HST STIS image (Gardner et al. 1999) and the VLT Test Camera images in UBVRI (Fontana et al. 1999). We use the NICMOS pipeline zero-points and have computed the VLT effective zero-points for the combined optical images from the ESO photometric solution. The filter functions and instrument response curves are kindly made available by ESO and STScI. The detection is performed using SExtractor (Bertin & Arnauts 1996) in a combined $J+H$ frame. Near total magnitudes are measured in J,H,K and PSF matched STIS images within a matched aperture set. A total of $\approx 200$ galaxies are detected to H=27 within the central 50”$\times$ 50” field. The VLT UBVRI images have PSF’s $\approx 3$ times wider that the NICMOS exposures, and require a different procedure. We measure magnitudes within a $0\farcs 9$ aperture and correct them to near-total magnitudes using versions of the $J+H$ band image degraded to the PSF of each of the VLT passbands to ensure the colors are defined within the same apertures for all bands.
Photometric redshifts are estimated using the methods described in Benítez (1999). The empirical template set of Coleman, Wu, & Weedman (1980) is used, and augmented with bluer local starbursts (Kinney et al. 1996), to determine the redshift probability distribution for each galaxy with Bayesian weighting. The fraction of objects with highly reliable photometric redshifts (p$>$0.99) is $57/69$ to H$<$25 ($24/25$ for objects classified as earlier than Scd). Fainter than this magnitude limit the reliability of the photometric redshift estimates quickly degrades. We also allow for the quoted 10% uncertainty in the NICMOS zero-points for our spectral fits.
For redder early types, more detailed fits to single burst synthetic models are justified to accommodate the known spread of metallicity and age. For this, the compilation of Leitherer et al. (1996) is used, including the synthetic galaxy spectra of Bruzual & Charlot (1993). We also check that the colors of these objects are not compatible with dust reddened star–forming galaxies by generating reddened versions of the late and Im types in our template set.
Profile fitting is performed for bright galaxies in the IR (H$<$23.7). For this we simulate NICMOS images using the methods described in Bouwens, Broadhurst & Silk (1998) to account for the pixelization, dithering and point spread functions and the available weight map for the NICMOS dithering pattern.
High-Redshift Red Galaxies
==========================
The $H<25$ sample contains five clear elliptical galaxies, of which four lie above z=1. Their SED’s and surface brightness profiles are compared with the models in Fig 1 and listed in Table 1. All the objects spectrally classified as ellipticals are also found to have spheroidal profiles (Fig 1). The most impressive object (NIC3/ET4, Fig 1, Table 1) is extremely red, consistent with minimal passive evolution at an estimated $z=1.94\pm0.15$, with a clear de-Vaucouleurs profile and uniform color over its surface ($\sim$ 300 pixels). The effective radius 1.6(2.2)kpc/h is consistent with its inferred luminosity $M_{AB}=-21.96(-22.6)$ in restframe R for $\Omega_m=1(0.1)$ (Binggelli, Sandage & Tarenghi 1984). This IR conspicuous object is absent or only marginally detected in the U,B,V,R,I and STIS bands (Fig 1). Dust reddened spectra extinguished with standard reddening (Calzetti, Kinney, & Storchi-Bergmann 1994) are unable to generate the steep break from the relative flat IR. Another elliptical identified at $z=1.66 \pm 0.25$ (NIC3/ET3, Fig 1, Table1) with a clear de-Vaucouleurs profile, is the luminous object reported by Treu et al. (1998) in the earlier NICMOS test images and confirmed by Stiavelli et al. (1999).
We also find several IR-bright high-redshift galaxies of later spectral type, 10 objects with z$>$1 and H$<$25. Two typical examples are shown in Fig 1. The profiles of these objects are generally exponential. The SED’s are very conspicuous with a turn up towards the UV, and a distinctive bump in the IR separated by a trough, so there is little uncertainty in redshift (e.g. Fig 1). The best fit single burst minimum ages are typically $\sim 1$ Gyr. Comparison with the SEDs of CWW show these objects span the spectral types of local Sab to Sbc galaxies. The full sample of 24 galaxies with H$<$25 (open histogram, Fig 2) have a redshift distribution peaked at $z~1.5$ with high-z tail similar to that found for the HDF (Connolly et al. 1997, Fernández–Soto, Lanzetta & Yahil 1998, Benítez 1999).
Density Evolution
=================
A reliable measurement of the space density of massive galaxies at high redshift provides a basic check of hierarchical models of galaxy formation, where the most massive haloes form late (Lacey & Cole 1993) and in which ellipticals and massive spheroids may be viewed as the end product of the merging of smaller systems.
The presence of so many luminous red galaxies at high-z in such a small field is surprising and indicates that relatively evolved luminous galaxies are commonplace at z$>$1. Using the Pozzetti, Bruzual, & Zamorani (1996) local luminosity functions for E/SO-Sbc as input, we recover 4(9) ellipticals from our realistic NICMOS simulations for $z>1$, $H<25$, with $\Omega_m=1(0.2)$ and no evolution. Adding the later-type IR-luminous galaxies allows a useful comparison with models (Fig 2). The redshifts and luminosities observed above z=1 are consistent with no density or luminosity evolution of standard Hubble sequence galaxies for $0.2<\Omega_m<1$. Note, passive evolution is minimized here for consistency with the red SEDs by choosing $H_0=50$km/s/Mpc and $z_f=15$. The hierarchical model parametrization of Kauffmann, Charlot & White (1996) for $\Omega=1$ underpredicts the numbers as expected. Hierarchical models with $\Omega <1$ may allow a declining density but must be checked for self consistency. The optical-IR colors prefer older higher redshift models and are consistent with no size evolution, but do not distinguish clearly between the above cosmological models.
The integrated luminosity density to H=25 at 1$<$z$<$2.5 is comparable with local estimates integrated over E/SO-Sbc types (see Fig 2). The predicted luminosity density for $z>1$ at restframe 6000$\AA$ is $4.3\times10^{19}$ergs/cm$^{2}$/s/Hz compared to $3.9\pm1.4\times
10^{19}$ ergs/cm$^{2}$/s/Hz, $2.41\pm0.9\times 10^{19}$ ergs/cm$^{2}$/s/Hz and $2.06\pm0.8\times 10^{19}$ ergs/cm$^{2}$/s/Hz, for $\Omega_m=1$, $\Omega_m=0.1$ and $\Omega_m+\Omega_{\Lambda}=0.25+0.75$ respectively. A factor of two decrease in luminosity density is implied relative to the popular low $\Omega_m$ models assuming passive or no stellar evolution.
Our results seem to extend to $z\approx 2$ the results of other authors notably Lilly et al. (1995) and Kodama, Bower & Bell (1999) who find no evidence for a dramatic evolution in the density of early types galaxies to $z\sim1$. The contrast of our results with the absence of distant ellipticals claimed by others (Kauffmann & Charlot 1998; Zepf 1997; Franceschini et al. 1996; Barger et al. 1998; Menanteau et al. 1998) can be put in context by appreciating that the NICMOS data is complete to $H\approx 27$, which is 3 magnitudes fainter than the deepest ground based data and 6 magnitudes fainter than most work on this question. Two of the four ellipticals found here in the IR would barely register in the I band (I$\sim$27) and are certainly not bright enough for morphological classification in the optical even from space. Recent claims of a decline in the elliptical density for z$>$1 rely on significant passive evolution, where brightening enhances the expected numbers. The claim by Kauffmann, Charlot, & White (1996), based on the CFRS (Lilly et al. 1995) is subject to redshift incompleteness, in particular of red high-z galaxies, which are very hard to measure spectroscopically. Some of the red galaxies reported here are bright enough in the IR to be detected in previous deep Keck images (Moustakas et al. 1997; Hogg et al. 1997) but morphological classification is not possible in ground based images and deep complementary optical data are required to establish the break. The NICMOS images directly show that luminous early type galaxies are present in significant numbers at high redshift with SEDs too red to be recognized as ellipticals in the optical even in the deepest WFPC2 images.
An obvious problem with this kind of work is the presence of clustering in small field, rendering the results uncertain. To explore this we have searched for spikes in the redshift distribution using the approach described in Benítez (1999). The minimal number of structures along the light–of-sight to H=25 is three, at $z=0.55$ (consistent with the $0.58$ redshift spike found in the HDF-S by Glazebrook et al. 1999), $z=1.35$ and $z=1.95$. Each of these structures corresponds approximately to the redshift of one or more elliptical galaxies, and the two higher redshift peaks contain $\approx 1/3$ of all the $z>1, H<25$ galaxies. Our statistical analysis shows that it is virtually impossible that all the early type galaxies at $1<z\lesssim 2$ belong to a single spike. A rough estimate of the expected clustering noise can be obtained by integrating the faint galaxy correlation function over the NICMOS field for the range 1$<$z$<$2. Connolly, Szalay & Brunner 1998 measure an amplitude $A(10'')=0.12 (\omega \propto \theta^{-0.8})$ for the angular correlation of HDF-N galaxies within $\Delta z=0.4$ slices to $z\sim 1.5$. Assuming that $A\propto 1/\Delta z$, the total variance in the number of galaxies in a redshift slice $\Delta z$ is $\sigma^2(\Delta z)\sim N(\Delta z)[1+N(\Delta z)(0.4/\Delta z)0.07]$. For the interval $1<z<2$ with 35 galaxies this yields a variance of $\sigma=8.3$ galaxies, which is 40% greater than Poisson. If the early types are clustered like the integral of all types then we estimate their number to be $14\pm4$ in the range $1<z<2$.
Very High-Redshift Blue Galaxies
================================
The lack of spectral evolution of early type galaxies demonstrates that star-formation does not extend over time, but is confined to an early brief period, so that we should expect the precursors of these galaxies to appear extremely luminous and blue at higher redshift. The width of this period is constrained to be short by the tight E/SO color-magnitude and M/L sequences (e.g. Bower, Lucey & Ellis 1992, VanDokkum et al. 1998). More directly, the lack of blue ellipticals in the HDF (e.g. Franceschini et al. 1998) means this epoch lies at $z>5$, and therefore may appear in the near-IR for redshifts in range 5$<$z$<$15 as I or J band ‘dropouts’, provided the IMF of early-type galaxies is not unusually deficient in hot stars.
Our simulated NICMOS images show that significant numbers of elliptical galaxies are expected in the IR if they form luminous stars high redshift and were not significantly extinguished during the first Gyr. The simulations include an exponential burst at $z_f=15$ with an e-folding time of 1Gyr and predicts 6, 10 and 15 objects for z$>$5 to H=26 for, $\Omega_m=1$, $\Omega=0.1$, $\Omega_m+\Omega_{\Lambda}=0.25+0.75$. In comparison, no I or J band ‘dropout’ galaxies are observed brighter than this limit.
At lower redshift, five compact optically blue dropout galaxies are found, all consistent with a single redshift of $z\approx 3.8$. Fits to their spectra are shown in Fig 3. These objects have low SFR of 6-23$M_{\odot}/yr$ and very compact sizes in the H-band, $\sim$0.5 kpc/h at restframe $3500\AA$ ($\Omega_m$=1), and therefore appear unlikely to be the precursors of luminous elliptical galaxies if no evidence of SF is to remain by z$\sim$1.
Conclusions
===========
We have detected a substantial population of early type galaxies at z$>$1. These include $r^{1/4}$ profile galaxies in the redshift range z$>$1, all of which have surprisingly red colors despite their large redshifts. Later spectral-types are also present in this redshift range, with a conspicuous IR bump. These galaxies are all luminous and well evolved, with minimum single-burst ages $>$1Gyr for a standard IMF.
The clear detection of early type galaxies and spheroidal components of other high-z disk galaxies points to the existence of a roughly similar density of luminous Hubble sequence galaxies at $z\sim2$, compared to the local Universe. The declining density of luminous galaxies expected in hierarchical schemes can be made consistent with the data only with for low $\Omega_m$ models. Low $\Omega_m$ helps accommodate the absence of spectral evolution, but in any cosmology the lack of bright IR-blue precursors of elliptical galaxies requires a very high redshift of formation z$>$10, for a standard IMF and extinction.
Although the field studied is small, the number of red high-z galaxies detected is surprising and strongly suggests that the bulk of luminous galaxies were formed at $z\gtrsim2$. In the short–term a deep NICMOS pointing in the HDF will provide similar data and help average over clustering. Progress on this question awaits deep diffraction-limited IR imaging on large telescopes and improved IR facilities in space.
We would like to thank Bob Williams and the Hubble Deep Field South Team as well as Alvio Renzini and the VLT-UT1 Science Verification Team for making available this useful data.
Barger, A. J., et al. 1998, , 501, 522 Benítez, N. 1999, submitted to ApJ, astro-ph/9811129. Bertin, E., & Arnouts, S. 1996, , 117, 393. Binggeli, B., Sandage, A., & Tarenghi, M. 1984, , 89, 64.
Bower, R.G., Lucey, J.R., & Ellis, R.S. 1992, , 254, 601
Bouwens, R.J., Broadhurst, T.J., & Silk, J. 1998, , 506, 557. Bruzual A., G. & Charlot, S. 1993, , 405, 538 Calzetti, D., Kinney, A.L., Storchi-Bergmann, T. 1994, , 429, 582. Coleman, G.D., Wu, C.-C., & Weedman, D.W. 1980, , 43, 393. Connolly, A. J., Szalay, A. S., Dickinson, M., Subbarao, M. U., & Brunner, R. J. 1997, , 486, L11 Connolly, A. J., Szalay, A. S. & Brunner, R. J. 1998, , 499, L125 Dunlop, J., Peacock, J., Spinrad, H., Dey A., Jimenez, R., Stern, D., & Windhorst, R.A. 1996, , 381, 581. Fernández-Soto, A., Lanzetta, K.M. & Yahil, A. 1998, to appear in ApJ, astro-ph/9809126 Fontana, A., D’Odorico, S., Fosbury, R., Giallongo, E., Hook, R., Poli, F., Renzini, A., Rosati, P., & Viezzer, R. 1999, to appear in A&A, astro-ph/9901359 Franceschini, A., Silva, L., Fasano, G., Granato, L., Bressan, A., Arnouts, S., & Danese, L. 1998, , 506, 600 Fruchter et al. 1999, to appear in AJ
Gardner et al. 1999, to appear in AJ
Glazebrook et al. 1999, in preparation
Hogg, D. W., Neugebauer, G., Armus, L., Matthews, K., Pahre, M. A., Soifer, B. T., & Weinberger, A. J. 1997, , 113, 2338 Kauffmann, G. 1996, , 281, 487. Kauffmann, G. & Charlot, S. 1998, , 297, L23. Kauffmann, G., Charlot, S., & White, S.D.M. 1996, , 283, L117. Kinney, A.L., Calzetti, D., Bohlin, R.C., McQuade, K., Storchi-Bergmann, T., & Schmitt, H.R. 1996, , 467, 38. Kodama, T., Bower, R.G. & Bell, E.F. 1999, submitted to MNRAS, astro-ph/9810138
Lacey, C. & Cole, S. 1993, , 271, 676 Leitherer, C., et al. 1996, 108, 996.
Lilly, S.J., Tresse, L., Hammer, F., Crampton, D., & Le Févre, O. 1995, , 455, 108
Madau, P., Pozzetti, L., & Dickinson, M. 1998, , 498, 106
Menanteau, F., Ellis, R.S., Abraham, R.G., Barger, A.J. & Cowie, L.L., submitted to MNRAS, astro-ph/9811465
Moustakas, L. A., Davis, M., Graham, J. R., Silk, J., Peterson, B. A., & Yoshii, Y. 1997, , 475, 445
Pozzetti, L., Bruzual, G. & Zamorani, G. 1996, , 281, 953. Spinrad, H., Dey, A, Stern, D., Dunlop, J., Peacock, J., Jimenez, R., & Windhorst, R. 1997, , 484, 581. Stanford, S.A., & Eisenhardt, P.R., & Dickinson, M. 1998, , 492, 461. Stiavelli, M., et al. 1999, submitted to , astro-ph/9812102. Treu, T., et al. 1998, , 340, L10 Van Dokkum, P. G., Franx, M., Kelson, D. D., & Illingworth, G. D. 1998, , 504, L17
Zepf, S. 1997, , 390, 377.
[cccccccc]{} NIC3/ET1 & 181.6 & 868.87 & $21.24\pm0.01$ & $1.41\pm0.20$ & $0\farcs34$ & 3.3 & $>1.5$ NIC3/ET2 & 439.2 & 352.34 & $24.50\pm0.06$ & $1.55\pm0.20$ & $0\farcs31$ & $>4.0$ & $>1.5$ NIC3/ET3 & 607.4 & 371.52 & $21.79\pm0.01$ & $1.66\pm0.25$ & $0\farcs26$ & 4.2 & $>2$ NIC3/ET4 & 646.8 & 898.32 & $23.18\pm0.02$ & $1.94\pm0.15$ & $0\farcs39$ & $>5.4$ & $>3.5$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A thermodynamic theory is developed to describe the behavior of the entanglement between the coin and position degrees of freedom of the quantum walk on the line. This theory shows that, in spite of the unitary evolution, a steady state is established after a Markovian transient stage. This study suggests that if a quantum dynamics is developed in a composite Hilbert space (*i.e.* the tensor product of several sub-spaces) then the behavior of an operator that only belongs to one of the sub-spaces may camouflage the unitary character of the global evolution.'
author:
- Alejandro Romanelli
title: The quantum walk temperature
---
Introduction
============
The concept of isolated system plays a fundamental role in the formulation of the quantum mechanics. This concept is an idealization that was constructed as an aid to understand some phenomena displayed by real systems which may be regarded as approximately isolated. However, since about 50 years ago, the study of quantum decoherence has acquired a central position in the formulation of the quantum mechanics. In fact, concepts such as thermodynamic equilibrium seem impossible to coordinate with the idea of isolated system because the quantum state for such a system follows a unitary evolution and it cannot reach a final equilibrium state at $t\rightarrow\infty $.
In this context we ask ourselves if it is possible to introduce the concept of temperature for an isolated quantum system which evolves in a composite Hilbert space. In this paper the system known as the quantum walk on the line [@QW] has been chosen as a model to answer this question . The quantum walk (QW) is a natural generalization of the classical random walk in the frame of quantum computation and quantum information processing and it is receiving permanent attention [childs,Linden,Alejo3]{}. It has the property to spread over the line linearly in time as characterized by the standard deviation $\sigma
(t)\sim t$, while its classical analog spreads out as $\sigma
(t)\sim t^{1/2}$. This property, as well as quantum parallelism and quantum entanglement, could be used to increase the efficiency of quantum algorithms [@Shenvi; @Childs; @et]. Recently we have been investigating [@alejo2010; @alejo2011; @armando2011] the asymptotic behavior of the QW on the line, focusing on the probability distribution of chirality independently of position. We showed that this distribution has a stationary long-time limit that depends on the initial conditions. This result is unexpected in the context of the unitary evolution of the QW because such a behavior is usually associated to a Markovian process. In this paper we further explore the behavior of the chirality distribution and define a thermodynamic equilibrium between the degrees of freedom of position and chirality. This equilibrium allows to introduce a temperature concept for this unitary closed system. We obtain a master equation with a time-dependent population rate, that describes the transient behavior of the reduced density operator of the QW towards thermodynamic equilibrium. The QW’s reduced density operator shows an surprising behavior: Its behavior looks diffusive but however the global evolution of the system is unitary.
The paper is organized as follows. In the next section the standard QW model is developed, in the third section the entanglement temperature is defined. Then in the fourth and fifth sections the entanglement temperature is obtained for localized initial conditions and for distributed initial conditions respectively. In the six section the transient behavior towards thermal equilibrium is studied, and finally in the last section some conclusions are drafted.
QW on the line
==============
The composite Hilbert space of the QW is the tensor product $\mathcal{H}_{s}\otimes \mathcal{H}_{c}$ where $\mathcal{H}_{s}$ is the Hilbert space associated to the motion on the line and $\mathcal{H}_{c}$ is the chirality (or coin) Hilbert space. In this composite space the walker moves, at discrete time steps $t\in
\mathbb{N}$, along a one-dimensional lattice of sites $k\in
\mathbb{Z}$. The direction of motion depends on the state of the chirality, with the eigenstates $R$ and $L$. The wave vector can be expressed as the spinor $$|\Psi (t)\rangle =\sum\limits_{k=-\infty }^{\infty }\left[
\begin{array}{c}
a_{k}(t) \\
b_{k}(t)\end{array}\right] |k\rangle , \label{spinor}$$where the upper (lower) component is associated to the left (right) chirality.
Then $P_{kL}(t)=\left\vert a_{k}(t)\right\vert ^{2}$ and $P_{kR}(t)=\left\vert b_{k}(t)\right\vert ^{2}$ denote the probability of finding the walker at $\left( k,t\right) $ and the coin in state $R$ and $L$, respectively. The probability of finding the walker at $\left( k,t\right) $ is $$P(k,t)=\left\langle \Psi _{k,t}\right. \left\vert \Psi _{k,t}\right\rangle
=\left\vert a_{k}(t)\right\vert ^{2}+\left\vert b_{k}(t)\right\vert ^{2},
\label{prob}$$and $\sum_{k}P(k,t)=1$.
The QW is ruled by a unitary map whose standard form is [@Romanelli09; @Alejo2; @Alejo1; @Alejo4] $$\begin{aligned}
a_{k}(t+1)& =a_{k+1}(t)\,\cos \theta \,+b_{k+1}(t)\,\sin \theta ,\, \notag
\\
b_{k}(t+1)& =a_{k-1}(t)\,\sin \theta \,-b_{k-1}(t)\,\cos \theta .
\label{mapa}\end{aligned}$$where $\theta \in \left[ 0,\pi /2\right] $ is a parameter defining the bias of the coin toss ($\theta =\frac{\pi }{4}$ for an unbiased or Hadamard coin). The global left and right chirality probabilities is defined as $$\begin{aligned}
P_{L}(t)& \equiv \sum_{k=-\infty }^{\infty }P_{kL}(t)=\sum_{k=-\infty
}^{\infty }\left\vert a_{k}(t)\right\vert ^{2},\, \notag \\
P_{R}(t)& \equiv \sum_{k=-\infty }^{\infty }P_{kR}(t)=\sum_{k=-\infty
}^{\infty }\left\vert b_{k}(t)\right\vert ^{2}, \label{chirality}\end{aligned}$$with $P_{R}(t)+P_{L}(t)=1$. The global chirality distribution (GCD) is defined as the distribution formed by the couple $\left[
\begin{array}{c}
P_{L}(t) \\
P_{R}(t)\end{array}\right] $. It is shown in Ref. [@alejo2010] that the GCD satisfies the following map $$\begin{aligned}
{\left[
\begin{array}{c}
P_{L}(t+1) \\
P_{R}(t+1)\end{array}\right] }& ={\left(
\begin{array}{cc}
\cos ^{2}\theta & \sin ^{2}\theta \\
\sin ^{2}\theta & \cos ^{2}\theta
\end{array}\right) }\left[
\begin{array}{c}
P_{L}(t) \\
P_{R}(t)\end{array}\right] \notag \\
& +\mathrm{Re}\left[ Q(t)\right] \sin {2}\theta \left[
\begin{array}{c}
1 \\
-1\end{array}\right] , \label{master}\end{aligned}$$where $$Q(t)\equiv \sum_{k=-\infty }^{\infty }a_{k}(t)b_{k}^{\ast }(t). \label{qdet}$$The two dimensional matrix in Eq.(\[master\]) can be interpreted as a transition probability matrix for a classical two dimensional random walk as it satisfies the necessary requirements, namely, all its elements are positive and the sum over the elements of any column or row is equal to one. On the other hand, it is clear that $Q(t)$ accounts for the interferences. When $Q(t)$ vanishes the behavior of the GCD can be described as a classical Markovian process. However in the generic case $Q(t)$ together with $P_{L}(t)$ and $P_{R}(t)$ are time depend functions that have long-time limiting values [@alejo2010] which are determined by the initial conditions of Eq.(\[mapa\]). Eq.(\[master\]) can be solved in this limit. We define $$\begin{aligned}
\Pi _{L}& \equiv
\begin{array}{c}
\lim \text{ }P_{L}(t) \\
t\rightarrow \infty~~~~
\end{array},\, \notag \\
\Pi _{R}& \equiv
\begin{array}{c}
\lim \text{ }P_{R}(t) \\
t\rightarrow \infty~~~~
\end{array},\, \notag \\
Q_{0}& \equiv
\begin{array}{c}
\lim \text{ }Q(t) \\
t\rightarrow \infty~~
\end{array},\, \label{asym}\end{aligned}$$and then we obtain the asymptotic stationary solution for the GCD as $${\left[
\begin{array}{c}
\Pi _{L} \\
\Pi _{R}\end{array}\right] }=\frac{1}{2}\left[
\begin{array}{c}
1+2\mathrm{Re}(Q_{0})/\tan \theta \\
1-2\mathrm{Re}(Q_{0})/\tan \theta
\end{array}\right] . \label{estacio}$$ This interesting result for the QW shows that the long-time probability to find the system with left or right chirality has a limit. Therefore, although the dynamical evolution of the QW is unitary, the evolution of its GCD has an asymptotic limit characteristic of a diffusive behavior. This situation is further surprising if we compare our case with the case of the QW on finite graphs [@Aharonov] where it is shown that there is no converge to any stationary distribution.
Entanglement and temperature
============================
The concept of entanglement is an important element in the development of quantum communication, quantum cryptography and quantum computation. In this context several authors have investigated the relation between asymptotic entanglement and the initial conditions of the QW [Carneiro,abal,Annabestani,Omar,Pathak,Petulante,Venegas,Endrejat,Ellinas1,Ellinas2,Maloyer]{}. Other authors [@Venegas1; @Chandrashekar] have proposed to use the QW as a tool for quantum algorithm development and as an entanglement generator potentially useful to test quantum hardware.
The unitary evolution of the QW generates entanglement between the coin and position degrees of freedom. To characterize this entanglement we start with the von Neumann entropy which is quantum analogue of the Gibbs entropy $$S_{N}(\rho )=-\mathrm{tr}(\rho \log \rho ). \label{uno}$$where $\rho =|\Psi (t)\rangle \langle \Psi (t)|$ is the density matrix of the quantum system. Due to the unitary dynamics of the QW the system remains in a pure state and this entropy vanishes. However for these pure states the entanglement between the chirality and the position can be quantified by the associated von Neumann entropy for the reduced density operator [Carneiro,abal]{} that defines the entropy of entanglement $$S(\rho )=-\mathrm{tr}(\rho _{c}\log \rho _{c}), \label{dos}$$ where $$\rho _{c}=\mathrm{tr}(\rho ), \label{dos}$$ and the partial trace is taken over the positions. Using the wave function Eq.(\[spinor\]) and its normalization properties we obtain the reduced density operator [@Carneiro] $$\rho_{c} =\left(
\begin{array}{cc}
P_{L}(t) & Q(t) \\
Q(t)^{\ast } & P_{R}(t)\end{array}\right) . \label{rho}$$ The density operator $\rho_{c}$ has the eigenvalues $$\lambda _{\pm}=\frac{1}{2}\left[ 1\pm \sqrt{1-4\left(
P_L(t)\,P_R(t)-\left\vert Q(t)\right\vert ^{2}\right) }\right]. \label{lam}$$ Then the entanglement entropy Eq.(\[dos\]) is expressed through these two eigenvalues as $$S(\rho)=-\lambda_{+}\log \lambda_{+}-\lambda_{-}\log \lambda_{-}.
\label{ttres}$$In the asymptotic regime $t\rightarrow\infty$ the eigenvalues go to a stationary limit, $\lambda _{\pm}\rightarrow\Lambda _{\pm}$ and $$\Lambda _{\pm}=\frac{1}{2}\left[ 1\pm \sqrt{1-4\left(\Pi_L\,\Pi_R-\left\vert
Q_0\right\vert ^{2}\right) }\right]. \label{lam0}$$ Further from Eq.(\[estacio\]) follows the relation $$\Pi_L\,\Pi_R= \frac{1}{4}-{\left(\frac{\mathrm{Re}(Q_{0})}{\tan \theta}\right)}^{2}, \label{piqu}$$ which is substituted in Eq.(\[lam0\]), and then the asymptotic eigenvalues are expressed as $$\Lambda _{\pm}=\frac{1}{2}\pm \sqrt{\chi}, \label{lam1}$$ with $$\chi\equiv \left\vert Q_0\right\vert ^{2}+{\left({\mathrm{Re}(Q_{0})}/{\tan
\theta}\right)}^{2}. \label{defi}$$ Note that the values of the interference term $Q_{0}$ are constrained to satisfy the condition $$0<\Lambda _{+}\Lambda _{-}<1, \label{condition0}$$ and then $$0<\chi<\frac{1}{4}. \label{condition}$$
![Dimensionless thermodynamic function normalized by $\log(2)$ as a function of the dimensionless parameter $\protect\chi$. From the top to the bottom they are: in thick line $\protect\beta\protect\epsilon$, in dashed line the entropy $S_0$, in thin line the energy $\protect\beta U$ and in dashed-dot line the Helmholtz free energy $\protect\beta A$.[]{data-label="f1"}](fig1.eps)
The entanglement entropy has an asymptotic limit too $$S_0=-\Lambda _{+}\log \Lambda _{+}-\Lambda _{-}\log \Lambda _{-},
\label{s0}$$ that only depends on the initial conditions through the interference term $Q_{0}$. Therefore we are led to consider that after some transient time the QW achieves a thermodynamic equilibrium between the position and chirality degrees of freedom. In order to make a fuller description of this equilibrium it is necessary to connect the eigenvalues of $\rho_c$ with its associated Hamiltonian operator $H_c$. To obtain this connection we shall use the quantum Brownian motion model of Ref.[@Kubo]. We considered the system associated with the chirality degrees of freedom and characterized by the density matrix $\rho_c$ in thermal contact (entanglement) with the bath system associated with the position degrees of freedom, the lattice. In this context $\rho_c$ satisfies the equation $$\frac{\partial\rho_{c}}{\partial
t}=\frac{1}{i\hbar}[H_c,\rho_{c}]+\Gamma \rho_{c}, \label{scho}$$ where $[H_c,\rho_{c}]$ is the commutator and $\Gamma \rho_{c}$ represents the Brownian motion of $\rho_c$ induced by the noise (fluctuating forces) exerted on $\rho_c$ by the lattice (position degrees of freedom). In the equilibrium (stationary) situation we must have ${\partial\rho_{c}}/{\partial t}= 0$ and $\Gamma
\rho_{c}=0$ [@Kubo], that is $$[H_c,\rho_{c}]=0. \label{scho2}$$ Therefore, in the asymptotic regime, the density operator $\rho_{c}$ must be an explicit function of the Hamiltonian operator which must be time independent. Now we call $\{\Phi_{+},\Phi_{-}\}$ the eigenfunctions of the density matrix, and then the operators $H_{c}$ and $\rho_{c}$ are both diagonal in this basis. Therefore, the eigenvalues $\Lambda _{+}$ and $\Lambda _{-}$ depend of the corresponding eigenvalues of $H_c$. We take these eigenvalues to be $\{-\epsilon,\epsilon\}$ without any loss of generality, they represent the possible values of the entanglement energy. This interpretation agrees with the fact that $\Lambda _{+}$ and $\Lambda _{-}$ are the probabilities that the system is in the eigenstate $\Phi_{+}$ or $\Phi_{-}$. The precise dependence between $\Lambda _{\pm}$ and $\pm \epsilon$ is determined by the type of ensemble we construct. The main proposal of this paper is that this equilibrium corresponds to a quantum canonical ensemble. Therefore we propose that $$\Lambda _{\pm}\equiv\frac{e^{\pm\beta\epsilon}}{e^{\beta\epsilon}+e^{-\beta\epsilon}}, \label{lam2}$$ which defines the entanglement temperature $T\equiv1/\beta$. Then the probability that the state chosen at random from the ensemble $\{\Phi_{+}
$, $\Phi_{-}$}, possesses an energy $\epsilon$ is determined by the Boltzmann factor $e^{-\beta\epsilon}$. Let us call $\widetilde{\rho_{c}}$ the diagonal expression of the density operator $\rho_{c}$, then $$\widetilde{\rho_{c}}=\left(
\begin{array}{cc}
\Lambda _{+} & 0 \\
0 & \Lambda _{-}\end{array}\right) =\frac{1}{e^{\beta\epsilon}+e^{-\beta\epsilon}}\left(
\begin{array}{cc}
e^{\beta\epsilon} & 0 \\
0 & e^{-\beta\epsilon}\end{array}\right) . \label{rhoni}$$ This operator is formally the same density operator that corresponds to an electron with possesses an intrinsic spin and a magnetic moment in a external magnetic field [@pathria].
Starting from Eq.(\[rhoni\]) it is possible to build the thermodynamics for the QW entanglement. The partition function of the system is then given by $$\mathcal{Z}=e^{\beta\epsilon}+e^{-\beta\epsilon}=2\cosh({\beta\epsilon}).
\label{betaA}$$ Accordingly, and also using Eqs.(\[lam0\],\[lam2\]), the temperature is given by $$T=2\epsilon/\ln\left(\frac{1+2\sqrt{\chi}}{1-2\sqrt{\chi}}\right),
\label{betae}$$ the Helmholtz free energy by $$A=-\frac{1}{\beta}\ln[2\cosh({\beta\epsilon})]=\frac{T}{2}\ln\left(\frac{1}{4}-\chi\right), \label{betaA}$$ the internal energy by $$U=-\epsilon\tanh(\beta\epsilon)=-2\epsilon\sqrt{\chi}, \label{betaU}$$ and finally the entropy by $$S_0=\beta U-\beta A, \label{s02}$$ where this last thermodynamic definition for the entropy of course agrees with the previous Shannon expression in Eq.(\[s0\]). To finished this section in Fig. \[f1\] we present the dependence of these thermodynamic magnitudes with the interference parameter $\chi$.
Localized initial conditions
============================
![Isothermal curves as functions of the dimensionless angles $\protect\gamma$ and $\protect\varphi$. Due to the rotation symmetry in the angle $\varphi$, only four zones are distinguished: two “cold” and two “hot”. The “hot” zones (orange on line) have six isotherms; from inside to outside their temperatures are: $T/T_{0}=6.5,~3.2,~2.2,~1.6,~1.3$ and $1.1$. The “cold” zones (blue on line) have five isotherms; from outside to inside their temperatures are: $T/T_{0}=0.9,~0.8,~0.7,~0.68$ and $0.66$. The straight (green) lines corresponded to $T/T_{0}=1$, see Eq.(\[varphygamma\]).[]{data-label="f2"}](fig2.eps)
![The isothermal curves of Fig. \[f2\] shown on the Bloch sphere. The QW initial chirality determines the entanglement temperature, see Eqs.(\[betae\],\[psi0\],\[chi2\]).[]{data-label="f3"}](fig3.eps)
As seen in the previous section the thermodynamics of entanglement only depends on the interference term $Q_{0}$ which in turn only depends on the initial conditions, as shown in [@alejo2010].
In order to investigate this dependence on the initial conditions of the system we consider first the localized case. The initial state of the walker is assumed to be sharply localized at the origin with arbitrary chirality, thus $$|\Psi (0)\rangle =\left(
\begin{array}{c}
\cos ({\gamma}/{2}) \\
\exp i\varphi \text{ }\sin ({\gamma}/{2})\end{array}\right) |0\rangle , \label{psi0}$$where $\gamma \in \left[ 0,\pi \right]$ and $\varphi \in \left[ 0,2\pi \right] $ define a point on the unit three-dimensional Bloch sphere. The expression of $Q_{0}$ was obtained in Ref. [@abal], fixing the bias of the coin toss $\theta =\pi /4$, following the method developed by Nayak and Vishwanath [@nayak]$$Q_{0}=\frac{1}{2}(1-\frac{1}{\sqrt{2}})\left[ \cos \gamma \text{ }+\sin
\gamma \text{ }(\cos \varphi +i\sqrt{2}\sin \varphi )\right] .
\label{q0entengl}$$ Using this result in Eq.(\[defi\]) the dependence of $\chi$ with the initial conditions is given by $$\chi=\chi_{0}\left(1+ \cos \varphi \sin 2\gamma \right) , \label{chi2}$$ where $\chi_{0}={3}/{4}-1/\sqrt{2}$. It is useful to define a characteristic temperature (in units of $\epsilon$) $$T_{0}=2/ \left[\ln\left(\frac{1+2\sqrt{\chi_{0}}}{1-2\sqrt{\chi_{0}}}\right)\right], \label{t0}$$ in order to express any other temperature as a proportionality with $T_0=1/\beta_0$. Then from Eq.(\[betae\]) we obtain an expression for $\beta$ as a function of the angles $\gamma$ and $\varphi$ $$\cos\varphi \sin
2\gamma=\left(\frac{\tanh\beta}{\tanh\beta_0}\right)^2-1.
\label{varphygamma}$$ Figures \[f2\] and \[f3\] show the level curves (isotherms) for the entanglement temperature as a function of the QW initial position. In Fig. \[f2\] the initial position is defined through the angles $\gamma$ and $\varphi$ and in Fig. \[f3\] it is defined through the position on the Bloch sphere (see Eq.(\[psi0\])). Both figures show four regions, two of them corresponding to temperatures $T>T_0$ (orange color on line) and the other two to temperatures $T<T_0$ (blue color on line). The longest isotherms (green color on line) correspond to the temperature $T=T_{0}$ and their initial conditions are $\gamma=0,~\pi/2,~\pi$ and $\varphi=\pi/2,~3\pi/2$.
Distributed initial conditions
==============================
![Isothermal curves as functions of the dimensionless angles $\protect\gamma$ (initial conditions) and $\protect\theta$ (bias of the coin). Four curves (with different colors on line) are presented; each curve has two branches placed symmetrically on both sides of $\protect\gamma=\protect\pi/2$, where $T=\infty$. The values of $T$ are given by Eqs.(\[betae2\]). From left to right the values of $T$ are, in units of $\protect\epsilon$: 0.5 (purple), 1. (blue), 2. (green) and 5. (orange). The diagram has two discontinuities in $\theta=\protect\pi/2$ and in $\protect\gamma=\protect\pi/2$ (see Eqs.(\[betae2\],\[re2\])).[]{data-label="f4"}](fig4.eps)
In previous works [@Eugenio; @alejo2010] we have studied the QW with extended initial conditions. Now the entanglement temperature is studied in such a case. The following extended Gaussian distributions is proposed: $$a_{k}^{0}\equiv{\left[ \frac{1}{\sigma_0\sqrt{2\pi }}\exp \left( -\frac{k^{2}}{2\sigma_0^{2}}\right) \right]}^{\frac{1}{2}}\cos({\gamma}/{2}) \text{,}
\label{aes}$$$$b_{k}^{0}\equiv e^{i \varphi }{\left[ \frac{1}{\sigma_0\sqrt{2\pi }}\exp
\left( -\frac{k^{2}}{2\sigma_0^{2}}\right) \right]}^{\frac{1}{2}}\, \sin({\gamma}/{2})\text{,} \label{bes}$$where $\sigma_0$ is the initial standard deviation, $\gamma \in \left[ 0,\pi \right]$ determines the initial proportion of the left and right chirality and $\varphi\in \left[ 0,2\pi \right] $ is a global phase. Using these initial conditions, Eqs.(\[aes\], \[bes\]), the asymptotic value of $Q(t)$, see Eqs.(\[qdet\],\[asym\]), was obtained [@alejo2010] as $$Q_0=\frac{1}{2}\cos\gamma\,\tan\theta\,, \label{q1}$$ with the restrictions $$\sigma_0\gg1, \label{sig1}$$ and $$\cos\varphi=\frac{\tan\theta}{\tan\gamma}. \label{phi2}$$ Replacing Eq.(\[q1\]) in Eq.(\[defi\]) we obtain $$\chi= \left( \frac{\cos\gamma}{2\cos\theta}\right)^{2}, \label{chides}$$ and then using Eq.(\[betae\]) we have $$\beta\epsilon=\frac{1}{2}\ln\left(\frac{\left|\cos\theta\right|
+\left|\cos\gamma\right|}{\left|\cos\theta\right|-\left|\cos\gamma\right|}\right), \label{betae2}$$ where taking into account Eqs.(\[condition\],\[chides\]) the initial condition satisfies the constraint $${\left|\cos\gamma\right|}<{\left|\cos\theta\right|}. \label{re2}$$
The functions $Q_0$, $\chi$ and $\beta$ vanish for $\gamma=\pi/2$ ( see Eqs.(\[q1\],\[chides\],\[betae2\])) and simultaneously the entanglement entropy Eq.(\[ttres\]) has its maximum value $S_0
= 1$. This maximum value is achieved when the entanglement temperature is $T=\infty$. Under these conditions the system behaves as a classical Markov process [@alejo2010]. On the other hand, the initial conditions $\gamma$ and $\varphi$ are not independent (see Eq.(\[phi2\])) and for each value of $\gamma$ there is only one value of $T$, then for fixed $\theta$, it is not possible to have isotherms as functions of $\gamma$ and $\varphi$. Instead the entanglement temperature depends on $\theta$ and $\gamma$ from Eq.(\[betae2\]), *i.e-* the choice of the bias of the coin toss $\theta$ or of the initial proportion of the chirality $\gamma$ could lead to the same entanglement temperature. Fig. \[f4\] shows the isotherms as functions of $\gamma$ and $\theta$.
Transient behavior
==================
![Envelope of the probability $\protect\lambda _{+}-\Lambda
_{+}$ as a function of the dimensionless time $t$ for two different initial conditions. Each initial condition is established by the couple $(\varphi, \gamma)$ in Eq. (\[psi0\]). Their values are $(\protect\pi /8, \protect\pi /4)$ for the full black line and $(\protect\pi/4, \protect\pi /3) $ for the dashed red line. In both cases, the temperature is $T=0.79~T_{0}$.[]{data-label="f5"}](fig5.eps)
In the QW a stationary entanglement is established between the chirality and position degrees of freedom after a transient time. This fact allowed us to introduce the concept of entanglement temperature. The transient behavior of the system is studied using the original map Eq.(\[mapa\]) in a numerical code with initial conditions given by Eq.(\[psi0\]). These numerical calculations are summarized in Figs. \[f5\] and \[f6\]. Fig. \[f5\] presents the difference between the transient ($\lambda _{+}$) and the stationary ($\Lambda _{+}$) eigenvalues of the density matrix as a function of time (see Eqs.(\[lam\],\[lam0\])). The figure only presents the envelope of the curves because the real eigenvalues dynamics is very intricate; it presents quick oscillations with high density of paths where it is only possible distinguish its global contour. However, the average evolution of the system is determined by the envelope dynamics. Each envelope has two branches placed symmetrically on both sides of $\lambda _{+}-\Lambda _{+}=0$. Two pair of curves are presented in full and dashed lines with color black and red respectively. In both cases, the envelopes decay for $t\rightarrow \infty $ as a power law $1/t^{c}$ with $c=0.490$ for the dashed red line and $c=0.486$ for the full black line. The envelopes of $\lambda _{\pm}$ will be called $\widetilde{\lambda}_{\pm}$ respectively.
![The power law exponent as a function of the dimensionless angle $\protect\gamma $. The initial conditions $(\varphi, \gamma)$ correspond to the isotherms $T=1.1~T_{0}$.[]{data-label="f6"}](fig6.eps)
It was numerically verified for several initial conditions given by Eq.(\[psi0\]) that the transient behavior of $\widetilde{\lambda}_{\pm}-{\Lambda}_{\pm}$ can be adjusted by a power law function of time. Fig. \[f6\] shows the power law exponent $c$ as a function of the initial condition $\gamma $ for the same temperature. Remember that, for $T$ and $\gamma $ given, $\varphi $ is determined by Eq.(\[varphygamma\]). Therefore, the exponent $c$ has a dependence with the initial conditions, however this dependence is not determined by the asymptotic temperature value.
Additionally, the transient behavior of $\widetilde{\lambda}_{\pm}-{\Lambda}_{\pm}$ was numerically studied, using initial conditions given by Eqs.(\[aes\],\[bes\]) with $\sigma\gg1$. In these cases, the system showed a negligible transient dynamics, in concordance with the calculation developed in Ref. [@alejo2010]. Therefore, for these initial conditions, the reduced density matrix is essentially always in thermodynamic equilibrium.
With the aim to understand the transient behavior of the system we develop an analytic theory implementing a parallelism between the reduced density operator Eq.(\[rhoni\]) and the density operator of an electron in a external magnetic field. With this picture on mind we propose the following master equation for the probabilities ${\lambda} _{+}$ and ${\lambda}_{-}$ $$\begin{aligned}
\frac{d{\lambda}_{-}}{dt}& ={\lambda}_{+}\,w_{+-}\,-{\lambda}_{-}\,w_{-+},\, \notag \\
\frac{d{\lambda}_{+}}{dt}& ={\lambda}_{-}\,w_{-+}\,-{\lambda}_{+}\,w_{+-}, \label{masterfin}\end{aligned}$$ where $w_{+-}$ and $w_{-+}$ are transition probabilities per unit of time, that can be understood as population rates. $w_{+-}$ corresponds to the transition ${\lambda}_{+}\rightarrowtail{\lambda}_{-}$ and $w_{-+}$ corresponds to the transition ${\lambda}_{-}\rightarrowtail{\lambda}_{+}$. These rates are time-dependent functions, and their behaviors are known in the limit $t\rightarrow\infty$ when ${d{\lambda}_{\pm}}/{dt}\longrightarrow 0$. In this limit, the stationary solution of Eq.(\[masterfin\]) must be the couple $\Lambda _{-}$ and $\Lambda _{+}$, given by Eq.(\[lam0\]). Then the asymptotic values of the population rates satisfy $$\frac{w_{b}}{w_{a}}= \frac{\Lambda _{-}}{\Lambda _{+}}, \label{rate}$$ where $w_{a}$ and $w_{b}$ are defined by $$\begin{aligned}
w_{a}& \equiv
\begin{array}{c}
\lim \text{ }w_{-+} \\
t\rightarrow \infty~~~
\end{array},\, \label{rates} \\
w_{b}& \equiv
\begin{array}{c}
\lim \text{ }w_{+-} \\
t\rightarrow \infty~~~
\end{array}.\, \label{rates1}\end{aligned}$$Eq.(\[rate\]) expresses a condition of detailed balance which says that the rate of occurrence for any transition equals the rate for the inverse transition. Using our knowledge about the transient and asymptotic behaviors, the following population rates are proposed $$\begin{aligned}
w_{+-}=w_{b}+\xi(t),\, \label{rate2}\\
w_{-+}=w_{a}-\xi(t),\, \label{rate3}\end{aligned}$$ where $$\xi(t)=\frac{K}{t^{c}}[\omega\sin(\omega
t+\delta)+(\frac{c}{t}-w_{a}-w_{b})\cos(\omega t+\delta)],
\label{xi}$$ with $c>0$, $K$, $\omega$ and $\delta$ constants. The general solution of Eq.(\[masterfin\]) with these population rates is $$\begin{aligned}
{\lambda}_{+}=\Lambda _{+}+\frac{K}{t^{c}}\cos(\omega t+\delta)+d\,
e^{-(w_{a}+w_{b})t},\, \label{ratefin}\\
{\lambda}_{-}=\Lambda_{-}-\frac{K}{t^{c}}\cos(\omega t+\delta)-d\,
e^{-(w_{a}+w_{b})t},\, \label{ratefin1}\end{aligned}$$ where $d$ is an additional constant. Note that $e^{-(w_{a}+w_{b})t}\rightarrow0$ faster than ${1/t^{c}}$ for $t\rightarrow\infty$, then Eqs.(\[ratefin\],\[ratefin1\]) verify the asymptotic behavior obtained numerically. All the constants $K$, $c$, $\omega$, $\delta$ and $d$ depend on the initial conditions, however their values should be compatible with the positive character of the functions $\lambda_{+}$, $\lambda_{-}$, $w_{+-}$ and $w_{-+}$. In particular, to describe correctly the numerical results, $K$ takes a finite value for localized initial conditions and $K$ takes negligible value for distributed initial conditions. In summary, the Brownian motion equation for our reduced density matrix Eq.(\[scho\]), takes the form of a master equation Eq.(\[masterfin\]).
Finally it is important to point out that the asymptotic behavior found here is similar to the behavior of the simple cellular automaton known as *sandpile* [@btw]. Such behaviors are characteristic of extended dynamical systems with spatial degrees of freedom. They naturally evolve to self-organized states with correlations that decay with a power law.
Conclusion
==========
The unitary evolution of the QW in a composite Hilbert space is studied. In particular the entanglement between chirality and position degrees of freedom is investigated. After a transient time the system establishes a stationary entanglement between the coin and the position that allows to develop a thermodynamic theory. The asymptotic reduced density operator is used to introduce the entanglement thermodynamic functions in the canonical equilibrium. These thermodynamic functions characterize the asymptotic entanglement. It is shown that the QW initial condition determines the system’s temperature, as well as other thermodynamic function. A map for the isotherms is analytically built for arbitrary localized initial conditions. Additionally, it is shown that choosing appropriately the bias of the coin-toss, it is possible to obtain a predetermined entanglement temperature.
The transient dynamics of the reduced density operator outside the thermodynamic equilibrium is also studied. We show numerically that this transient behavior can be adjusted with a power law, whose exponent depends on the initial conditions. We built a master equation to describe this behavior where the population rates have a time dependence. The accuracy of the master equation solution is numerically verified and it is shown that the reduced density has a cellular automaton behavior.
The behavior of the reduced density operator looks diffusive but it has a dependence with the initial conditions, the global evolution of the system is unitary. Then, if an observer only had information related with the chirality degrees of freedom, it would be very difficult for him to recognize the unitary character of the quantum evolution. In general, from this simple model we can conclude that if the quantum system dynamics is developed in a composite Hilbert space, then the behavior of operators that only belong to one sub-space could camouflage the unitary character of the global evolution.
I acknowledge stimulating discussions with Víctor Micenmacher, Guzmán Hernández, Raúl Donangelo and Armando Pérez and the support from PEDECIBA and ANII.
[99]{} Y. Aharonov, L. Davidovich, and N. Zagury, Phys. Rev. A **48**, 1687 (1993); D. A. Meyer, J. Stat. Phys. **85**, 551 (1996); J. Watrous, Proc. STOC’01 (ACM Press, New York, 2001), p.60; A. Ambainis, Int. J. Quant. Inf. **1**, 507 (2003); J. Kempe, Contemp. Phys. **44**, 307 (2003); V. Kendon, Math. Struct. Comp. Sci. **17**, 1169 (2006); V. Kendon, Phil. Trans. R. Soc. A **364**, 3407 (2006); N. Konno, *Quantum Walks*, in Quantum Potential Theory, Lect. Notes Math., Vol. 1954, edited by U. Franz and M. Schürmann (Springer, 2008); A. Wojcik, T. Luczak,P. Kurzynski, A. Grudka, M. Bednarska, Phys. Rev. Lett. **93**, 180601 (2004); M.C. Bañuls, C. Navarrete, A. Pérez, E. Roldán, and J.C. Soriano, Phys. Rev. A **73**, 062304 (2006), A. Romanelli, A. Auyuanet, R. Siri, G. Abal, and R. Donangelo, Physica A **352**, 409 (2005).
A.M. Childs, Phys. Rev. Lett. **102**, 180501 (2009).
A. Romanelli, Phys. Rev. A, **80**, 042332 (2009).
N. Linden, and J. Sharam, Phys. Rev. A, **80**, 052327 (2009).
N. Shenvi, J. Kempe, K. BirgittaWhaley, Phys. Rev. A **67**, 052307 (2003).
A. Childs, E. Deotto, E. Farhi, S. Gutmann, and D. A. Spielman, in Proc. 35th ACM Symposium on Theory of Computing (STOC 2003), pp. 59–68, 2003, arXiv preprint quant-ph/0209131.
A. Romanelli, Phys. Rev. A **81**, 062349 (2010).
A. Romanelli, Physica A, **390**, 1209 (2011).
A. Pérez, and A. Romanelli, in preparation (2011).
A. Romanelli, Phys. Rev. A **80**, 042332 (2009).
A. Romanelli, A.C. Sicardi Schifino, R. Siri, G. Abal, A. Auyuanet, and R. Donangelo, Physica A, **338**, 395 (2004).
A. Romanelli, A.C. Sicardi Schifino, G. Abal, R. Siri, and R. Donangelo, Phys. Lett. A **313**, 325 (2003).
A. Romanelli, R. Siri, G. Abal, A. Auyuanet, and R. Donangelo, Physica A, **347**, 395 (2005).
D. Aharonov, A. Ambainis, J. Kempe, and U. Vazirani, *Proc. of the 33rd Annual ACM STOC*, ACM, NY, 50 (2001). arXiv preprint quant-ph/0001.2090v2 (2002).
I. Carneiro, M. Loo, X. Xu, M. Girerd, V. M. Kendon, and P. L. Knight, New J. Phys. **7**, 56 (2005).
G. Abal, R. Siri, A. Romanelli, and R. Donangelo, Phys. Rev. A **73**, 042302, 069905(E) (2006).
M. Annabestani, M. R. Abolhasani and, G. Abal, J. Phys. A: Math. Theor. **43**, 075301 (2010).
Y. Omar, N. Paunkovic, L. Sheridan, and S. Bose, Phys. Rev. A, **74**, 042304 (2006)
P. K. Pathak, and G. S. Agarwal, Phys. Rev. A, **75**, 032351 (2007)
C. Liu, and N. Petulante, Phys. Rev. A **79**, 032312 (2009).
S. E. Venegas-Andraca, J.L. Ball, K. Burnett, and S. Bose, New J. Phys., **7**, 221 (2005).
J. Endrejat, H. Büttner, J. Phys. A: Math. Gen. **38**, 9289 (2005).
A.J. Bracken, D. Ellinas, and I. Tsohantjis, J. Phys. A: Math. Gen. **37**, L91 (2004).
D. Ellinas, and A.J. Bracken, Phys. Rev. A **78**, 052106 (2008).
O. Maloyer, and V. Kendon, New J. Phys., **9**, 87 (2007).
S. E. Venegas-Andraca , and S. Bose, arXiv preprint quant-ph/0901.3946v1 (2009).
S. Goyal, and C. M. Chandrashekar, arXiv preprint quant-ph/0901.0671 (2009).
R. Kubo, M. Toda, and N. Hashitsume *Statistical Physics II, Nonequilibrium Statistical Mechanics*, Springer-Verlag, Berlin Heidelberg New York Tokyo; ISBN 3 540 11461 0, (1985).
10 A. Nayak and A. Vishwanath, e-print quant-ph/0010117
G.J.de Valcárcel, E. Roldán and A. Romanelli, New Journal of Physics **12**, 123022 (2010).
R. K. Pathria *Statistical Mechanics* , secon edition,ISBN 0 7506 2469 8, (1997).
H. J. Jensen, *Self-Organized Criticality*, first edition, ISBN 0 521 48371 9 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper investigates the design and application of write-once memory (WOM) codes for flash memory storage. Using ideas from Merkx [@m84], we present a construction of WOM codes based on finite Euclidean geometries over $\mathbb{F}_2$. This construction yields WOM codes with new parameters and provides insight into the criterion that incidence structures should satisfy to give rise to good codes. We also analyze methods of adapting binary WOM codes for use on multilevel flash cells. In particular, we give two strategies based on different rewrite objectives. A brief discussion of the average-write performance of these strategies, as well as concatenation methods for WOM codes is also provided.'
author:
- |
Kathryn Haymaker and Christine A. Kelley\
Department of Mathematics\
University of Nebraska-Lincoln\
Lincoln, NE 68588\
title: 'Geometric WOM codes and coding strategies for multilevel flash memories [^1]'
---
[**Keywords:**]{} write once memory, flash memory, finite geometries, concatenated codes.
Introduction {#intro}
============
Non-volatile flash memories are becoming increasingly popular due to their potential for high throughputs and low power consumption. Flash memory storage is a technology that is based on organizing the memory into blocks of cells in which each cell can be charged up to one of $q$ levels. While increasing the charge of a cell is easy, decreasing the charge is costly since the entire block containing the cell must be erased and rewritten. Such an operation involves reprogramming roughly $10^5$ cells. Moreover, frequent block erasures also reduce the lifetime of the flash device. It is therefore desirable to be able to write as many times as possible before having to erase a block [@rs82; @fs84; @kysvw10; @csbb07]. Like any storage device, the flash cells are also prone to errors due to charge leakage or the writing process. Thus, the coding design goals for flash memories include maximizing the number of writes between block erasures, correcting cell charge leakage errors, and correcting errors that occur during the writing process.
An information theoretic approach to writing on memories with defects was first considered by Kuznetsov and Tsybakov [@kt74], and later surveyed in [@kh94]. The write-once memory (WOM) model, introduced by Rivest and Shamir [@rs82], and other constrained memory models (WUM, WIM, WEM) can be considered as particular cases of the general defective channel [@kh94; @az89; @c88]. Due to the asymmetric costs associated with increasing and decreasing cell levels, the flash memory model can be viewed as a generalization of the WOM model. As a result, WOM codes have been proposed for flash cells having two levels (i.e., capable of storing one bit of information per cell) [@j07; @jb08; @yvsw08]. Error-correcting codes for the general defective channel and for WOM have also been considered, although addressing errors while incorporating rewriting capabilities is difficult, and many codes in the literature are optimized primarily for one of these goals [@csbb07; @jlw09; @jb09; @jb08; @zc91; @j07].
This paper is organized as follows. The rest of this section provides notation and background on codes for flash memories. In Section 2, after summarizing Merkx’s construction of WOM codes from finite projective geometries, we present a new construction of WOM codes using finite Euclidean geometries. In Section 3 we explore methods of adapting binary WOM codes for multilevel flash cells, and introduce two strategies that achieve this with respect to different goals. We also examine the average write analysis of these strategies for two specific WOM codes. Finally, we summarize ways to combine WOM codes with classical error correcting codes using concatenation in Section 4. We conclude the paper in Section 5 with some future directions.
Preliminaries
-------------
We now give some definitions and notation that will be used in this paper. A write-once memory (WOM) is a storage device over a binary alphabet where a zero can be increased to a one, but a one cannot be changed back to a zero. An information message is encoded and stored in a string of cells in the memory, referred to as a *cell state vector*[^2]. The cells in the cell state vector form the codeword and can be updated, or *rewritten*, to represent a different message. Only the most recently written message is retained.
A write-once memory code is composed of a set $V$ of information words, called *variable vectors*, and a set $S$ of cell state vectors with $S\subseteq \mathbb{F}_2^n$, corresponding to the codewords of the WOM code. Many different cell state vectors can represent the same information message. In addition, the WOM code is equipped with an encoding and decoding function. The encoding function takes as inputs both the current state of the memory and the new information message to be stored. Specifically, it maps the current cell state vector to an updated cell state vector that represents the new information message and is component-wise greater than or equal to the previous state. The decoding function maps the resulting cell state vector to the updated information message. The amount of information messages that can be encoded at each time step need not be the same, however, as the following notation conveys.
Let $\langle v_1, \ldots, v_t\rangle/n$ denote a $t$-write WOM code on $n$ cells, where $v_i$ is the number of messages that can be represented on the $i^{th}$ write. In the fixed information case, i.e., when $v_1=\cdots = v_t$, such a WOM code will be denoted by $\langle v\rangle^t/n$.
The *rate* of a WOM code is $$R=\frac{ \log_2(v_1\cdots v_t)}{n}.$$
When $q = 2$, the flash cell is called a single level cell (SLC) since the cell can only represent one nonzero value, and a multilevel cell (MLC) when $q > 2$ as it can store values $\{0,1,2,\ldots, q-1\}$. Note that an SLC can store one bit of information per cell whereas an MLC can store multiple bits of information per cell. Fiat and Shamir considered a generalized version of a WOM, in which the storage cells have more than two states with transitions given by a directed acyclic graph [@fs84]. The idea of extending to multilevel cells was further explored by Jiang in [@j07], in which he considered generalizing error-correcting WOM codes. Techniques for rewriting codes on $q$-ary cells include floating codes, which were introduced by Jiang, Bohossian, and Bruck [@jbb07], and more generally, trajectory codes, which are described in [@jlsb09]. Although these are similar objects, we will use the term [*flash codes*]{}, introduced in [@yvsw08], to refer to a rewriting code on multilevel cells.
When $q>2$, $\langle v\rangle^t_q/n$ will denote a $t$-write *flash code* for use on cells having $q$ levels, where the code has block length $n$ and $v$ messages can be represented at each write. The capacity of a flash memory is the maximum number of writes possible for $n$ number of cells, $v$ number of information messages to be represented in each write, and $q$ number of levels per cell.
Fu and Han Vinck [@fh99] showed that the maximum total number of information bits that can be stored per cell over $t$ writes is at most $$\log_2(1+(q-1)t).$$
The next example, from [@rs82], gives the canonical example of a WOM code.
The Rivest and Shamir WOM code is shown in Table \[r-s-WOM\] [@rs82]. It maps two information bits to three coded bits and is capable of tolerating two writes. Note that any of the four messages may be written at either write. The table is interpreted as follows: on the first write, the encoding function takes the current all-zero state and the new information message and maps it to the representation of that message in the ‘first write’ column. On the second write, the encoding function takes the current cell state and the new information message and outputs the cell state vector opposite the new message in the ‘second write’ column. For example, the message sequence $01 \rightarrow 11$ would be recorded as $100 \rightarrow 110$. If the new information message is the same as the information represented by the current cell state vector, the memory remains unchanged. Decoding is as follows: the cell state vector $(a_1, a_2, a_3)$ can be decoded as $((a_2+a_3) \mbox{ (mod 2) }, (a_1+a_3) \mbox{ (mod 2)})$.
\[r-s-WOM\]
$1^{st}$ write $2^{nd}$ write
---- ------------------- -----------------
00 000 111
01 100 011
10 010 101
11 001 110
: $\langle 4 \rangle^2/3$ WOM-code by Rivest and Shamir.
\[rivest-shamir-WOM\] $\hfill \Box$
Finite Geometry WOM-codes
=========================
In this section, we apply ideas from [@m84] to design WOM codes based on finite Euclidean geometries. We first provide some relevant definitions.
The *$m$-dimensional Euclidean geometry over $\mathbb{F}_2$*, denoted by $EG(m,2)$, is an incidence structure with $2^m$ points and $2^{(m-1)}(2^{m}-1)$ lines. The points in $EG(m,2)$ may be regarded as all $m$-tuples over $\mathbb{F}_2$, and each pair of points defines a line.
Note that the set of points in $EG(m,2)$ forms an $m$-dimensional vector space over $\mathbb{F}_2$. A $\mu$-flat in $EG(m,2)$ is a $\mu$-dimensional subspace of the finite geometry, defined next. We will use the term [*hyperplane*]{} to refer to a subspace of dimension $m-1$ in either $PG(m,2)$ or $EG(m,2)$.
Let $X$ be the set of points in $EG(m,2)$. A *$\mu$-flat* in $EG(m,2)$ passing through a point $a_0$ consists of points of the form $a_0+ \beta_1 a_1+\cdots + \beta_{\mu}a_{\mu}$, where $a_0, \ldots, a_{\mu}\in X$ are linearly independent and $\beta_1, \ldots, \beta_{\mu}\in \mathbb{F}_2$.
The number of $\mu$-flats in $EG(m,2)$ is $$2^{(m-\mu)} \prod_{i=1}^{\mu} \frac{2^{(m-i+1)}-1}{2^{(\mu-i+1)}-1}.$$ Moreover, each $\mu$-flat in $EG(m,2)$ is a coset of an $EG(\mu, 2)$, and thus contains $2^{\mu}$ points.
The *finite projective geometry of dimension $m$ over $\mathbb{F}_2$*, denoted $PG(m,2)$, is an incidence structure with $2^{m+1}-1$ points and $\frac{(2^{m+1}-1)(2^m-1)}{3}$ lines. The points are the nonzero $(m+1)$-tuples $(a_0, a_1, \ldots, a_m)\in \mathbb{F}_2^{m+1}$, and a line through two distinct points $a_0$ and $a_1$ contains exactly the set of points $\{a_0, a_1, a_0+a_1\}$.
For more details, see [@lc04] and [@ms79].
Merkx constructed a family of WOM codes based on the $m$-dimensional finite projective geometries over $\mathbb{F}_2$ [@m84]. The construction exploits a connection between the binary Hamming codes and $PG(m,2)$ that allows the WOM codes to be decoded via syndrome decoding. Specifically, the minimum weight codewords of the $[2^{m+1}-1, 2^{m+1}-m, 3]$ Hamming code $\mathcal{C}$ generate $\mathcal{C}$ and correspond to the incidence vectors of lines in $PG(m,2)$. In Merkx’s construction, the messages correspond to points in the geometry. The WOM codewords, i.e. the cell state vectors, are a subset of $\mathbb{F}_2^{m+1}\setminus \mathcal{C}$, and thus, since the Hamming code is perfect, these codewords are always one error from a binary Hamming codeword. The location of the error indicates the point in the geometry that corresponds to the information message.
\[fig:Fano\_Example\]
The $PG(2,2)$ WOM code of [@m84] is a $\langle 7\rangle^4/7$ code. Each position of a codeword corresponds to a point of the Fano Plane, and each codeword is the incidence vector of a substructure of the geometry that highlights a particular point being represented. Codewords are incidences of the following: on the first write, a point on the Fano Plane; on the second write, a line missing a point; on the third write, a line with a point off of it; on the final write, either the union of two lines or the plane missing a point. Thus to decode the WOM code, Merkx observed that syndrome decoding identifies the information message. Figure 1 shows the write sequence $3\rightarrow 5 \rightarrow 7 \rightarrow 3$ using the $\langle 7 \rangle^4/7$ code from the Fano Plane. The arrow indicates the information point and the corresponding cell state vector representing that information is listed below each write. Note that the sequence of cell state vectors is monotonically increasing in each component as the writes progress. \[PG2-2-WOM\]
$\hfill \Box$
The following proposition, by Cohen, Godlewski, and Merkx in [@cgm86], formulates more precisely the parameters of the WOM codes that result from this construction method.
The number of writes that can be attained with a length $2^m-1$-WOM code, storing $m$ bits on each write, is $2^{m-2}+2$.
WOM codes from $EG(m,2)$
------------------------
We now extend Merkx’s idea and design WOM codes from $EG(m,2)$. Since Hamming codes are punctured Reed-Muller codes, and are given by geometric designs over the binary field, a construction similar to the method above can be applied to $EG(m,2)$. Minimum weight codewords also generate the $r^{th}$ order Reed-Muller code $\mathcal{R}(r,m)$, of length $2^m$, and correspond to $(m-r)$-flats in the Euclidean geometry $EG(m,m-r)$. Analogous to the Merkx construction, we will use the connection between minimum weight words in $\mathcal{R}(m-2, m)$ and the planes in $EG(m,2)$ to construct our WOM code. The codewords are designed to be Hamming distance one away from a codeword of $\mathcal{R}(m-2, m)$, and thus are incidence vectors of configurations of points in the Euclidean geometry. Such substructures include a point, a plane with a point missing, and a plane with a point off of it. These WOM codes may be decoded using any Reed-Muller decoding technique.
The next two examples illustrate this construction for $m=3$ and $m=4$.
\[EG3-2-WOM\] Using $EG(3,2)$, the resulting code is an $\langle 8,8,8,4\rangle/8$ WOM code. In other words, the code attains four writes on eight cells, where eight possible messages can be stored in the first three writes, and four messages can be stored in the fourth write. Recall that $EG(3,2)$ has eight points, $28$ lines, and $56$ planes. Each message corresponds to one of the points in the geometry. On the first write, a message $i \in \{1,\ldots, 8\}$ is represented by a weight one cell state vector, where the one is in the $i^{th}$ coordinate. On the second write, a weight three cell state vector indicates a plane with a point missing, where the missing point is the information message. On the third write, the ones in the cell state vector correspond to a plane with a point off of it, where the point off the plane is the message. Observe that on each of the first three writes, it is possible to represent any of the eight messages. Finally, on the fourth write, only messages corresponding to positions of the cell state vector with entry zero can be represented (except for the message represented in the third write, which can always remain on the fourth write, if needed). If $i$ is one of these messages, then to represent $i$ on the fourth write, the cell state vector will have a one in every coordinate except position $i$.
As an example, the message sequence $1\rightarrow 3 \rightarrow 2\rightarrow 7$ is demonstrated in Figure \[eg32writes\].
$\hfill \Box$
In constructing the WOM code from $EG(3,2)$, it is not possible to represent more than four messages on the fourth write. Indeed, after the third write, the cell state vector contains five ones and three zeros, so at most $ \log_2(3) $ information bits can be conveyed by the remaining zero-valued positions. The message that is stored in the third write can always be represented on the fouth write, simply by leaving the memory state unchanged. Thus, one of at most four messages can be represented on the fourth write.
Using $EG(4,2)$, the resulting WOM code has parameters $$\langle 16, 16, 16, 12, 8, 8, 8, 4 \rangle / \ 16.$$ Recall that $EG(4,2)$, shown in Figure 3, has 16 points and 140 planes, and can be partitioned into two parallel $3$-flats. The first four writes are the same as in Example \[EG3-2-WOM\], by using the $EG(3,2)$ code on a $3$-flat that contains the points corresponding to the first four information messages. After the fourth write, the points in that $3$-flat are all programmed to one, and the $EG(3,2)$ WOM code may be applied to the points of the remaining $3$-flat to encode the final four writes.
$\hfill \Box$
\[eg42\]
The $EG(m,2)$ WOM code achieves $4(m-2)$ writes and has parameters
$$\langle \underbrace{ \strut 2^m, 2^m, 2^m, 2^m - 4, 2^{m-1}, 2^{m-1}, 2^{m-1}, 2^{m-1} -4, \ldots, 8, 8, 8, 4 }_{4(m-2)} \rangle / 2^m.$$
[*Proof:*]{} The cell state vector has length $2^m$, equal to the number of points in $EG(m,2)$. Recall that each cell state vector in the $EG(m,2)$ WOM code will be Hamming distance one away from a codeword of the Reed-Muller code $\mathcal{R}(m-2, m)$. We proceed by induction on the dimension of the finite geometry. The base case is the $EG(3,2)$ WOM code. Now suppose that there exists an $EG(k,2)$ WOM code with the parameters described in Example \[eg32writes\]. Consider the finite Euclidean geometry $EG(k+1, 2)$. Note that $EG(k+1, 2)$ can be partitioned into two parallel hyperplanes, i.e. two disjoint copies of $EG(k,2)$. Since any four points lie on a common hyperplane (in fact, many), there exists a hyperplane that contains the points that correspond to the first four information messages to be written. These messages can be encoded using the $EG(3,2)$ WOM code on a cube within this hyperplane containing those points. After the first four writes, all points in the hyperplane are set to one, and the $EG(k,2)$ code can be used on the remaining hyperplane. Thus, this $EG(k+1, 2)$ WOM code allows for $4((k+1)-2)$ writes, and has the parameters listed above, with $m=k+1$. $\hfill \Box$
Since codewords of the WOM code are Hamming distance one from a codeword of the corresponding Reed-Muller code, performing syndrome decoding on a stored cell state vector will provide the location of the position of the “error". The code is designed so that this position corresponds to an information message, i.e., a point in the geometry. Thus, syndrome decoding identifies the message, and can be used to decode the $EG(m,2)$ WOM code.
Comparison
----------
Table \[geom-WOM-table\] shows the rates of the proposed $EG(m,2)$ WOM codes and the $PG(m,2)$ WOM codes from [@m84] for small values of $m$. As expected from the geometric structure, the efficiency of the $EG$ WOM codes is less than that of the $PG$ codes. Indeed, when $m = 2$ and $3$, the $PG(m,2)$ WOM codes have been shown to be optimal [@cgm86]. However, the construction presented here yields a new family of WOM codes that have simple encoding and decoding algorithms, and shows that variable information WOM codes may also be obtained from incidence structures.
In general, designing efficient WOM codes from incidence structures requires low weight incidence vectors, and intersections of these structures that can point to specific messages. In the case of $EG(m,2)$, the $(m-2)^{th}$ order Reed-Muller code was chosen so that the corresponding minimum weight codewords would be planes and therefore have low weight. Since any two distinct planes intersect in $0$ or exactly $2$ points, taking unions of multiple planes does not uniquely designate any one particular point when multiplicity is considered. The authors are interested in using other structures that may be exploited in designing WOM codes where multiplicity can be incorporated, and are currently working on designing WOM codes from general bipartite graphs using insights gained from the the rewriting rules of the geometric constructions.
--------- ---- ------
PG(2,2) 7 1.60
EG(3,2) 8 1.38
PG(3,2) 15 1.82
EG(4,2) 16 1.66
PG(4,2) 31 1.60
EG(5,2) 32 1.50
--------- ---- ------
: Comparison of rates of small dimension projective and Euclidean geometry WOM codes.[]{data-label="geom-WOM-table"}
Using binary WOM codes on multilevel cells
==========================================
The development of flash memory cells on $q>2$ levels has renewed interest in efficient coding strategies for ‘generalized’ write-once memories, i.e., those with greater than two states per cell. Applying binary WOM codes for use on multilevel cells provides a basis for comparison for efficient multilevel coding schemes. In this section we examine construction methods for adapting binary WOM codes for use on multilevel cells.
$x$ $\mu^{1}(x)$ $\mu^{2}(x)$ $\mu^{3}(x)$ $\mu^{4}(x)$
----- -------------- -------------- -------------- --------------
00 000 111 111 222
01 100 011 211 122
10 010 101 121 212
11 001 110 112 221
: Rivest-Shamir code adapted to $q=3$ levels.[]{data-label="rivest-shamir-WOM-table"}
One way to use binary codes[^3] on $q$-level cells is to read the cells modulo 2. One naive approach is to let the set of codewords consist of all cell-state vectors that reduce modulo 2 to a binary codeword. A more efficient application of a $\langle v\rangle^t/n$ code to $q$-level cells is to increase the charge of all cells to $1$ after the $t^{th}$ write, and then employ the code again. We will refer to this scheme as the [*complement scheme*]{}, since reduction modulo 2 either reveals a WOM codeword or the complement of a codeword. More precisely, in the complement scheme, let $x$ denote the information message, and $c^i(x)$ be a codeword that represents $x$ on the $i^{th}$ write. We reuse the binary WOM code by taking $c^{t+i}(x)=c^i(x)+{\bf 1}$, for $i<t$, where ${\bf 1}$ is the all ones vector. Similarly, after $mt$ writes, the cell values are increased to $m$, and we set $c^{mt+k}(x)=c^{k}(x)+m\cdot{\bf 1}$ for $k=1, \ldots, t-1$. Note that this scheme guarantees $(q-1)t$ writes. Table \[rivest-shamir-WOM-table\] shows Example \[rivest-shamir-WOM\] adapted to $q=3$-level cells in this way.
We will use this simple scheme as a basis for comparison when considering the following methods of adapting binary WOM codes to $q$-levels.\
[**Construction:**]{} Consider a $\langle 2^k\rangle^t/n$ WOM code. Let $x$ be a binary information sequence of length $k$, and let $U(x)=\{u: u=c^{i}(x) \text{ for some } i=1,\ldots, t\}$. Let $s$ be a length $n$ cell state vector representing the message $x$. Given $s$, suppose we want to write a new message $y \ne x$. Let $V$ be the set of $n$-tuples with all entries even (possibly 0) and less than $q$. We present two strategies.
- [**Strategy A**]{}: To minimize the number of cells that are increased, search the set $U(y)+V$ for the representation whose difference from $s$ requires the fewest cells to increase. Thus, look for $s'\in U(y)+V$ such that $s'\ge s$ (componentwise, all entries in $s'$ are at least as much as those in $s$) and further that $s$ and $s'$ differ in the least number of places, i.e. the Hamming weight, $wt_H(s'-s)$ is minimized. The new cell state vector is $s'$ and represents the new message $y$. In searching the set $U(y)+V$ as the cell values approach $q$, we omit the values of $s'$ that would cause a block erasure.
- [**Strategy B**]{}: To minimize the magnitude of the resulting cell state vector $s'$, search the set $U(y)+V$ for the representation whose difference from $s$ is such that the maximum cell entry of $s'$ is minimized. If there is a tie, arbitrarily choose one that requires the fewest number of cells to increase. Thus, look for $s'\in U(y)+V$ such that $s'\ge s$ and that the maximum entry in $s'$ is the smallest.
\[ WOM\_new\]
For specific codes, the strategies can be described more explicitly. For example, the following flash code encoding map is based on Example \[rivest-shamir-WOM\], and uses reduction modulo 2 to identify the decoding map from the cell state vectors to the variable vectors. Following Strategy A, the rewriting rule is as follows. Let $s$ be the current cell state vector representing the message $x$, and $y$ the new message to be written.
- If $x,y \in \mathbb{F}_2^2 \setminus \{00\}$,
- If $s$ mod 2 $= c^{1}(x)$, add the weight one vector $w=c^{2}(y)-c^{1}(x)$ to the current state, to obtain the new cell state vector $s'$ = $s+w$.
- If $s $ mod 2$ = c^{2}(x)$ write $w=c^{1}(z)$, where $z\in
\mathbb{F}_2^2\setminus\{00, x, y\}$, to obtain $s'=s+w$.
- If $x = 00$, write $c^{1}(y)$.
- If $y = 00$, then if $s$ mod 2 $= c^{1}(x)$, add $c^{1}(x)$ to $s$; otherwise add ${\bf 1} - c^{2}(x)$ to $s$.
Following Strategy B, the rewriting rule depends on the actual magnitude (in $\{0,\ldots, q-1\}$) of each cell entry.
The general rule is to increase a subset of the cells such that the new vector reduces to either $c^{1}(y)$ or $c^{2}(y)$ modulo 2 and no one cell is allowed to gain too much charge.
Using the rules above for the Rivest-Shamir WOM code in Example \[rivest-shamir-WOM\], suppose the following information sequence is to be stored in a given set of cells with $q = 4$ levels. $$11 \rightarrow 00
\rightarrow 01 \rightarrow 10 \rightarrow 11 \rightarrow
01$$
Following Strategy A, the sequence of cell state vectors is as follows $$A: 001 \rightarrow 002 \rightarrow 102 \rightarrow 103 \rightarrow
203 \rightarrow 213$$
Following Strategy B, the sequence of cell state vectors is as follows $$B: 001 \rightarrow 111 \rightarrow 211 \rightarrow 212 \rightarrow
312 \rightarrow 322$$
$\hfill \Box$
To further illustrate the different strategies, consider writing the sequence $1\rightarrow2\rightarrow 1 \rightarrow 3$ using the $PG(2,2)$ WOM code in Example \[PG2-2-WOM\], where the labeling on the Fano Plane is as in Figure 1. Following Strategies A and B, the sequence of cell state vectors is as follows:
$$A: (1000000) \rightarrow (1001000) \rightarrow (1002000) \rightarrow (1002001)$$
$$B: (1000000) \rightarrow (1001000) \rightarrow (1001101) \rightarrow (1101111)$$
$\hfill \Box$
Analysis of Strategies A and B
------------------------------
The expected number of writes for floating codes was studied in [@flm08; @cflm10] and can be more important than the worst case analysis in determining which codes to use in practice. Code constructions in [@jbb07] have a guarantee of $(q-1)+\lfloor \frac{q-1}{2}\rfloor$ writes for a $k=2$-dimensional message space and $n=2$ cells. The same paper also proved the existence of floating codes that achieve $(q-1)n-o(n)$ writes as $n\rightarrow \infty$ for fixed $k$ and $q$. Asymptotically optimal codes for the average case with $k=2$ have been constructed where the expected number of writes grows like $n(q-1)-o(q)$ [@cflm10]. Both cases include the assumption that only one cell level changes at each write, which is reasonable when $n\gg 2^k$. However, since Strategies A and B are intended to be used for [*any*]{} WOM code, not just those that meet this criterion, we do not use this assumption.
The guaranteed number of writes using Strategy B for the $\langle 4\rangle^2/3$ Rivest-Shamir WOM code on $q$ level cells is $2(q-1)$. This can be seen by examining a sequence of messages that cause a maximum number of cell increases under Strategy B. For example, the alternating sequence of messages $00\rightarrow 01\rightarrow 00\rightarrow 01\rightarrow 00 \rightarrow \dots \rightarrow 01\rightarrow 00$ has cell state vector sequence $000 \rightarrow 100\rightarrow 111\rightarrow 211 \rightarrow 222 \rightarrow \dots \rightarrow (q-1)(q-2)(q-2)\rightarrow (q-1)(q-1)(q-1)$. Observe that for every two writes, the cell state vector does not increase a cell level more than once, and both representations of a given message are used. Thus, the guaranteed number of writes using Strategy B is $2(q-1)$.
The guaranteed number of writes using Strategy A for the $\langle 4\rangle^2/3$ Rivest-Shamir WOM code on $q$ level cells is also $2(q-1)$. Again we consider a sequence of messages that causes the maximum number of cell increases. For example, the alternating sequence of messages $00\rightarrow 01\rightarrow 00\rightarrow 01\rightarrow 00 \rightarrow \dots \rightarrow 01\rightarrow 00 \rightarrow 01$ has cell state vector sequence $000 \rightarrow 100\rightarrow 200\rightarrow 300 \rightarrow 400 \rightarrow \dots \rightarrow (q-2)00\rightarrow (q-1)00 \rightarrow (q-1)11\rightarrow (q-1)22\rightarrow \dots \rightarrow (q-1)(q-1)(q-1)$. Observe that the first $q-1$ writes follow the Strategy A protocol to increase the fewest number of cells, but that once any cell attains the maximum charge, the Strategy continues to write using the next best representation choice for each message. Thus, a total of $2(q-1)$ writes are guaranteed.
The following theorem shows that the guaranteed number of writes for both Strategies A and B is at least as good as the complement scheme for any general binary WOM code.
Let $\mathcal{C}$ be a $\langle v\rangle^t/n$ binary WOM code. Then, the guaranteed number of writes by applying either Strategy A or Strategy B to $\mathcal{C}$ on $q$-level flash cells is at least $(q-1)t$. \[guaranteed\]
[*Proof:*]{} We prove by induction on $q$. For $q=2$, the WOM code already guarantees $t$ writes. So assume the hypothesis holds for $q=r$. That is, for any sequence of messages, we are guaranteed at least $(r-1)t$ writes using Strategy A or B. Now let us consider the case when $q=r+1$. Then for any sequence of $(r-1)t$ messages, using Strategy A or Strategy B, by the induction hypothesis we will reach a cell state vector $(c_1,c_2,\dots,c_n)$, with entries $c_i \le r-1$, $i=1,2,\dots,n$. We can now artificially increase each cell levels to $r-1$ at the end of $(r-1)t$ writes to yield a cell state vector $(r-1,r-1,\dots,r-1)$. Without loss of generality, the cell state vector $(r-1, r-1, \cdots,r-1)$ can be thought of as being the all-zero vector $(0,0,\cdots,0)$. It is now easy to see that either Strategy A or Strategy B will allow us to write at least $t$ more times using the original $t$ writes of the binary WOM code $\mathcal{C}$. Thus, a total of $rt$ writes is guaranteed for either Strategy when $q=r+1$, thereby proving the result. $\hfill \Box$\
To see if the lower bound of $(q-1)t$ writes is met in Theorem \[guaranteed\], the weight distributions of the different representations for each message in the original WOM code have to be taken into account. For example, for two write WOM codes where the minimal weight representation for each message is unique, the guaranteed number of writes is $2(q-1)$ as above. The authors are currently looking at how to classify when a WOM code meets this lower bound using the weight distributions of the message representations.
Strategies A and B applied to the Rivest-Shamir code each guarantee two writes when $q = 2$ and four writes when $q = 3$, whereas the expected number of writes using the Strategies for this code (assuming a uniform distribution on the message space) is approximately 2.47 for $q = 2$ and 4.89 for $q = 3$ for each case. Note that the simple application of the Rivest-Shamir code to $q$-level cells using the complement scheme requires $q \ge 3$ to get more than two guaranteed writes. Figure \[ave\_writes\] compares the average number of writes of the complement scheme, Strategy A, and Strategy B on $q$-level cells when applied to the binary Rivest-Shamir WOM code from Example \[rivest-shamir-WOM\]. In Monte Carlo simulations, $10^5$ random message sequences were generated and the number of writes was recorded for the three different methods. As shown in Figure \[ave\_writes\], the strategies applied to the Rivest-Shamir code exhibit a noticeable gain over the the complement scheme that is growing as $q\rightarrow \infty$. However, the average number of writes for each Strategy is still quite far from the capacity limit on the number of writes possible for representing four messages per write using three cells on $q$-levels (see Section 1).
Strategies A and B did not exhibit much gain over the complement scheme when the $PG(2,2)$ code in Example \[PG2-2-WOM\] was simulated for small $q$. We believe that this is due to the near-optimality of the $PG(2, 2)$ WOM code. Further, we believe that in general, the more optimal a code is, the less it will benefit from the strategies, since the reapplication of the code under the complement scheme already generates an efficient code.
In [@cflm10], two coding schemes are presented that have a similar flavor to Strategies A and B, but apply in the different setting of random floating codes. In that work, the authors propose two random coding schemes: a “Simple scheme" that randomly chooses to increase a single cell by one, and a “Least scheme" that chooses a message representation that increases the coordinate with the lowest charge level. In contrast, Strategies A and B in this paper apply to any WOM code without the assumption that only one cell increases at each write. We also expect that the performance of codes under these strategies will differ more for certain classes of codes and when non-uniform distributions on the message space are considered. Further analysis of the performances of the strategies for different WOM codes is underway, including quantifying their average performance using both uniform and nonuniform distributions on the message space.
Concatenated error-correcting flash codes
=========================================
In this section we consider ways that code concatenation may be used to obtain new WOM or flash codes. Let $[n,k,d]_q$ denote a classical $q$-ary linear code of block length $n$, dimension $k$, and minimum distance $d$. Two classical codes may be concatenated as follows.
Let $\mathcal{A}$ be an $[n_1,k_1,d_1]_{q^{k_2}}$ code and $\mathcal{B}$ be an $[n_2,k_2,d_2]_q$ code. Then the concatenated code $\mathcal{C} = \mathcal{A}\boxtimes\mathcal{B}$ is an $[n_1n_2,k_1k_2,d_1d_2]_q$ code with outer code $\mathcal{A}$ and inner code $\mathcal{B}$. The $k_1$ information symbols (each chosen from a $q^{k_2}$-ary alphabet) are first encoded into $n_1$ symbols using $\mathcal{A}$. Each of the encoded symbols is then represented by $k_2$ $q$-ary symbols. Each group of these $k_2$ symbols is then encoded into $n_2$ $q$-ary symbols using $\mathcal{B}$. Thus, $n_1n_2$ encoded symbols are obtained to form a codeword in $\mathcal{C}$. \[ concatenation\_classical\_classical\]
The above concatenation may be seen by the following mapping $$\mathbb{F}_{q^{k_2}}^{k_1}\ \ \underrightarrow{\ \ \ \mathcal{A}\ \ \ } \ \ \mathbb{F}_{q^{k_2}}^{n_1}\ \ \underrightarrow{\mbox{$q$-ary representation}} \ \ \mathbb{F}_q^{n_1k_2} \ \ \underrightarrow{\ \ \ \mathcal{B}\ \ \ }\ \ \mathbb{F}_q^{n_1n_2}$$
Concatenating classical codes with binary WOM or flash codes yields codes with both error correction and rewrite capabilities.
Several researchers have observed that an outer $\langle 2^k\rangle^t/n$ WOM code $\mathcal{A}$ when concatenated with an inner $[m,1]_{2}$ repetition code $\mathcal{B}$ yields a $\langle 2^k\rangle^t/nm$ binary WOM code $\mathcal{C} = \mathcal{A}\boxtimes\mathcal{B}$, where $C$ can correct $\lfloor\frac{m-1}{2}\rfloor$ errors [@zc91; @ysvw10; @j07]. We expand on these ideas to obtain codes for multilevel flash cells.
A code $C_W\boxtimes C_R$, where $C_W$ is a WOM code and $C_R$ is a length-$m$ repetition code, can be employed as an error-correcting code on $q$-level cells with the following strategy: on the first write, the binary codeword is written on the cells. An error can be detected by majority decision among each set of $m$ consecutive positions. For subsequent writes and error correction, we will read the $q$-ary vector as a binary codeword from $C_W$, by reducing the values in the cells modulo 2. In particular, if a one was erroneously written on the first write in a cell that should have contained a zero, we correct the error by increasing the level of the cell to $2$, which is viewed as a 0 (modulo 2). The error has been corrected in the binary word that is read, and the code can correct $\lfloor\frac{m-1}{2}\rfloor$ errors on each write. Subsequent writes are achieved by increasing chosen cell levels to obtain the desired parity, modulo 2.
The following theorem uses this method to obtain an error-correcting WOM code. Note that errors can occur in either direction and are assumed to be of magnitude one.
Let $C_W$ be a $\langle 2^k\rangle^t/n$ WOM code and let $C_R$ be the $[m,1,m]_2$ repetition code. The code $C_W \boxtimes C_R$ is an $\langle 2^k \rangle^t/mn$ $\lfloor\frac{m-1}{2}\rfloor$-error-correcting WOM-code on SLCs. Moreover, applied to $q$-level cells and using the reduced binary vector representation, ${C_W} \boxtimes C_R$ is a $\langle 2^k \rangle_q^{t'}/mn$ flash code, where $t'=\lceil \frac{(q-1)t}{3}\rceil$ and $\lfloor\frac{m-1}{2}\rfloor$ errors can be corrected at each write. \[basic\_thm\]
[*Proof:*]{} For $q = 2$ the resulting code is a $\langle 2^k \rangle^t/mn$ $\lfloor \frac{m-1}{2}\rfloor$-error correcting WOM code. For any $q$, the length $mn$-code has dimension $k$. We show that the worst-case number of rewrites is $\lceil \frac{(q-1)t}{3}\rceil$. Note that $C_W\boxtimes C_R$ is still a binary code, but we use it on the $q$-ary cells by reading the information stored in the cells via the reduced binary vectors. Up to $\lfloor\frac{m-1}{2}\rfloor$ errors can be detected and corrected at each write. Note that in this scheme, error correction consists of increasing the charge level of the cell by one to correct the parity in that entry of the reduced binary vector. In the worst case, an error occurs in the same position on every write, and so that position sees an increase of three levels at each write. However, in the absence of errors we could achieve $(q-1)t$ writes due to the rewriting capability of $C_W$ and the reapplication of the WOM code on $q$-level cells. Thus, the worst-case number of writes in the error case is $\lceil \frac{(q-1)t}{3}\rceil$. $\hfill \Box$
As an example of the reading process, if $q=4, n=1, m=3$, the sequence $(332)$ in a cell-state vector would be read as $(110)$ in $C_W\boxtimes C_R$, and decoded to $(111)$ using majority rule. As an example of the error-correction process, consider a cell that is meant to be increased to $0$ (modulo 2); if an error causes the cell to instead be read as $1$ (modulo 2), then to correct it the charge is increased again. Thus that cell has seen a total increase of three levels on that write cycle. A similar idea of increasing the cell levels to correct for errors has also been considered in [@j07; @jlw09].
Let $C_W$ be the $\langle 4 \rangle^2/3$ WOM code defined in Example \[rivest-shamir-WOM\] and let $C_R$ be the $[3,1,3]_2$ repetition code. Then the code $C_W \boxtimes C_R$ is a $\langle 4 \rangle^2/9$ single error-correcting WOM code on SLCs (first observed in [@zc91]). Moreover, on $q$-level cells, the code $C_W \boxtimes C_R$ is a $\langle 4\rangle_q^{\lceil\frac{2(q-1)}{3}\rceil}/9$ single error-correcting flash code. $\hfill \Box$
Let $C_W$ be the $\langle 7\rangle^4/7$ code based on $PG(2,2)$ from [@m84] and let $C_R$ be the $[3,1,3]_2$ binary repetition code. Then the code $C_W \boxtimes C_R$ is a $\langle 7 \rangle^4/21$ single error-correcting WOM code on SLCs. Moreover, on $q$-level cells, the code $C_W \boxtimes C_R$ is a $\langle 7 \rangle_q^{\lceil \frac{4(q-1)}{3}\rceil}/21$ single error-correcting flash code. $\hfill \Box$
We next show how to obtain a flash code with increased error-correction by concatenating an inner flash code with an outer classical code.
Let $C_1$ be an $[n_1, k_1]_{q^{k_2}}$ code that corrects $e$ errors, and $C_2$ a $\langle 2^{k_2}\rangle_q^t/n_2$ $E$-error-correcting WOM code. Then $C_1\boxtimes C_2$ is a $\langle 2^{k_1k_2}\rangle^t_q/(n_1n_2)$ WOM code capable of correcting ${\bf (E+1)(e+1) - 1}$ errors. \[ classical\_WOM\_thm\]
[*Proof:*]{} The length and dimension of $C_1\boxtimes C_2$ is immediate. Note that this code achieves $t$ writes since the inner flash code is capable of $t$ writes. The minimum number of errors that must occur for a decoding failure is $(E+1)(e+1)$, where $E+1$ errors occur among each of $e+1$ distinct length-$k_2$ $q$-ary expansions of symbols in $C_1$. Any smaller number of errors can be corrected by the length $n_1n_2$ concatenated code. $\hfill \Box$
For comparison, we show the concatenation of a inner binary repetition code with a classical binary outer code for use on $q$-level flash cells.
Let $C$ be an $[n,k,d]_2$ $e$-error-correcting code and let $C_R$ be the $[2m+1,1,2m+1]_2$ binary repetition code. Then the code $C \boxtimes C_R$ for $q$-level cells results in a $\langle 2^k \rangle^t_q/((2m+1)n)$ flash code that corrects $(me+m+e)$ errors and guarantees $t=\lceil \frac{q-1}{3}
\rceil$ writes. \[classical\_classical\_thm\]
[*Proof:*]{} The length and dimension follow from the construction. Concatenating two binary codes results in a binary code, but we use reduction modulo 2 to adapt the code to $q$-ary cells. Errors that result in a change in parity of a cell can be corrected by increasing the level of the cell by one. In the worst case, an error occurs in the same cell at every write. In order to correct it, the cell level is increased by one so that it has the same parity as the entry before the error occurred. Thus this code guarantees $\lceil \frac{q-1}{3} \rceil$ writes. Note that the outer code can correct up to $e$ errors and the inner code can correct up to $m$ errors. Thus, the concatenated code can tolerate $(m+1)(e+1)-1=me+m+e$ errors. $\hfill \Box$
Observe that this use of a classical code on multilevel cells gives better error-correction capabilities than the code in Theorem \[basic\_thm\] but can tolerate fewer rewrites since the only rewrite capabilites come from the number of levels.
Let $C$ be an $[n,k,d]_2$ $e$-error-correcting code and let $C_R$ be the $[3,1,3]$ binary repetition code. Then the code $C \boxtimes
C_R$ for $q$-level cells yields a $\langle 2^k\rangle^t_q/(3n)$ flash code that corrects ${\bf 2e+1}$ errors and gets $t=\lceil \frac{q-1}{3}
\rceil$ writes. $\hfill \Box$
Conclusions
===========
We showed how the structure of finite Euclidean geometries can be used to obtain new variable information WOM codes. We also introduced several strategies for adapting WOM codes to multilevel cells that allow for a greater rewrite capability than classical codes adapted for multilevel coding schemes. Combined with concatenation, these codes also have the ability to correct multiple errors. We are currently investigating the use of other incidence structures, including finite geometries over $\mathbb{F}_q$, for developing new coding schemes for multilevel flash memories.
Acknowledgments {#acknowledgments .unnumbered}
---------------
Some of this work was completed while the authors were at the Fall 2011 program on [*Combinatorial, Algebraic and Algorithmic Aspects of Coding Theory*]{} at the Bernoulli Center, EPFL, Switzerland. The authors thank the organizers and administrative staff for their hospitality, and the anonymous reviewers for their helpful comments.
F. Merkx, “Womcodes constructed with projective geometries," [*Traitment du Signal*]{}, vol. 1, no. 2-2, pp. 227-231, 1984.
R. L. Rivest and A. Shamir, “How to reuse a “write-once“ memory,” [*Inform. Control*]{}, vol. 55, pp.1-19, Oct./Nov./Dec. 1982.
A. Fiat and A. Shamir, “Generalized write-once memories," [*IEEE Trans. Inform. Theory*]{}, vol. 30, pp. 470-480, September 1984.
S. Kayser, E. Yaakobi, P. H. Siegel, A. Vardy, and J. K. Wolf, “Multiple write WOM-codes", [*Proc. of the 48th Annual Allerton Conf. on Comm., Control and Computing*]{}, Sept. 2010.
Y. Cassuto, M. Schwartz, V. Bohossian, and J. Bruck, “Codes for multi-level flash memories: Correcting asymmetric limited-magnitude errors," in [*Proc. IEEE ISIT*]{}, Nice, France, June 24-29, 2007, pp. 1176-1180.
A.V. Kuznetsov and B.S. Tsybakov, “Coding in a memory with defective cells," [*Problemy Peredachi Informatsii*]{}, vol. 10, no. 2, pp. 52-60, Apr.-June 1974.
A.V. Kuznetsov, A. J. Han Vinck, “On the general defective channel with informed encoder and capacities of some constrained memories," [*IEEE Trans. Inform. Theory*]{}, vol. 40, no. 6, November 1994.
R. Ahlswede, Z. Zhang, “Coding for Write-Efficient Memory," in [*Information and Computation*]{}, vol. 83, no. 1, October 1989.
G. Cohen, “On the capacity of write-unidirectional memories," [*Bull. Instit. Mathemat. Academia Sinica*]{}, vol.16, no. 4, pp. 285-293. Dec. 1988.
A. Jiang, H. Li, Y. Wang, “Error scrubbing codes for flash memories," in [*Proc. Canadian Workshop on Information Theory (CWIT)*]{}, Ottawa, Canada, May 2009, pp. 32-35.
A. Jiang, J. Bruck, “Information representation and coding for flash memories," in [*Proc. IEEE Pacific Rim Conference on Communications, Computers and Signal Processing (PACRIM)*]{}, pp. 920-925, Victoria, B.C., Canada, August 2009.
A. Jiang, J. Bruck, “Joint coding for flash memory storage," in [*Proc. IEEE ISIT*]{}, 2008, pp. 1741-1745.
G. Zëmor, and G. D. Cohen, “Error-correcting WOM-codes," [*IEEE Trans. Inform. Theory*]{}, vol. 37, pp. 730-734, May 1991.
A. Jiang, “On the generalization of error-correcting WOM codes," in [*Proc. IEEE ISIT*]{}, 2007.
A. Jiang, V. Bohossian, and J. Bruck, “Floating codes for joint information storage in write asymmetric memories," in [*Proc. IEEE ISIT*]{}, Nice, France, June 2007, pp. 1166-1170.
A. Jiang, M. Langberg, M. Schwartz, and J. Bruck, “Universal rewriting in constrained memories," in [*Proc. IEEE ISIT*]{}, Seoul, Korea, 2009, pp. 1219-1223.
E. Yaakobi, A. Vardy, P. Siegel, J. Wolf, “Multidimensional flash codes," in [*Proc. of the Annual Allerton Conference*]{}, 2008.
F. Fu and A.J. Han Vinck, “On the capacity of generalized write-once memory with state transitions described by an arbitrary directed acyclic graph," [*IEEE Trans. Inform. Theory*]{}, vol. 45, no. 1, pp. 308-313, 1999.
S. Lin, D. J. Costello, Jr., [*Error Control Coding: Fundamentals and Applications*]{}, 2nd Ed. Upper Saddle River, NJ: Prentice-Hall, 2004.
F. J. MacWilliams, N. J. A. Sloane, [*The Theory of Error-correcting Codes*]{}, Amsterdam, The Netherlands: Elsevier B.V., 1979.
G. Cohen, P. Godlewski, F. Merkx, “Linear binary codes for write-once memories," [*IEEE Trans. Inform. Theory*]{} vol. 32, no. 5, Sept. 1986.
Q. Huang, S. Lin, and K. Abdel-Ghaffar, “Error-correcting codes for flash coding", in [*Proc. IEEE Information Theory and Applications*]{}, San Diego, February 2011.
H. Finucane, Z. Liu, M. Mitzenmacher, “Designing floating codes for expected performance," [*Proc. 46th Annual Allerton Conference on Communication, Control, and Computing*]{}, Monticello, IL, September 2008.
F. Chierichetti, H. Finucane, Z. Liu, M. Mitzenmacher, “Designing floating codes for expected performance,"[*IEEE Trans. Inform. Theory*]{}, vol. 56, pp. 968-978, March 2010.
E. Yaakobi, P. H. Siegel, A. Vardy, and J. K. Wolf, “Multiple error-correcting WOM codes", [*Proc. IEEE ISIT*]{}, Austin, TX, June 2010.
[^1]: This work was supported in part by the National Security Agency under Grant Number H98230-11-1-0156. The United States Government is authorized to reproduce and distribute reprints not-withstanding any copywrite notation herein.
[^2]: This terminology was introduced in [@j07] in reference to the structure of flash memory, but it is convenient to use in the WOM case as well.
[^3]: The idea of reducing the cell state vectors modulo 2 was also used in [@hla11] to adapt [*classical*]{} codes for use on multilevel cells.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A simple two color selection based on $B$-, $z$-, and $K$- band photometry is proposed for culling galaxies at $1.4\simlt
z\simlt2.5$ in $K$-selected samples and classifying them as star-forming or passive systems. The method is calibrated on the highly complete spectroscopic redshift database of the K20 survey, verified with simulations and tested on other datasets. Requiring $BzK=(z-K)_{AB}-(B-z)_{AB}>-0.2$ allows to select actively star-forming galaxies at $z\simgt1.4$, independently on their dust reddening. Instead, objects with $BzK<-0.2$ and $(z-K)_{AB}>2.5$ colors include passively evolving galaxies at $z\simgt1.4$, often with spheroidal morphologies. Simple recipes to estimate the reddening, SFRs and masses of $BzK$-selected galaxies are derived, and calibrated on $K<20$ galaxies. These $K<20$ galaxies have typical stellar masses $\sim10^{11}M_\odot$, and sky and volume density of $\sim1$ arcmin$^{-2}$ and $\sim10^{-4}$Mpc$^{-3}$ respectively. Based on their UV (reddening-corrected), X-ray and radio luminosities, the $BzK$–selected star-forming galaxies with $K<20$ turn out to have average $SFR\approx 200\ M_{\odot} yr^{-1}$, and median reddening $E(B-V)\sim0.4$. This SFR is a factor of 10 higher than that of $z\sim1$ dusty EROs, and a factor of 3 higher than found for $z\sim2$ UV selected galaxies, both at similar $K$ limits. Besides missing the passively evolving galaxies, the UV selection appears to miss some relevant fraction of the $z\sim2$ star-forming galaxies with $K<20$, and hence of the (obscured) star-formation rate density at this redshift. The high SFRs and masses add to other existing evidence that these $z=2$ star-forming galaxies may be among the precursors of $z=0$ early-type galaxies. A V/V$_{max}$ test suggests that such a population may be increasing in number density with increasing redshift. Theoretical models cannot reproduce simultaneously the space density of both passively evolving and highly star-forming galaxies at $z=2$. In view of Spitzer Space Telescope observations, an analogous technique based on the $RJL$ photometry is proposed to complement the $BzK$ selection and to identify massive galaxies at $2.5\simlt z\simlt 4.0$. By selecting passively evolving galaxies as well as actively star-forming galaxies (including strongly dust reddened ones), these color criteria should help in completing the census of the stellar mass and of the star-formation rate density at high redshift.
author:
- 'E. Daddi, A. Cimatti, A. Renzini, A. Fontana, M. Mignoli, L. Pozzetti, P. Tozzi, G. Zamorani'
title: |
A new photometric technique for the joint selection of star-forming\
and passive galaxies at [$1.4\simlt \lowercase{z}\simlt2.5$]{}$^1$
---
Introduction
============
Tracing and understanding the history of cosmic star formation and the growth of the cosmic stellar mass density are currently the objects of major observational efforts (e.g., Dickinson et al. 2003; Fontana et al. 2003; Rudnick et al. 2003), yet much remains to be done before reaching fully satisfactorily conclusions. Obtaining a complete census of the populations of high-redshift galaxies, and their physical characterization (stellar mass and star-formation rate, SFR) is necessary to observationally map the processes that lead to galaxy formation and evolution. This requires direct spectroscopic identification (e.g., Fontana et al. 2004; Glazebrook et al. 2004), but building large samples of spectroscopically confirmed galaxies at high redshift is a time consuming process. This is especially true for magnitude-limited samples, as high-$z$ galaxies represent a small fraction of the galaxy counts at faint magnitudes in the optical and infrared bands. Therefore, techniques for pre-selecting high redshift targets are required to focus the spectroscopic multiplex capability on the interesting objects. Photometric redshifts from deep multicolor datasets offer an alternative to massive spectroscopic efforts (e.g. Bolzonella et al. 2002; Firth et al. 2002; Poli et al. 2003), but the results may suffer from biases that are difficult to quantify, and their accuracy depends critically on the quality of the photometry and the colors and redshifts of the objects.
An efficient alternative to either photometric redshifts or magnitude limited surveys is offered by simpler single or two-color criteria, then followed by targeted spectroscopy. The best known example is the dropout technique for selecting Lyman-Break Galaxies (LBG, Steidel et al. 1996; 2003) from their $U_{\rm n}GR_{\rm s}$ rest-frame UV colors, which opened the research field on normal star-forming galaxies at $z\sim3$. Recently, the same technique has been extended to work at lower redshifts $1.4<z<2.5$ (Erb et al. 2003; Adelberger et al. 2004) and large samples of UV-selected objects have been spectroscopically confirmed at $z\sim2$ (Steidel et al. 2004).
Star forming galaxies can be selected as LBGs only if they are UV bright (i.e. actively star forming) and not heavily reddened by dust. Currently, the best possible alternatives to find dust-enshrouded high redshift star-forming objects include detecting them from their far-IR or sub-mm emission due to cold dust (Franceschini et al. 2001; Smail et al. 2002; Champan et al. 2003), observing the emission at X-ray or radio wavelengths that are not extincted by dust (e.g., Norman et al. 2004; Haarsma et al. 2000), and selecting very red objects in near-infrared samples that are less affected by dust extinction (Cimatti et al. 2002a, 2003).
Besides targeting star-forming galaxies, color criteria have also been used to search for passively evolving galaxies at high redshifts. A simple criterion is the one used for Extremely Red Objects (ERO), selected according to their very red optical to near-IR colors, e.g. $R-K>5$ or $I-K>4$ (Thompson et al 1999; Daddi et al. 2000a; 2000b; Firth et al. 2002; Roche et al. 2002; 2003; Miyazaki et al. 2003; McCarthy 2004). Spectroscopy showed that EROs include both old passive galaxies at $0.8<z<2$ and dusty star-forming systems at similar redshifts (Cimatti et al. 2002a, 2003; Yan et al. 2004). The bulk of EROs, however, is made of galaxies at redshifts $z\sim 1$ with only a small fraction being at $z\simgt2$, a crucial redshift range for the evolution of galaxies. In order to identify old stellar systems at $z\sim2$ or beyond, color criteria based on $J$ and $K$ imaging have been proposed (e.g., Pozzetti & Mannucci 2000; Totani et al. 2001; Franx et al. 2003; Saracco et al. 2004). As confirmed by van Dokkum et al. (2003), objects with red spectral energy distributions at $z\simgt2$, can be selected requiring very red colors $J-K>2.3$ (Vega scale).
Although the selection of objects with extremely red colors has been quite successful, one could expect that moderately old, or moderately reddened, objects exist at high redshift that would be missed by both the “red-color" techniques and by the UV techniques. In addition, it appears unsatisfactory to use so many different color criteria in order to build representative samples of galaxies as a function of redshifts, as the physical relations between these different classes remain unknown, and the selection biases not fully understood.
The recently completed K20 survey (Cimatti et al. 2002a,b,c; Daddi et al. 2002; Pozzetti et al. 2003; Fontana et al. 2004) has reached an unprecedented spectroscopic completeness ($>92\%$) for a sample of $K<20$ galaxies. In particular, for one of the two K20 fields (included in the CDF-S/GOODS-S field) the recent deep observations (Daddi et al. 2004, hereafter D04; Cimatti et al. 2003, 2004; Vanzella et al. 2004[^1]) allowed to reach a spectroscopic completeness of $>94\%$. This includes the presence of a significant fraction of high redshift, $z>1.4$, galaxies (redshift desert coverage).
In this paper we take advantage of the K20 spectroscopic sample to define a simple two-color criterion based on the $B$-, $z$- and $K$-band photometry which, with a minimal contamination from lower redshift galaxies, is capable of identifying the full range of high-redshift $z>1.4$ galaxies in our $K$-selected sample, including both actively star-forming (D04) and old passive objects (Cimatti et al. 2004), and to distinguish between the two classes.
For galaxies identified with this criterion (at least to $K=20$) it is found that a $K$-band selection is close to a galaxy stellar-mass selection, while a $K$-selected sample of star-forming galaxies allows to reach completeness down to a given star-formation rate limit almost independently of dust reddening. Therefore, the technique offers a powerful tool to explore with the minimum possible biases the histories of cosmic star-formation and cosmic stellar-mass build-up at $z\sim2$. We discuss in detail to which extent the cosmic stellar-mass and star-formation rate density can be estimated with the properties of galaxies in the proposed two-color diagram. The X-ray and radio properties of $K$-selected star-forming galaxies are also investigated in order to provide an independent estimate of their star-formation rates.
The paper is organized as follow. The spectroscopic and imaging datasets used in the paper are described in Section \[sec:data\]. The $BzK$ selection and classification technique for $1.4<z<2.5$ galaxies is empirically calibrated in Section \[sect:BzK\], checked against stellar population models in Section \[sec:modeling\] and compared to HST morphological classification in Sect. \[sect:HST\]. The SFR and mass content of $z\sim2$ galaxies are described in Section \[sec:SFR\] and \[sec:masses\], together with methods to obtain ensemble averages from the $BzK$ photometry alone. Section \[sec:other\] compares the samples selected with the $BzK$ technique to those of other criteria, including UV selected $z\sim2$ galaxies, EROs and $J-K$ red galaxies. We extend the technique for use at higher redshifts using Spitzer Space Telescope (SST) imaging in Section \[sec:SIRTF\]. The results are discussed in Section \[sec:discussion\] and summary and conclusions are in Section \[sec:summary\].
We use the Salpeter IMF extending between 0.1 and 100 $M_\odot$ and a WMAP cosmology with $\Omega_\Lambda, \Omega_M = 0.73, 0.27$, and $h = H_0$\[km s$^{-1}$ Mpc$^{-1}$\]$/100=0.71$.
The data {#sec:data}
========
The K20/GOODS Field
-------------------
The K20 survey has obtained spectra for $545$ objects selected in the $K$-band over two widely separated fields for a total area of 52 arcmin$^2$, including a 32 arcmin$^2$ region of the GOODS-South field (Cimatti et al. 2002b). Of the 347 objects with $K<20$ in this area, 328 have been spectroscopically identified at the moment by complementing the K20 spectroscopy with a few additional redshifts from the ESO/GOODS public spectroscopy (Vanzella et al. 2004). The identified targets include 292 extragalactic objects and 36 stars, while the residual 19 objects have only photometric redshifts. Among these 311 galaxies, 19 have $z_{\rm spec}>1.4$ (6% of the sample) and 13 have $z_{\rm phot}>1.4$, or $\sim 10\%$ lie at an estimated redshift beyond 1.4. As already pointed out in previous K20 papers (Cimatti et al. 2002c; D04), no K20 galaxies were expected to lie at these high redshifts ($z\sim2$) based on current semi-analytical models of galaxy formation.
In addition to spectroscopy, deep and high quality imaging and photometry is available for this field, including ground-based $BVRIzJHK$ imaging with very good seeing (generally 0.4–0.7$''$) obtained with FORS1, FORS2 and ISAAC at the VLT (the same imaging dataset used in D04), together with the HST+ACS $bviz$ data released by the GOODS Team (Giavalisco et al. 2004). Shallower $U$-band imaging from Arnouts et al. (2001) was also used, obtained at the ESO 2.2m telescope. Photometric redshifts were computed for all galaxies with [*hyperz*]{} (Bolzonella et al. 2000) using the available multiwavelength photometry. These are updated with respect to D04 and Cimatti et al. 2002 because of the inclusion of the final ACS photometry ($bviz$) from GOODS. Fig. \[fig:BzK\_photoz\] shows the comparison of these photometric vs. the spectroscopic redshifts.
![Comparison of spectroscopic and photometric redshifts for the 311 extragalactic objects in the K20/GOODS area. Filled triangles show objects having $BzK>-0.2$ (Eq. \[eq:BzK\]), filled circles are old galaxies at $z>1.4$, circled symbols are X-ray detected sources and small symbols are $z<1.4$ (or $z>2.5$) galaxies. The photometric redshifts of old $z>1.4$ galaxies are slightly but systematically underestimated, probably because the Coleman et al. (1980) templates are too red with respect to these $\sim1$–2 Gyr old passive galaxies. Using the full library of BC03 SSP models with no dust reddening, the photometric redshifts of old $z>1.4$ galaxies become much more accurate and without systematic effects. []{data-label="fig:BzK_photoz"}](f1.eps){width="8.8cm"}
The $B$-, $z$- and $K$-band photometry, on which most of the paper is focused, is based on the Bessel $B$-band, F850LP $z$-band and the $K_{\rm s}$-band, (referred to as the $K$ band in the rest of this paper). The F850LP zeropoint was rescaled to match the photometry of the $z$-band imaging based on VLT+FORS1 Gunn-$z$ filter, that is considerably less deep than the GOODS $z$-band imaging. No color term was considered given the overall similarity of the two $z$-band filters whose effective wavelengths differ only by about 1%. Our ground based VLT Bessel $B$-band imaging is instead significantly deeper than the GOODS ACS imaging with F435W when measuring on apertures $\simgt1"$ comparable to the size of the K20 galaxies. The $K$-band data was obtained with VLT+ISAAC. Fig. \[fig:pfilter\] shows the total efficiency of the $BzK$ photometric systems as a function of wavelength. In order to allow a fine tuning of the photometry of objects from surveys using slightly different $BzK$ filter sets we make publicly available[^2] the $B-z$ and $z-K$ magnitudes of the stars identified in the GOODS area of the K20 survey. By matching the colors of the stellar sequence to the K20 one, it is possible to accurately apply the selection criteria described in Section \[sect:BzK\].
![The total transmission curves (including the detectors QE and atmospheric transmission) of the $BzK$ filters used to define the criteria for selecting $z>1.4$ galaxies. For the $z$-band filter, the HST+ACS curve is shown, with the VLT filter being very similar and only slightly less extended to the red. The best fit BC03 model to the SED of a galaxy at $z=1.729$ is shown for reference, on linear flux scale both for $F_\nu$ and $F_\lambda$. []{data-label="fig:pfilter"}](f2.eps){width="8.8cm"}
{width="16cm"}
In addition to optical-IR data, the publicly available deep 1 Msec Chandra X-ray observations of the area (Giacconi et al. 2002), and deep VLA radio maps (Kellermann et al. 2004), are used. Some details on the properties of the radio data and data analysis methods were summarized in Cimatti et al. (2003).
The K20/Q0055 Field
-------------------
The K20 spectroscopic dataset from the 19 arcmin$^2$ area centered on the QSO 0055-269 at $z=3.656$ (Q0055 hereafter) was also used as a valuable additional sample (Cimatti et al. 2002b). The spectroscopic completeness in the area is lower (89%). Of the 198 objects, 176 have a spectroscopic redshift identification at the moment, including 167 extragalactic objects and 9 stars. Of these, 13 lie at $z>1.4$ (8% of the galaxies) and 12 have $z_{\rm phot}>1.4$, for a total of 15% expected at $z>1.4$. The imaging (described in full detail in Cimatti et al. 2002b) has worse seeing (generally $\sim1''$) and is shallower than in the K20/GOODS area, as the data were mainly obtained at the ESO NTT 3.5m telescope (with SUSI2 and SOFI instruments), except for the Gunn $z$-band obtained with the VLT+FORS1 which has a similar depth and seeing to the one of the K20/GOODS area (although shallower than the ACS $z$-band imaging). For the above reasons, the photometry is less accurate for this field. In particular for the $BzK$ bands it turns out that in general the galaxies still have very good photometry in the $K$- and $z$-bands but the reddest galaxies have quite poor $B$-band photometry in the Q0055 field, as the 5$\sigma$ limits in the photometric apertures are $\sim26.0$ and $\sim27.6$ AB magnitudes for the K20/Q0055 and K20/GOODS regions respectively. X-ray, radio and HST data are not available for this field.
Near-IR color selection and classification of $1.4\simlt z\simlt2.5$ galaxies {#sect:BzK}
=============================================================================
The $BzK$ Criterion
-------------------
Fig. \[fig:BzK\] shows the $B-z$ vs. $z-K$ colors of the 311 galaxies and the 36 stars in the K20/GOODS sample. The classification of galaxies at $z>1.4$ as star-forming objects relies on \[OII\]$\lambda3727$ emission ($1.4<z\simlt1.7$), while those at $z>1.7$ have UV spectra showing the typical features of star-forming galaxies including e.g. the CIV absorption system at 1550 Å (D04, De Mello et al. 2004). It is found that $z>1.4$ star-forming galaxies occupy a narrow range and well defined region in this plane, well separated by lower redshift galaxies, with the bluest $B-z$ color at fixed $z-K$. By defining: BzK(z-K)\_[AB]{}-(B-z)\_[AB]{}, \[eq:BzK\] it follows that $z>1.4$ star-forming galaxies are all selected by the criterion:
BzK-0.2, \[eq:cond1\] i.e., to the left of the solid line in Fig. \[fig:BzK\]. In Fig. \[fig:BzK\] are also marked the spectroscopically confirmed passive systems at $z>1.4$. The classification of these old galaxies relies on the detection of significant continuum breaks and absorption features in the rest-frame 2500–3000 Å region (Cimatti et al. 2004). Being the reddest objects in both $B-z$ and $z-K$ colors, old stellar systems at $z>1.4$ can also be readily isolated in a BzK diagram using:
BzK<-0.2 (z-K)\_[AB]{}>2.5. \[eq:ES0\]
![The $BzK$ diagram for the Q0055 field of the K20 survey. Symbols are as in Fig. \[fig:BzK\], except that here circled symbols represent objects with a Type 1 AGN classification based on the optical spectra (X-ray data are not available for the Q0055 region). []{data-label="fig:Q0055_BzK"}](f4.eps){width="8.8cm"}
![ The redshift histogram of the 57 K20 galaxies selected with the criteria defined in Section \[sect:BzK\]. The shaded areas are for objects with photometric redshift only. The bottom panel shows the redshifts for all galaxies, center panel for the old objects and top panel for the star-forming ones. The contamination of galaxies at $z<1.4$ is only 12% of the sample and often consists of $z\sim1$ X-ray luminous galaxies, likely AGN. []{data-label="fig:zhisto"}](f5.eps){width="8.8cm"}
![The $BzK$ diagram for galaxies in the GDDS survey with $K<20.6$ (Abraham et al. 2004). Large symbols are for galaxies with spectroscopic redshift $1.4<z<2.2$, small squares for $z<1.4$ galaxies. []{data-label="fig:GDDS"}](f6.eps){width="8.8cm"}
![Top: the $BzK$ diagram for galaxies with $20<K<22$ in the GOODS-ISAAC region selected with $1.4<z_{\rm phot}<2.5$. Bottom: the photometric redshift distribution of galaxies with $20<K<22$ in the GOODS-ISAAC region selected with the $BzK$ criteria. It is not clear whether the narrow spike at $z_{\rm phot}\simeq 1.5$ is real or just an artifact of photometric redshifts. []{data-label="fig:IS2"}](f7.eps){width="8cm"}
All objects with $z_{\rm phot}>1.4$ are also selected by the above criteria, as evident from Fig. \[fig:BzK\]. Thus, the overall $BzK$-selected sample includes 25 star-forming galaxies at $z>1.4$ having $BzK\geq-0.2$ (15 $z_{spec}$ and 10 $z_{\rm phot}$) and 7 old galaxies at $z>1.4$ having $BzK<-0.2$ and $z-K>2.5$ (4 $z_{spec}$ and 3 $z_{\rm phot}$). The above criteria are therefore quite efficient in singling out $z>1.4$ galaxies, as the lower redshift [*interlopers*]{} are only 13% of the resulting samples, i.e. 5 objects (including 3 Chandra sources at $0.8<z<1.2$ and 2 star-forming galaxies at $1.2<z<1.4$). It is not unexpected that X-ray luminous objects, i.e. AGN, may contaminate these samples as a similar selection technique to the one devised here was proposed to identify luminous QSOs (Sharp et al. 2002). A QSO at $z=2.8$ (not highlighted in Fig. \[fig:BzK\]) is instead not selected by the method. Stars have colors that are clearly separated from the regions occupied by galaxies (and in particular by those at $z>1.4$), and can be efficiently isolated with the criterion: $(z-K)<0.3(B-z)-0.5$ (dotted diagonal line in Fig. \[fig:BzK\]).
The Q0055 dataset was used as an independent verification for the validity of the $BzK$ selection. Fig. \[fig:Q0055\_BzK\] shows the resulting $B-z$ versus $z-K$ diagram analogue to Fig. \[fig:BzK\]. Also in this field 18/23 of the galaxies with either spectroscopic or photometric redshift $1.4<z\simlt2.5$ are selected by the method. A few $z>1.4$ objects remain marginally out of the $BzK$ selection regions. Most of these have either $1.4<z<1.5$ or very poor $B$-band photometry and all are consistent with lying in the selection regions within 1–$1.5\sigma$. Comparison of Fig. \[fig:BzK\] and \[fig:Q0055\_BzK\] clearly shows that the objects in the $BzK>-0.2$ region are more scattered out in the K20/Q0055 field than in the K20/GOODS area because of the worse quality of the photometry. Two objects, including a galaxy and an AGN at $z>3$ (not highlighted in Fig. \[fig:Q0055\_BzK\]), are not selected by the criteria, that appear to have its main efficiency at $1.4<z\simlt2.5$, as justified from modeling in Section \[sec:modeling\]. The contamination from low-redshift galaxies is also here quite reduced. There are 2 objects having $BzK>-0.2$ that lie at $z<1.4$ and result to be a $z=1.367$ star-forming galaxy and an AGN at $z=1.119$. The criteria appear quite successful on the Q0055 dataset as well, once accounting also for the overall lower quality of the dataset, as discussed above.
These criteria thus allow a very efficient and highly complete selection of the 55 galaxies with either spectroscopic or photometric redshift $1.4<z<2.5$ in the K20 survey. The 57 galaxies selected with the $BzK$ criteria to $K<20$ correspond to a surface density of about $1.1\pm0.15$ arcmin$^2$ and have a redshift distribution mainly spread over $1.4<z<2.5$ (see Fig. \[fig:zhisto\]).
$BzK$-selected Galaxies in the GDDS and GOODS Fields
----------------------------------------------------
The $BzK$ selection was applied to other available samples in order to further verify its validity and test it at magnitudes fainter than $K\sim20$.
The Gemini Deep Deep Survey (GDDS; Abraham et al. 2004) performed spectroscopy of galaxies selected to $K<20.6$ (Vega). They spectroscopically observed a fraction of their $K$-selected sample favoring the objects with the redder colors. Two of the GDDS fields (SA12 and SA15) have $BzK$ photometry. After converting the GDDS $B$ and $K$ band photometry from Vega to AB scale, Fig. \[fig:GDDS\] shows that 7/9 GDDS galaxies with $z>1.4$ can be selected with the $BzK$ technique, with only two contaminants from $z<1.4$. Two galaxies with $z>1.4$ are just outside the $BzK$ selection regions. The photometric errors in the $BzK$ magnitudes in such a catalog are on average significantly larger than for the K20/GOODS galaxies with $z>1.4$. Within the errors, also the two outliers are consistent with the $BzK$ criterion.
As a further check, the galaxies within the deep ISAAC imaging of GOODS were considered at depths $K>20$ and up to $K=22$, i.e. two magnitudes fainter than reached by the K20 survey, for the same 32 arcmin$^2$ region covered by the K20 survey. Only the objects with well determined SEDs were included, requiring errors smaller than 0.3, 0.15, 0.15 mags for the $B$-, $z$-, and $K$-bands, respectively. This ensures reasonably reliable photometric redshift determinations, that were obtained using [*hyperz*]{} in a similar way as for the brighter K20 galaxies and using the same $UBVRIzJHK$ imaging datasets. The $BzK$ colors of the 125 galaxies selected to have $1.4<z_{\rm phot}<2.5$ are shown in Fig. \[fig:IS2\], where they indeed concentrate in the region with $BzK>-0.2$. About $10$% of them are just marginally outside the $BzK$ selection region. No red, passively evolving galaxies and very few red $BzK>-0.2$ galaxies are identified in the sample, most likely because of the adopted, stringent criterion on the photometric errors. Fig. \[fig:IS2\] also shows that the photometric redshift distribution of the 159 galaxies selected with the $BzK$ criteria (for a lower limit to their sky density of $\simgt5$ arcmin$^{-2}$ at $K=22$) is indeed centered at $z\sim2$. Only about 10% of the galaxies are at $z_{\rm phot}<1.4$, while 15% of them has $z_{\rm phot}>2.5$. The galaxies in the deep samples of Fig. \[fig:IS2\] have a median of $K(Vega)=21.2$.
These two checks are satisfactory and support the idea that the method is valid also for samples selected at magnitudes somewhat fainter than $K=20$. Such a validity should be further tested with future surveys.
{width="16cm"}
![$BzK$ versus redshift for galaxies in the K20/GOODS area. Symbols are as in Fig. \[fig:BzK\], except that here all objects with photometric redshift only are shown as empty squares. Model tracks are also over-plotted showing the expected $BzK$ color as a function or redshift. Constant star-formation rate models (solid lines) are shown for ages between 0.1 and 2 Gyr and reddening $E(B-V)=0.3$ (but note that the $BzK$ color is nearly reddening independent at $z>1$). Also shown is color evolution for evolving stellar populations formed in an instantaneous burst at redshifts $z=2$, 3 and 6 (dotted lines) and the variation of color with redshift for the (non evolving) templates of E, Sbc, Scd, and Im local galaxies from Coleman et al. (1980, CWW). []{data-label="fig:BzKvsz"}](f9.eps){width="8.8cm"}
Stellar population modeling {#sec:modeling}
===========================
Bruzual & Charlot (2003) stellar population synthesis models were used to further elucidate the physical meaning and the validity of these, phenomenologically established, $BzK$ criteria.
Galaxy Models in the $BzK$ Diagram
----------------------------------
All the four panels of Fig. \[fig:4p\] reproduce the $B-z$ and $z-K$ range of Fig. \[fig:BzK\] for different set of models.
The top-left panel shows the location of constant star-formation (CSF) models, computed for ages from $10^{-3}$ to 2 Gyr, $1.4<z<2.5$, various levels of reddening (using the Calzetti et al. 2000 extinction law), and solar metallicity. The figure confirms that galaxies in such a redshift range with ongoing star formation are indeed expected to lie in the $BzK>-0.2$ region of the diagram. The duration of the star-formation (age) has little influence on the $B-z$ color, while $z-K$ increases with age, an effect due to the development of strong Balmer/4000 Å breaks falling beyond the $z$-band for $z\approx2$. Very young bursts with age less than 10 Myr (and no underlying older stellar populations) would be located just below the threshold. Galaxies with similar properties are not present in the K20 database ($K<20$) and are at most only a minority in the GOODS photometric redshift sample to $K=22$ (Fig. \[fig:IS2\]). In order to include also such objects one should formally require $BzK\simgt-0.8$. In the K20 sample, this would increase significantly the contamination by $z<1.4$ galaxies.
The reddening vector is virtually parallel to the $BzK=-0.2$ limiting line, implying that dust affects the $B-z$ and $z-K$ colors by [*the same amount*]{} for $1.4<z<2.5$ galaxies. This appears to be the reason why the $BzK>-0.2$ criterion is successful at identifying $z>1.4$ star forming galaxies, virtually regardless of their dust reddening. It can be noticed that for $E(B-V)=0$ the tracks lie near the edge of the diagram, where no galaxies are found in the K20 sample (see Fig. 3), suggesting that purely unreddened, star-forming galaxies are rare at least in a $K$-selected sample at the relatively bright $K<20$ limit. Instead, in the $K<22$ sample from GOODS (Fig. \[fig:IS2\]), some fainter objects start to occupy also the region with $0<(z-K)_{AB}<1$.
The top-right panel of Fig. \[fig:4p\] shows the location of simple stellar population (SSP) models in the $BzK$ diagram, for ages of 0.1 to 2 Gyr, no reddening, and solar metallicity. For young ages ($\simlt0.2$) Gyr the tracks are marginally redder but similar to those of star-forming galaxies. As ages grow to $\simgt 1$ Gyr the tracks occupy the region where passive $z>1.4$ galaxies are detected, as expected. There is an intermediate age regime at $\sim0.5$ Gyr for SSP models in which such objects would be missed by both criteria of Section \[sect:BzK\] for $z\approx 2$. SSP models with young ages (i.e. comparable to the duration of major star formation events in real galaxies) and no reddening might be an unrealistic schematization, as real young galaxies are likely to be to some extent still star-forming and dust-reddened.
As a more reasonable rendition and in order to explore the redshift and aging effects, some evolutionary models were computed with star formation histories more extended in time, as described in Daddi et al. (2000b). Left-bottom panel of Fig. \[fig:4p\] shows the $BzK$ colors in $1.4<z<2.5$ for galaxies with various formation redshifts and exponentially declining SFRs ($\tau=0.3$ and 1 Gyr), with no reddening and solar metallicity. For high formation redshifts ($z_f>5$) the implied color evolution at $z>1.4$ is such that objects move directly from the star-forming galaxy region ($BzK>-0.2$) to the passive galaxy region ($BzK<-0.2$ and $z-K>2.5$) without crossing the bluer regions populated by $z<1.4$ objects.
The last panel of Fig. \[fig:4p\] finally shows that the color of normal galaxies at $0<z<1.4$, computed using the Coleman et al (1980) templates of E-Sbc-Scd-Irregular galaxies, are indeed expected to fall outside the range defined for $z>1.4$ and bracket quite well the range of colors observed for K20 galaxies at $z<1.4$ (see Fig. \[fig:BzK\]). This additional test strengthens the validity of the $BzK$ selection to isolate galaxies at $z>1.4$.
Modeling $BzK$ versus Redshift
------------------------------
The key quantity in the selection and classification of $z>1.4$ galaxies is the $BzK$ term defined in Eq. \[eq:BzK\]. Fig. \[fig:BzKvsz\] shows the $BzK$ evolution as a function of redshift for the CSF models described above. Objects with $BzK\simgt0$ start to appear in significant numbers only beyond $z>1.4$. The CSF models enter the region $BzK>-0.2$ at $z\simgt1.2$, although models with ages of 1–2 Gyr can marginally fulfill the $BzK>-0.2$ condition even at much lower redshifts. CSF models evolve out of the $BzK>-0.2$ region at $z>2.6$–3.2, depending on age, because the Ly$\alpha$ forest starts entering the $B$-band at those redshifts, thus producing a reddening of the $B-z$ colors. Also shown in the figure are $BzK$ colors expected for passively evolving (SSP) galaxies formed at $z=2$, 3 and 6, and templates of local galaxies (Coleman et al. 1980). Again, one notices that in general passively evolving galaxies are contaminating $BzK>-0.2$ samples only for very young ages close to the formation redshift. The Coleman et al. (1980) templates bracket the $BzK$ color range observed for $0<z<1$ galaxies, as well as that of higher redshift $z\sim2$ galaxies (although they have too old stellar populations to be truly representative of $z\sim2$ galaxies).
The Effects of Metallicity and Extinction Laws
----------------------------------------------
We also explored the effects of using alternative choices for the metallicity and the extinction law. Models with metallicity significantly below solar seem inappropriate even for $z\sim2$ star-forming systems with $K<20$, as these objects show deep photospheric absorption spectra indicative of solar or higher metallicity (De Mello et al. 2004). Old systems are consistent with being fully assembled spheroids, that are known to have nearly solar or higher metallicity today. We checked that using above-solar metallicities the CSF galaxy tracks are basically unchanged, while the tracks of passive galaxies change according to the well known age/metallicity degeneracy. We also investigated the effect of using extinction laws other than that of Calzetti et al. (2000). Using the extinction law proposed by Silva et al. (1998) yields results fully consistent with those obtained above with the Calzetti et al. law. The SMC extinction curve produces a higher reddening to the $B-z$ color than to the $z-K$ color, so that for very high reddening ($E(B-V)\simgt0.6$) the model tracks would enter the $BzK<-0.2$ region of passive galaxies. However, with the SMC extinction law the colors of the reddest galaxies with $BzK>-0.2$ are difficult to reproduce.
The Nature of the Reddest Galaxies with $BzK>-0.2$
--------------------------------------------------
Most of the K20 galaxies with no spectroscopic redshift available and $1.4<z_{\rm phot}<2.5$ have very red $(z-K)_{AB}>2.5$ colors and $BzK>-0.2$ (Fig. \[fig:BzK\]), qualifying thus as star-forming galaxies based on the proposed classification criteria. The top-left panel of figure \[fig:4p\] confirms that such objects are fully consistent with being heavily reddened, star-forming galaxies. Their full multicolor SEDs cannot generally be fitted by models for old/passive galaxies with no star-formation and reddening, implying that some amount of young-hot stars is required for them to show the relatively high $B$-band fluxes and blue $B-z$ colors. This is quite reasonable, as they appear to follow the trend of increasing reddening for the spectroscopically established star-forming galaxies at $z>1.4$. Nevertheless, some of the objects in that region may actually be post-starburst galaxies, having passed their strongest episode of star-formation, in which case part of their red colors could be due to an aged burst of star-formation (see bottom-left panel of Fig. \[fig:4p\]). Additional evidence that, typically, these are indeed actively star-forming galaxies will be derived from their average X-ray and radio properties in Section \[sec:Xr\].
HST/ACS Morphology of the $z>1.4$ Galaxies {#sect:HST}
==========================================
HST imaging provides a fundamental complement to investigate the nature of the $BzK$ galaxies and to elucidate their evolutionary status. In Fig. \[fig:morphoBzK\] ACS z-band imaging of the $z>1.4$ galaxies in the K20/GOODS sample are presented. The $z$-band is centered at rest-frame wavelengths from 2500 Å to 3700 Å for the objects in $1.4<z<2.5$. Objects with $BzK\geq-0.2$ appear generally irregular/merging-like and have very large sizes, with an average half-light radius of about 6 kpc at $z\sim2$ (D04). Objects with $BzK<-0.2$ and $z-K>2.5$ have instead generally a compact and regular morphology. With only a few exceptions, there is a very good agreement of the [*early-type/late-type*]{} morphological appearance with both the $BzK$ color classification and with the spectroscopic classification as passive or star-forming galaxies. This supports the evidence that the $BzK$ criteria allow to efficiently isolate high redshift galaxies in a $K$-selected sample and to distinguish passive and star-forming objects.
Star-formation Rates {#sec:SFR}
====================
The critical question is then to investigate the level of star formation activity present in the $BzK$ (star-forming) galaxies. This is done in this Section where we measure the SFRs of galaxies at $1.4<z<2.5$ in the K20/GOODS region (hence with $K<20$), selected and classified as star-forming with the criterion $BzK>-0.2$, and we explore whether the $BzK$ photometry alone can still provide an estimate of the SFRs of these galaxies.
SFRs from the Rest-Frame UV-Continuum Luminosity {#sect:ebv}
------------------------------------------------
Estimating the SFR of a galaxy from its multicolor optical photometry is generally based upon relations between the intrinsic UV continuum luminosity and SFR (e.g. Madau et al. 1998) and estimating the extinction by dust, necessary to derive the intrinsic UV continuum luminosity from the observed one. Both steps are in general quite uncertain and rely on assumptions about star-formation history, dust reddening law, metallicity (as well as on the IMF, fixed to Salpeter between 0.1 and 100 $M_\odot$ in this work, but in a way that is generally easy to factor-out so that the results can be easily scaled to other choices of IMF).
Neglecting any reddening correction, the SFRs are of order of $\approx10$–40 $M_\odot$yr$^{-1}$ for the spectroscopically confirmed galaxies (D04), and even smaller for the reddest ones with only photometric redshifts. As well known, such estimates are severely affected by dust extinction. In this section, we attempt to infer the level of dust reddening from the photometric properties, limiting this analysis to the case of CSF models with solar metallicity and a Calzetti et al. (2000) extinction law (as already done in D04). These assumptions appear reasonably justified for our galaxies, as discussed in Section \[sec:modeling\], and allow a comparison with a broad variety of literature work based on the same assumptions.
With the above assumptions and following the approach of D04, best-fitting SFRs and $E(B-V)$ have been derived for the 24 purely star-forming galaxies with $z>1.4$ in the K20/GOODS area, from their full observed SED from $U$ to $K$ (an object with AGN signatures in the spectrum and high X-ray to optical luminosity ratio has been excluded from the SFR analysis). The derived SFRs are typically in 100–600 $M_\odot$yr$^{-1}$ and the reddening range is $0.2\simlt
E(B-V)\simlt1$.
![ The $B-z$ color is plotted vs. the reddening $E(B-V)$ for the $BzK$-selected star-forming galaxies in the K20/GOODS region. The best fitting $E(B-V)$ values from the full SED analysis are shown for individual objects. Circles with error bars: objects with spectroscopic redshifts; triangles: objects with only photometric redshifts. Also shown is the $B-z$ color vs. $E(B-V)$ for constant star-formation rate, 500 Myr old models and various redshifts within the range $1.4<z<2.5$ (filled circles connected by vertical lines). The 500 Myr age is the typical SED best-fitting age for K20 $z\sim2$ star-forming galaxies (D04). The diagonal line shows the relation defined in Eq. \[eq:redde\]. []{data-label="fig:redde"}](f11.eps){width="8.8cm"}
Then, we have explored if, within the same assumptions, the $BzK$ photometry alone could allow an estimate of the SFR content of the 24 galaxies equivalent to the one derived from SED fitting. At $z\sim 2$ the $B$-band samples quite well the rest-frame ultraviolet at $\sim 1500$ Å (actually, the rest-frame 1250–1800 Å range for $1.4<z<2.5$), and the UV luminosity at 1500 Å of a star-forming galaxy is a calibrated measure of the ongoing SFR (e.g. Madau et al. 1998). Fig. \[fig:redde\] suggests that $E(B-V)$ estimated from SED fitting (and with the knowledge of the galaxy redshift) correlates very well with the observed $B-z$ color (see also Fig. \[fig:4p\], top left panel), following the relation:
E(B-V)=0.25(B-z+0.1)\_[AB]{} \[eq:redde\] that indeed for the models and assumptions described above holds as an average over redshift and age. The rms dispersion of the residuals is only 0.06 in $E(B-V)$ for the objects with measured spectroscopic redshift (mainly due to a single outlier, with dispersion dropping to 0.026 with such object removed), and the relation still holds quite well also for objects with photometric redshift only. This tight relation between $E(B-V)$ and $B-z$ is related to the assumption of the “grey” and self-similar Calzetti et al. (2000) extinction law and to the fact that the UV shape of CSF models has little dependence on age.
The observed flux at 1500 Å rest-frame, de-reddened using Eq. \[eq:redde\], can be used to estimate the 1500 Å luminosity once the redshift is known. For the BC03 models this can be converted into a SFR using the relation:
(M\_\^[-1]{}) = L\_[1500 Å]{} \[[erg s\^[-1]{} Hz\^[-1]{}]{}\] / (8.8510\^[27]{}). \[eq:madau\]
For objects lacking a spectroscopic (or even photometric) redshift identification, this conversion can be done only assuming an average redshift for the sample galaxies. For individual objects, if they lie at lower (higher) redshift than the average both the luminosity distance and 1500 Å luminosity (based on just the observed $B$-band flux) will be overestimated (underestimated), typically by factors up to $\sim 2$. However, this procedure would still allow to derive a fairly correct ensemble average of the SFR in the survey when a statistical sample of galaxies is considered together.
With the above recipes we have derived a $BzK$ color estimate of the SFR of each galaxy in the following way: 1) the observed $B$-band flux is used as a measure of the 1500 Å UV continuum flux; (2) $E(B-V)$ is derived from the observed $(B-z)_{AB}$ color as in Eq. \[eq:redde\]; and (3) the average redshift of the sample $<\! z\!>=1.9$ is used for all objects to derive the SFR following Eq. \[eq:madau\].
For the 14 K20 star-forming galaxies with $1.4<z_{\rm spec}<2.3$, the SED fitting and the simple $BzK$ color estimates described above are all in agreement within a factor of 2, given our range of redshifts, with a rms fluctuation of only 20%. The total SFR from the 14 galaxies results to be 3600 $M_\odot$yr$^{-1}$ from SED fitting and 3400 $M_\odot$yr$^{-1}$ with the $BzK$-based estimate, in excellent agreement between them and indicating a quite high average SFR$\sim250$ $M_\odot$yr$^{-1}$ for these galaxies. The SFRs derived for objects with only photometric redshifts are more uncertain, but also in that case the agreement among the two estimates is reasonable. The SFR estimated for the 10 galaxies with $BzK\geq-0.2$ and $z_{\rm phot}>1.4$ (likely $z\sim2$ star-forming galaxies with high reddening) is $\sim 1700\; M_\odot$yr$^{-1}$ in total, corresponding to SFR$\sim170$ $M_\odot$yr$^{-1}$ per object. Averaging over the two samples yields SFR$\sim210$ $M_\odot$yr$^{-1}$ for the typical objects among this population of $K$-selected starbursts.
These results show that, within the assumptions made, the total SFR content of $1.4<z<2.5$ galaxies can be estimated from the $BzK$ photometry alone with an accuracy similar to that reachable by fitting to the whole SEDs with known spectroscopic redshifts. However, we notice that assuming exponentially declining star-formation models the amount of reddening and SFR can be significantly reduced.
In order to derive more stringent clues, X-ray and radio data were also used to derive independent estimates of the SFR unaffected by dust extinction and to test the above results.
SFR from the X-ray Luminosity {#sec:Xr}
-----------------------------
Alternative measures of the SFR of galaxies can be obtained from their X-ray and radio properties, as the X-ray and radio luminosities of star-forming galaxies (with no major AGN contribution) are proportional to the SFR (e.g., Condon et al. 1992; Ranalli et al. 2003; Nandra et al. 2003). The X-ray and radio properties also offer an additional opportunity (besides optical spectra) to check for the presence of AGN contamination.
Two of the K20 objects at $z>1.4$ with $BzK>-0.2$ are listed as detections in the catalog based on the 1 Msec Chandra Deep Field South observations (Giacconi et al. 2002). One of these is the object with AGN line optical spectrum, and we already mentioned that this was excluded from the star-forming galaxy sample. Another galaxy at $z=2.223$ with a faint soft X-ray detection is present in the sample. This is also detected as a faint radio source at 1.4 GHz and it is consistent with being a vigorous starburst with SFR$\simgt500$ $M_\odot$yr$^{-1}$ (D04).
The X-ray emission in the observed soft (0.5–2 keV) and hard (2–10 keV) bands have been measured at the position of the remaining 23 star-forming objects to check for other possible detections. No other individual detection is found above the $3\sigma$ level. The stacked X-ray signal from the 23 individually undetected sources was then obtained to constrain the average X-ray emission of the $z\sim2$ star-forming galaxies. In the soft band $\sim96\pm23$ net counts are recovered, after background subtraction. We performed Monte Carlo simulations by placing at random positions in the X-ray image (excluding regions around known sources) and found that the chance probability of recovering such a strong signal is $1.7\times10^{-5}$. The average 4.4 soft counts per objects are close to the detection limits of the 2 Msec Chandra observations in the HDF North (Alexander et al. 2003). Performing the analysis separately on galaxies with or without spectroscopic redshift identification, it is found that the two samples have not statistically different X-ray properties and both samples are positively detected at the $\sim 3\sigma$ level in the soft band. On the other hand no significant detection is found from the stacked hard band data, constraining the hardness ratio of the population to be HR$<-0.54$ at the 2-sigma level. This is consistent with the low hardness ratio expected for starbursts galaxies. AGN are found generally to have $-0.5<HR<0.5$ (Szokoly et al. 2004). The low average HR for our sources thus disfavors that the detected soft X-ray emission is due to low-level AGN activity. A similar conclusion is supported by the low average X-ray-to-optical flux ratio of ${\rm
log} (f_{0.5-2 keV}/f_R) \sim -1.5$ and by the lack of AGN signatures in the spectra (see also D04). The X-ray emission is therefore most likely due to star-formation. Using $\Gamma=2.1$ appropriate for starbursts (e.g. Brusa et al. 2002), the counts correspond, for $<\!z\!>=1.9$, to a rest frame 2–10 keV luminosity of $L_{2-10 keV}=8.6\times10^{41}$ ergs s$^{-1}$, which translates into an average SFR$\sim170$ $M_\odot$yr$^{-1}$ (Ranalli et al. 2003; Nandra et al. 2002). When adding back to the sample the individually X-ray detected (non AGN) object one obtains an average X-ray luminosity corresponding to an average SFR$\sim190$ $M_\odot$yr$^{-1}$, in quite good agreement with the estimate from the reddening corrected UV luminosities and constant star-formation rate models.
SFR from Radio Luminosity {#sec:Rr}
-------------------------
Deep radio maps at 1.4 GHz and 5 GHz (Kellermann et al. 2004) were used to measure the radio properties of the $z\sim2$ star-forming galaxies in our sample. The radio data reach rms flux densities of about 8 $\mu$Jy at both 1.4 GHz and 5 GHz. Two of the star-forming $z\sim2$ galaxies are individually detected at 1.4 GHz at better than the 3$\sigma$ level. One of the two is the vigorous starburst with SFR$\simgt500$ $M_\odot$yr$^{-1}$ also detected in the X-ray and discussed in D04. The other object has a 1.4 GHz flux density of $\sim
25\,\mu$Jy and is therefore a $\sim 3\sigma$ detection. No individual object is detected at 5 GHz.
The average flux density of non-detections has been evaluated in a similar fashion as was done for the EROs by Cimatti et al. (2003), and using the same dataset. The radio flux densities were measured at the nominal optical position for each of the galaxies averaging the flux density in the beam (3.5$''$) over a range of 1$''$ radius in order to correct for possible residual coordinate mismatch. An average signal is measured of $7.4\pm1.8\mu$Jy at 1.4 GHz and $1.5\pm1.8\mu$Jy at 5 GHz, for the 22 $z>1.4$ $K$-selected star-forming galaxies that are individually undetected both in radio and X-ray. The above flux densities do not strongly constrain the radio continuum slope $\alpha$, but are consistent with the value of $-\alpha\sim0.6$–0.8 typical of starburst galaxies (Condon et al. 1992). For consistency with the work at $z=2$ of Reddy & Steidel (2004), a slope of $\alpha=-0.8$ is adopted to derive an average 1.4 GHz rest-frame luminosity of $16\times10^{22}$W Hz$^{-1}$, corresponding to an average SFR$\sim160$ $M_\odot$yr$^{-1}$ per object, using the relation given by Yun et al. (2004) and corrected for a binning error as in Reddy & Steidel (2004)[^3]. Including again into the sample the two starburst galaxies detected at 1.4 GHz we obtain an average SFR$\sim270$ $M_\odot$yr$^{-1}$ per object, again with reasonable consistency with both the optical and the X-ray estimates. Also in this case, no statistically significant difference is found for the average radio flux density of objects with or without known spectroscopic redshift.
![The stellar masses for the K20/GOODS objects at $z>1.4$ are shown as a function of $K$-band magnitudes. The two plots presented refer to each of the two methods discussed by Fontana et al. (2004) to estimate the galaxy stellar masses (see text). Symbols are as in Fig. \[fig:BzK\]. []{data-label="fig:MassCal"}](f12.eps){width="8.8cm"}
To summarize, all the available SFR indicators agree with each other and confirm the presence of $K$-band luminous $<\!z\!>\simeq 2$ star-forming galaxies with typical SFR$\approx200$ $M_\odot$yr$^{-1}$ and a median reddening of $E(B-V)\sim0.4$ (cf. D04).
The Stellar Mass of [K]{}-Selected Galaxies {#sec:masses}
===========================================
As part of the K20 project the stellar mass $M_*$ of each galaxy was estimated from the known redshifts and full multicolor photometry (Fontana et al. 2004; F04 hereafter). Using the F04 results, in this section we explore the possibility of estimating the stellar-mass content of $K$-selected galaxies at $z>1.4$ from the $BzK$ photometry alone.
Fig. \[fig:MassCal\] shows the results of two different stellar mass estimates from F04: one based on synthetic stellar population models fitting to the whole $UBVRIzJHK$ SED, and one based on fitting just the $R-K$ color. The latter approach is designed to provide an estimate of the maximal mass of each galaxy (see F04). The masses estimated with the SED-fit technique are in reasonable agreement with those for the objects at $1.7<z<2.3$ analyzed in D04.
Fig. \[fig:MassCal\] shows a plot of $M_*$ from both methods as a function of the observed $K$ magnitude for the sample of 31 out of the 32 objects with $z>1.4$ in the K20/GOODS sample. One object was excluded because exhibiting a clearly AGN dominated spectrum. Best-fitting linear relations between the stellar-mass and the observed $K$-band flux were obtained, in the form: (M\_\*/10\^[11]{} M\_) = -0.4(K\^[tot]{}-K\^[11]{}) \[eq:M1\] where $K^{11}$ is the $K$-band magnitude corresponding on average to a mass of $10^{11} M_\odot$. For the SED fit and single-color method we find $K^{11}=19.51$ and $K^{11}=20.14$ (Vega scale), respectively.
It can be noted that at $z>1.4$ the single-color method yields masses a factor of 1.7 higher, on average, than the SED-fit technique (see also F04). The rms dispersions observed for these relations are $\sigma(\Delta {\rm log} M_*)=0.25$ and 0.15 for the best fit and single color method, respectively. We searched for further correlations between the residuals in the masses $\Delta {\rm log} M_*$ as derived from Eq. \[eq:M1\] versus the F04 values, and the colors available from the $BzK$ photometry. No significant trend was noticeable for the masses derived with the single-color method. Instead, the residuals in the SED fitting derived masses do positively correlate with the $z-K$ color, with:
M\_\* = 0.218\[(z-K)\_[AB]{}-2.29\], \[eq:delta\] a term that would reduce the rms dispersion to 0.20 if added to the right-hand side of Eq. \[eq:M1\].
These relations allow to estimate masses with average uncertainties on single objects of about 40% and 60% relative to the single-color and the SED-fit method, respectively. This is an encouragingly good accuracy, given the large intrinsic differences in the luminosity distance and actual rest-frame wavelength sampled by the observed $K$-band, for objects within the $1.4<z<2.5$ range. Intrinsic differences in the $M/L$ ratio for given magnitudes and/or colors also contribute to increase the scatter. However, when averaging over large samples of galaxies these statistical fluctuations may be largely mitigated.
Note also from Fig. \[fig:MassCal\] that the stellar masses of dusty star-forming and old/passive galaxies are estimated to be on average quite similar at given observed $K$-band magnitude. This seems to happen by chance: the star-forming galaxies have lower mass to light ratio but their $K$-band light is attenuated by an amount which produces similar observed magnitudes to old objects with comparable stellar-masses. The mass of substantially obscured stellar populations within the galaxies would however obviously fail to be accounted for in such estimates.
It should be reminded that relations \[eq:M1\] and \[eq:delta\] were derived for $K<20$ galaxies, and it remains to be assessed whether they are also valid at fainter $K$ magnitudes.
The BK vs. Other High-$z$ Galaxy Selection Criteria {#sec:other}
===================================================
In this section, the properties of $BzK$-selected galaxies at $1.4<z<2.5$ having $K<20$ are compared to those of samples selected according to other color or multi-color selection criteria. We will consider the $U_{\rm n}GR_{\rm s}$ selection of $z=2$ galaxies, the Extremely Red Objects (ERO) selection based on the $R-K>5$ threshold, and the infrared-selected galaxies found with the criterion $J-K>2.3$ proposed by Franx et al. (2003) to isolate $z>2$ evolved galaxies.
![The $U_{\rm n}GR_{\rm s}$ two-color diagram for $1.4<z<2.5$ galaxies in the K20/GOODS region (top panel) that are selected by the criteria defined in Section \[sect:BzK\], and (bottom) fainter galaxies with $1.4<z_{\rm phot}<2.5$ and $20<K<22$ from the same K20/GOODS area (the sample also shown in Fig. \[fig:IS2\]). The $U_{\rm n}GR_{\rm s}$ colors were derived from the SED fitting. Symbols are as in Fig. \[fig:BzK\] (in the bottom panels all redshifts are photometric). In both panels, the color regions defined for the UV identification of $z\sim2$ are also shown (Steidel et al. 2003 for $z\sim3$; Erb et al. 2003, solid line, and Adelberger et al. 2004, dotted line, for $z\sim2$).[]{data-label="fig:LBG_z2"}](f13.eps){width="8cm"}
$BzK$- vs. UV-selected Galaxies at $z\sim2$ {#sec:UV}
-------------------------------------------
Very recently, the UV technique for selecting LBGs has been extended to $z<3$ using a $U_{\rm n}GR_{\rm s}$ two-color diagram which isolates star-forming galaxies at $1.4<z<2.5$ (Erb et al. 2003; Steidel et al. 2004; Adelberger et al. 2004), a redshift range fully matching that of the $BzK$ selection.
![ The differential contribution to the SFR density at $z\simeq 2$ from $BzK$ galaxies as a function of their $K$-band magnitude (for K20 galaxies in the K20/GOODS region). Note that this contribution is still increasing at the $K=20$ limit of the survey, suggesting a non-negligible contribution from unaccounted $K>20$ galaxies. Error bars are purely Poissonian. []{data-label="fig:SFRplot"}](f14.eps){width="8.8cm"}
The UV-selection requires the UV continuum to be relatively flat, thus limiting the overall dust extinction to $E(B-V)\simlt0.3$ (Adelberger & Steidel 2000), while many of the K20 galaxies at $z>1.4$ are more reddened (see Fig. \[fig:redde\]). We estimated how many of the galaxies at $1.4<z<2.5$ could also be selected by the UV criterion, in our sample. As no $U_{\rm n}GR_{\rm s}$ photometry is available to us for the K20 galaxies, synthetic $U_{\rm n}GR_{\rm s}$ magnitudes have been derived from the BC03 models providing the best fit to the observed $UBVRIzJHK$ SEDs, for the objects in the K20/GOODS region. Fig. \[fig:LBG\_z2\] shows the resulting synthetic $(G-Rs)$ vs. $(U_n-G)$ colors. Of the 32 $K<20$ objects at $1.4<z<2.5$ in Fig. 3, only 2 (6%) would be selected by the Erb et al. (2003) criteria, and 9 (28%) by the Adelberger et al. (2004) criteria, corresponding to a surface density of about 0.3 arcmin$^{-2}$. We can compare these numbers with those in the UV-selected surveys. The Adelberger et al. (2004) BM and BX criteria result cumulatively in a sky density of 9 arcmin$^{-2}$ candidate $z\sim2$ galaxies (Steidel et al. 2004). Roughly 90% of these is found within $1.4<z<2.5$ and about 8% of the $z\approx2$ objects has $K<20$. Based on the above numbers, we should have found $\sim20$ galaxies UV-selectable with the Adelberger et al. (2004) criterion in $1.4<z<2.5$ in the K20/GOODS region, while only 9 are recovered. It is not clear what is the reason of this possible discrepancy, that may be in part due to small number statistics and/or to cosmic variance due to clustering (D04).
Some large fraction (perhaps as high as $\sim 70\%$) of the $K<20$ galaxies at $z>1.4$ fail to be selected by the UV criteria to identify galaxies at $z\sim2$. The lost fraction includes not only the old passive systems but also a high proportion of actively star-forming, highly reddened galaxies. As a consequence, the UV-selection fails to recover most of the stellar mass in $K<20$ galaxies at $z=2$, as expected given that it was not devised with the aim of probing the galaxy mass density, but rather the star-formation rate (Adelberger et al. 2004). However, a significant amount of the SFR density is also missed. Using the SFRs estimated from reddening-corrected UV luminosities (Section \[sect:ebv\]), it is found that the 9 $K<20$ starbursts in our sample that satisfy the $U_{\rm n}GR_{\rm s}$ criteria produce only $\approx 15$% of the SFR density at $z=2$ from $K<20$ galaxies, with the residual $\approx 85\%$ being missed because of dust reddening in excess of $E(B-V)\approx0.3$. The missed objects include all the K20 galaxies with the most extreme starbursts with SFR$>200
M_\odot$ yr$^{-1}$.
Comparing the typical star-formation rate level per galaxy, the bright $K<20$ starbursts in the K20 survey appear to be forming stars more vigorously than the UV-selected not strongly reddened galaxies with $K<20$ (Shapley et al. 2004), with average SFRs larger by a factor of $\approx3$. Extending the comparison including also fainter UV selected galaxies, the radio and X-ray measurements (both in agreement with the extinction-corrected estimates based on the UV continuum), imply that the average radio and X-ray luminosities, hence the SFRs, of our $K<20$ $z=2$ star-forming objects are higher than those of the average of all UV selected galaxies (Reddy & Steidel 2004) by a factor of $\sim 4$. The $U_{\rm n}GR_{\rm s}$ selected galaxies studied by Reddy & Steidel (2004) have a significantly higher space density than the $K<20$ galaxies and sample regimes with much lower SFRs.
Contributions to the $z\sim2$ Star-formation Rate Density
---------------------------------------------------------
In this section we derive the contribution of the $BzK$-selected galaxies to the [integrated]{} star formation rate density (SFRD) at $z\sim2$, and compare it to an estimate of the SFRD derived from the UV-selected galaxies.
For the volume in the redshift range $1.4<z<2.5$, a SFRD of $0.044\pm0.008$ $M_\sun$ yr$^{-1}$ Mpc$^{-3}$ is derived from the 24 K20 star-forming galaxies fulfilling the $BzK>-0.2$ criterion (and of course $K<20$), where the error is derived from bootstrap resampling. This may well be an underestimate of the error, given the small number of galaxies used in the estimate, and to them belonging to a population likely to be strongly clustered (D04). For example, assuming that these galaxies are as clustered as $z\sim1$–3 red galaxies ($r_0\simlt10$ 1 Mpc; e.g., Daddi et al. 2001; 2003), the error would become 35% larger on the lower side and a factor of 3 larger on the upper-side.
This estimate of the SFRD contributed by the $BzK>-0.2$ objects with $K<20$ is comparable to the global SFRD at $z\approx2$ as estimated from other surveys (in the same units: $\sim0.08$, Connolly et al. 1997, corrected for extinction by Steidel et al. 1999; $\sim0.055$, Heavens et al. 2004), or as predicted by $\Lambda$CDM semi-analytical models (0.05–0.10 $M_\sun$ yr$^{-1}$ Mpc$^{-3}$, Somerville et al. 2001) and by $\Lambda$CDM hydrodynamical simulations ($\sim0.055$ $M_\sun$ yr$^{-1}$ Mpc$^{-3}$, Hernquist & Springel 2003). However, our present estimate must be incomplete because it does not include the contributions of all the $K>20$ galaxies, and in particular of those still fulfilling the $BzK>-0.2$ condition. Fig.\[fig:SFRplot\] shows that the SFRD is not yet converging by $K\sim20$, and a significant additional contribution from $K>20$ galaxies is therefore expected.
The [total]{} SFRD produced by UV-selected galaxies at $z\sim2$ in the Steidel et al. (2004) sample has not been published yet, but a crude estimate can be derived in comparison to the $BzK$ star-forming galaxies at $K<20$ by considering that the $z\sim 2$ $U_{\rm n}GR_{\rm s}$-selected candidates down to $R_{\rm s}=25.5$ have $\approx10$ times higher sky density (Steidel et al. 2004), that $\sim
90$% of them are in the redshift range covered by the $BzK$ selection ($1.4<z<2.5$), and that they have $\sim 4$ times smaller average SFRs. This would yield an integrated contribution by UV-selected galaxies (with $R_{\rm s}<25.5$) a factor of 2–2.5 times larger than that of bright $K<20$ $BzK$ galaxies. Hence, taking into account that some galaxies are picked by both criteria, the UV selection may miss of order of $\sim 20$–30% of the total SFRD provided by galaxies selected by at least one the two criteria, while the $BzK$ criterion limited to $K<20$ allows to select only a similar fraction.
On a broader perspective, discussing the potentials of the two color criteria ($BzK$ and $U_{\rm n}GR_{\rm s}$) is perhaps more interesting than comparing the existing samples drawn with them. As mentioned above, much of the limitation of the present application of the $BzK$ criterion actually comes from the fairly bright limiting $K$ magnitude, rather than from the color criterion itself. One may expect that applying it to fainter $K$ magnitudes a higher fraction of the total SFRD could be recovered. It is not presently known, however, if some fraction of $z=2$ galaxies would be missed by the $BzK$ criterion, especially at $K>20$, where it may start loosing some very young starburst, as suggested by the top-left panel in Fig. 8. If they exist at faint $K$ magnitudes, such young galaxies could be more easily selectable in the UV. These points should be tested by future surveys[^4].
The UV-selection appears to miss the most actively star-forming galaxies not because of the limiting $R$ magnitude, but because they are much too reddened \[$E(B-V)\simgt0.3$\] for satisfying the $U_{\rm
n}GR_{\rm s}$ color selection. The existence of highly reddened star-forming galaxies also at $K>20$ would imply for the UV-selection additional losses of star-forming galaxies (hence of part of the SFRD). A preliminary analysis based on the GOODS/ISAAC sample at $K<22$ (Fig. \[fig:IS2\]) suggests that at $K>20$ star-forming galaxies with $z_{phot}\sim2$ and progressively bluer colors start to appear, and occupy the bluest $U_{\rm n}GR_{\rm s}$ region where most UV-selected galaxies also lie (Fig. \[fig:LBG\_z2\], see also Fig. 11 of Adelberger et al. 2004 for comparison). This may be consistent with a general blueing trend at fainter magnitudes, indicative of a trend to lower reddening. The fractional SFRD lost by the UV selection may then therefore decrease with increasing $K$ magnitude limit. It should be mentioned, however, that the paucity of red star-forming (as well as of passive) galaxies in the sample of Fig. \[fig:IS2\] is at least in part due to the requirement of accurate photometry for the photometric redshifts determination and that, e.g., candidate star-forming galaxies with $BzK>-0.2$, $z_{\rm phot}\approx2$, and $2<(z-K)_{AB}<4$ are found also down to $K=22$. In addition, a population of faint red galaxies at $z\simgt2$ appear to exist even down to $K=24$. They tend to be more clustered than LBGs, as expected for the precursors of early-type galaxies (Daddi et al. 2003). In summary, an application of the $BzK$ technique to much fainter $K$-selected surveys could shed light on the amount of reddened star-formation at rates lower than probed by the K20 survey, and better establish the fractions of the SFRD recovered by each of the two criteria.
$BzK$-selected Galaxies and Extremely Red Objects
-------------------------------------------------
A simple method to select relatively high redshift galaxies relies on requiring very red optical to near-IR colors, typically $R-K>5$ (e.g. Elston, Rieke & Rieke 1988; Hu & Ridgway 1994; Thompson et al. 1999; Daddi et al. 2000a; Roche et al. 2002, 2003; for a comprehensive review see McCarthy 2004). K20 survey spectroscopy unveiled for the first time the nature of EROs in a sizable sample, and showed that EROs include similar fractions of old and dusty star-forming systems (Cimatti et al. 2002a, 2003; see also Yan et al. 2004).
The redshift range is one of the main differences among the samples produced with the two $BzK$ and ERO methods. Fig. \[fig:EROS2\] shows that EROs with $K<20$ are found at $0.8\simlt z\simlt2.5$, coherently with the rationale for their selection (e.g. Daddi et al. 2000a). Many K20 galaxies exist in the same redshift range that are not EROs. In contrast, the strength of the $BzK$ selection (Eq. \[eq:cond1\] and \[eq:ES0\]) is that it provides a fairly complete sample of galaxies in the redshift range $1.4<z<2.5$. The $z-K>2.5$ condition of Eq. \[eq:ES0\] that allows to recover old passive galaxies is basically equivalent to the ERO criterion $R-K>5$, apart from the higher low-redshift cutoff ($z\simgt1.4$ instead of $\simgt 0.8)$. A very similar $z>1.4$ cutoff for passive galaxies would be obtained requiring $R-K\simgt6$, albeit with a larger contamination by both lower redshift and reddened star-forming galaxies. Fig. \[fig:EROS2\] shows that only about 50% of $BzK$ selected galaxies at $z>1.4$ have EROs colors, while only 35% of all the EROs are selected with the $BzK$ criteria, i.e. lie at $z>1.4$.
![The $R-K$ (Vega) color versus redshift (top) and versus $BzK$ (bottom), for galaxies in the K20/GOODS region. Symbols are as in Fig. \[fig:BzKvsz\]. []{data-label="fig:EROS2"}](f15.eps){width="8cm"}
![The $J-K$ (Vega) color versus redshift (top) and versus $BzK$ (bottom), for galaxies in the K20/GOODS region. Symbols are as in Fig. \[fig:BzKvsz\]. []{data-label="fig:JK2"}](f16.eps){width="8cm"}
Particularly interesting is the comparison of the physical properties of dusty star-forming EROs to those of $z\sim2$ starbursts. The two samples include a few common objects, but it appears that on average the SFR of the $z=2$ star-forming galaxies is one order of magnitude higher than that of star-forming EROs. Indeed, the average X-ray 2–10 keV luminosity of dusty EROs in the K20 survey with $<\!z_{\rm spec}\!>
=1.053$ (Brusa et al. 2002) is a factor of $\sim10$ smaller than that measured for the $z=2$ $BzK$-selected starbursts. Similarly, the average 1.4 GHz luminosity of the same EROs is a factor of $\sim6$ smaller than found at $z=2$ (Cimatti et al. 2003). As these estimates are limited to EROs with known spectroscopic redshift, they exclude the highest redshift $z_{\rm phot}\sim2$ EROs in common with the $BzK$ selected sample. Forming stars much more vigorously, the reddened starburst galaxies seen at $z\sim2$ appear therefore to be of a different nature with respect to dust-reddened galaxies at $z\sim1$.
Similarly to the method proposed by Pozzetti & Mannucci (2000; PM2000 hereafter), requiring $BzK\geq-0.2$ would allow to distinguish dusty EROs from the old ones, extending the diagnostic in $1.4<z<2.5$ (the PM2000 criterion is formally valid only up to $z\sim2$). The two criteria are however substantially different and complementary to each other: while the PM2000 criterion relies on detecting the signature of the 4000 Å break of old galaxies, the $BzK$ criterion aims at detecting the UV tail in the SEDs due to the youngest stars, even in the presence of substantial reddening. We verified that 8/9 of the EROs that are also star-forming galaxies at $z>1.4$ with $BzK>-0.2$, mostly objects with photometric redshifts only, are correctly classified as star-forming galaxies by the PM2000 criterion.
It would be extremely interesting to apply the $BzK$ diagnostic to EROs samples in order to statistically distinguish $z<1.4$ EROs from those at $z\simgt1.4$, either old or star-forming ones. This would allow to solve the long standing issue of whether the EROs overdensities observed in the field of AGN/QSO at $z\simgt1.5$ are true spatial associations (i.e. clusters or proto-clusters) or are just due to lensing effects (e.g., Cimatti et al. 2000; Best et al. 2003; Wold et al. 2003). In case of true spatial associations one would also know if such enhancements are due to dusty star-forming or passive galaxies with important implications for galaxy formation in clusters.
Finally, we notice that the color properties of star-forming galaxies at $z\sim2$, having blue $B-z$ colors and the reddest $z-K$ color, are fully matching those of the mysterious population of [*red outlier*]{} galaxies (Moustakas et al. 1997). These galaxies were found to show blue $V-I$ colors with respect to their large $I-K>4$ colors and their nature had remained so far unclear.
$BzK$- vs. $(J-K)$-selected Galaxies
------------------------------------
Recently, Franx et al. (2003) have proposed the criterion $J-K>2.3$ to select evolved galaxies at $z>2$, and spectroscopic evidence that $z>2$ galaxies are indeed selected by this criterion has been provided for a sample of star-forming galaxies and AGN (van Dokkum et al. 2003; 2004). Based on a single color, the $J-K>2.3$ criterion is similar to that for EROs ($R-K>5$), and like this one it may eventually result to select both reddened star-forming galaxies and evolved/passive ones. The properties of $J-K>2.3$ galaxies in the K20 survey were then examined and compared with those recovered by the $BzK$ criteria. Fig. \[fig:JK2\] shows the $BzK$ vs $J-K$ diagram and the redshift vs $J-K$ diagram for galaxies in the K20/GOODS region. Only 4 galaxies with $J-K>2.3$ have a spectroscopic redshift, and lie in the range $0.6<z<1.3$. Additional 3 galaxies have a photometric redshift $z\sim1.5$, and only 6 out of 13, of the $J-K>2.3$ objects, have $z_{\rm phot}\simgt2$. Therefore, it seems that the contamination of $z<2$ galaxies among $J-K$ red galaxies could be higher than in the van Dokkum et al. (2003) sample. In part this could be due to photometric errors, as the lowest $z$ contaminants have $2.3<J-K<2.5$, hence close to the edge of the $J-K>2.3$ region. Note however that both the van Dokkum et al. spectroscopic sample and the K20/GOODS $J-K>2.3$ sample are quite small.
Only 9 out of 32 objects in the K20/GOODS sample at $1.4<z<2.5$ fulfill the $J-K>2.3$ condition, but the fraction rises to 5/11 for the $z>2$ galaxies. All the galaxies having $J-K>2.3$ and $z_{phot}>1.9$ have also $BzK>-0.2$, and would be classified as reddened star-forming galaxies rather than purely passive systems. This agrees with the recent results by van Dokkum et al. (2004) and Förster Schreiber et al. (2004). The highest redshift passive systems at $1.6<z<2$ in the K20 survey (Cimatti et al. 2004) have $J-K\sim1.7-2$.
In summary, it appears that the $BzK$ selection has the advantage of allowing in principle to recover the bulk of the galaxy population for the redshift range $1.4<z<2.5$ for which it is tuned, including the reddest and bluest ones, and to distinguish the passive from the star-forming ones. However, while the $BzK$ selection is efficient only up to $z\sim2.5$, the $J-K>2.3$ criterion can allow to pick up the reddest galaxies up to much higher redshifts $z\simlt4$ (Franx et al. 2003).
![A possible reddening independent selection criterion for $2.5\simlt z\simlt 4.0$ star-forming galaxies is obtained with $RJL\equiv J-L_{3.6}-1.4(R-J)>0$ (AB magnitudes). The above quantity is plotted for the galaxies in the K20 survey (the $L$-band magnitudes were derived from the best-fitting SED) and for constant star-formation rate models and 0.2, 0.5, 1 and 2 Gyr ages. Circled points show X-ray detected sources. []{data-label="fig:RJL"}](f17.eps){width="8.8cm"}
We finally notice that most of $BzK$ galaxies have $J-K>1.7$ (Fig. \[fig:JK2\]). At this threshold the clustering of faint $K$-selected galaxies at $z\simgt2$ was observed to become quite strong, compared to bluer galaxies (Daddi et al. 2003). This is consistent with $BzK$-like galaxies contributing to such a clustering enhancement (as suggested by D04), together with the reddest $J-K>2.3$ galaxies.
Extending the Technique to Select $\lowercase{z}>2.5$ Galaxies with Spitzer Photometry {#sec:SIRTF}
======================================================================================
The $BzK$ criterion is based on the rest-frame colors of (reddened) star-forming and (unreddened) passive galaxies, and then tuned to select those at $z\sim2$. Therefore, by choosing a different set of bands that sample the same rest-frame wavelengths one can forge a new criterion tuned to select the same kind of galaxies at a higher redshift.
By multiplying by 1.5 the central wavelengths of the $BzK$ bands one obtains values that roughly correspond to the central wavelengths of the $RJL$ bands, and correspondingly a criterion based on the quantity:
RJL (J-L\_[3.6]{})\_[AB]{} - 1.4(R-J)\_[AB]{} \[eq:RJL\] can be used to select galaxies in the redshift range $2.5\simlt
z\simlt 4$, i.e., complementary to the $BzK$ criterion that can select galaxies up to $z\sim 2.5$. The $RJL$ quantity has its peak for star-forming galaxies in the above redshift interval, again due to the Balmer break being located between the $J$ and $L$-band at 3.6 $\mu$m. Given that the ratio of the central wavelengths of the $RJL$ bands to the $BzK$ ones is not exactly constant, a factor $1.4$ in Eq. \[eq:RJL\] is necessary to make $RJL$ reddening independent in $2.5<z<4$ (using the reddening law of Calzetti et al. 2000). Model tracks for the $1.4(R-J)_{AB}$ versus $(J-L_{3.6})_{AB}$ colors in the range $2.5<z<4.0$ are nearly identical to those for the $(B-z)_{AB}$ versus $(z-K)_{AB}$ colors in $1.4<z<2.5$, already discussed in Section \[sec:modeling\] and shown in the panels of Fig. \[fig:4p\]. By requiring $RJL\simgt 0$ one should thus in principle cull $z\sim2.5$–4 star-forming galaxies in $L$-band limited samples, independently on their reddening, while objects having $RJL<0$ and $J-L_{3.6}\simgt2$–2.5 should turn out to be passive objects at $z>2.5$ (if such a population of galaxies exists). As a consistency check, synthetic $L$-band magnitudes were extrapolated for K20 galaxies from their best fitting SEDs to test for contamination by $z<2$ galaxies in $RJL\simgt 0$ selected samples, that results to be small (Fig. \[fig:RJL\]), though the K20 sample does not cover the redshift range $z>2.5$ that should be sampled by the $RJL>0$ criterion. A few $z>2$ galaxies in the K20 sample start to show $RJL>0$ following the models trend. A possible problem of this $RJL$ selection technique might be contamination by low redshift (e.g. $z\simlt0.5$) galaxies that could arise if significant contribution by dust (e.g. from AGN) is starting to appear in the $L$-band. No contribution of this kind was considered in the derivation of $K20$ synthetic $L$-band magnitudes. It is unclear how often this can happen for faint low-redshift galaxies. Filtering of low-redshift interlopers may be desirable for the application of this technique if the above contamination should result to be relevant.
If placed at $z=3.5$ a typical $z=2$ star-forming galaxy in the K20 survey would have $L_{3.6}\sim 22$–23 (AB). This is much brighter than the limits that should have been reached by the GOODS-SST observations at 3.6 $\mu$m (Dickinson et al. 2002). As an example, for $E(B-V)\sim0.6$ at $z\sim3.5$ one expects colors $(R-L_{3.6})_{AB}\approx5$ and $(J-L_{3.6})_{AB}\approx3$, implying the need to reach quite faint magnitudes in the optical/near-IR in order to detect such galaxies, i.e., $R\sim27$–28 and $J\sim25$–26 (AB scale magnitudes). A first check of this criterion should be possible with the GOODS ACS+ISAAC+SST dataset, that is expected to be deep enough in all the $RJL$ bands. In particular, it will be possible to test whether galaxies exist in the range $z\approx 2.5-4.0$ that are picked up by the $RJL$ selection but missed by the $U_{\rm n}GR_{\rm s}$ UV-selection, in analogy to what found for the $BzK$ selection.
Discussion {#sec:discussion}
==========
Early-Type Galaxies in Formation
--------------------------------
The masses of the $z\sim2$ $BzK$ galaxies in the K20 survey are overall quite high, with a median of $\sim 10^{11} M_\odot$, and in the local universe objects with such high masses are almost uniquely found among early-type galaxies. In D04 we had in fact suggested that the properties of $K$-selected galaxies at $z=2$ are consistent with those expected for the star-forming precursors of massive spheroids. This was also supported by the high SFRs, large sizes, merging-like morphologies and strong redshift space spikes that hint for strong clustering. Additional evidence in this direction comes from their strong photospheric and interstellar lines, indicative of solar metallicity or above, typical of massive spheroids (de Mello et al. 2004).
The properties of the $BzK$-selected galaxies with $K<20$ suggest that there is a whole population of vigorous starbursts within the $1.4<z<2.5$ range that can qualify as spheroids in the making. Such properties are quite different from those of star-forming galaxies at lower redshifts, e.g., the SFRs are $\sim 10$ times higher than those of $z\sim1$ dusty EROs, which in turn may be much less strongly clustered than $BzK$-selected galaxies (Daddi et al. 2002, 2004).
High redshift dusty star-forming galaxies selected at submm/mm wavelengths (see Blain et al. 2002 for a review) are often considered as the precursors of the present-day massive spheroids. Although a detailed comparison between $BzK$-selected and submm/mm-selected galaxies is beyond the scope of this paper, we notice that the latter systems have a redshift distribution largely overlapping with that of the former ones (median redshift of $z\sim2.4$; Chapman et al. 2003), even higher SFRs (Blain et al. 2002), similarly high masses (Genzel et al. 2003) and possibly similarly high clustering (Blain et al. 2004). However, the space density of submm/mm-selected galaxies is a factor of 10–100 lower than that of the $BzK$-selected ones. Submm/mm sources may be extreme subsets of $BzK$ galaxies.
Entering the Spheroid-Formation Epoch? A V/V$_{\rm max}$ Test {#sec:vmax}
-------------------------------------------------------------
Many lines of evidence suggest that ellipticals and bulges formed the bulk of their stars at $z\simgt 2.5-3$, both in clusters (e.g., Bower, Lucey, & Ellis 1992) and in the field (Bernardi et al. 1998, 2003), with much evidence having been accumulated from both low- and high-redshift ($z\sim 1$) observations of passively evolving spheroids (see e.g., Renzini 1999 for an extensive review; see also Thomas et al. 2002). This is further reinforced by the recent discovery of passive early-type galaxies at $z\sim2$, with UV-luminosity-weighted ages of 1–2 Gyr, implying formation redshifts beyond $z\simgt 2.5$–4 (Cimatti et al. 2004). Therefore, a natural question is whether the $z\sim 2$ $BzK$-selected starbursts belong to the major epoch or to the low redshift tail of spheroid formation.
In order to tentatively distinguish between these alternatives, a $V/V_{max}$ test (Schmidt 1968) was performed on the flux limited population of $K$-selected star-forming galaxies at $z>1.4$. For each object, $V/V_{\rm max}$ is computed as the ratio between the volume within the range $1.4<z<z_{obj}$ and that within $1.4<z<z_{max}$, where $z_{\rm max}$ is the maximum redshift for which the object would still be detected with $K<20$. To compute $z_{\rm max}$ we use the observed $J-K$ color of each galaxy to estimate the $K$-correction. In the case of a non evolving population the distribution of $V/V_{\rm
max}$ values should have an average of 0.5. For the 24 star-forming galaxies with $z>1.4$ in the K20/GOODS region we derive $<\!V/V_{\rm max}\!>\ = 0.594\pm0.048$. Considering only the objects in $1.4<z<2.5$, thus limiting $z_{\rm max}<2.5$, results in $<\!V/V_{\rm max}\!>\ =
0.64\pm0.05$. A $<\!V/V_{\rm max}\!>$ greater than 0.5 suggests that the comoving number density of these galaxies is increasing with redshift, but given the small sample the effect is only at the 2–3$\sigma$ level. Limits of these calculations are also that they are based in part on the use of photometric redshifts and that the results could be well affected by cosmic variance due to the clustering (D04).
The significance and rate of evolution could be however higher than recovered here, because biases are likely to work against the detection of objects at higher and higher redshifts. For example, the above calculation assumes no intrinsic evolution in the luminosity of the galaxies. On the other hand, if objects at lower redshift (e.g., at $z<2$) were forming stars with similar rates also at higher redshift (e.g. $z>2$), then the former would be intrinsically more luminous than the latter ones, and the amount of evolution would be higher than estimated above. Moreover, the strong bias due to surface brightness dimming with increasing redshift was not considered. Given the typical large sizes and low surface brightness of many of these star-forming galaxies (D04; Fig. \[fig:morphoBzK\]) one expects this effect to be relevant and to bias to low values the $V/V_{max}$ estimates.
Consideration of these effects would further enhance the significance and amplitude of the evolution, suggesting that by $z\sim 1.4$ we may have just started entering the epoch of widespread starburst activity, i.e., of major formation of galactic spheroids. An application of the $V/V_{max}$ test with upcoming larger redshift surveys should shed more light on this important point.
The Space Density of Vigorous $z\sim2$ Starbursts: Comparison with Models
-------------------------------------------------------------------------
We have shown that a substantial population of vigorous starburst galaxies with average $SFR\approx200$ $M_\odot$ yr$^{-1}$ exists at $1.4<z<2.5$. The number density of such $BzK$-selected starbursts can be further compared to predictions of theoretical galaxy formation models.
For example, the GIF semi-analytical models (Kauffmann et al. 1999; Kaviani, Haehnelt & Kauffmann 2003) predict that within $1.5\simlt z\simlt2$ the population of galaxies with masses in the range $10.5<{\rm
log}(M/M_\odot)<11.3$ (similar to the one derived for the K20 $z>1.4$ objects) are either passive galaxies with no ongoing SF or very active starbursts with SFR$\simgt50$ $M_\sun$ yr$^{-1}$. This is in very good agreement with our observations, but in these models the number density of objects with SFR$>100 M_\sun$ yr$^{-1}$ is $\sim0.8\times10^{-5}$ Mpc$^{-3}$ and $\sim1.3\times10^{-5}$ Mpc$^{-3}$ respectively at $z=1.46$ and $z=2.12$, which is a factor of 10–20 below the observed number densities. The space density of passive and massive galaxies is also similarly underpredicted.
Somerville et al. (2004) provide a mock catalog of galaxies with $K<20$ based on an updated semi-analytical model with enhanced starburst activity. While predicting the highest number density of $z>1.4$ galaxies with SFR$>100 M_\sun$ yr$^{-1}$ compared to all other models of this class, still it falls short by about an order of magnitude with respect to the present findings. Also the space density of passively evolving galaxies with $K<20$ at $z>1.4$ appears to be underpredicted by a similar factor by this model (Cimatti et al. 2004).
In general, $\Lambda$CDM semi-analytical models fail to account for the sheer number of $z\sim2$ galaxies with $K<20$ (Cimatti et al. 2002c; D04; Somerville et al. 2004). An exception is the hierarchical model by Granato et al (2004) based on the assumption of a coeval growth of QSOs and spheroids, which succeeds in producing the high space density of near-IR bright $z=2$ galaxies (see also Silva et al. 2004). However, in its current realization this model predicts that the $z>1.4$ tail of $K$-selected galaxies is predominantly populated by passively-evolving spheroids (see Fig. 8 in Silva et al. 2004), at variance with the observed prevalence of vigorous starbursts.
The recent $\Lambda$CDM hydrodynamical simulations by Nagamine et al. (2004) appear instead quite successful in reproducing the space density of $M>10^{11}M_\odot$ massive galaxies at $z=2$, as observed in the K20 survey, at least in two out of three of its different realizations. In one of their three simulation sets the authors also recover 2 galaxies with old stellar populations and red $G-Rs$ colors, consistent with the colors of passive spheroids at $z>1.4$ found in K20 (Fig. \[fig:LBG\_z2\]) and well matching to the K20 space density of passive sources when accounting for the different volumes. It is not clear instead if these models can reproduce the observed high density of vigorous starbursts with SFR$>100 M_\sun$ yr$^{-1}$, as observed in our survey.
As recently pointed out by Cimatti et al. (2004) and Grazebrook et al. (2004), in the traditional semianalytic models the formation of massive spheroids appears to be delayed to much too low redshifts. On the other hand, the Granato et al. (2004) models appear to move in the right direction, by pushing the formation to higher redshift with strong bursts of star-formation, then quenched by strong AGN activity. However, in doing so they may exceed somewhat, as compared to our findings they appear to underpredict the number of starbursting galaxies still present at $z\sim 2$. The present results, and an application of the $BzK$ selection to substantially larger samples, may help in further tuning theoretical models toward a more realistic description of galaxy formation and evolution.
Summary and Conclusions {#sec:summary}
=======================
$\bullet$ We have introduced a new criterion for selecting galaxies within the redshift range $1.4\simlt z\simlt 2.5$ which is based on the $BzK$ photometry and allows to identify both active star-forming as well as passively-evolving galaxies, and to distinguish between the two classes. The criterion has been tested empirically – using the spectroscopic redshifts and spectral types from the K20 survey ($K<20$) including 32 $z>1.4$ objects out of 504 with a spectroscopic redshift – and justified by simulations showing that active and passive synthetic stellar populations actually follow this selection criterion and are correctly identified. Albeit smaller in size, other spectroscopic samples such as the GDDS and photometric redshift of faint galaxies from the GOODS samples (as currently available) confirm that the criterion is effective in selecting galaxies in the mentioned redshift range and also for limiting $K$-band magnitudes somewhat fainter than $K=20$. We have shown that this $BzK$ criterion provides a very efficient way of selecting galaxies at $z\approx2$, that is not biased against passive galaxies and star forming galaxies that are highly reddened.
$\bullet$ The classification of $K<20$, $z>1.4$ galaxies as actively star forming or passive was then complemented by HST/ACS morphologies from the GOODS database, showing that indeed the spectral and morphological classifications are generally consistent: star-forming galaxies show clumpy, asymmetric morphologies typical of starbursts and mergers, while passive galaxies show symmetric surface brightness distribution in general typical of early-type galaxies.
$\bullet$ It is shown that the $BzK$ photometry can be used to estimate the internal reddening for the K20 galaxies classified as star-forming, and their intrinsic luminosity at 1500 Å. This allows an estimate of their dust-extinction corrected SFRs. The X-Ray and radio luminosities of these galaxies provide SFR estimates in very good agreement with the ones from the de-reddened 1500 Å luminosity.
$\bullet$ A significant population of $z=2$ galaxies with $K<20$, average SFR$\sim200$ $M_\odot$ yr$^{-1}$, and median reddening $E(B-V)\sim0.4$ is uncovered as a result, with a high volume density of $\sim10^{-4}$ Mpc$^{-3}$ and sky density of $\sim 1$ arcmin$^{-2}$. These vigorous starbursts produce a SFRD of $\sim 0.044$ $M_\odot$ yr$^{-1}$ Mpc$^{-3}$, representing a sizeable fraction of the total SFRD at $z=2$ as currently estimated.
$\bullet$ For $BzK$-selected galaxies at $1.4\simlt z\simlt 2.5$ the stellar mass derived from their redshift and multicolor photometry is tightly correlated to the observed $K$-band magnitude (with a $1\sigma$ dispersion of $\sim 50\%$), at least down to $K=20$.
$\bullet$ The $BzK$ selection and the above correlations ($BzK$ vs. $E(B-V)$ and $BzK$ vs. stellar mass) provide a fairly accurate and economic method that might be statistically applied to the very large samples of galaxies coming from the current or imminent wide-area surveys and/or for galaxy samples beyond the present spectroscopic capabilities.
$\bullet$ A comparison with the UV-selected galaxies at $z\sim
2$ (Steidel et al. 2004), including those at the same $K$ limit, shows that $BzK$-selected star-forming galaxies have typically higher reddening and SFRs. Among our $K<20$ sample, the galaxies satisfying the UGR selection criterion contribute roughly $\sim 15$% of the SFRD at $z\sim 2$ produced by the whole K20 sample. On the other hand the surface density of the Steidel et al. UV-selected galaxies down to $R=25.5$ is $\sim 10$ times higher than that of $K<20$ star-forming galaxies in the same redshift range, and their contribution to the SFRD at $z\sim 2$ is a factor $\simgt 2$ higher than that of the $K<20$, $BzK$-selected galaxies.
$\bullet$ The $BzK$ galaxies at $z\sim 2$ are characterized by a much higher SFR (by a factor $\sim 8$ on average) compared to dusty, star-forming EROs ($R-K>5$) at $z\sim 1$, and $K<20$. We conclude that these vigorous starbursts at $z\sim 2$ are of a different nature compared to highly reddened $z\sim 1$ galaxies.
$\bullet$ A $BzK$ analysis of the infrared-selected galaxies with $J-K>2.3$ (Franx et al. 2003) detected within the K20 survey shows that those with $z\simgt2$ are likely to be reddened star-forming objects, rather than passively evolving galaxies. A fraction $\sim 50$% of $J-K>2.3$ galaxies in the K20 survey is estimated to lie at relatively low redshifts $z\sim1\pm0.5$.
$\bullet$ $BzK$- and submm/mm-selected galaxies appear to share properties such as the redshift distribution, high SFRs and high masses, but the former ones have higher space density while the latter ones have higher SFRs. An interesting hypothesis is that submm/mm selected star-forming galaxies might represent extreme subsets of $BzK$ galaxies, at least when lying at $1.4\simlt z\simlt2.5$.
$\bullet$ Being based on the rest-frame shape of the spectra of starburst and passive galaxies, the $BzK$ criterion can be modified to select the same kinds of galaxies within a higher redshift range. In this mood, we propose a $RJL$ criterion to select galaxies within the range $2.5<z<4$, which would complement the $BzK$ selection of $1.4<z<2.5$ galaxies. With uniquely deep $L$-band ($3.6\, \mu$m) data that is becoming available from the Spitzer Space Telescope, this criterion should allow a selection of massive galaxies at $z\sim 3$ that may efficiently complement the traditional LBG selection.
$\bullet$ The high masses, SFRs, and metallicities of the bright $BzK$-selected galaxies at $z\sim 2$, together with a hint for a strong clustering of them, qualify these galaxies as possible precursors of $z\sim 1$ passively evolving EROs and $z=0$ early-type galaxies. A $V/V_{\rm max}$ test indicates that the space density of these galaxies may increase with redshift in the range $1.4\simlt z\simlt 2.5$. Current theoretical simulations of hierarchical galaxy formation generally fail to account simultaneously for the space density of both passively evolving and star-forming galaxies at $z=2$. Hydrodynamical simulations can reproduce our observations better than semyanalitical models.
Some of the above conclusions may be affected by cosmic variance, given the relatively small size of the explored field. To cope with this limitation a project is underway to cull $BzK$-selected galaxies over a $\sim1000$ arcmin$^2$ field, $\sim 20$ times larger than the full K20 survey area, by combining $K$-band data from ESO telescopes with optical data from SuprimeCam at the SUBARU telescope (Kong et al. in preparation), and to follow them up spectroscopically with VIMOS at the VLT. The validity of the $BzK$ selection at faint $K>20$ magnitudes will be further tested in great detail with a planned VLT/FORS2 survey (GMASS project) targeting among others $BzK$ selected galaxies fainter than $K=20$.
We are very grateful to Ken Kellermann and John Kelly for having provided access to their VLA radio maps of CDFS, and for having measured radio fluxes for our sources; to Piero Rosati, Mario Nonino and the CDFS team for allowing us to use their $BVRI$ FORS images of the CDFS field; to Gian Luigi Granato, Rachel Somerville and Kentaro Nagamine for providing details of their models and for useful discussions; to Micol Bolzonella for the assistence with the [*hyperz*]{} software; to Alice Shapley for sending us the transmission curves of the $U_{\rm n}GR_{\rm s}$ system in a digital form, and for discussions. Finally, we would like to thank the referee, Charles Steidel, for constructive comments and suggestions that resulted in a significant improvement of this paper. This research was funded in part with an ASI grant (IR-059-02). E.D. and A.R. gratefully acknowledge financial support from the ESO Office for Science.
Abraham, R. G., et al. 2004, , 127, 2455 Adelberger K. L., Steidel C. C., 2000, ApJ, 544, 218 Adelberger, K. L., Steidel, C. C., Shapley, A. E., Hunt, M. P., Erb, D. K., Reddy, N. A., & Pettini, M. 2004, ApJ, 607, 226 Alexander, D. M., et al. 2003, , 126, 539 Arnouts, S., et al. 2001, , 379, 740 Best, P. N., Lehnert, M. D., Miley, G. K., & R[" o]{}ttgering, H. J. A. 2003, , 343, 1 Bernardi, M., Renzini, A., da Costa, L. N., Wegner, G., Alonso, M. V., Pellegrini, P. S., Rit[é]{}, C., & Willmer, C. N. A. 1998, , 508, L143 Bernardi, M., et al. 2003, AJ, 125, 1882 Blain, A. W., Smail, I., Ivison, R. J., Kneib, J.-P., & Frayer, D. T. 2002, , 369, 111 Blain, A. W., Chapman, S. S., Smail, I., & Ivison, R. 2004, ApJ, 611, 725 Bolzonella, M., Miralles, J.-M., & Pell[' o]{}, R. 2000, , 363, 476 Bolzonella, M., Pell[' o]{}, R., & Maccagni, D. 2002, , 395, 443 Bower, R. G., Lucey, J. R., & Ellis, R. S. 1992, , 254, 601 Brusa M., Comastri A., Daddi E., et al., 2002, ApJ, 581, L89 Bruzual, G. & Charlot, S. 2003, , 344, 1000 Calzetti D., Armus L., Bohlin R. C., et al., 2000, ApJ, 533, 682 Cimatti, A., Villani, D., Pozzetti, L., & di Serego Alighieri, S. 2000, , 318, 453 Cimatti, A., et al. 2003, , 412, L1 Cimatti A., Daddi E., Renzini A., et al., 2004, Nature, 430, 184 Chapman, S. C., Blain, A. W., Ivison, R. J., & Smail, I. R. 2003, , 422, 695 Coleman, G. D., Wu, C.-C., & Weedman, D. W. 1980, , 43, 393 Condon J. J., 1992, ARA&A, 30, 575 Daddi E., Röttgering H., Labbé I, et al., 2003, ApJ, 588, 50 Daddi E., Cimatti A., Renzini A., et al., 2004, ApJ, 600, L127 (D04) de Mello D., Daddi E., Renzini A., et al., 2004, ApJ, 608, L29 Dickinson, M., Papovich, C., Ferguson, H. C., & Budav[' a]{}ri, T. 2003, , 587, 25 Dickinson, M., Giavalisco, M., & The GOODS Team 2003, in The Mass of Galaxies at Low and High Redshift, ed. R. Bender & A. Renzini (Berlin: Springer), 324 Elston, R., Rieke, G. H., & Rieke, M. J. 1988, , 331, L77 Erb, D. K., Shapley, A. E., Steidel, C. C., Pettini, M., Adelberger, K. L., Hunt, M. P., Moorwood, A. F. M., & Cuby, J. 2003, , 591, 101 Firth, A. E., et al. 2002, , 332, 617 Fontana, A., et al. 2003, , 594, L9 Fontana A., Pozzetti L., Donnarumma I., et al., 2004, A&A, 424, 23 (F04) Förster Schreiber N. M., van Dokkum P.G., Franx M., et al. 2004, ApJ in press (astro-ph/0408077) Franceschini, A., Aussel, H., Cesarsky, C. J., Elbaz, D., & Fadda, D. 2001, , 378, 1 Franx M., Labb[' e]{} I., Rudnick G., et al., 2003, ApJ, 587, L79 Genzel, R., Baker, A. J., Tacconi, L. J., Lutz, D., Cox, P., Guilloteau, S., & Omont, A. 2003, , 584, 633 Giacconi R., Zirm A., Wang J., et al., 2002, ApJS, 139, 369 Giavalisco M., et al., 2004, ApJ, 600, L93 Glazebrook, K., et al. 2004, Nature, 430, 181 Granato, G. L., De Zotti, G., Silva, L., Bressan, A., & Danese, L. 2004, , 600, 580 Haarsma, D. B., Partridge, R. B., Windhorst, R. A., & Richards, E. A. 2000, , 544, 641 Heavens, A., Panter, B., Jimenez, R., & Dunlop, J. 2004, , 428, 625 Hu, E. M. & Ridgway, S. E. 1994, , 107, 1303 Kauffmann, G., Colberg, J. M., Diaferio, A., & White, S. D. M. 1999, , 307, 529 Kaviani, A., Haehnelt, M.G. & Kauffmann, G. 2003, MNRAS, 340, 739 Kellermann K.I., Fomalont E.B., Rosati P., Shaver P., 2004, in preparation Madau, P., Pozzetti, L., & Dickinson, M. 1998, , 498, 106 McCarthy, P.J. 2004, ARA&A, 42, 477 Miyazaki, M., et al. 2003, , 55, 1079 Moustakas, L. A., Davis, M., Graham, J. R., Silk, J., Peterson, B. A., & Yoshii, Y. 1997, , 475, 445 Nandra K., Mushotzky R. F., Arnaud K., et al., 2002, ApJ, 576, 625 Nagamine K., Cen R., Ostriker J.P., Springer V., et al., 2004, ApJ submitted (astro-ph/0406032) Norman, C., et al. 2004, ApJ, 607, 721 Pirzkal N., Xu C., Malhotra S., et al. 2004, ApJS in press (astro-ph/0403458) Poli, F., et al. 2003, , 593, L1 Pozzetti, L. & Mannucci, F. 2000, , 317, L17 Pozzetti L., Cimatti A., Zamorani G., et al., 2003, , 402, 837 Ranalli P., Comastri A., Setti G., 2003, A&A, 399, 39 Reddy, N. A. & Steidel, C. C. 2004, , 603, L13 Renzini, A. 1999, in The Formation of Galactic Bulges, ed. C.M. Carollo et al. (Cambridge: CUP), 9 (astro-ph/9902108) Roche, N. D., Dunlop, J., & Almaini, O. 2003, , 346, 803 Roche, N. D., Almaini, O., Dunlop, J., Ivison, R. J., & Willott, C. J. 2002, , 337, 1282 Rudnick, G., et al. 2003, , 599, 847 Saracco, P., et al. 2004, A&A, 420, 125 Schmidt, M. 1968, , 151, 393 Silva, L., Granato, G. L., Bressan, A., & Danese, L. 1998, , 509, 103 Silva, L., De Zotti, G., Granato, G. L., Maiolino, R., & Danese, L. 2004, submitted to A&A (astro-ph/0403166) Smail, I., Ivison, R. J., Blain, A. W., & Kneib, J.-P. 2002, , 331, 495 Somerville R. S., Primack J. R., Faber S. M., 2001, MNRAS, 320, 504 Somerville R. S., Moustakas L. A., Mobasher B., et al., 2004, ApJ, 600, L135 Sharp R.G., Sabbey C.N., Vivas A.K., et al., 2002, MNRAS 337, 1153 Shapley A.E., Erb D.K., Pettini M., et al. 2004, ApJ, 612, 108 Springel, V. & Hernquist, L. 2003, , 339, 312 Steidel, C. C., Giavalisco, M., Dickinson, M., & Adelberger, K. L. 1996, , 112, 352 Steidel, C. C., Adelberger, K. L., Giavalisco, M., Dickinson, M., & Pettini, M. 1999, , 519, 1 Steidel, C. C., Adelberger, K. L., Shapley, A. E., Pettini, M., Dickinson, M., & Giavalisco, M. 2003, , 592, 728 Steidel, C. C., Shapley, A. E., Pettini, M., Adelberger, K. L., Erb, D. K., Reddy, N. A., & Hunt, M. P. 2004, , 604, 534 Szokoly, G. P., et al. 2003, ApJS in press (astro-ph/0312324) Thomas, D., Maraston, C., & Bender, R. 2002, Reviews in Modern Astronomy, 15, 219 Thompson, D., et al. 1999, , 523, 100 Totani, T., Yoshii, Y., Iwamuro, F., Maihara, T., & Motohara, K. 2001, , 558, L87 van Dokkum P. G., et al., 2003, ApJ, 587, L83 van Dokkum P. G., Franx M., Förster Schreiber N.M. et al. 2004, ApJ, 611, 703 Yan, L., Thompson, D., & Soifer, B. T. 2004, , 127, 1274 Yun, M. S., Reddy, N. A., Scoville, N. Z., Frayer, D. T., Robson, E. I., & Tilanus, R. P. J. 2004, , 601, 723 Wold, M., Armus, L., Neugebauer, G., Jarrett, T. H., & Lehnert, M. D. 2003, , 126, 1776 Vanzella E., Cristiani S., Dickinson M., et al.2004, A&A submitted (astro-ph/0406591)
[^1]: Publicly available ESO observations obtained as part of the GOODS project: http://www.eso.org/science/goods/
[^2]: http://www.arcetri.astro.it/$\sim$k20/releases/
[^3]: Reddy, private communication
[^4]: We are grateful to the referee (C. Steidel) for having informed us prior of publication that in his sample of $z\sim2$ galaxies (c.f. Sect. \[sec:UV\]) the $BzK>-0.2$ selection appears to miss a fraction of star-forming galaxies with $K>21$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In any communication network, the maximum number of link-disjoint paths between any pair of communicating nodes, S and T, is limited by the S-T minimum link-cut. Multipath routing protocols have been proposed in the literature to make use of these S-T paths in enhancing the survivability of the S-T information flow. This is usually accomplished by using a subset of these paths to forward redundant data units or combinations (if network coding is allowed) from S to T. Therefore, this enhancement in survivability reduces the useful S-T information rate. In this paper we present a new way to enhance the survivability of the S-T information flow without compromising the maximum achievable S-T information rate. To do this, bottleneck links (in the min-cut) should only forward useful information, and not redundant data units. We introduce the idea of extra source or destination connectivity with respect to a certain S-T max-flow, and then we study two problems: namely, pre-cut protection and post-cut protection. Although our objective in both problems is the same, where we aim to maximize the number of protected paths, our analysis shows that the nature of these two problems are very different, and that the pre-cut protection problem is much harder. Specifically, we prove the hardness of the pre-cut protection problem, formulate it as an integer linear program, and propose a heuristic approach to solve it. Simulations show that the performance of the heuristic is acceptable even on relatively large networks. In the post-cut problem we show that all the data units, forwarded by the min-cut edges not incident to T, can be post-cut-protected.'
author:
- |
Osameh M. Al-KofahiAhmed E. Kamal\
Department of Electrical and Computer Engineering, Iowa State University, Ames, IA 50010
bibliography:
- 'main.bib'
title: 'Max-Flow Protection using Network Coding'
---
Introduction {#intro}
============
The survivability of an information flow between two terminal nodes, S and T, can be enhanced by using part of the available network resources (bandwidth) to forward redundant information from S to T. Depending on the used survivability mechanism, the redundant information can be used to recover from data corruption if, for example, a Forward Error Correcting code (FEC) is used, or it can be used to recover from network component failures, if a proactive protection mechanism is used. In proactive protection, traditionally $k$ edge-disjoint S-T paths are used to forward $k$ copies of the same data unit from S to T, which guarantees the successful delivery of data if at most $k-1$ link failures occurred in the network. This is usually accomplished by means of a multipath routing protocol, such as MDVA [@SJ01] in wired networks or AOMDV [@MS01] in ad hoc wireless networks. The maximum number of edge-disjoint S-T paths is limited by the minimum S-T link-cut, which is defined as the smallest set of links that, when removed, all the S-T paths become disconnected. Let $h$ denote the value of the S-T min-cut. Then, if we want to forward data units from S to T and protect them against $q$ failures, we cannot send more than $k = \lfloor\frac{h}{q+1}\rfloor$ data units since $q+1$ copies of each data unit should be forwarded.
It is clear that traditional proactive protection approaches are very demanding and waste a lot of resources. Even if $q=1$, at least $50\%$ of the used network resources will be wasted to deliver the redundant information, which reduces the useful S-T information rate by at least $50\%$. Network coding [@RR00] can be used to overcome this problem in traditional proactive protection schemes. The basic idea of network coding is that it allows intermediate network nodes to generate combinations from the original data units, instead of just forwarding them as is. Therefore, to recover $k$ data units at the destination node T, $k$ linearly independent combinations in the $k$ data units should be delivered to T. That is, if we want to forward data units from S to T and protect them against $q$ failures, we can send at most $k = h - q$ data units. Note that this is done by designing a network code that creates $k+q$ combinations at intermediate network nodes such that any $k$ of them are solvable, which means that it is enough to receive only $k$ combinations to recover the $k$ data units at T. This simple analysis shows that the useful information rate of network coding-based protection is better than that of traditional protection approaches as long as $h > q + 1$, which is usually the case. Examples of network coding-based protection can be found in [@A09; @OA09; @OA08; @AC07].
Network coding-based protection and traditional protection schemes, provide end-to-end protection of the whole S-T paths used to forward useful data from S to T. In these approaches, the more we enhance the S-T flow survivability, the more we reduce the useful S-T information rate. This is because such approaches treat all network links equally, i.e., bottleneck links (that belong to the min-cut) as well as non-bottlenecks are used to forward redundant data units or combinations. Usually, most of the links in a network are not bottleneck links, which means that link failures are more likely to affect non-bottleneck links than links in the min-cut. Therefore, we can enhance the survivability of the S-T information flow without reducing the useful S-T rate below the max-flow, if we provide protection to the non-bottleneck links only. We call this kind of protection *Max-flow protection* because the max-flow can still be achieved under these conditions as long as no link in the min-cut fails. Note that max-flow protection can be transparently combined with end-to-end protection if needed. In this paper, we focus our analysis on the problem of max-flow protection only, and we do not consider combining it with traditional protection schemes. To the best of our knowledge the problem of max-flow protection has not been studied before.
The rest of this paper is organized as follows. Section \[Sec:prelim\] presents the terminology and definitions that will be used throughout the paper. The problems of pre-cut and post-cut protection are presented in Section \[Sec:problem\]. In Section \[Sec:preCut\] we study the pre-cut protection problem and prove its hardness. The problem of pre-cut protection is formulated as an Integer Linear Program (ILP) in Section \[Sec:ILP\]. A 3-phase heuristic approach to solve the pre-cut protection problem is described in Section \[Sec:Hrstc\]. Section \[Sec:postCut\] discusses the post-cut protection problem. Finally, Section \[Sec:conc\] concludes the paper.
Preliminaries {#Sec:prelim}
=============
We represent a network by a directed acyclic graph G=(V,E), where V is the set of network nodes and E is the set of available links, where each link is assumed to have unit capacity. The network has a source node (S) that wants to send data to a destination (T), where the S-T max-flow is assumed to be $h$. We assume that a multipath routing protocol is used, e.g., [@SJ01] or [@MS01], and the source is fully utilizing the available connectivity by sending $h$ data units to the destination simultaneously. To simplify the analysis, we assume that the network has a single cut. In the rest of this section we define the meaning of extra connectivity with respect to the S-T max-flow. After that we discuss some of the properties of nodes with extra connectivity.
Terminology
-----------
Let $f^{(A)}(B)$ denote the max-flow from the nodes in set A to the nodes in set B on a directed graph, which can be calculated by computing the max-flow between a virtual source/sink pair, such that the virtual source is connected to the nodes in A with infinite capacity edges and the virtual sink is connected to the nodes in B with infinite capacity edges also. Let $h = f^S(T) $ be the S-T max-flow. We define the following:
1. A node with Extra Source Connectivity (wESC) is a node, $u$, that satisfies the following conditions:
- $f^{S}(u,T) > h$, and $f^{(S,u)}(T) = h$.
2. A node with Extra Destination Connectivity (wEDC) is a node, $v$, that satisfies the following conditions:
- $f^{S}(v,T) = h$, and $f^{(S,v)}(T) > h$.
3. A node with No Extra Connectivity (wNEC) has:
- $f^{S}(v,T) = f^{(S,v)}(T) = h$.
Of course, a node with both extra source and extra destination connectivity cannot exist, because this contradicts the assumption that the max-flow equals h. Consider the graph G in Figure \[origG\]. The S-T max-flow in G is 4, which implies that four data units can be forwarded from S to T on four link-disjoint paths. Assume we found the following paths, $P_1 = \{S \rightarrow A \rightarrow E \rightarrow J \rightarrow T\}$ that forwards data unit $w$, $P_2 = \{S \rightarrow B \rightarrow F \rightarrow G \rightarrow T\}$ that forwards data unit $x$, $P_3 = \{S \rightarrow F \rightarrow H \rightarrow T\}$ that forwards data unit $y$, and $P_4 = \{S \rightarrow D \rightarrow I \rightarrow T\}$ that forwards data unit $z$. Each path $P_i$ contains a cutting edge $C_i$, which , if deleted, will result in reducing the max-flow by exactly 1 unit of flow because path $P_i$ will be disconnected and cannot be reestablished in any way. In our example, $P_1$ contains $C_1=\{(J,T)\}$, $P_2$ contains $C_2=\{(G,T)\}$, $P_3$ contains $C_3=\{(F,H)\}$, and $P_4$ contains $C_4=\{(I,T)\}$. Note that the min-cut may not always be unique, but in this paper we assume that the graph under consideration has only one cut.
Properties of nodes wESC/wEDC {#SubSec:properties}
-----------------------------
Consider a path that contains a node, u, wESC and a node, v, wEDC. Note that node u must be closer (in number of hops on the path) to the source than v, otherwise the max-flow assumption will be contradicted, as shown in Figure \[NodesOnTheSamePath\].
In general, removing the min-cut edges (i.e., the edges in $\cup_{i=1}^{h}C_i$) partitions the network into two partitions $A$ and $A'$, such that $S \in A$ and $T \in A'$. Note that, after deleting the min-cut edges, each of the partitions $A$ and $A'$ is a connected component (at least weakly), and that partition $A$ contains nodes wESC, but partition $A'$ contains nodes wEDC.
Any node $u \in A, u \neq S$ is a node wESC.
We prove this by contradiction. Let $u \in A, u \neq S$, but $u$ is not a node wESC. Then, $f^S(u,T) = h$, which means that node $u$ cannot receive additional flow from S if the S-T max-flow is established. This implies that either node $u$ is behind the min-cut (i.e., $u \in A'$), which contradicts the starting assumptions, or that there is another min-cut between S and $u$, which contradicts the single min-cut assumption.
In a similar fashion, we can prove the following for any node $v \in A', v \neq T$.
Any node $v \in A', v \neq T$ is a node wEDC.
In our following discussion we refer to $A$ as the *pre-cut* portion of the network, and to $A'$ as the *post-cut* portion of the network. Figure \[genCase\] summarizes the previous discussion.
Problem description {#Sec:problem}
===================
The cutting-edges, cannot be protected unless we trade bandwidth for survivability (i.e., unless we use an S-T path to carry redundant information to the destination), which reduces the useful S-T information rate. This tradeoff not only protects the cutting-edges, but also protects any edge carrying data in the network. However, the non-cutting-edges (or a subset of them) can be protected without reducing the S-T information rate, if the graph contains nodes wESC and/or wEDC. For example, nodes E, F, I and J in Figure \[origG\] are nodes wESC, and node H is a node wEDC. There are four possible ways to utilize the extra source connectivity in Figure \[origG\]; 1) protect data units $x$ and $y$ by sending $x+y$ to F through C, 2) protect $w$ by sending a duplicate to E through C and F, 3) protect $w$ by sending a duplicate to J through C, F, E and G 4) protect $z$ by sending a duplicate to I through C and F. The first option is better than the other three since sending $x+y$ to F enhances the chances of two data units ($x$ and $y$) to reach T, compared to duplicating $w$ or $z$ alone, which protects a single data unit only. Figure \[infoFlowG\] shows the first option, and it also shows how to utilize the extra destination connectivity from node H, where H sends a duplicate of $y$ to T through node K.
In this work, we propose a different way to handle the “survivability vs. bandwidth” trade-off. We propose a new approach to provide protection to the S-T information flow without reducing the useful S-T data rate. Basically, we avoid protecting the bottlenecks in the network (the min-cut links), and we try to efficiently utilize (by using network coding if possible) the available network connectivity before and/or after the bottleneck to provide protection to the non-min-cut links in the graph. We divide the problem into two sub-problems as follows:
1. Pre-cut protection: Our objective is to maximize the number of pre-cut-protected S-T paths. We show that this problem is NP-hard, and we provide a heuristic to solve it. To evaluate our heuristic we compare its performance to an ILP.
2. Post-cut protection: Similar to the previous objective, we aim to maximize the number of post-cut-protected S-T paths. Let $e_i$ be the closest cutting edge to the destination T on path $P_i$. We show that all the paths that do not have T as the head node of $e_i$ , where $1 \leq i \leq h$, can be post-cut-protected together against at least one failure.
Pre-Cut: Nodes with Extra Source Connectivity {#Sec:preCut}
=============================================
As discussed in Section \[Sec:prelim\], all nodes wESC are located in the pre-cut portion of the network. Assume that the set ${\mathcal{X}}$ contains all the nodes wESC, ${\mathcal{X}}=A \backslash S = \{u_1, u_2, \dots, u_{|{\mathcal{X}}|}\}$. Then, the following is true:
$$\label{equ:UnrealESC}
(\sum^{|{\mathcal{X}}|}_{i=1} f^S(u_i,T)) - |{\mathcal{X}}|f^S(T) \geq f^S(u_1, u_2, \dots, u_{|{\mathcal{X}}|}, T) - f^S(T)$$
This is because the extra source connectivity may be shared between the nodes in ${\mathcal{X}}$. Therefore, the right hand side of the inequality is what really determines the available extra source connectivity (ESC). This implies that not all nodes wESC in ${\mathcal{X}}$ can receive redundant flows from S to be used to protect the S-T max-flow, and thus, a subset $X \subseteq {\mathcal{X}}$ should be intelligently selected to receive the available extra source flow and utilize it in the best way possible. Note that the number of nodes in $X$ cannot exceed the extra available connectivity, i.e.: $$\label{equ:ESC}
ESC = f^S(u_1, u_2, \dots, u_{|{\mathcal{X}}|}, T) - f^S(T) \geq |X|$$
The selection of $X$ depends on how the S-T max-flow is routed on the graph. Consider the graph in Figures \[ex1\_a\] and \[ex1\_b\], the S-T max-flow in this network is 2, and there is only one S-T min-cut in the graph, which contains the edges (A,T) and (C,T). Nodes A, B and C are nodes wESC, and the total available extra source connectivity equals $f^S(A,B,C,T) - f^S(T) = 4 - 2 = 2$. Assume that the max-flow is routed as shown in Figure \[ex1\_a\] (the dashed lines), in this case $X_1=\{B,C\}$ since the extra source connectivity is consumed by B and C. Moreover, note that only the path forwarding **b** can be pre-cut-protected by sending copies of **b** on $(S,B)$ and $(S,C)$. Now consider the routing shown in Figure \[ex1\_b\], in this case $X_2=\{A,C\}$. Unlike the previous case, both paths can be pre-cut-protected by sending a second copy of **a** to A, and a second copy of **b** to C through B. Obviously, the second routing option is better since it allows the protection of both paths (equivalently both data units), in this sense we say $X_2$ is better than $X_1$.
It was shown in the previous example that routing the max-flow and selecting $X$ are inseparable problems, and that routing the S-T max-flow corresponds to selecting $X$. Let us define the extra source connectivity to a node $u$ with respect to the routing of the S-T max-flow in the network as:$$\footnotesize EC(u) = f^S(u,T) - f^S(T)$$ We say that an S-T path is pre-cut-protected if a segment of this path in the pre-cut portion of the network is protected. That is, a path is pre-cut-protected if it contains a node wESC with respect to the routing of the S-T max-flow. Therefore, maximizing the number of pre-cut-protected paths means maximizing the number of paths containing nodes wESC.
For large networks, trying-out all possible routing choices to find the best one that will maximize the number of paths containing nodes wESC is computationally expensive. The following theorem proves that this problem is in fact an NP-hard problem. The full-proof is omitted due to space limitations, and only a sketch of the proof is provided.
Routing the S-T max-flow to maximize the number of S-T paths containing nodes wESC is NP-hard.
To prove this theorem, we reduce the Maximum Coverage problem with Group budget constraints (MCG) [@CA04] to our problem. In the MCG problem, we have a collection of sets $C=\{C_1, C_2, ..., C_m\}$ that are not necessarily disjoint, where each set is a subset of a given ground set $H$. In addition, $C$ is partitioned into disjoint groups $\{G_1,G_2,...,G_n\}$, where each $G_j$ consists of a group of sets in $C$. The problem asks to select $k$ sets from $C$ to maximize the cardinality of their union, such that at most one set from each group is selected. Note that the cover size in the MCG problem is limited by the group budget constraints, and that the number of paths containing nodes wESC is limited by the available extra source connectivity in our problem. To prove the theorem we reduce any instance of the MCG problem to a directed graph with a single cut that translates the group budget constraints into constraints on the available extra source connectivity (similar to the one in Figure \[MCG\_reduction\]). It is now easy to prove that solving the MCG problem solves our problem and vice versa.
Note that if network coding was not allowed, then from equation [(\[equ:ESC\])]{} we cannot protect more than ESC data units. Therefore, to utilize the extra source connectivity in a more efficient manner we should apply network coding whenever possible. Network coding can be used if a node wESC, say $u$, lies on more than one S-T path, and has $EC(u) \geq 1$. For example, let $u$ be a node wESC that lies on two S-T paths, and that has $EC(u) = 1$. A network code can be designed to deliver three combinations in two data units to $u$, such that any two combinations are solvable, i.e., two data units are protected from S to $u$ against a single link failure. Note that the number of failures that can be tolerated is at most $EC(u)$. Therefore, the nodes in $X$ should have the following properties:
1. Each node $u_i \in X$ must have $f^S(u_i) > f^{u_i}(T)$.
2. The combinations received by a node $u_i \in X$ must be solvable if at most $e = f^S(u_i) - f^{u_i}(T)$ failures occurred on the $f^S(u_i)$ paths from S to $u_i$.
The first condition requires the flow from the source to each node $u_i \in X$ to be larger than the flow from that node to the destination. This condition is necessary to introduce redundancy in the forwarding process from S to the nodes in $X$. The second condition can be satisfied by designing a network code that delivers, for each node $u_i$, a set of $f^S(u_i)$ combinations, such that any $f^{u_i}(T)$ combinations of them are solvable. These two conditions allow a node $u_i$ to act as pre-cut decoding node, which can recover the data units sent from S to T through $u_i$, if at most $e = f^S(u_i) - f^{u_i}(T)$ failures occurred on the S-$u_i$ link-disjoint paths, and then send these native data units to T.
In the next section we present an integer linear program (ILP) formulation of our problem. Solving the ILP will select the routes for the S-T max-flow, and will maximize the number of pre-cut-protected paths (the number of S-T paths containing nodes wESC).
Integer Linear program Formulation {#Sec:ILP}
==================================
We need to maximize the number of S-T paths that contain nodes wESC, regardless of the number of those nodes. We assume that the S-T max-flow equals $h$, and that the flow can take integer values only. Since we are interested in the number of paths containing nodes wESC, we treat each of the $h$ units of flow as a commodity. That is, we have $h$ commodities, each of which is responsible for selecting a single S-T path. The ILP find the routes for these $h$ commodities on a graph with unit-capacity links, such that the number of paths containing nodes wESC is maximized. Let us begin by defining our notation:
- Let $\sigma_i$ be a binary variable that equals 1 if path $i$ ($P_i$) goes through at least one node wESC, and 0 otherwise. That is, $\sigma_i = 1$ if $P_i$ is pre-cut-protected, and $0$ otherwise.
- $f^i_{(a,b)}$ is the value of the flow from commodity $i$ on link (a,b). The links forwarding $f^i$ determines $P_i$.
- $u^i_j$ is the amount of flow $f^i$ entering node j. Although $u^i_j$ is not constrained to be binary, it will be either 1 or 0 since the source sends only one unit of flow $f^i$.
- $g^j_{(a,b)}$ is the amount of extra flow $g^j$ that is sent from the source to node $j$ on link (a,b). A node that consumes (not forwards) this flow will be included in $X$.
- $x_j$ is the amount of flow $g^j$ entering node $j$. Although $x_j$ is not constrained to be binary, it will be either 1 or 0 since the source sends only one unit of flow $g^i$.
- $\zeta^i_j$ is a binary variable that equals 1 if node $j$ is on $P_i$ and is wESC, i.e., $\zeta^i_j = u^i_jx_j$.
- $d_j$ is the minimum hop distance of node j from the source, which is a constant that can be computed for each node before solving the ILP, e.g., using Dijkstra’s shortest path algorithm.
- $\delta^i_j$ is a variable that equals $d_j$ if $\zeta^i_j = 1$, i.e., $\delta^i_j = d_j \zeta^i_j$.
- $\Omega$ is a very large positive constant.
- $w$ is a weighing factor for $\sum \sigma_i$, and is larger than the length of the longest possible path from the source to any node in the network, and can be set to $|E|$. This way the ILP maximizes the length of the protected paths if it does not reduce the number of protected data units.
our objective function is: $$\label{obj}
Maximize ~~~ w\sum^h_{i=1} \sigma_i + \sum_{\forall j}\delta_j$$
Subject to,
$$\label{FlowIsH}
{\sum_{\forall(S,b) \in E} f^i_{(S,b)} = 1,~\forall i, ~ where~ 1\leq i \leq h}$$
$$\label{FlowIsCnsrvd}
\sum_{\forall(a,b) \in E}f^i_{(a,b)} - \sum_{\forall(b,a) \in E} f^i_{(b,a)} = 0, ~\forall i,~\forall b \in V \backslash \{S,T\}$$
$$\label{NodeIsInP}
u^i_j - \sum_{\forall(a,j) \in E}f^i_{(a,j)} = 0 , \forall i, j$$
$$\label{gIs1}
\sum_{\forall(S,b) \in E} g^j_{(S,b)} \leq 1,~\forall j$$
$$\label{xIsgIn}
x_j - \sum_{\forall(a,j) \in E} g^j_{(a,j)} = 0,~\forall j$$
$$\label{gIsCnsrvd}
\sum_{\forall(k,v) \in E}g^j - \sum_{\forall(v,k) \in E} g^j = 0, ~\forall v \in V \backslash \{S,j\}$$
$$\label{capCnstrnt}
\sum_{\forall j}g^j_{(a,b)} + \sum_{\forall i} f^i_{(a,b)} \leq 1, ~\forall (a,b) \in E$$
$$\label{zetaCnstrnt}
\zeta^i_j - \frac{u^i_j + x_j}{2} \leq 0, ~\forall i, j$$
$$\label{sigmaCnstrnt}
\sigma_i - \frac{\sum_{\forall j \in V} \zeta^i_j}{\Omega} < 1$$
$$\label{deltaCnstrnt}
\delta^i_j - d_j\zeta^i_j = 0$$
Constraint [(\[FlowIsH\])]{} forces the S-T flow to be h, and constraint [(\[FlowIsCnsrvd\])]{} conserves all commodities on all nodes except S and T. [(\[NodeIsInP\])]{} make $u^i_j = 1$ if node $j$ is on path $i$. The extra flow that can be sent to a node wESC is bounded by 1 as shown in constraint [(\[gIs1\])]{}. Constraint [(\[xIsgIn\])]{} sets $x_j$ to 1 if node j receives any extra flow. The extra flow ($g^j$) is conserved at all nodes except the source and node j by constraint [(\[gIsCnsrvd\])]{}. Constraint [(\[capCnstrnt\])]{} guarantees that the link capacity of unit of flow is not exceeded. Constraint [(\[zetaCnstrnt\])]{} sets $\zeta^i_j$ to 1 if node j is on path $P_i$ and is a node wESC. Constraint [(\[sigmaCnstrnt\])]{} prevents $\sigma_i$ from being 1 if $P_i$ has no node wESC. The value of $\delta^i_j$ is set to $d_j$ if $\zeta^i_j = 1$ by constraint [(\[deltaCnstrnt\])]{}. Note that forcing the extra flow $g^j$ sent to node $j$ to be at most 1 does not affect the ILP optimality, since a path is considered pre-cut-protected if it has a node wESC regardless of the amount of extra flow received at that node. In the next section we present a heuristic approach to solve the problem of maximizing the number of paths containing nodes wESC. Moreover, we compare the heuristic results to the results from the ILP.
Heuristic approach {#Sec:Hrstc}
==================
Our heuristic works in three phases; the first one greedily selects an initial set $X'$; the second one modifies the flow on the graph (if needed) to guarantee that the S-T max-flow is achieved, and the third one utilizes any remaining connectivity and produces the final set $X$. The first phase works in iterations, where a single node is added to $X'$ in each iteration. Each time we add the node that can send the most flow to the destination, while being able to receive more flow from the source, to satisfy the two conditions stated at the end of Section \[Sec:preCut\]. If no more nodes satisfy this criteria and the S-T flow is still less than $h$, the second phase is entered. The second phase finds as much augmenting paths as possible from S to T so that the S-T max-flow is maximized. Finally, the third phase checks the nodes in the pre-cut portion of the graph to see if there are any remaining nodes wESC, and makes use of this extra connectivity.
Phase 1: Selecting the initial set $X'$
---------------------------------------
Recall that if all the min-cut edges are deleted, then the graph will be divided into two partitions $A$ (pre-cut), and $A'$ (post-cut). Note that the routing of the S-T flow in the post-cut portion of the network is independent from the routing of the S-T flow in the pre-cut portion of the network. Therefore, and since the selection of the final set $X$ depends on the routing of the S-T max-flow in the pre-cut portion of the graph, we can simplify the graph under consideration and just focus on the sub-graph, $H$, induced by the nodes in $A$ with a little modification. Specifically, given a directed graph $G(V_G,E_G)$, let $t(u,v)$ represent the tail node of edge (u,v), i.e., node $u$. Also, let $F_S$ be the set of tail nodes on the min-cut edges,i.e., $F_S$ contains the nodes in:$$\bigcup_{i=1}^h t(C_i)$$ where the S-T max-flow $= h$, and $C_i$ is the cutting edge on path $i$ as defined in Section \[Sec:prelim\]. We transform graph $G$ to $H(V_H,E_H)$ as follows:
1. Delete the nodes in $\{V_G \backslash A\}$
2. $V_H = \{A, T'\}$, where $T'$ is a dummy destination node.
3. $E_H = \{(u,v)|u,v \in A\} \bigcup \{(u,T')| \forall u \in F_S\}$. Note that $\{S, F_S\} \subset A$.
Each iteration of phase 1 adds the node that can send the most flow to T (or equivalently $T'$), while being able to receive more flow from the source. Let this node be $x$, then it satisfies the following conditions:
1. $f^x(T') \geq f^u(T'), \forall u \in V_H$
2. $f^S(x) > f^x(T')$
After identifying node $x$, the flow is sent in two steps; in the first step, $(f^x(T') + 1)$ units of flow are sent from S to $x$, and in the second one $f^x(T')$ units of flow are sent from $x$ to $T'$. This way, node $x$ can receive redundant information to protect the $f^x(T)$ path segment from S to $x$. Only one extra unit of flow is sent to $x$ so that the extra source connectivity is fairly divided between the nodes in $X'$ at the end of phase1.
From a network flows perspective, to forward the flow as described in the previous paragraph, $(f^x(T') + 1)$ units of flow should be sent on $(f^x(T') + 1)$ augmenting paths from S to $x$, and $f^x(T')$ units of flow should be sent on $f^x(T')$ augmenting paths from $x$ to $T'$. Note that since we are working on a residual graph, the paths found from $x$ to $T'$ may contain backward edges, which were used initially to forward flow from S to $x$. If this happens then the flow sent from S to the nodes in $X'$ may be changed and some nodes in $X'$ may not still be nodes wESC. To resolve this issue, we can delete all the edges on the paths found from S to $x$ in each iteration. However, this may reduce our ability to find augmenting paths from S to the nodes in $H$, and thus, may reduce the number of nodes that can be added to $X'$. Therefore, to be able to find augmenting paths without causing any of these problems we work with two copies of $H$. The first one, which we call $H^S$, is used to find paths from S to the nodes in $H$, and the second one, which we refer to as $H^T$, is used to find paths from the nodes in $H$ to $T'$. The links in $H^S$ and $H^T$ are related to each other as follows:
- After the first step is done, and $(f^x(T') + 1)$ paths were found from S to $x$ and augmented on $H^S$. Every edge (u, v) in $E_{H^T}$ that corresponds to a backward edge (v, u) in $E_{H^S}$ is deleted.
- Similarly, after the second step is completed, and $f^x(T')$ paths were found from $x$ to $T'$ and augmented on $H^T$. Every edge (u, v) in $E_{H^S}$ that corresponds to a backward edge (v, u) in $E_{H^T}$ is deleted.
In an iteration, if two or more candidate nodes have the same flow to $T'$, the tie is broken in favor of the largest minimum hop distance from the source, i.e., the one with the largest $d^S(u)$ is chosen to be added to $X'$. After that, if two or more nodes have the same flow and minimum hop distance a node is chosen randomly. Taking this into consideration, phase 1 ends when no more nodes can be added to $X'$.
Phase 2: Maximizing the S-T flow
--------------------------------
The resulting S-T flow from phase 1 equals $\sum_{\forall x \in X'} f^x(T')$, which might be less than or equal to $h$ (the max-flow). This is because the extra available connectivity is shared between the nodes in $V_H$. For example, consider the graph in Figure \[phase2\_ex\], where the S-T max-flow is 2. Phase 1 resulted in adding only one node (F) to $X'$. Assume that node F receives two units of flow from S along the two paths $P_1 = \{S \rightarrow C \rightarrow F\}$ and $P_2 = \{S \rightarrow B \rightarrow A \rightarrow D \rightarrow E \rightarrow F\}$, and sends one unit of flow to T on the direct edge (F, T). The resulting residual graph after augmenting these paths is shown in Figure \[phase2\_ex\_b\], where the backward edges resulting from the augmentation process are shown in boldface. At this point, no more nodes wESC can be added to $X'$ (because the two conditions in the previous subsection are not met for any node), but the S-T flow so far is only equal to 1. Therefore, phase 2 should be entered to maximize the S-T flow. Assume that phase 2 found the path $P_3 = \{S \rightarrow A \rightarrow B \rightarrow E \rightarrow D \rightarrow T\}$, and augmented the flow. After this step, no more S-T paths can be found on the residual graph, which means that the S-T flow is maximized, the resulting residual graph is shown in Figure \[phase2\_ex\_c\]. Note that after phase 2, node F still has two link-disjoint paths from S.
Phase 3: Utilizing the remaining ESC
------------------------------------
This phase simply checks if it is possible to send extra flow to any node in $H$ (that lies on at least one path) after the first two phases are finished. If a node $u$ is found to be able to receive extra flow $e'$ from S, then if it is not already in $X'$ it should be added to $X'$. The number of data units node $u$ sends to $T'$ equals $f^u(T')$. The number of data units or combinations it can receive from S is $k = f^u(T') + e'$ if it is not in $X'$, and is $k = f^u(T') + e' + 1$ if it is already in $X'$. If $f^u(T') = 1$ no coding is needed and we need to just send copies of the same forwarded data unit on all the paths to $u$. However, if $f^u(T') > 1$ a network code should be designed to deliver $k$ combinations to $u$ such that any $f^x(T)$ of them are solvable. Algorithm \[ALG:SelectingX\] summarizes the three phases.
Graph $H(V_H,E_H)$, $h$ = S-T max-flow Set $X$ containing nodes wESC $X' = \phi$, $ST\_flow = 0$, $Phase\_done = 0$ Create matrices $Flow_S[V_H]$, $Flow_T[V_H]$ Create graphs $H^S$ and $H^T$, where $V_{H^S} = V_{H^T} = V_{H}$ and $E_{H^S} = E_{H^T} = E_{H}$. Compute $f^S(u)$ on graph $H^S$, $\forall u \in V_{H^S}$ Compute $f^u(T')$ on graph $H^T$, $\forall u \in V_{H^T}$ Select node $x$, where $f^x(T') \geq f^u(T'), \forall u \in V_H$, and $f^S(x) > f^x(T')$ $Phase\_done = 1$ Find $f^x(T')+1$ augmenting paths from S to $x$ on $H^S$ Delete all forward edges in $H^T$ if they are reversed in $H^S$ Find $f^x(T')$ augmenting paths from $x$ to $T'$ Delete all forward edges in $H^S$ if the are reversed in $H^T$ $X' = X' \cup \{x\}$ $ST\_flow = ST\_flow + f^x(T')$ $Flow_S[x] = f^x(T') + 1$ $Flow_T[x] = f^x(T')$ Reverse $(u,v)$ in $H$ $Phase\_done = 0$ $Phase\_done = 1$ Find an S-$T'$ augmenting path in $H$ $ST\_flow++$ $Phase\_done = 0$ Compute $p = f^S(u)$ on the current residual graph of $H$ Find $p$ augmenting paths from S to $u$ on $H$ $Flow_S[u] = Flow_S[u] + p$ Compute $q = f^{T'}(u)$ on $H^T$ $Flow_T[u] = Flow_T[u] + q$ $X' = X' \cup \{u\}$ $X'$
Evaluation
----------
In this section we compare the results from our heuristic to the results from the ILP presented in Section \[Sec:ILP\]. The heuristic was compared to the ILP in five different cases. Each case represents a different network size, where the number of network nodes $V$ was changed to take the values $\{5, 10, 15, 20, 25\}$. In each case eighty random network instances were generated, and fed to the heuristic and the ILP. Figure \[Fig:PerfRatio\] shows the ratio between the average number of protected paths by the heuristic and the average number of protected paths by the ILP for the eighty runs. The figure shows that the performance of our heuristic is acceptable, where in the worst case at $V=20$ it was around 77% of the optimal on average.
To gain a better insight on the operation of the heuristic compared to the ILP we measured the S-T max-flow, counted the number of pre-cut-protected paths from the heuristic, and the number of pre-cut-protected paths resulting from the ILP in each time the heuristic and the ILP were executed (on the same network instance).
The histograms for the cases of V = 10, 15, 20, and 25 are shown in Figures \[SubFig:V\_10\], \[SubFig:V\_15\], \[SubFig:V\_20\], and \[SubFig:V\_25\] respectively. In general, the results from the heuristic are close to those from the ILP. Note that in some cases, the number of times the heuristic is able to protect $X_1$ paths may be larger than the number of times the ILP is able to protect the same number of paths $X_1$. However, this does not invalidate the heuristic because it comes at the price of protecting a larger number of paths $X_2 > X_1$ a fewer number of times. For example, in Figure \[SubFig:V\_20\], the heuristic was able to protect $X_1 = 2$ paths more than the ILP, but the ILP was able to protect $X_2 = 4$ paths more than the heuristic.
Coding
------
The resulting S-$T'$ flow from the heuristic (or the ILP) can be decomposed into two parts; the first, a one-to-many flow from S to the nodes in $X$, and the second is a many-to-one flow from S and the nodes in $X$ to $T'$. The many-to-one flow is not and cannot be coded, since it is composed from the $h$ native data units that are forwarded from S and the nodes in $X$ (possibly after decoding) to $T'$, on $h$ disjoint paths. However, the one-to-many flow from S to the nodes in $X$ can, and should be coded to utilize the extra source connectivity in the most efficient manner. Note that this one-to-many flow is different from normal multicast flow since different data is sent to different nodes. Therefore, a standard multicast network code cannot be used. In fact, the coding in our case is simpler, and needs to be done at a limited number of network nodes as we will show in the following discussion.
After the heuristic is done and the flow is constructed in the pre-cut portion of the graph. A node $u \in X$ can receive $k+e = Flow_S[u]$ units of flow from S and can send $k = Flow_T[u]$ units of flow to $T'$ (these values were computed in the heuristic). This implies that there are $k+e$ edge-disjoint paths from S to $u$, and $k$ edge disjoint paths from $u$ to $T'$ (or equivalently to T). Note that $k$ represents the number of S-T paths (or data units) going through node $u$, and that $e$ represents the paths used to carry redundant information to $u$.
Let $N_{x_i}$ be the set of 1-hop neighbors of the source on all the $k+e$ paths from S to $x_i$. Assume that all the nodes in $N_{x_i}$ have received the same set of $k$ data units from the source (the $k$ data units on the $k$ S-T paths). To construct a network code that delivers $k+e$ combinations to $x_i$ such that any $k$ of them are solvable using the received data units, we need to assign the proper coding vectors to the nodes in $N_{x_i}$. The coding vectors can be assigned from an $k \times (k+e)$ matrix that has no singular $k \times k$ submatrices, i.e., any $k \times k$ submatrix is invertible. A class of matrices that satisfies this requirement is the Cauchy matrices [@FN77]. Therefore, we can simply assign to each node in $N_{x_i}$ a column from a $k \times (k+e)$ Cauchy matrix, such that no two nodes are assigned the same column.
However, such a coding scheme requires decoding at the nodes in $X$ in each transmission round. An alternative way that will require a fewer number of decoding operations would be to use a systematic code. In a systematic code, $k$ out of the $k+e$ combinations will be trivial combinations, where each of which carries one of the $k$ native data units. In this case, decoding is necessary at a node $x_i \in X$, only if one of the native data units was lost due to a failure on one of the $k$ S-T paths going through node $x_i$. A simple way to do this is presented in [@JJ03]. Basically, let ${\mathcal{M}}_i$ denote a $k \times (k+e)$ Cauchy matrix with columns representing the coding vectors of the nodes in $N_{x_i}$. We can view ${\mathcal{M}}_i$ as two side-by-side matrices ${\mathcal{M}}_i = ({\mathcal{M}}_{k_i}|{\mathcal{M}}_{e_i})$, where ${\mathcal{M}}_{k_i}$ is a $k \times k$ matrix , and ${\mathcal{M}}_{e_i}$ is a $k \times e$ matrix. Let ${\mathcal{M}}'_i$ be the $k \times (k+e)$ matrix resulting from multiplying ${\mathcal{M}}_{k_i}^{-1}$ by ${\mathcal{M}}_i$: $${\mathcal{M}}'_i = {\mathcal{M}}_{k_i}^{-1} \times {\mathcal{M}}_i = (I_k|{\mathcal{M}}_{k_i}^{-1} \times {\mathcal{M}}_{e_i}) = (I_k|{\mathcal{M}}_{e_i}') =$$ $$\left[
\begin{array}{cccc|ccc}
1 & 0 & \dots & 0 & \alpha'_{0,k} & \dots & \alpha'_{0,k+e-1} \\
0 & 1 & \dots & 0 & \alpha'_{1,k} & \dots & \alpha'_{1,k+e-1} \\
\vdots & \vdots & \ddots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & \dots & 1 & \alpha'_{k-1,k} & \dots & \alpha'_{k-1,k+e-1} \\
\end{array} \right]$$
Since the original matrix ${\mathcal{M}}_i$ has no singular submatrices, then the resulting matrix ${\mathcal{M}}'_i$ has no singular submatrices also. Note that although the non-singularity property is preserved, the matrix is no longer a Cauchy matrix. Therefore, given that the source has already transmitted the $k$ data units to the nodes in $N_{x_i}$, assigning the columns of ${\mathcal{M}}'_i$ to the nodes in $N_{x_i}$ will create $k+e$ combinations such that any $k$ of them are solvable. Moreover, the code is systematic, where out of the $k+e$ combinations there are $k$ trivial combinations, each of which is composed of a single native data unit.
A special case is when $e=1$. In this case, after the source finishes transmitting the $k$ data units to the nodes in $N_{x_i}$ (where $|N_{x_i}| = k+1$), one of the nodes in $N_{x_i}$ can sum all the received data units and send this sum along with the $k$ native data units on $k+1$ paths to $x_i$.
Post-Cut: Nodes with Extra Destination Connectivity {#Sec:postCut}
===================================================
Nodes with extra destination connectivity (wEDC) can be found in the post-cut portion of the network only. Nodes wEDC (or a subset of them) can act as post-cut encoding nodes, which create and send redundant combinations to the destination node T to enhance the survivability of the information flow. Note that this case is different from the one considered previously, because all the nodes wEDC are sending their data units to the same destination. Let $hd(u,v)$ denote the head node of edge $(u,v)$, and let $F_T$ be the closest set of nodes wEDC to S, or equivalently the farthest set of node wEDC from T, then $F_T$ contains the nodes in:
$$\bigcup_{i=1}^h hd(C_i)$$ where $C_i$ is the cutting edge on path $P_i$. Note that if an edge $C_i$ has $T$ as a head node, then $T \in F_T$, which means that the data unit on the cutting edge $C_i$ is delivered to the destination directly after the cut and cannot be protected. That is, the flow that can be protected from the nodes in $F_T$ is reduced by the number of edges in the cut incident to T. Let $F'_T = F_T \backslash T$, note that $0 \leq |F'_T| \leq h$ (0 when all the nodes in $F_T$ are direct neighbors to T, and $h$ when none of them is a direct neighbor to T), where $h=f^S(T)$. Also, note that since the nodes in $F'_T$ are the head nodes of edges in the min-cut, then we have $f^S(F'_T) = |F'_T|$. That is, each node in $F'_T$ has only one data unit to forward to T, and $|F'_T|$ is the maximum post-cut flow that can be protected. Let $e$ denote the total available extra destination connectivity from the nodes in $F'_T$, then $e$ is calculated as follows $$e = f^{F'_T}(T) - |F'_T|.$$ Note that if $F'_T \neq \emptyset$, then $e \geq 1$. If network coding is not allowed, then no more than $e$ data units can be protected. However, if network coding is allowed, we prove that all the data units in $F'_T$ can be protected against at least a single failure:
Let $F'_T$ be the set of head nodes of the closest min-cut edges to T, such that $T \notin F'_T$. Then if network coding is allowed, the data units at the nodes in $F'_T$ can be protected together against a single failure.
It was shown in [@OA09] that a many-to-one flow, similar to the flow from $F'_T$ to T, can be protected against a single link failure (using network coding) if and only if any subset of $k$ source nodes can reach the common destination node through at least $k+1$ edge-disjoint paths.
Therefore, to prove the theorem we need to prove that any $k$ nodes in $F'_T$ can reach T through at least $k+1$ edge-disjoint paths. We prove this by contradiction. Assume that there is a set, $Q$, of $k$ nodes in $F'_T$ that can reach T through only $k$ edge-disjoint paths, i.e., $f^Q(T) = k$. Then there are $k$ cutting edges on the $k$ paths from the nodes in $Q$ to T, which contradicts the assumption of the single min-cut. Therefore, any $k$ nodes in $F'_T$ must be able to reach the destination node T through at least $k+1$ link disjoint paths, which concludes the proof.
If $e=1$ we can use the coding tree approach presented in [@OA09]. However, if $e > 1$, then to be able to recover the $|F'_T|$ data units if at most $e$ failures occurred in the post-cut portion of the graph, we need two conditions to be satisfied. First, any set of $k$ nodes in $F'_T$ must be able to reach the destination through at least $k+e$ link-disjoint paths. Second, we need to assign coding vectors to the $f^{F'_T}(T)$ combinations such that any $|F'_T|$ vectors from them are linearly independent. Note that if $e > 1$, then the linear independence of any $|F'_T|$ vectors does not necessarily mean that we can recover the $|F'_T|$ data units from any $|F'_T|$ combinations. This is because, when $e$ is larger than 1, the first condition is not necessarily satisfied. To clarify this issue, consider the example in Figure \[Fig:EDC\_solv\]. In the figure, $F'_T = \{A,B,C,D\}$, $f^{F'_T}(T) = 6$ and $e = 2$. The black nodes represent the 6 paths from $F'_T$ to T, and $c_i$ is the combination carried on $P_i$. The links represent the ability of the nodes in $F'_T$ to reach the different paths. If a path $P_i$ can be reached by $k$ nodes in $F'_T$ then $c_i$ is a function of $k$ data units. Note that since $e=2$, the first condition stated above is not satisfied, because nodes C and D can reach T through only three paths not four. To satisfy the second condition, the coding vectors can be chosen as the columns of a $4 \times 6$ Cauchy matrix. Now consider the four combinations $\{c_1,c_2,c_3,c_4\}$. Since $c_1$, $c_2$ and $c_3$ are functions of only two data units A and B (i.e., the coefficients of C and D are zeros), then the three combinations are linearly dependent. However, note that any two combinations of them are linearly independent, because in a Cauchy matrix any square submatrix has full rank (since it is another Cauchy matrix). That is, although the four combinations are in four data units (because of $c_4$), only three are linearly independent and only two are solvable.
If each node $v \in F'_T$ has $f^v(T)$ paths to T that are link disjoint from the paths from all other nodes in $F'_T$ to T, network coding will not be necessary and each node in $F'_T$ can send $f^v(T)$ copies of its data on its $f^v(T)$ paths to T. However, network coding becomes necessary if the paths from the nodes in $F'_T$ to T share links. The first links to be shared are in the link-cut between $F'_T$ and $T$ that is closest to $F'_T$. Let $f^{F'_T}(T) = n$, then there are $n$ edge-disjoint paths $\{P'_1, \dots, P'_n\}$ from $F'_T$ to $T$. Let $C'_i$ denote the cutting edge on path $P_i$ from a node in $F'_T$ to $T$ that is closest to $F'_T$ (if path $P'_i$ has more than one cutting edge). Recall that $C'_i$ is a cutting edge only if the maximum achievable $F'_T$-T flow is reduced by 1. Let $Z$ be the set of coding nodes, which contains the tail nodes of all the $n$ cutting edges as follows:$$\bigcup_{i=1}^nC'_i$$
Note that $|Z| \leq n$, and that network coding is not necessary at any of the downstream nodes after $Z$, since the combinations created at the nodes $Z$ will be forwarded to T on $|F'_T|$ edge-disjoint paths. Let $|F'_T| = m$, then a network code can be constructed by assigning each edge $C'_i$, where $1 \leq i \leq n$, a distinct column from an $m \times n$ Cauchy matrix. The solvability of any $m$ combinations depends on how the nodes in $F'_T$ are connected to $T$ as shown in the previous example. Specifically, let $r$ be the minimum number of solvable combinations in any $m$ combinations, and let $q$ denote the number of failures in the post-cut portion of the graph. Then we are guaranteed the full recovery of the $m$ data units if $q=1$ (by Theorem 2), and we are guaranteed the partial recovery of at least $r$ data units if $q = e$ (by the definition of $r$).
Conclusions {#Sec:conc}
===========
We presented a new protection approach, called max-flow protection, which can enhance the survivability of the whole S-T max-flow. The basic idea is not to protect links in the min-cut, but try to protect all other links if possible. We divided the problem into two problems; pre-cut protection and post-cut protection. Pre-cut protection is NP-hard. Therefore, the problem is formulated as an ILP, and a heuristic is proposed to solve it. We showed that all data units that are not delivered directly to T after the min-cut can be post-cut-protected. Finally, simple network codes are proposed to maximize the number of pre- and post-cut protected paths.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[ Yu Zeng-Hui $^{a,c}$ Herbert Pietschmann $^{a}$ Ma Wen-Gan $^{b,c}$ Han Liang $^{c}$ Jiang Yi $^{c}$ ]{}\
\
[$^{b}$CCAST (World Laboratory), P.O.Box 8730, Beijing 100080,P.R.China]{}\
[$^{c}$Department of Modern Physics, University of Science and Technology]{}\
[of China (USTC), Hefei, Anhui 230027, P.R.China]{}\
title: ' Top-quark pair production via polarized and unpolarized protons in the supersymmetric QCD [^1]'
---
ABSTRACT
0.3in
[The QCD corrections to the top-quark pair production via both polarized and unpolarized gluon fusion in $pp$ collisions are calculated in the Minimal Supersymmetric Model(MSSM). We find the MSSM QCD corrections can reach $4 \%$ and may be observable in future precise experiments. Furthermore, we studied the CP violation in the MSSM, our results show that the CP violating parameter is sensitive to the masses of SUSY particles (It becomes zero, when the c.m. energy is less than twice the masses of both gluino and stop quarks.) and may reach $10^{-3}$.]{}\
[ PACS number(s): 13.65.+i, 13.88.+e, 14.65.-q, 14.80.Dq, 14.80.Gt]{}
=0.36in
=0.36in
[**I. Introduction**]{}
The minimal supersymmetric model(MSSM) [@s1] is one of the most interesting extensions of the Standard Model (SM). Therefore testing the MSSM has attracted much interest. As is well known, the MSSM predicts supersymmetric(SUSY) partners to all particles expected by the SM, and searching for their existence is very important.
Since the top-quark was already found experimentally by the CDF and D0 Collaborations at Fermilab [@s2], we believe that more and more experimental events including top-quark will be collected in future experiments. That gives us a good chance to study the physics in top-quark pair production from $pp$ or $p \bar p$ collisions with more precise experimental results. Because of the heavy mass of the top quark this process provides a test of the SM and possible signals of new physics at high energy.
The dominant subprocesses of top-quark pair production in $pp$ or $p\bar p$ colliders are quark-antiquark annihilation and gluon-gluon fusion. The lowest order of those two subprocesses has been studied in Ref. [@s3]. There it was found that the former subprocess ($q\bar q$ annihilation) is more dominant in $p\bar p$ collisions when the c.m. energy($\sqrt{s}$) is near the threshold value $2 m_{t}$, whereas subprocess via $gg$ fusion will be more and more important with increasing c.m. energy, and can become the most dominant one when the c.m. energy is much larger than $2 m_{t}$.
In Ref. [@s4], the QCD corrections to top-quark pair production in $p\bar p$ collisions have been studied in the frame of the SM. It may seem natural that the QCD corrections of those processes in the frame of the MSSM are important for distinguishing those two models. Recently, the SUSY QCD corrections to top pair production via $q \bar q$ annihilation were given in Ref. [@s5]. The SUSY QCD corrections via unpolarized gluon-gluon fusion were presented by C.S.Li. et. al [@s6].
It is obvious that the correction from the SUSY QCD is related to the masses of top-quark and SUSY particles. Assuming the SUSY breaking scale at about 1 TeV, the masses of SUSY particles would be smaller than 1 TeV. So we can hope that corrections from SUSY particles are significant, since the heavy mass of the top quark ($m_{t}= ~175.6\pm
5.5 ~ GeV$ (world average)) may be comparable to some of the light SUSY particle masses. Therefore the SUSY QCD correction would give us some significant information about the existence of SUSY particles indirectly.
Recently, the spin structure of the nucleon has been intensely studied by polarized deep inelastic scattering experiments at CERN and SLAC. This knowledge allows us to find a clear signal beyond the SM, if we collect enough events in the process of top-quark pair production from polarized $pp$ or $p \bar{p}$ collisions. In the SM QCD, there is no CP violation mechanism, whereas in the SUSY QCD, the situation may be different. If we introduce phase angle of quark SUSY partners, we can get CP violation in the MSSM QCD [@s7]. Once we get enough statistics of top-quark pairs from $pp$ or $p \bar{p}$ colliders at higher energy, it will be possible to test CP violation. On the other hand, the spin-dependent parton distributions can be obtained from their polarized structure function data in Ref.[@Geh][@GLK][@STR]. There one found that the shape of polarized gluon and quark distributions in the nucleon depends on its polarization. Therefore the CP violation effects through the process of top-quark pair production via $gg$ fusion may be observed in polarized $pp$ or $p \bar{p}$ collisions.
In this work we concentrate on the SUSY QCD corrections to the process $pp \rightarrow g g \rightarrow t \bar{t} X$ both in polarized and unpolarized colliding beams. In section 2, we give the tree level contribution to subprocess $g g \rightarrow t \bar{t}$. In section 3 we give the analytical expressions of the SUSY QCD corrections to $gg \rightarrow t \bar{t}$. In section 4 the numerical results of the subprocess $gg \rightarrow
t \bar{t}$ and the process $pp \rightarrow g g \rightarrow t \bar{t} X$ are presented. The conclusion is given in section 5 and some details of the expressions are listed in the appendix.
[**II. The Tree-Level Subprocess**]{}
The graphical representation of the process $g(\lambda_{1},k_{1})
g(\lambda_{2},k_{2}) \rightarrow t(p_{1}) \bar{t} (p_{2})$ is shown in Fig.1 (a). The Mandelstam variables are defined as usual $$\begin{array} {lll}
\hat{s} & =(p_{1}+p_{2})^2=(k_{1}+k_{2})^2
\end{array}
\eqno {(2.1)}$$ $$\begin{array} {lll}
\hat{t} & =(p_{1}-k_{1})^2=(k_{2}-p_{2})^2
\end{array}
\eqno {(2.2)}$$ $$\begin{array} {lll}
\hat{u} & =(p_{1}-k_{2})^2=(k_{1}-p_{2})^2
\end{array}
\eqno {(2.3)}$$ so $\hat{s}+\hat{t}+\hat{u}=2 m_{t}^2$. The amplitude of tree-level diagrams with polarized gluons can be written as:[@s3] ($a, b$ are color indices of external gluons , $i, j$ are colors of external top-quarks and $T^{a}=\frac{\lambda_{a}}{2}$ are the Gell-Mann matrices.) $$\begin{array} {lll}
M_{0}^{(l)} &~ = ~ g_{s}^2 \epsilon ^{\mu,a} (\lambda_{1},k_{1}) \epsilon ^{\nu,b} (\lambda_{2},k_{2}) \bar{u}_{i}(p_{1})
\Gamma^{(l)} v_{j}(p_{2}), ~~~(l=s,t,u) \\
\end{array}
\eqno {(2.4)}$$ with $$\begin{array} {lll}
\Gamma^{(s)} &~= ~ \frac{T^{c}_{ij} f_{abc}}{s} [(\rlap/k_{1} - \rlap/k_{2}) g_{\mu\nu}+
(2 k_{2} + k_{1})_{\mu} \gamma_{\nu} - (2 k_{1} + k_{2})_{\nu} \gamma_{\mu}]
\end{array}
\eqno {(2.5)}$$ $$\begin{array} {lll}
\Gamma^{(t)} &~= ~ \frac{-i T_{im}^{a} T_{mj}^{b}}{t-m_{t}^{2}} \gamma _{\mu}
(\rlap/k_{2}-\rlap/p_{2}+m_{t})\gamma _{\nu }
\end{array}
\eqno {(2.6)}$$ $$\begin{array} {lll}
\Gamma^{(u)} &~= ~ \frac{-i T_{im}^{b} T_{mj}^{a}}{u-m_{t}^{2}}\gamma _{\nu }(\rlap/k_{1}-\rlap/p_{2}+m_{t})
\gamma _{\mu }
\end{array}
\eqno {(2.7)}$$ We chose a form in which only physical polarizations of gluons remained: $$\begin{array} {lll}
\epsilon^{\mu \ast} (\lambda_{1},k_{i}) \epsilon^{\nu} (\lambda_{2},k_{i})~=~
\frac{\delta_{\lambda_1,\lambda_2}}{2} (-g^{\mu\nu}+
\frac{n^{\mu} k_{i}^{\nu}+n^{\nu} k_{i}^{\mu}}{n \cdot k_{i}}-
\frac{n^2 k_{i}^{\mu} k_{i}^{\nu}}{(n \cdot k_{i})^2} +
i \lambda_1 \epsilon^{\sigma \mu \rho \nu} \frac{k_{i \sigma} n_{\rho}}
{n\cdot k_{i}}).
\end{array}
\eqno {(2.8)}$$ where $n=k_1+k_2$, $\lambda_{1,2}= \pm 1$. >From that, we can get the cross section at the tree-level with both polarized and unpolarized gluons.
[**III. SUSY QCD corrections (non-SM) to the subprocess $gg\rightarrow t\bar{t}$**]{}
1\. Relevant Lagrangian in the MSSM.
The difference between the MSSM QCD and the SM QCD corrections stems from the interactions of SUSY particles. Thus we can divide SUSY QCD corrections into a standard and a non-standard part. The Lagrangian density of the non-SM part of the SUSY QCD interaction is written as: $$\begin{array} {lll}
L&=L_{1}+L_{2}+L_{3}+L_{4}
\end{array}
\eqno {(3.a.1)}$$ Where $$\begin{array} {lll}
L_{1}&=-i g_{s} A^{\mu}_{a} T_{jk}^{a} (\tilde{q}^{j}_{L}
\partial_{\mu} \tilde{q}^{k}_{L}-
\tilde{q}^{k}_{L} \partial_{\mu} \tilde{q}^{j}_{L})+(L \rightarrow R)
\end{array}
\eqno {(3.a.2)}$$ $$\begin{array} {lll}
L_{2}&=-\sqrt{2} \hat{g_{s}} T_{jk}^{a} (\bar{\tilde{g}}_{a} P_{L} q^{k}
\tilde{q}^{j\ast}_{L} +
\bar{q}^{j} P_{R} \tilde{g}_{a} \tilde{q}^{k}_{L} -
\bar{\tilde{g}}_{a} P_{R} q^{k} \tilde{q}_{R}^{j\ast} -
\bar{q}^{j} P_{L} \tilde{g}_{a} \tilde{q}^{k}_{R})
\end{array}
\eqno {(3.a.3)}$$ $$\begin{array} {lll}
L_{3}&=\frac{i}{2} g_{s} f_{abc} \bar{\tilde{g^{a}}} \gamma_{\mu} \tilde{g^{b}} A_{\mu}^{c}
\end{array}
\eqno {(3.a.4)}$$ $$\begin{array} {lll}
L_{4}&=\frac{1}{6} g_{s}^{2} A_{\mu}^{a} A^{\mu}_{a} (\tilde{q}_{L}^{\ast}
\tilde{q}_{L}+ \tilde{q}_{R}^{\ast} \tilde{q}_{R})+\frac{1}{2} g_{s}^2
d_{abc} A_{\mu}^{a} A^{\mu b} (\tilde{q}_{L}^{i\ast} T_{ij}^{c}
\tilde{q}^{j}_{L}+\tilde{q}_{R}^{i\ast} T_{ij}^{c} \tilde{q}^{j}_{R})
\end{array}
\eqno {(3.a.5)}$$ q stands for quark, $\tilde{q}$ for corresponding squark, $\tilde{g}$ for gluino, $P_{L}$ and $P_{R}$ for left, right helicity projections, respectively. The mixing between the left- and right-handed stop quarks $\tilde{t}_{L}$ and $\tilde{t}_{R}$ can be very large due to the large mass of the top quark, and the lightest scalar top-quark mass eigenstate $\tilde{t}_1$ can be much lighter than the top-quark and all the scalar partners of the light quarks. Therefore the left-right mixing for the SUSY partners of the top quark plays an important role. Here we only considered the SUSY QCD effect from stop-quark, because we assume that other scalar SUSY quarks are much heavier than the stop-quark and hence decoupled. Furthermore we introduce the phase angle $\phi_{A}$ in the stop mixing matrix. Defining $\theta$ as mixing angle of stop-quark, we have $$\begin{array} {lll}
\tilde{t}_{L}&=e^{\frac{-i \phi_{A}}{2}} (\tilde{t}_{1}\cos{\theta} +
\tilde{t}_{2} \sin{\theta} )
\end{array}
\eqno {(3.a.6)}$$ $$\begin{array} {lll}
\tilde{t}_{R}&=e^{\frac{i \phi_{A}}{2}} (-\tilde{t}_{1} \sin{\theta} +
\tilde{t}_{2} \cos{\theta} )
\end{array}
\eqno {(3.a.7)}$$ where we suppose $m_{\tilde{t}_{1}} \leq m_{\tilde{t}_{2}}$.
2\. Analytical results of the MSSM QCD corrections.
The one-loop SUSY QCD correction diagrams are shown in Fig.1(b). In the following we present only the amplitude expressions of s-channel and t-channel. The amplitude of u-channel can be obtained from the t-channel expression by the following variable exchanges: $t \leftrightarrow u$, $k_1 \leftrightarrow k_2$, $\epsilon_{\mu}^{a} (k_1) \leftrightarrow
\epsilon_{\nu}^{b} (k_2)$ and $T^{a} \leftrightarrow T^{b}$. The one-loop diagrams can be divided into three groups: the self-energy diagrams of gluon and top-quark shown in Fig.1(b.1); $gtt$ and $ggg$ vertex correction diagrams shown in Fig.1(b.2); box diagrams shown in Fig.1 (b.3). The ultraviolet divergence is controlled by dimensional regularization ($n=4-\epsilon$). The strong coupling-constants are renormalized by using the modified Minimal Subtraction ($\bar{MS}$) scheme at charge-renormalization scale $\mu_{R}$. This scheme violates SUSY explicitly and the $q\tilde{q}\tilde{g}$ Yukawa coupling $\hat{g}_s$, which should be the same with the $qqg$ gauge coupling $g_s$ in supersymmetry, takes a finite shift at one-loop order. Therefore we take this shift between $\hat{g}_s$ and $g_s$ as shown in Eq.(3.b.1) into account in our calculation, in order to have the physical amplitudes independent of the renormalization scheme and we subtract the contribution of the false, non-supersymmetric degrees of freedom (also called $\epsilon$ scalars) [@Martin]. $$\hat{g}_s=g_s \left [ 1+ \frac{\alpha_s}{4 \pi}
(\frac{2}{3} C_A-\frac{1}{2} C_F) \right] ,
\eqno {(3.b.1)}$$ where $C_A=3$ and $C_F=4/3$ are the Casimir invariants of SU(3) gauge group. The heavy particles(top quarks, gluino, stop-quarks, etc.) are removed from the $\mu_{R}$ evolution of $\alpha_{s}(\mu_{R}^{2})$, then they are decoupled smoothly when momenta are smaller than their masses[@s8]. We define masses of heavy particles as pole masses.
The renormalized amplitude corresponding to all SUSY QCD one-loop corrections (as shown in Fig.1) can be split into the following components: $$\begin{array} {lll}
\delta M = \delta M_{s}+ \delta M_{v} + \delta M_{box} + \delta M_{d}.
\end{array}
\eqno {(3.b.2)}$$ where $\delta M_{s}$, $\delta M_{v}$, $\delta M_{box}$ and $\delta M_{d}$ are the one-loop amplitudes corresponding to the self-energy, vertex, box correction diagrams and the decoupling part, respectively. The $\delta M_{d}$ stems from the decoupling of the heavy flavors from the running strong coupling, and is given explicately by (see also [@Martin] [@s8]): $$\begin{array} {lll}
\delta M_{d} ~ =
M_{0} (\frac{\alpha_{s}(\mu)}{\pi})
[\frac{1}{24} \log (\frac{\mu_{R}^{2}}{m_{\tilde{t}_{1}}^2})
+ \frac{1}{24} \log (\frac{\mu_{R}^{2}}{m_{\tilde{t}_{2}}^2})+
+ \frac{1}{6} \log (\frac{\mu_{R}^{2}}{m_{t}^2})+
\frac{1}{2} \log (\frac{\mu_{R}^{2}}{m_{\tilde{g}}^2})]
\end{array}
\eqno {(3.b.3)}$$
3.Self-energy corrections to the amplitude.
The amplitude of self-energy diagrams $\delta M_{s}$ (Fig.1.(b.1)) can be decomposed into $\delta M_{s}^{g}$ (gluon self-energy) and $\delta M_{s}^{q}$ (top-quark self-energy), i.e. $$\begin{array} {lll}
\delta M_{s} &=~ \delta M_{s}^{g} + \delta M_{s}^{q} \\
&=~ \delta M_{s}^{g(s)} + \delta M_{s}^{g(t)} +
& \delta M_{s}^{g(u)}+ \delta M_{s}^{q(t)} + \delta M_{s}^{q(u)}.
\end{array}
\eqno {(3.c.1)}$$ The amplitudes $\delta M_{s}^{g(s)}$, $\delta M_{s}^{g(t)}$ and $\delta M_{s}^{g(u)}$ are for s-, t- and u-channel, respectively. They can be expressed as: $$\begin{array} {lll}
\delta M_{s}^{g(s)} = \frac{1}{2} M_{0}^{(s)}
[\Pi(k_{1}^{2})+\Pi(k_{2}^{2})+2 \Pi(s)],
\end{array}
\eqno {(3.c.2)}$$ $$\begin{array} {lll}
\delta M_{s}^{g(t)} = \frac{1}{2} M_{0}^{(t)}
[\Pi(k_{1}^{2})+\Pi(k_{2}^{2})],
\end{array}
\eqno {(3.c.3)}$$ $$\begin{array} {lll}
\delta M_{s}^{g(u)} = \frac{1}{2} M_{0}^{(u)}
[\Pi(k_{1}^{2})+\Pi(k_{2}^{2})] ,
\end{array}
\eqno {(3.c.4)}$$ where $M_{0}$ is the tree-level amplitude defined in Eq (2.4). $$\begin{array} {lll}
\Pi(k^{2}) ~ = &~ -\frac{\alpha_{s}}{4 \pi}
(T_{F}(\bar{B}_{0}+4 \bar{B}_{1}+4 \bar{B}_{21})[k,m_{\tilde{t}_{1}},m_{\tilde{t}_{1}}]+\\
& T_{F}(\bar{B}_{0}+4 \bar{B}_{1}+4 \bar{B}_{21})[k,m_{\tilde{t}_{2}},m_{\tilde{t}_{2}}]-
4 C_{A} (\bar{B}_{1}+\bar{B}_{21})[k,m_{\tilde{g}},m_{\tilde{g}}]-\frac{1}{3} C_{A}).
\end{array}
\eqno {(3.c.5)}$$ where $C_{F}=\frac{4}{3}$, $T_{F}=\frac{1}{2}$ , $C_{A}=3$ are invariants in the SU(3) color group, $B_{i}$ and $B_{ij}$ are Passarino-Veltman two-point functions [@s19][@s20]. The definitions of $\bar{B}_{0}$, $\bar{B}_{1}$ and $\bar{B}_{21}$ are listed in Appendix A. The amplitude $\delta M_{s}^{q(t)}$ is written as: $$\begin{array} {lll}
\delta M_{s}^{q(t)} & = & \frac{-i g_{s}^2 T_{ik}^{a} T_{lj}^{b}}
{(t-m_{t}^{2})^{2}} \epsilon ^{\mu,a} (k_1) \epsilon ^{\nu,b} (k_2)
\bar{u}_{i}(p_1) \gamma_{\mu} \\
&& (\rlap/k_{2}-\rlap/p_{2}+m_{t}) \left[ \hat{\Sigma}_{kl} (k_{2}-p_{2})
\right] (\rlap/k_{2}-\rlap/p_{2}+m_{t}) \gamma_{\nu} v_{j}(p_2).
\end{array}
\eqno {(3.c.6)}$$ Here we define $$\begin{array} {lll}
\hat{\Sigma}_{kl} (p) ~=~ C_{F} (H_{L} \rlap/p P_{L} +
H_{R} \rlap/p P_{R}- H^{S}_{L} P_{L} - H^{S}_{R} P_{R})
\delta_{kl}
\end{array}
\eqno {(3.c.7)}$$ with $$\begin{array} {lll}
H_{L} ~=~ \frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} x_{1} x_{3}
B_{1}[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]+
(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_{2}}, x_i \rightarrow y_i)+
\frac{1}{2} (\delta Z_{L} +\delta Z_{L}^{\dag}),
\end{array}
\eqno {(3.c.8)}$$ $$\begin{array} {lll}
H_{R} ~=~\frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} x_{2} x_{4}
B_{1}[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]+
(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_{2}}, x_i \rightarrow y_i)+
\frac{1}{2} (\delta Z_{R} +\delta Z_{R}^{\dag}) ,
\end{array}
\eqno {(3.c.9)}$$ $$\begin{array} {lll}
H^{S}_{L} ~=~\frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} x_{2} x_{3} m_{\tilde{g}}
B_{0}[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]+
(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_{2}}, x_i \rightarrow y_i)+
\frac{1}{2} m_{t} (\delta Z_{L} +\delta Z_{R}^{\dag})+\delta m_{t},
\end{array}
\eqno {(3.c.10)}$$ $$\begin{array} {lll}
H^{S}_{R} ~=~\frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} x_{1} x_{4} m_{\tilde{g}}
B_{0}[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]+
(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_{2}}, x_i \rightarrow y_i)+
\frac{1}{2} m_{t} (\delta Z_{R} +\delta Z_{L}^{\dag})+\delta m_{t}.
\end{array}
\eqno {(3.c.11)}$$ Where we abbreviate $\phi=\phi_{A}$, $x_{1}=\cos \theta e^{-i \phi}$, $x_{2}=\sin \theta e^{i \phi}$, $x_{3}=\cos \theta e^{i \phi}$, $x_{4}=\sin \theta e^{-i \phi}$, $y_{1}=\sin \theta e^{-i \phi}$, $y_{2}=- \cos \theta e^{i \phi}$, $y_{3}=\sin \theta e^{i \phi}$, $y_{4}=- \cos \theta e^{-i \phi}$, and $\theta$ is mixing angle of stop-quarks, see Eq (3.a.6 $\sim$ 7).
The explicit expressions of the top-quark wave-function renormalization constants have the following forms: $$\begin{array} {lll}
\delta Z_{L} &=&
- \frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} (x_{1} x_{3} Re[B_{1}]- \frac
{m_{\tilde{g}}}{m_{t}} (x_{1} x_{4}-x_{2} x_{3}) Re[B_{0}]\\
&+& m_{t}^{2} (x_{1} x_{3}+x_{2} x_{4}) Re[B_{1}^{'}] \\
&-& m_{t} m_{\tilde{g}} (x_{2} x_{3}+x_{1} x_{4}) Re[B_{0}^{'}])
[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]|_{p^{2}=m_{t}^{2}},
\end{array}
\eqno {(3.c.12)}$$ $$\begin{array} {lll}
\delta Z_{R} &=&
- \frac{\hat{g}_{s}^{2}}{8 \pi ^{2}} (x_{2} x_{4} Re[B_{1}]+
m_{t}^{2} (x_{1} x_{3}+x_{2} x_{4})
Re[B_{1}^{'}] \\
&-& m_{t} m_{\tilde{g}} (x_{2} x_{3}+x_{1} x_{4}) Re[B_{0}^{'}])
[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]|_{p^{2}=m_{t}^{2}},
\end{array}
\eqno {(3.c.13)}$$ $$\begin{array} {lll}
\delta m_{t} &=&
\frac{\hat{g}_{s}^{2}}{16 \pi ^{2}} ((x_{1} x_{3}+x_{2} x_{4})
m_{t} Re[B_{1}]\\
&-&(x_{2} x_{3}+x_{1} x_{4}) m_{\tilde{g}} Re[B_{0}])
[p,m_{\tilde{g}},m_{\tilde{t}_{1}}]|_{p^{2}=m_{t}^{2}},
\end{array}
\eqno {(3.c.14)}$$ We use the following abbreviations: $B_{i,ij}^{'}[p,m_{1},m_{2}]=\frac{\partial B_{i,ij}[p,m_{1},m_{2}]}
{\partial p^2}$.
4.Vertex-corrections to the amplitude.
The amplitudes for vertex diagrams can be expressed as: $$\begin{array} {lll}
\delta M_{v}^{(l)} ~ = ~ g_{s} \epsilon ^{\mu,a} (k_1) \epsilon ^{\nu,b} (k_2)
\bar{u}_{i}(p_1) \Lambda^{(l)}v_{j}(p_2),~~~(l=s,t,u), \\
\end{array}
\eqno {(3.d.1)}$$ where $$\begin{array} {lll}
\Lambda^{(s)} ~=& ~ -\frac{T^{c}_{ij}}{s}
\left[ \Lambda_{\mu\nu\rho}^{(3g)}(k_1,k_2) \right] \gamma_{\rho}\\
&-\frac{f_{abc}}{s} [(k_{1} - k_{2})_{\rho} g_{\mu\nu}+
(2 k_{2} + k_{1})_{\mu} g_{\nu\rho} \\
&-(2 k_{1} + k_{2})_{\nu} g_{\mu\rho}]
\left[ \Lambda_{\rho,(ij)}^{c}(p_{1},p_{2}) \right],
\end{array}
\eqno {(3.d.2)}$$ and $$\begin{array} {lll}
\Lambda^{(t)} ~=& ~ \frac{-i}{t-m_t^2}
\left\{ T^{b}_{mj} \left[ \Lambda_{\mu,(im)}^{a} (p_{1},k_{1}-p_{1})\right]
(\rlap/k_{2}-\rlap/p_{2}+m_{t})\gamma _{\nu } \right. \\
& \left. + T^{a}_{im} \gamma _{\mu } (\rlap/k_{2}-\rlap/p_{2}+m_{t})
\left[ \Lambda_{\nu,(mj)}^{b}(k_2-p_2,p_2) \right] \right\}.
\end{array}
\eqno {(3.d.3)}$$ The functions $\Lambda_{\mu\nu\rho}^{(3g)}$ and $\Lambda_{\mu,(ij)}^{a}$ are listed in Appendix B.
5\. Box-corrections to the amplitude.
The box diagram corrections in the t-channel (Fig.1(b.3)) are given as follows: $$\begin{array} {lll}
\delta M_{box}^{(t)} ~ =& ~2 g_{s}^2 \epsilon ^{\mu,a} (k_1) \epsilon ^{\nu,b} (k_2) \bar{u}_{i}(p_1)
((T^{c} T^{a} T^{b} T^{c})_{ij} F_{\mu\nu}^{(t1)}\\
&- i f_{bcd} (T^{c} T^{a} T^{d})_{ij} F_{\mu\nu}^{(t2)}-
f_{acm} f_{bmd} (T^{c} T^{d})_{ij} F_{\mu\nu}^{(t3)} \\
&- [T^{c} (T^{a} T^{b}+T^{b} T^{a}) T^{c}]_{ij} F_{\mu\nu}^{(t4)})
v_{j}(p_2),
\end{array}
\eqno {(3.e.1)}$$ Where $f_{abc}$ is defined as $[T^{a},T^{b}]=if_{abc}T^{c}$. The form factors $F_{\mu\nu}^{(ti)}(i=1 - 4)$ correspond to the kernel of the four Feynman diagrams in Fig.1(b.3) respectively and are given explicitly in Appendix C.
6\. Total cross section.
Collecting all terms in Eq (3.b.2), we can get the total cross section: $$\begin{array} {lll}
\sigma (\lambda_1, \lambda_2) &= \sigma_{0}(\lambda_1, \lambda_2)
(1+\delta \sigma (\lambda_1, \lambda_2))\\
&=\frac{1}{16 \pi s^2 }
\int_{t^{-}}^{t^{+}} dt {\sum_{spins}^{}}
[|M_{0}|^{2}+2 Re(M_{0}^{\dagger} \delta M)]
\end{array}
\eqno {(3.f.1)}$$ where $t^\pm=(m_t^2-\frac{1}{2}s)\pm\frac{1}{2}s \beta_t$, $\beta_t=\sqrt{1-4m_t^2/s}$, and the spin sum is performed only over the final top-quark pair when we considered polarized gluons.
[**IV. Numerical results**]{}
We denote $\hat{\sigma}_{0}$ for the Born cross section and $\hat{\sigma}$ for the cross section including one-loop SUSY QCD corrections of subprocess $gg \rightarrow t \bar{t}$, and define its relative correction as $\hat{\delta} = \frac{\hat{\sigma} - \hat{\sigma}_{0}}{\hat{\sigma}_{0}}$. For polarized gluon fusions, $\hat{\sigma}_{++}$, $\hat{\sigma}_{--}$ and $\hat{\sigma}_{+-}$ are the cross sections with positive, negative and mixed polarization of the gluons, respectively. In order to inspect the CP violating effects we introduce the CP-violation parameter for the subprocess defined by $\hat{\xi}_{CP}=\frac{\hat{\sigma}_{++} - \hat{\sigma}_{--}}
{\hat{\sigma}_{++} + \hat{\sigma}_{--}}$. The possible SUSY QCD effects in $gg \rightarrow t\bar{t}$ should be observed in $pp$ colliders. By analogy we can define also the relative correction and CP violating parameter for the process $pp \rightarrow gg \rightarrow t \bar{t}$ as $\delta =
\frac{\sigma - \sigma_{0}}{\sigma_{0}}$ and $\xi_{CP}=\frac{\sigma_{++} -
\sigma_{--}}{\sigma_{++} + \sigma_{--}}$, respectively. The SUSY QCD contribution to the process $p(P_{1},x)p(P_{2},y) \rightarrow gg \rightarrow
t \bar {t} X$ (x,y are polarizations of protons) can be obtained by convoluting the subprocess with gluon distribution functions. $$\begin{array} {lll}
\sigma (s) &= \int dx_{1} dx_{2} G(x_{1},Q) G(x_{2},Q)
\hat{\sigma}(\hat{s},\alpha_{s}(\mu))
\end{array}
\eqno {(4.1)}$$ with $k_{1}=x_{1}P_{1}$, $k_{2}=x_{2}P_{2}$ and $\tau=x_{1}x_{2}=\hat{s}/s$. $G(x_{i},Q)(i=1,2)$ are gluon distribution functions of protons. We take $Q=\mu_{R}=2m_{t}$.
In order to get results of top quark pair production from polarized $pp$ collisions, we need to consider the polarized gluon distributions in protons. The cross sections of polarized $pp \rightarrow gg \rightarrow t \bar{t} X$ can be written as $$\begin{array} {lll}
\sigma (x,y) &= \Sigma_{\lambda_{1},\lambda_{2} =\pm}
\int dx_{1} dx_{2} G^{x\lambda_{1}}(x_{1},Q) G^{y\lambda_{2}}(x_{2},Q)
\hat{\sigma}_{\lambda_{1},\lambda_{2}}(\hat{s},\alpha_{s}(\mu))
\end{array}
\eqno {(4.2)}$$ where x and y are the polarizations of incoming protons and $\lambda_{1},~~ \lambda_{2}$ are the polarizations of gluons inside protons. $G^{x\lambda_{1}}(x,Q),~G^{y\lambda_{2}}(x,Q)=G^{\pm}(x,Q)$ for equal (+) and opposite (-) polarization, $G^{+}(x,Q)$ and $G^{-}(x,Q)$ are polarized gluon distribution functions in the proton.
We used unpolarized proton structure functions of Glück et al. [@Glk] in our numerical calculations. For the polarized proton structure functions, we use the evolution equations of Glück et al. [@GLK] with input parameters from the paper of Stratmann et al. [@STR] (Next-To-Leading-Order). Since the structure functions are one of the least certain input in our calculation, we checked the result against other set, i.e. the polarized structure functions $G^{\pm}(x,Q)$ of Brodsky et. al [@Geh] (Using Leading-Order only). This tests the stability of our results against the particular form of the input structure functions. The two different sets of input are compared in Fig.2, which gives the relative SUSY QCD correction ($\delta$) and $\xi_{CP}$ versus c.m. energy $\sqrt{s}$ for the process $pp \rightarrow gg \rightarrow t \bar{t} X$. Though the SUSY QCD corrections from the two sets of structure functions are not too different, for $\delta$, there is some noticable change for $\xi_{CP}$. Because $\xi_{CP}$ depends strongly on the c.m. energy of the subprocess $gg \rightarrow t\bar{t}$ (shown in Fig.3(b)), a small modification of structure functions may lead to a large change of $\xi_{CP}$. Thus we can infer that the NLO-QCD calculation is required and the precise numerical prediction does depend on the reliability of the structure functions.
The SUSY QCD relative corrections are about $2 \% \sim 4 \%$ and decrease with increasing c.m. energy from Fig.2. These correction effects are within reach of future precision experiments and provide a possible discrimination of the SM and the MSSM effects. From Fig.2(c) we can see that the CP violation parameter $\xi_{CP}$ can be $10^{-3}$. Therefore, CP violation in this process stemming from the SUSY QCD can in principle be tested in future precision experiments. That would help us to learn more about the sources of CP violation.
In order to explore the effects of the SUSY QCD correction for future arrangements of optimal experimental conditions, we also investigate the subprocess $gg \rightarrow t \bar{t}$.
The relative SUSY QCD correction and CP violating parameter versus c.m. energy ($\sqrt{\hat{s}}$) for different polarization gluons are plotted in Fig.3 ($a \sim c$) with $m_{\tilde{g}}~=~200~GeV $, $m_{\tilde{t}_1}~=~ 250~GeV $, $m_{\tilde{t}_2}~ =~ 450 ~GeV $, and $\theta~ =~ \phi~ =~ 45^{\circ}$. In Fig.3(a) $\hat{\delta}_{++}$ and $\hat{\delta}_{--}$ are drawn in solid line and dashed line, respectively. $\hat{\xi}_{CP}$ as a function of c.m. energy is depicted in Fig.3(b) and $\hat{\delta}_{+-}$ as function of $\sqrt{\hat{s}}$ is plotted in Fig.3(c). Each curve in Fig.3(a) has an obvious peak near the position of the threshold of top pair production. That large enhancement is the combined effect of the threshold, when $\sqrt{\hat{s}}$ is just larger than $2 m_t = 350~GeV$, and the resonance when $\sqrt{\hat{s}}
\sim 2 m_{\tilde{g}} = 400~GeV$. The small spikes around the position of $\sqrt{\hat{s}}=900~GeV$, there $\sqrt{\hat{s}} \sim 2 m_{\tilde{t}_2} =
900~GeV$, shows also the resonance effect. Although Fig.3(a) shows that $\hat{\delta}_{++}$ and $\hat{\delta}_{--}$ approach equal values when the c.m. energy is far beyond its threshold value $2 m_{t}$, the quantitative difference between $\hat{\delta}_{+-}$ and $\hat{\delta}_{++}$ still exists in the whole energy range plotted in these figures. Fig.3(b) shows also that $\hat{\xi}_{CP}$ will be zero, if the c.m. energy is below the threshold of SUSY particles in the loop( i.e. $\sqrt{\hat{s}} \le 2 m_{\tilde{g}}~=~400~GeV $ in Fig.3(b)). This is reasonable because only beyond this point can we have absorptive terms which give contributions to $\hat{\xi}_{CP}$. $\hat{\xi}_{CP}$ has obvious resonance effect in the regions around $\sqrt{\hat{s}} \sim 2 m_{\tilde{g}}~=~400 GeV$ and $\sqrt{\hat{s}} \sim
2 m_{\tilde{t}_i}, (i=1,2) =500~GeV,~ 900~GeV$. We also find that the two stop quarks give opposite contributions to $\hat{\xi}_{CP}$ and when their masses are degenerate $\hat{\xi}_{CP}$ will vanish. When the c.m. energy $\sqrt{\hat{s}}$ is larger than $1~TeV$, $\hat{\xi}_{CP}$ will be near zero, because the contributions from the two stop quarks will cancel each other. Therefore a quantitative strong change of $\hat{\xi}_{CP}$ as function of c.m. energy can be an indication for the signals of stop quarks and gluino.
$\hat{\sigma}(\pm,\pm)$ and $\hat{\xi}_{CP}$ as functions of $m_{\tilde{g}}$ are shown in Fig.4 (a) and Fig.4 (b), respectively. In Fig.4 we take $\sqrt{\hat{s}} = 500 GeV$, $m_{\tilde{t}_1} = 100 GeV $, $m_{\tilde{t}_2}
= 450 GeV $, $\theta = \phi =45^{\circ}$. We can see from Fig.4(b) that $\hat{\xi}_{CP}$ changes its sign when $m_{\tilde{g}}$ is near $m_{t}=175~GeV$. The curves in Fig.4(a)(b) show again the resonance effect when $\sqrt{\hat{s}} \sim 2 m_{\tilde{g}} = 500~GeV$, note that for each line there is a steep change of the value of $\hat{\sigma}(\pm,\pm)$ or $\hat{\xi}_{CP}$ around the position of $m_{\tilde{g}}=250 GeV$.
Dependences of relative correction $\hat{\delta}_{\pm \pm}$ and $\hat{\xi}_{CP}$ for the subprocess $gg \rightarrow t \bar{t}$ on $m_{\tilde{t}_{1}}$ are plotted in Fig.5 (a) and Fig.5 (b). $\hat{\delta}_{\pm \pm}$ and $\hat{\xi}_{CP}$ as functions of $m_{\tilde{t}_{2}}$ are shown in Fig.6 (a) and Fig.6 (b), respectively. In all figures of Fig.5 and Fig.6, we take the common parameter set with $\sqrt{\hat{s}} = 500~GeV$, $m_{\tilde{g}}=200~GeV$ and $\theta = \phi
=45^{\circ}$. In Fig.5, we set $m_{\tilde{t}_{2}}=450~GeV$, whereas $m_{\tilde{t}_{1}}=100~GeV$ in Fig. 6. We find that $\hat{\xi}_{CP}$ in fact increases with mass splitting of stop-quarks (i.e. $m_{\tilde{t}_{2}}-m_{\tilde{t}_{1}}$) and when $m_{\tilde{t}_{1}}=m_{\tilde{t}_{2}}$, $\hat{\xi}_{CP}$ is equal to zero. The resonance effect of stop quarks, when $\sqrt{\hat{s}} \sim 2~m_{\tilde{t}_{i}}
(i=1,2)$, is superimposed on the curves in Fig.5 (a)(b) and Fig. 6(a)(b) around the positions of $m_{\tilde{t}_{1}}=250~GeV$ in Fig.5(a)(b) and $m_{\tilde{t}_{2}}=250~GeV$ in Fig.6(a)(b), respectively. Around those points the relatively sharp changes of the values of $\hat{\xi}_{CP}$ and the relative corrections are shown in these figures.
Finally, the dependence of $\hat{\delta}_{\pm \pm}$ and $\hat{\xi}_{CP}$ on the phase $\phi$ is shown in Fig.7 (a) and (b). In Fig.7, we take $\sqrt{\hat{s}} = 500~GeV$, $m_{\tilde{g}}=200GeV$, $\theta =45^{\circ}$ and $m_{\tilde{t}_{1}}=150~GeV$. We find that $\hat{\xi}_{CP}$ is directly proportional to $\sin{(2\phi)}$ and reaches its maximal value when $\phi=\frac{\pi}{4}$.
[**IV. Conclusion**]{}
In this work we have studied the one-loop supersymmetric QCD corrections to the subprocess $gg \rightarrow t\bar{t}$ and process $pp \rightarrow gg \rightarrow t \bar{t}X$. The calculations show that the SUSY QCD effects are significant. The absolute values of the corrections are about $2\%\sim 4 \%$, so they may be observable in future precision experiments. Furthermore, we find $\xi_{CP}$ depends strongly on masses of SUSY particles and can reach $10^{-3}$ when we take plausible SUSY parameters.
The results show that there is an obvious difference between the corrections for the protons polarized with parallel spin and that with anti-parallel spin. Hence there is a possibility to study spin-dependence in the frame of the MSSM QCD.
We also presented and discussed the results of the subprocess $gg \rightarrow
t \bar{t}$. We find that when the c.m. energy passes through the value $2 m_{\tilde{g}}$ or $2 m_{\tilde{t}_i}$ ($i=1,2$), the value of the CP violating parameter $\hat{\xi}_{CP}$ changes strongly. If c.m. energy is less than both $2 m_{\tilde{g}}$ and $2 m_{\tilde{t}_{i}}~(i=1,2)$, $\hat{\xi}_{CP}$ will be zero. If in future experiments a sharp change in $\hat{\xi}_{CP}$ is found with $\sqrt{\hat{s}}$ running from low c.m. energy to high c.m. energy, it would be interpreted as a signal of SUSY particles. Furthermore, because the CP violating parameter $\hat{\xi}_{CP}$ is sensitive on the mass of gluino(as shown in Fig.4 (b)) and the mass splitting of stop-quarks $m_{\tilde{t}_{2}}-m_{\tilde{t}_{1}}$ (as shown in Fig.5 and Fig.6), we can also get information of SUSY particles from precise measurements of $\hat{\xi}_{CP}$.
The authors would like to thank Prof. A. Bartl for useful discussions and comments. One of the authors, Yu Zeng-Hui, would like to thank Prof. H. Stremnitzer for his help.
Appendix
A. Loop integrals:
We adopt the definitions of two-, three- and four-point one-loop Passarino-Veltman integral functions of reference[@s19][@s20].
1.The two-point integrals are: $$\{B_0;B_{\mu};B_{\mu\nu}\}(p,m_1,m_2)=
{\frac{(2\pi\mu)^{4-n}}{i\pi^2}}\int d^n q
{\frac{\{1;q_{\mu};q_{\mu}q_{\nu}\}}{[q^2-m_1^2][(q+p)^2-m_2^2]}},
~~~~~(A.a.1)$$ The function $B_{\mu}$ is proportional to $p_{\mu}$: $$B_{\mu}(p,m_{1},m_2)=p_{\mu} B_{1}(p,m_1,m_2)
~~~~~(A.a.2)$$ Similarly we define: $$B_{\mu\nu}=p_{\mu}p_{\nu} B_{21}+g_{\mu\nu} B_{22}
~~~~~(A.a.3)$$ We denote $\bar{B}_{0}= B_{0}-\Delta$, $\bar{B}_{1}= B_{1}+\frac{1}{2}\Delta$ and $\bar{B}_{21}= B_{21}-\frac{1}{3}\Delta$. with $\Delta= \frac{2}{\epsilon}
-\gamma +\log (4\pi)$, $\epsilon=4-n$. ${\mu}$ is the scale parameter.
2\. The three-point integrals are: $$\{C_0;C_{\mu};C_{\mu\nu};C_{\mu\nu\rho}\}(p,k,m_1,m_2,m_3)=$$ $$-{\frac{(2\pi\mu)^{4-n}}{i\pi^2}}\int d^n q
{\frac{\{1;q_{\mu};q_{\mu}q_{\nu};q_{\mu}q_{\nu}q_{\rho}\}}
{[q^2-m_1^2][(q+p)^2-m_2^2][(q+p+k)^2-m_3^2]}},
~~~~~(A.a.4)$$ We define form-factors as follows: $$C_{\mu}=p_{\mu} C_{11} + k_{\mu} C_{12}$$ $$C_{\mu\nu}=p_{\mu} p_{\nu} C_{21}+k_{\mu}k_{\nu} C_{22}+
(p_{\mu}k_{\nu}+k_{\mu}p_{\mu}) C_{23}+ g_{\mu\nu} C_{24}$$ $$C_{\mu\nu\rho}=p_{\mu}p_{\nu}p_{\rho} C_{31}+
k_{\mu}k_{\nu}k_{\rho} C_{32}+
(k_{\mu}p_{\nu}p_{\rho} +
p_{\mu}k_{\nu}p_{\rho} +
p_{\mu}p_{\nu}k_{\rho}) C_{33}+$$ $$(k_{\mu}k_{\nu}p_{\rho} +
p_{\mu}k_{\nu}k_{\rho} +
k_{\mu}p_{\nu}k_{\rho}) C_{34}+
(p_{\mu} g_{\nu\rho}+p_{\nu} g_{\mu\rho}+
p_{\rho} g_{\mu\nu}) C35+$$ $$(k_{\mu} g_{\nu\rho}+k_{\nu} g_{\mu\rho}+
k_{\rho} g_{\mu\nu}) C36
~~~~~~~~~~~~~~~(A.a.5)$$
3\. The four-point integrals are: $$\{D_0;D_{\mu};D_{\mu\nu};D_{\mu\nu\rho};D_{\mu\nu\rho\alpha}\}
(p,k,l,m_1,m_2,m_3,m_4)=$$ $${\frac{(2\pi\mu)^{4-n}}{i\pi^2}}\int d^n q
{\frac{\{1;q_{\mu};q_{\mu}q_{\nu};q_{\mu}q_{\nu}q_{\rho};q_{\mu}q_{\nu}q_{\rho}q_{\alpha}\}}
{[q^2-m_1^2][(q+p)^2-m_2^2][(q+p+k)^2-m_3^2][(q+p+k+l)^2-m_4^2]}},
~~~~~(A.a.6)$$ Again we define form-factors of D functions: $$D_{\mu}=p_{\mu} D_{11}+k_{\mu} D_{12}+l_{\mu} D_{13}$$ $$D_{\mu\nu}=p_{\mu}p_{\nu} D_{21} +k_{\mu}k_{\nu} D_{22} +l_{\mu}l_{\nu}
D_{23}+\\
\{pk\}_{\mu\nu} D_{24}+ \{pl\}_{\mu\nu} D_{25}+ \{kl\}_{\mu\nu}
D_{26}+g_{\mu\nu}D_{27}$$ $$D_{\mu\nu\rho}=p_{\mu}p_{nu}p_{\rho}D_{31} +
k_{\mu}k_{nu}k_{\rho}D_{32}+
l_{\mu}l_{nu}l_{\rho}D_{33}+
\{kpp\}_{\mu\nu\rho} D_{34}+$$ $$\{lpp\}_{\mu\nu\rho} D_{35}+
\{pkk\}_{\mu\nu\rho} D_{36}+
\{pll\}_{\mu\nu\rho} D_{37}+
\{lkk\}_{\mu\nu\rho} D_{38}+$$ $$\{kll\}_{\mu\nu\rho} D_{39}+
\{pkl\}_{\mu\nu\rho} D_{310}+
\{pg\}_{\mu\nu\rho} D_{311}+
\{kg\}_{\mu\nu\rho} D_{312}+
\{lg\}_{\mu\nu\rho} D_{313} ~~~~~~(A.a.7)$$
where $$\{pk\}_{\mu\nu}=p_{\mu}k_{\nu}+k_{\mu}p_{\nu}$$ $$\{pkl\}_{\mu\nu\rho}=p_{\mu}k_{\nu}l_{\rho}+l_{\mu}p_{\nu}k_{\rho}+
k_{\mu}l_{\nu}p_{\rho}$$ $$\{pg\}_{\mu\nu\rho}=p_{\mu}g_{\nu\rho}+p_{\nu}g_{\mu\rho}+p_{\rho}g_{\mu\nu}
~~~~~(A.a.8)$$
The numerical calculation of the vector and tensor loop integral functions can be traced back to the four scalar loop integrals $A_0$, $B_0$, $C_0$ and $D_0$ in Ref.[@s19][@s20] and the references therein.
B. Vertex corrections:
The 3-gluon-vertex can be written as: (a,b,c are the color indices of the external gluons) $$\begin{array} {lll}
\Lambda_{\mu\nu\rho}^{(3g)}(k_1,k_2)=
\frac{i g_{s}^3}{16 \pi^2} \left\{ Tr(T^{b} T^{c} T^{a})
\left[ \Lambda_{\mu\nu\rho}^{(1)}(k_1,k_2) \right] +
i f^{cmn} f^{anl} f^{blm} \left[ \Lambda_{\mu\nu\rho}^{(2)}(k_1,k_2)
\right] \right\},
\end{array}
\eqno {(A.b.1)}$$ the vertex functions $\Lambda_{\mu\nu\rho}^{(1)},
\Lambda_{\mu\nu\rho}^{(2)}$ are expressed as follows: $$\begin{array} {lll}
\Lambda_{\mu\nu\rho}^{(a)}(k_1,k_2)=& f_{1}^{(a)} g_{\mu\rho} k_{1 \nu}+
f_{2}^{(a)} g_{\mu\nu} k_{1 \rho}+
f_{3}^{(a)} g_{\nu\rho} k_{2 \mu}+
f_{4}^{(a)} g_{\mu\nu} k_{2 \rho}\\
&+ f_{5}^{(a)} k_{1\nu} k_{1 \rho} k_{2 \mu}+
f_{6}^{(a)} k_{1\nu} k_{2 \rho} k_{2 \mu}\\
&+(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_2}, x_i \rightarrow y_i),
\end{array}
\eqno {(A.b.2)}$$ where $a=1,2$, and the $f_{i}^{(1)},f_{i}^{(2)}$ are given in terms of the Passarino-Veltman functions with internal stop lines $C_{ij}^{(1)}(=C_{ij}[-k_1,-k_2,m_{\tilde{t}_{1}},
m_{\tilde{t}_{1}},m_{\tilde{t}_{1}}])$ and internal gluino lines $C_{ij}^{(2)}(=C_{ij}[-k_1,-k_2,m_{\tilde{g}},
m_{\tilde{g}},m_{\tilde{g}}])$. For simplicity, we abbreviate the definite part of C integral functions (using the definitions of [@s19][@s20]) as follows: $\bar{C}_{24}^{(a)}=C_{24}^{(a)}-\frac{1}{4}\Delta$, $\bar{C}_{35}^{(a)}=C_{35}^{(a)}+\frac{1}{6} \Delta$, $\bar{C}_{36}^{(a)}=C_{35}^{(a)}+\frac{1}{12} \Delta$. ($a=1,2$)
$$\begin{array} {lll}
f_{1}^{(1)} = -8 \bar{C}_{24}^{(1)}-8 \bar{C}_{35}^{(1)},
\end{array}$$ $$\begin{array} {lll}
f_{2}^{(1)} = -4 \bar{C}_{24}^{(1)}-8 \bar{C}_{35}^{(1)},
\end{array}$$ $$\begin{array} {lll}
f_{3}^{(1)} = -8 \bar{C}_{36}^{(1)},
\end{array}
\eqno {(A.b.3)}$$ $$\begin{array} {lll}
f_{4}^{(1)} = -4 \bar{C}_{24}^{(1)}-8 \bar{C}_{36}^{(1)},
\end{array}$$ $$\begin{array} {lll}
f_{5}^{(1)} = 4 C_{12}^{(1)}+12 C_{23}^{(1)}+8 C_{33}^{(1)},
\end{array}$$ $$\begin{array} {lll}
f_{6}^{(1)} = 4 C_{12}^{(1)}+8 C_{22}^{(1)}+4 C_{23}^{(1)}+8 C_{34}^{(1)},
\end{array}$$ and $$\begin{array} {lll}
f_{1}^{(2)} =& -8 m_{\tilde{g}}^{2} C_{0}^{(2)}-4 m_{\tilde{g}}^{2} C_{11}^{(2)}-
16 \bar{C}_{24}^{(2)}+12 \epsilon C_{24}^{(2)}-
8 \bar{C}_{35}^{(2)}+6 \epsilon C_{35}^{(2)}\\
& + 8 k_1 \cdot k_2 C_{12}^{(2)} +
16 k_1 \cdot k_2 C_{23}^{(2)} +
8 k_1 \cdot k_2 C_{33}^{(2)},
\end{array}$$ $$\begin{array} {lll}
f_{2}^{(2)} = -4 m_{\tilde{g}}^{2} C_{11}^{(2)}-
8 \bar{C}_{35}^{(2)}+6 \epsilon C_{35}^{(2)}+
8 C_{23}^{(2)} k_1 \cdot k_2+
8 C_{33}^{(2)} k_1 \cdot k_2,
\end{array}$$ $$\begin{array} {lll}
f_{3}^{(2)} =& 4 m_{\tilde{g}}^{2} C_{0}^{(2)}-4 m_{\tilde{g}}^{2} C_{12}^{(2)}+
8 \bar{C}_{24}^{(2)}-6 \epsilon C_{24}^{(2)}-
8 \bar{C}_{36}^{(2)}+6 \epsilon C_{36}^{(2)}\\
& + 8 k_1 \cdot k_2 C_{22}^{(2)}+
8 k_1 \cdot k_2 C_{34}^{(2)},
\end{array}
\eqno {(A.b.4)}$$ $$\begin{array} {lll}
f_{4}^{(2)} = &-4 m_{\tilde{g}}^{2} C_{0}^{(2)}-4 m_{\tilde{g}}^{2} C_{12}^{(2)}-
8 \bar{C}_{24}^{(2)}+6 \epsilon C_{24}^{(2)}-
8 \bar{C}_{36}^{(2)}+6 \epsilon C_{36}^{(2)}+\\
&+ 8 k_1 \cdot k_2 C_{12}^{(2)} +
8 k_1 \cdot k_2 C_{22}^{(2)} +
8 k_1 \cdot k_2 C_{23}^{(2)} +
8 k_1 \cdot k_2 C_{34}^{(2)},
\end{array}$$ $$\begin{array} {lll}
f_{5}^{(2)} = -8 C_{12}^{(2)}-24 C_{23}^{(2)}-16 C_{33}^{(2)}
\end{array}$$ $$\begin{array} {lll}
f_{6}^{(2)} = -8 C_{12}^{(2)}-16 C_{22}^{(2)}-8 C_{23}^{(2)}-16 C_{34}^{(2)}
\end{array}$$
Similarly, the $gtt$ vertex functions are composed of left-handed and right-handed contributions plus a counterterm: (We define $a$ as color index of external gluon and $i,j$ as colors of external top-quarks) $$\begin{array} {lll}
\Lambda_{\mu,(ij)}^{a}(p_1,p_2)= &
-\frac{g_{s} \hat{g}_{s}^{2}}{16 \pi^2} T^{a}_{ij} \{ (2 C_{F}-C_{A})
( \Lambda_{\mu}^{(1L)}(p_1,p_2) P_{L}+
\Lambda_{\mu}^{(1R)}(p_1,p_2) P_{R})\\
& +C_{A}
(\Lambda_{\mu}^{(2L)}(p_1,p_2) P_{L}+\Lambda_{\mu}^{(2R)}(p_1,p_2) P_{R})\}\\
&+(m_{\tilde{t}_1} \rightarrow m_{\tilde{t}_2}, x_i \rightarrow y_i)+
\Lambda_{\mu} ^{(CT)}\\
\end{array}
\eqno {(A.b.5)}$$ The expressions of $\Lambda_{\mu}^{(n)},n=1L,1R,2L,2R$ are given as following: $$\begin{array} {lll}
\Lambda_{\mu}^{(n)}(p_1,p_2)= &h_{1}^{(n)} \gamma_{\mu}+
h_{2}^{(n)} p_{1 \mu}+
h_{3}^{(n)} p_{2 \mu}+
h_{4}^{(n)} \rlap/p_{1} p_{1 \mu}\\
&+h_{5}^{(n)} \rlap/p_{1} p_{2 \mu}+
h_{6}^{(n)} \rlap/p_{2} p_{1 \mu}+
h_{7}^{(n)} \rlap/p_{2} p_{2 \mu}+
h_{8}^{(n)} \gamma_{\mu} \rlap/p_{1}\\
&+h_{9}^{(n)} \gamma_{\mu} \rlap/p_{2}+
+h_{10}^{(n)} \gamma_{\mu} \rlap/p_{1} \rlap/p_{2}
\end{array}
\eqno {(A.b.6)}$$
We define $$\begin{array} {lll}
C_{0}^{(3)},C_{ij}^{(3)}=C_{0},
C_{ij}[-p_1,-p_2,m_{\tilde{t}_{1}},m_{\tilde{g}},m_{\tilde{t}_{1}}]
\end{array}$$ $$\begin{array} {lll}
C_{0}^{(4)},C_{ij}^{(4)}= C_{0},
C_{ij}[-p_1,-p_2,m_{\tilde{g}},m_{\tilde{t}_{1}},m_{\tilde{g}}]
\end{array}$$ Then we can get $h_{i}^{(n)}$ as follows: (i=1,2,..., 10) $$\begin{array} {lll}
h_{1}^{(1L)}=-2 x_{2} x_{4} C_{24}^{(3)}
\end{array}$$ $$\begin{array} {lll}
h_{2}^{(1L)}=x_{2} x_{3} m_{\tilde{g}} (C_{0}^{(3)} +2 C_{11}^{(3)} )
\end{array}$$ $$\begin{array} {lll}
h_{3}^{(1L)}=x_{2} x_{3} m_{\tilde{g}} (C_{0}^{(3)} +2 C_{12}^{(3)} )
\end{array}$$ $$\begin{array} {lll}
h_{4}^{(1L)}=x_{2} x_{4} (C_{0}^{(3)} +3 C_{11}^{(3)} +2 C_{21}^{(3)} )
\end{array}
\eqno {(A.b.7)}$$ $$\begin{array} {lll}
h_{5}^{(1L)}=x_{2} x_{4} (C_{0}^{(3)} +C_{11}^{(3)} +2 C_{12}^{(3)} +2 C_{23}^{(3)} )
\end{array}$$ $$\begin{array} {lll}
h_{6}^{(1L)}=x_{2} x_{4} (C_{12}^{(3)} +2 C_{23}^{(3)} )
\end{array}$$ $$\begin{array} {lll}
h_{7}^{(1L)}=x_{2} x_{4} (C_{12}^{(3)} +2 C_{22}^{(3)} )
\end{array}$$ $$\begin{array} {lll}
h_{8}^{(1L)}=h_{9}^{(1L)}=h_{10}^{(1L)}=0
\end{array}$$ and $$\begin{array} {lll}
h_{1}^{(2L)}=&x_{2} x_{4} (-m_{\tilde{g}}^{2} C_{0}^{(4)}-2 C_{24}^{(4)}+
\epsilon C_{24}^{(4)})\\
&+x_{2} x_{4} p_{1}^{2} (C_{11}^{(4)}+C_{21}^{(4)})
+2 x_{2} x_{4} p_{1} \dot p_{2} (C_{12}^{(4)}+C_{23}^{(4)}) \\
&+ x_{2} x_{4} p_{2}^{2} (C_{12}^{(4)}+C_{22}^{(4)})
\end{array}$$ $$\begin{array} {lll}
h_{2}^{(2L)}=2 x_{2} x_{3} m_{\tilde{g}}^{2} C_{11}^{(4)}
\end{array}$$ $$\begin{array} {lll}
h_{3}^{(2L)}=2 x_{2} x_{3} m_{\tilde{g}}^{2} C_{12}^{(4)}
\end{array}$$ $$\begin{array} {lll}
h_{4}^{(2L)}=-2 x_{2} x_{4}(C_{11}^{(4)}+C_{21}^{(4)})
\end{array}
\eqno {(A.b.8)}$$ $$\begin{array} {lll}
h_{5}^{(2L)}=-2 x_{2} x_{4} (C_{12}^{(4)}+C_{23}^{(4)})
\end{array}$$ $$\begin{array} {lll}
h_{6}^{(2L)}=-2 x_{2} x_{4} (C_{11}^{(4)}+C_{23}^{(4)})
\end{array}$$ $$\begin{array} {lll}
h_{7}^{(2L)}=-2 x_{2} x_{4} (C_{12}^{(4)}+C_{22}^{(4)})
\end{array}$$ $$\begin{array} {lll}
h_{8}^{(2L)}=h_{9}^{(2L)}=x_{2} x_{3} m_{\tilde{g}}^{2} C_{0}^{(4)}
\end{array}$$ $$\begin{array} {lll}
h_{10}^{(2L)}=x_{2} x_{4} (C_{11}^{(4)}-C_{12}^{(4)})
\end{array}$$ $h_{i}^{(1R)}$ and $h_{i}^{(2R)}$ can be obtained by exchanging $x_{1} \leftrightarrow x_{2}$ and $x_{3} \leftrightarrow x_{4}$ in $h_{i}^{(1L)}$ and $h_{i}^{(2L)}$. ($i=1,2,...,10$)
The counter terms are given by: $$\begin{array} {lll}
\Lambda_{\mu}^{(CT)}= -C_{F} \frac{g_{s}}{2} T^{a}_{ij}
\gamma _{\mu} \left[ (\delta Z_{L}+ \delta Z_{L}^{\dagger}) P_{L}+
(\delta Z_{R}+ \delta Z_{R}^{\dagger}) P_{R}\right]
\end{array}
\eqno {(A.b.10)}$$ The wave function renormalization constants can be obtained from Eq.(3.c.12) and Eq.(3.c.13).
C. Box corrections:
Finally, we list the four form factors $F_{\mu\nu}^{ti}$ as given in Eq.(3.e.1) in terms of Passarino-Veltmann functions. First, we define $F_{k}^{tiL}$ and $F_{k}^{tiR}$ by: $$\begin{array} {lll}
F_{\mu\nu}^{(ti)} ~ =&~ \frac{i \hat{g}_{s}^{2}}{16 \pi^{2}} P_{R}[
\gamma _{\mu } \gamma _{\nu }F_{1}^{(tiR)}+
\gamma _{\nu } \gamma _{\mu }F_{2}^{(tiR)}\\
&+p_{1 \nu }\gamma _{\mu }F_{3}^{(tiR)} +
p_{2 \nu } \gamma _{\mu }F_{4}^{(tiR)}\\
&+p_{1 \mu }\gamma _{\nu } F_{5}^{(tiR)} +
p_{2 \mu }\gamma _{\nu } F_{6}^{(tiR)}\\
&+\gamma_{\mu} \gamma _{\nu } \rlap/k_{1} F_{7}^{(tiR)} +
\gamma_{\nu} \gamma _{\mu } \rlap/k_{1} F_{8}^{(tiR)}\\
&+\rlap/k_{1} p_{1 \mu} p_{2 \nu } F_{9}^{(tiR)} +
\rlap/k_{1} p_{2 \mu} p_{1 \nu } F_{10}^{(tiR)} \\
&+\gamma_{\mu} \rlap/k_{1} p_{1 \nu } F_{11}^{(tiR)} +
\gamma_{\mu} \rlap/k_{1} p_{2 \nu } F_{12}^{(tiR)} \\
&+\gamma_{\nu} \rlap/k_{1} p_{1 \mu } F_{13}^{(tiR)} +
\gamma_{\nu} \rlap/k_{1} p_{2 \mu } F_{14}^{(tiR)} \\
&+p_{1 \mu} p_{2 \nu } F_{15}^{(tiR)} +
p_{1 \mu} p_{1 \nu } F_{16}^{(tiR)} \\
&+p_{2 \mu} p_{1 \nu } F_{17}^{(tiR)} +
\rlap/k_{1} p_{1 \mu} p_{1 \nu } F_{18}^{(tiR)}\\
&+\rlap/k_{1} p_{2 \mu} p_{2 \nu } F_{19}^{(tiR)} +
p_{2 \mu} p_{2 \nu } F_{20}^{(tiR)} ]\\
&+(P_{R} \rightarrow P_{L}, F_{k}^{(tiR)} \rightarrow F_{k}^{(tiL)} ),
(k=1 \sim 20,~i=1 \sim 4).
\end{array}
\eqno{(A.c.1)}$$ In the following we only give the expressions of $F_{k}^{tiR}
(k=1,2,...,20~and ~i=1 \sim 4)$. The expressions of $F_{k}^{tiL}$ can be obtained from $F_{k}^{tiR}$ by exchanging $x_{1} \leftrightarrow x_{2}$ and $x_{3} \leftrightarrow x_{4}$. Furthermore, the form factors in the u-channel are given by $$\begin{array} {lll}
F_{\mu\nu}^{(ui)}(k_1,k_2,p_1,p_2)=F_{\nu\mu}^{(ti)}(k_2,k_1,p_1,p_2)
\end{array}
\eqno{(A.c.2)}$$ The expressions of $F_{k}^{(t1 R)}~(k=1 \sim 20)$ are given as below: $$\begin{array} {lll}
F_{1}^{(t1 R)}~ &=~F_{2}^{(t1 R)} \\
&=-2 x_{1} x_{4} m_{\tilde{g}} D_{27}^{(1)}
+2 x_{2} x_{4} m_{t} D_{311}^{(1)} \\
&~+2 (x_{1} x_{3}-x_{2} x_{4}) m_{t} D_{313}^{(1)} ,
\end{array}
\eqno{(A.c.3)}$$ $$\begin{array} {lll}
F_{3}^{(t1 R)}~ =~4 x_{1} x_{3} (D_{311}^{(1)}-D_{312}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{4}^{(t1 R)}~ =~ - 4 x_{1} x_{3} (D_{27}^{(1)}+D_{312}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{5}^{(t1 R)}~ =~ 4 x_{1} x_{3}(D_{27}^{(1)}+D_{311}^{(1)}-D_{313}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{6}^{(t1 R)}~ =~ - 4 x_{1} x_{3} D_{313}^{(1)} ,
\end{array}$$ $$\begin{array} {lll}
F_{7}^{(t1 R)}~ =F_{8}^{(t1 R)}~= 2 x_{1}
x_{3}(D_{313}^{(1)}-D_{312}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{9}^{(t1 R)}~ =&~ 4 x_{1} x_{3} (D_{13}^{(1)}-D_{12}^{(1)}
-D_{22}^{(1)}-D_{23}^{(1)}-D_{24}^{(1)} \\
&+D_{25}^{(1)}+2 D_{26}^{(1)}-D_{36}^{(1)}
+D_{38}^{(1)}-D_{39}^{(1)}+D_{310}^{(1)}),
\end{array}$$ $$\begin{array} {lll}
F_{10}^{(t1 R)}~ =~ 4 x_{1} x_{3} (D_{37}^{(1)}
-D_{39}^{(1)}+D_{38}^{(1)}-D_{310}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{11}^{(t1 R)} =F_{12}^{(t1 R)}=
F_{13}^{(t1 R)}= F_{14}^{(t1 R)}= 0 ,
\end{array}$$ $$\begin{array} {lll}
F_{15}^{(t1 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{13}^{(1)}-D_{23}^{(1)}
+D_{25}^{(1)}+D_{26}^{(1)}-D_{39}^{(1)}+D_{310}^{(1)}) \\
&-4 x_{1} x_{4} m_{\tilde{g}} (D_{0}^{(1)}+D_{11}^{(1)}
+D_{12}^{(1)}-D_{13}^{(1)}+D_{24}^{(1)}-D_{26}^{(1)}) \\
&+4 x_{2} x_{4} m_{t} (D_{11}^{(1)}-D_{13}^{(1)}
+D_{21}^{(1)}+D_{23}^{(1)}+D_{24}^{(1)}-2 D_{25}^{(1)}\\
&-D_{26}^{(1)}+D_{34}^{(1)}+D_{39}^{(1)}-2 D_{310}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{16}^{(t1 R)}~ =&~ 4 x_{1} x_{3} m_{t} (-D_{25}^{(1)}
+D_{26}^{(1)}-D_{35}^{(1)}+D_{37}^{(1)}-D_{39}^{(1)}+D_{310}^{(1)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (D_{11}^{(1)}-D_{12}^{(1)}
+D_{21}^{(1)}-D_{24}^{(1)}-D_{25}^{(1)}+D_{26}^{(1)})\\
&-4 x_{2} x_{4} m_{t} (D_{21}^{(1)}-D_{24}^{(1)}-D_{25}^{(1)}
+D_{26}^{(1)}+D_{31}^{(1)} \\
&-D_{34}^{(1)}-2 D_{35}^{(1)}+D_{37}^{(1)}-D_{39}^{(1)}+2 D_{310}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{17}^{(t1 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{37}^{(1)}-D_{39}^{(1)})
- 4 x_{1} x_{4} m_{\tilde{g}} (D_{25}^{(1)}-D_{26}^{(1)}) \\
&+ 4 x_{2} x_{4} m_{t} (D_{35}^{(1)}-D_{37}^{(1)}+D_{39}^{(1)}-D_{310}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{18}^{(t1 R)}~ =&~ 4 x_{1} x_{3}
(-D_{22}^{(1)}+D_{24}^{(1)}-D_{25}^{(1)}+D_{26}^{(1)}+D_{34}^{(1)} \\
&-D_{35}^{(1)}-D_{36}^{(1)}+D_{37}^{(1)}+D_{38}^{(1)}-D_{39}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{19}^{(t1 R)}~ =~ 4 x_{1} x_{3}
(-D_{23}^{(1)}+D_{26}^{(1)}-D_{39}^{(1)}+D_{38}^{(1)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{20}^{(t1 R)}~ =&~ 4 x_{1} x_{3} m_{t}
(-D_{23}^{(1)}-D_{39}^{(1)})+ 4 x_{1} x_{4} m_{\tilde{g}}
(D_{13}^{(1)}+D_{26}^{(1)}) \\
& +4 x_{2} x_{4} m_{t} (D_{23}^{(1)}-D_{25}^{(1)}
+D_{39}^{(1)}-D_{310}^{(1)}) ,
\end{array}$$ where we denote $D_{i}^{(1)},D_{ij}^{(1)},D_{ijk}^{(1)}=
D_{i},D_{ij},D_{ijk} [-p_{1},k_{1},k_{2},m_{\tilde{g}},m_{\tilde{t_{1}}},m_{\tilde{t_{1}}},
m_{\tilde{t_{1}}}]$.
The expressions of $F_{k}^{(t2 R)}~(k=1 \sim 20)$ are as follows: $$\begin{array} {lll}
F_{1}^{(t2 R)}~ =~ 2 x_{1} x_{3} m_{t} (D_{27}^{(2)}+D_{313}^{(2)})+
2 x_{1} x_{4} m_{\tilde{g}} D_{27}^{(2)}- 2 x_{2} x_{4} m_{t} (D_{27}^{(2)}+D_{311}^{(2)}),
\end{array}
\eqno{(A.c.4)}$$ $$\begin{array} {lll}
F_{2}^{(t2 R)}~ =~ 2 x_{1} x_{3} m_{t} D_{313}^{(2)}+
2 x_{1} x_{4} m_{\tilde{g}} D_{27}^{(2)}- 2 x_{2} x_{4} m_{t} D_{311}^{(2)},
\end{array}$$ $$\begin{array} {lll}
F_{3}^{(t2 R)}~ =~ 4 x_{1} x_{3} (-D_{27}^{(2)}-D_{311}^{(2)}+D_{312}^{(2)}),
\end{array}$$ $$\begin{array} {lll}
F_{4}^{(t2 R)}~ =~ 4 x_{1} x_{3} (D_{312}^{(2)}-D_{313}^{(2)}),
\end{array}$$ $$\begin{array} {lll}
F_{5}^{(t2 R)}~ =&~ 8 x_{1} x_{3} (D_{27}^{(2)}+D_{311}^{(2)}) +
2 x_{1} x_{3} m_{\tilde{g}}^{2} (D_{0}^{(2)}+D_{11}^{(2)})\\
&+2 x_{1} x_{3} m_{t}^{2} (-D_{11}^{(2)}-D_{13}^{(2)}-2 D_{21}^{(2)}-D_{23}^{(2)}
-D_{25}^{(2)}-D_{31}^{(2)}-D_{37}^{(2)})\\
&+2 (x_{2} x_{3}+x_{1} x_{4}) m_{\tilde{g}} m_{t} (D_{0}^{(2)}+D_{11}^{(2)})-
2 x_{2} x_{4} m_{t}^{2} (D_{11}^{(2)}-D_{13}^{(2)}+D_{21}^{(2)}-D_{25}^{(2)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{1} (D_{12}^{(2)}+2 D_{24}^{(2)}+D_{34}^{(2)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (D_{13}^{(2)}+D_{25}^{(2)}+D_{26}^{(2)}+D_{310}^{(2)})-
4 x_{1} x_{3} p_{1} \cdot p_{2} (D_{13}^{(2)}+2 D_{25}^{(2)}+ D_{35}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{6}^{(t2 R)}~ =&~
2 x_{1} x_{3} m_{\tilde{g}}^{2} D_{13}^{(2)}\\
&-2 x_{1} x_{3} m_{t}^{2} (D_{23}^{(2)}+D_{25}^{(2)}+D_{33}^{(2)}+D_{35}^{(2)})+
2 (x_{2} x_{3}+x_{1} x_{4}) m_{\tilde{g}} m_{t} D_{13}^{(2)}\\
&+2 x_{2} x_{4} m_{t}^{2} (D_{23}^{(2)}-D_{25}^{(2)})+
4 x_{1} x_{3} k_{1} \cdot p_{1} (D_{26}^{(2)}+D_{310}^{(2)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (D_{23}^{(2)}+D_{39}^{(2)})-
4 x_{1} x_{3} p_{1} \cdot p_{2} (D_{23}^{(2)}+D_{37}^{(2)})\\
&+ 8 x_{1} x_{3} D_{313}^{(2)},
\end{array}$$ $$\begin{array} {lll}
F_{7}^{(t2 R)}~ =~ 2 x_{1} x_{3} (D_{27}^{(2)}+D_{312}^{(2)}),
\end{array}$$ $$\begin{array} {lll}
F_{8}^{(t2 R)}~ =~ 2 x_{1} x_{3} D_{312}^{(2)},
\end{array}$$ $$\begin{array} {lll}
F_{9}^{(t2 R)}~ =~ 4 x_{1} x_{3} (D_{12}^{(2)}-D_{13}^{(2)}+D_{22}^{(2)}+
D_{24}^{(2)}-D_{25}^{(2)}-D_{26}^{(2)}+D_{36}^{(2)}-D_{310}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{10}^{(t2 R)}~ =~ 4 x_{1} x_{3} (D_{38}^{(2)}-D_{310}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{11}^{(t2 R)} =F_{12}^{(t2 R)}= 0 ,
\end{array}$$ $$\begin{array} {lll}
F_{13}^{(t2 R)}~ =&~ 2 x_{1} x_{3} m_{t} (-D_{12}^{(2)}+D_{13}^{(2)}-D_{24}^{(2)}+D_{25}^{(2)})+
2 x_{1} x_{4} m_{\tilde{g}} (D_{0}^{(2)}+D_{11}^{(2)})\\
&-2 x_{2} x_{4} m_{t} (D_{11}^{(2)}-D_{12}^{(2)}+D_{21}^{(2)}-D_{24}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{14}^{(t2 R)}~ =~ 2 x_{1} x_{3} m_{t} (D_{23}^{(2)}-D_{26}^{(2)})+
2 x_{1} x_{4} m_{\tilde{g}} D_{13}^{(2)}-
2 x_{2} x_{4} m_{t} (D_{25}^{(2)}-D_{26}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{15}^{(t2 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{12}^{(2)}-D_{13}^{(2)}-D_{23}^{(2)}+
D_{24}^{(2)}-D_{25}^{(2)}+D_{26}^{(2)}-D_{37}^{(2)}+D_{310}^{(2)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (D_{12}^{(2)}-D_{13}^{(2)}+D_{24}^{(2)}-D_{25}^{(2)})\\
&-4 x_{2} x_{4} m_{t} (D_{12}^{(2)}-D_{13}^{(2)}+2 D_{24}^{(2)}-2 D_{25}^{(2)}+D_{34}^{(2)}-D_{35}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{16}^{(t2 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{12}^{(2)}-D_{13}^{(2)}+D_{24}^{(2)}-2 D_{25}^{(2)}+
D_{26}^{(2)}-D_{35}^{(2)}+D_{310}^{(2)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (-D_{0}^{(2)}-2 D_{11}^{(2)}+D_{12}^{(2)}-D_{21}^{(2)}+D_{24}^{(2)})\\
&+4 x_{2} x_{4} m_{t} (D_{11}^{(2)}-D_{12}^{(2)}+2 D_{21}^{(2)}-2 D_{24}^{(2)}+D_{31}^{(2)}-D_{34}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{17}^{(t2 R)}~ =&~ 4 x_{1} x_{3} m_{t} (-D_{23}^{(2)}+D_{26}^{(2)}-D_{37}^{(2)}+
D_{39}^{(2)})+
4 x_{1} x_{4} m_{\tilde{g}} (-D_{13}^{(2)}-D_{25}^{(2)}+D_{26}^{(2)})\\
&+ 4 x_{2} x_{4} m_{t} (D_{25}^{(2)}-D_{26}^{(2)}+D_{35}^{(2)}-D_{310}^{(2)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{18}^{(t2 R)}~ =~ 4 x_{1} x_{3} (D_{22}^{(2)}-D_{24}^{(2)}-D_{34}^{(2)}
+D_{36}^{(2)}),
\end{array}$$ $$\begin{array} {lll}
F_{19}^{(t2 R)}~ =~ 4 x_{1} x_{3} (-D_{23}^{(2)}+D_{26}^{(2)}
+D_{38}^{(2)}-D_{39}^{(2)}),
\end{array}$$ $$\begin{array} {lll}
F_{20}^{(t2 R)}~ =&~ 4 x_{1} x_{3} m_{t} (-D_{23}^{(2)}+D_{26}^{(2)}-D_{33}^{(2)}+
D_{39}^{(2)})+ 4 x_{1} x_{4} m_{\tilde{g}} (-D_{23}^{(2)}+D_{26}^{(2)})\\
&+4 x_{2} x_{4} m_{t} (D_{23}^{(2)}-D_{26}^{(2)}+D_{37}^{(2)}-D_{310}^{(2)}) ,
\end{array}$$ where $D_{i}^{(2)},D_{ij}^{(2)},D_{ijk}^{(2)} ~=~ D_{i},D_{ij},D_{ijk}
[-p_{1},k_{1},-p_{2},m_{\tilde{g}},m_{\tilde{t_{1}}},m_{\tilde{t_{1}}},m_{\tilde{g}}]$.
The expressions for $F_{k}^{(t3 R)}~(k=1 \sim 20)$ are written as: $$\begin{array} {lll}
F_{1}^{(t3 R)}~ =&
2 x_{1} x_{3} m_{t} (D_{27}^{(3)}+2 D_{313}^{(3)})+
x_{1} x_{3} m_{t} m_{\tilde{g}}^{2} (D_{0}^{(3)}+D_{13}^{(3)})\\
&-x_{1} x_{3} m_{t}^{3} (D_{0}^{(3)}+2 D_{11}^{(3)}-D_{13}^{(3)}+D_{21}^{(3)}+2 D_{33}^{(3)}
+2 D_{35}^{(3)}-2 D_{37}^{(3)})\\
&+2 x_{1} x_{4} m_{\tilde{g}} D_{27}^{(3)}+
x_{1} x_{4} m_{\tilde{g}}^{3} D_{0}^{(3)}\\
&-x_{1} x_{4} m_{t}^{2} m_{\tilde{g}} (D_{0}^{(3)}+2 D_{11}^{(3)}-
2 D_{13}^{(3)}+D_{21}^{(3)}+2 D_{23}^{(3)}-2 D_{25}^{(3)})\\
&+2 x_{2} x_{4} m_{t} (D_{27}^{(3)}+2 D_{311}^{(3)}-2 D_{313}^{(3)}) +
x_{2} x_{4} m_{t} m_{\tilde{g}}^{2} (D_{11}^{(3)}-D_{13}^{(3)}) \\
&-x_{2} x_{4} m_{t}^{3} (D_{11}^{(3)}-D_{13}^{(3)}+2 D_{21}^{(3)}+2 D_{23}^{(3)}-4 D_{25}^{(3)}+
D_{31}^{(3)}-2 D_{33}^{(3)}-3 D_{35}^{(3)}+4 D_{37}^{(3)})\\
&+2 x_{1} x_{3} m_{t} k_{1} \cdot p_{1} (D_{12}^{(3)}-D_{13}^{(3)}-D_{23}^{(3)}+D_{24}^{(3)}
+D_{33}^{(3)}-D_{37}^{(3)}-D_{39}^{(3)}+D_{310}^{(3)})\\
&+2 x_{1} x_{4} m_{\tilde{g}} k_{1} \cdot p_{1} (D_{11}^{(3)}+D_{12}^{(3)}-
2 D_{13}^{(3)}+D_{23}^{(3)}+D_{24}^{(3)}-D_{25}^{(3)}-D_{26}^{(3)})\\
&+2 x_{2} x_{4} m_{t} k_{1} \cdot p_{1} (D_{11}^{(3)}-D_{13}^{(3)}+D_{21}^{(3)}+2 D_{23}^{(3)}\\
&+D_{24}^{(3)}-3 D_{25}^{(3)}-D_{26}^{(3)}-D_{33}^{(3)}+D_{34}^{(3)}-D_{35}^{(3)}+2 D_{37}^{(3)}
+D_{39}^{(3)}-2 D_{310}^{(3)})\\
&+2 x_{1} x_{3} m_{t} k_{1} \cdot p_{2} (-D_{13}^{(3)}-D_{26}^{(3)}+D_{33}^{(3)}-D_{39}^{(3)})\\
&+2 x_{1} x_{4} m_{\tilde{g}} k_{1} \cdot p_{2} (-D_{13}^{(3)}+D_{23}^{(3)}-D_{26}^{(3)})\\
&+2 x_{2} x_{4} m_{t} k_{1} \cdot p_{2} (D_{23}^{(3)}-D_{25}^{(3)}-D_{33}^{(3)}+
D_{37}^{(3)}+D_{39}^{(3)}-D_{310}^{(3)})\\
&+2 x_{1} x_{3} m_{t} p_{1} \cdot p_{2} (D_{13}^{(3)}+D_{25}^{(3)}-D_{33}^{(3)}+D_{37}^{(3)})\\
&+2 x_{1} x_{4} m_{\tilde{g}} p_{1} \cdot p_{2} (D_{13}^{(3)}-D_{23}^{(3)}+D_{25}^{(3)})\\
&+2 x_{2} x_{4} m_{t} p_{1} \cdot p_{2}
(-D_{23}^{(3)}+D_{25}^{(3)}+D_{33}^{(3)}+D_{35}^{(3)}-D_{37}^{(3)}),
\end{array}
\eqno{(A.c.5)}$$ $$\begin{array} {lll}
F_{2}^{(t3 R)}~ =~ - 2 x_{1} x_{4} m_{\tilde{g}} D_{27}^{(3)}
-2 x_{1} x_{3} m_{t} D_{313}^{(3)}
- 2 x_{2} x_{4} m_{t} (D_{27}^{(3)}+D_{311}^{(3)}-D_{313}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{3}^{(t3 R)}~ =&
4 x_{1} x_{3} (D_{27}^{(3)}+2 D_{311}^{(3)}+D_{312}^{(3)}-3 D_{313}^{(3)})+
2 x_{1} x_{3} m_{\tilde{g}}^{2} (-D_{0}^{(3)}+D_{11}^{(3)}-D_{13}^{(3)})\\
&+2 x_{1} x_{3} m_{t}^{2} (D_{13}^{(3)}-D_{21}^{(3)}+2 D_{25}^{(3)}-D_{31}^{(3)}+2 D_{33}^{(3)}+
3 D_{35}^{(3)}-4 D_{37}^{(3)})\\
&-2 x_{2} x_{3} m_{t} m_{\tilde{g}} D_{0}^{(3)}-
2 x_{1} x_{4} m_{t} m_{\tilde{g}} D_{0}^{(3)}-
2 x_{2} x_{4} m_{t}^{2} (D_{0}^{(3)}+D_{11}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{1} (D_{23}^{(3)}+D_{24}^{(3)}-D_{25}^{(3)}
-D_{26}^{(3)}\\
&-D_{33}^{(3)}+D_{34}^{(3)}-D_{35}^{(3)}+2 D_{37}^{(3)}+D_{39}^{(3)}
-2 D_{310}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (-D_{25}^{(3)}+D_{26}^{(3)}-D_{33}^{(3)}+D_{37}^{(3)}+
D_{39}^{(3)}-D_{310}^{(3)})\\
&+4 x_{1} x_{3} p_{1} \cdot p_{2} (D_{33}^{(3)}+D_{35}^{(3)}-2 D_{37}^{(3)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{4}^{(t3 R)}~ =&
4 x_{1} x_{3} D_{312}^{(3)}-
2 x_{1} x_{3} m_{\tilde{g}}^{2} D_{0}^{(3)}\\
&+2 x_{1} x_{3} m_{t}^{2} (D_{11}^{(3)}+D_{21}^{(3)}+2 D_{23}^{(3)}-2 D_{25}^{(3)})\\
&-2 x_{2} x_{3} m_{t} m_{\tilde{g}} D_{0}^{(3)}-
2 x_{1} x_{4} m_{t} m_{\tilde{g}} D_{0}^{(3)}\\
&-2 x_{2} x_{4} m_{t}^{2} (D_{0}^{(3)}+D_{11}^{(3)})+
4 x_{1} x_{3} k_{1} \cdot p_{1} (-D_{23}^{(3)}+D_{25}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (-D_{23}^{(3)}+D_{26}^{(3)})+
4 x_{1} x_{3} p_{1} \cdot p_{2} (D_{23}^{(3)}-D_{25}^{(3)}) ,
\end{array}$$ $$\begin{array} {lll}
F_{5}^{(t3 R)}~ =& 4 x_{1} x_{3} (-D_{311}^{(3)}+D_{313}^{(3)})+
2 x_{1} x_{3} m_{\tilde{g}}^{2} D_{0}^{(3)}\\
&-2 x_{1} x_{3} m_{t}^{2} (D_{13}^{(3)}+D_{25}^{(3)})+
2 x_{2} x_{3} m_{t} m_{\tilde{g}} (D_{0}^{(3)}+D_{11}^{(3)}-
D_{13}^{(3)})\\
&+2 x_{1} x_{4} m_{t} m_{\tilde{g}} (D_{0}^{(3)}+D_{11}^{(3)}-
D_{13}^{(3)})\\
&+2 x_{2} x_{4} m_{t}^{2} (D_{0}^{(3)}+2 D_{11}^{(3)}-D_{13}^{(3)}+
D_{21}^{(3)}-D_{25}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (D_{25}^{(3)}-D_{26}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{6}^{(t3 R)}~ =&
-8 x_{1} x_{3} D_{313}^{(3)}-
2 x_{1} x_{3} m_{\tilde{g}}^{2} D_{13}^{(3)}\\
&+2 x_{1} x_{3} m_{t}^{2} (D_{25}^{(3)}+2 D_{33}^{(3)}+D_{35}^{(3)}-2 D_{37}^{(3)})-
2 x_{2} x_{3} m_{t} m_{\tilde{g}} D_{13}^{(3)}\\
&-2 x_{1} x_{4} m_{t} m_{\tilde{g}} D_{13}^{(3)}-
2 x_{2} x_{4} m_{t}^{2} (D_{13}^{(3)}+D_{25}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{1} (D_{23}^{(3)}
-D_{25}^{(3)}- D_{33}^{(3)}+D_{37}^{(3)}+D_{39}^{(3)}
-D_{310}^{(3)})\\
&+4 x_{1} x_{3} k_{1} \cdot p_{2} (-D_{33}^{(3)}+D_{39}^{(3)})+
4 x_{1} x_{3} p_{1} \cdot p_{2} (D_{33}^{(3)}-D_{37}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{7}^{(t3 R)}~ =&~ -4 x_{1} x_{3} (D_{312}^{(3)}-D_{313}^{(3)})+
x_{1} x_{3} m_{\tilde{g}}^{2} (D_{0}^{(3)}-D_{12}^{(3)}+D_{13}^{(3)})\\
&+x_{1} x_{3} m_{t}^{2} (-D_{11}^{(3)}+D_{12}^{(3)}-D_{13}^{(3)}-D_{21}^{(3)}\\
&+2 D_{24}^{(3)}-2 D_{26}^{(3)}-2 D_{33}^{(3)}+D_{34}^{(3)}
-D_{35}^{(3)}+2 D_{37}^{(3)}
+2 D_{39}^{(3)}-2 D_{310}^{(3)})\\
&+(x_{2} x_{3}+x_{1} x_{4}) m_{t} m_{\tilde{g}} D_{0}^{(3)}
+ x_{2} x_{4} m_{t}^{2} (D_{0}^{(3)}+D_{11}^{(3)})\\
&+2 x_{1} x_{3} k_{1} \cdot p_{1} (-D_{22}^{(3)}
-D_{23}^{(3)}+2 D_{26}^{(3)} + D_{33}^{(3)}-D_{36}^{(3)}
-D_{37}^{(3)}+D_{38}^{(3)}-2 D_{39}^{(3)}+2 D_{310}^{(3)})\\
&+2 x_{1} x_{3} k_{1} \cdot p_{2} (D_{33}^{(3)}+D_{38}^{(3)}-2 D_{39}^{(3)})\\
&+2 x_{1} x_{3} p_{1} \cdot p_{2} (D_{25}^{(3)}-D_{26}^{(3)}-
D_{33}^{(3)}+D_{37}^{(3)}+D_{39}^{(3)}-D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{8}^{(t3 R)}~ =~ 2 x_{1} x_{3} (D_{27}^{(3)}+D_{312}^{(3)}-D_{313}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{9}^{(t3 R)}~ =~ 4 x_{1} x_{3} (D_{22}^{(3)}+D_{23}^{(3)}-D_{25}^{(3)}-D_{26}^{(3)}
+D_{36}^{(3)}-D_{38}^{(3)}+D_{39}^{(3)}-D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{10}^{(t3 R)}~ =~ 4 x_{1} x_{3} (D_{25}^{(3)}-D_{26}^{(3)}
-D_{37}^{(3)}-D_{38}^{(3)}+D_{39}^{(3)}+D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{11}^{(t3 R)}~ =&~ 2 x_{1} x_{3} m_{t} (-D_{25}^{(3)}+D_{26}^{(3)})+
2 x_{1} x_{4} m_{\tilde{g}} (-D_{11}^{(3)}
+D_{12}^{(3)})\\
&+2 x_{2} x_{4} m_{t} (-D_{11}^{(3)}+D_{12}^{(3)}
-D_{21}^{(3)}+ D_{24}^{(3)}+D_{25}^{(3)}
-D_{26}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{12}^{(t3 R)}~ =&~ 2 x_{1} x_{3} m_{t} (D_{13}^{(3)}+D_{26}^{(3)})+
2 x_{1} x_{4} m_{\tilde{g}} D_{12}^{(3)}\\
&+2 x_{2} x_{4} m_{t} (D_{12}^{(3)}-D_{13}^{(3)}
+D_{24}^{(3)}-D_{26}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{13}^{(t3 R)}~ =&~ 2 x_{1} x_{3} m_{t} (D_{12}^{(3)}-D_{13}^{(3)}+D_{24}^{(3)}-D_{26}^{(3)})+
2 x_{1} x_{4} m_{\tilde{g}} (D_{11}^{(3)}-D_{13}^{(3)})\\
&+2 x_{2} x_{4} m_{t} (D_{11}^{(3)}-D_{12}^{(3)}+D_{21}^{(3)}-D_{24}^{(3)}-
D_{25}^{(3)}+D_{26}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{14}^{(t3 R)}~ =&~ -2 x_{1} x_{3} m_{t} (D_{13}^{(3)}+D_{26}^{(3)})-
2 x_{1} x_{4} m_{\tilde{g}} D_{13}^{(3)}\\
&+2 x_{2} x_{4} m_{t} (-D_{25}^{(3)}+D_{26}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{15}^{(t3 R)}~ =& -4 x_{1} x_{3} m_{t}
(D_{12}^{(3)}-D_{23}^{(3)}+D_{24}^{(3)}+D_{25}^{(3)}
-D_{39}^{(3)}+D_{310}^{(3)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (-D_{11}^{(3)}-D_{12}^{(3)}+D_{13}^{(3)}-
D_{24}^{(3)}+D_{26}^{(3)})\\
&+4 x_{2} x_{4} m_{t} (-D_{11}^{(3)}+D_{13}^{(3)}-D_{21}^{(3)}-D_{23}^{(3)}\\
&-D_{24}^{(3)}+2 D_{25}^{(3)}+D_{26}^{(3)}
-D_{34}^{(3)}-D_{39}^{(3)}+2 D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{16}^{(t3 R)}~ =& -4 x_{1} x_{3} m_{t} (D_{12}^{(3)}-D_{13}^{(3)}+D_{24}^{(3)}
-D_{25}^{(3)}-D_{35}^{(3)}+D_{37}^{(3)}-D_{39}^{(3)}+ D_{310}^{(3)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (-D_{12}^{(3)}+D_{13}^{(3)}+D_{21}^{(3)}-
D_{24}^{(3)}-D_{25}^{(3)}+D_{26}^{(3)})\\
&+4 x_{2} x_{4} m_{t} (D_{21}^{(3)}-D_{24}^{(3)}- D_{25}^{(3)}+D_{26}^{(3)}\\
&+D_{31}^{(3)}-D_{34}^{(3)}-2 D_{35}^{(3)}+D_{37}^{(3)}-
D_{39}^{(3)}+2 D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{17}^{(t3 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{13}^{(3)}
+D_{26}^{(3)}-D_{37}^{(3)}+D_{39}^{(3)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (D_{13}^{(3)}
-D_{25}^{(3)}+D_{26}^{(3)})+
4 x_{2} x_{4} m_{t} (-D_{35}^{(3)}
+D_{37}^{(3)}-D_{39}^{(3)}+D_{310}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{18}^{(t3 R)}~ =&~ 4 x_{1} x_{3} (D_{22}^{(3)}-D_{24}^{(3)}
+D_{25}^{(3)}-D_{26}^{(3)}-
D_{34}^{(3)}+D_{35}^{(3)}+D_{36}^{(3)}\\
& -D_{37}^{(3)}-D_{38}^{(3)}+D_{39}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{19}^{(t3 R)}~ =~ 4 x_{1} x_{3} (D_{23}^{(3)}-D_{26}^{(3)}
-D_{38}^{(3)}+D_{39}^{(3)}),
\end{array}$$ $$\begin{array} {lll}
F_{20}^{(t3 R)}~ =&~ 4 x_{1} x_{3} m_{t} (D_{13}^{(3)}+D_{23}^{(3)}
+D_{26}^{(3)}+D_{39}^{(3)})\\
&+4 x_{1} x_{4} m_{\tilde{g}} (D_{13}^{(3)}
+D_{26}^{(3)})+
4 x_{2} x_{4} m_{t} (-D_{23}^{(3)}+D_{25}^{(3)}
-D_{39}^{(3)}+D_{310}^{(3)}),
\end{array}$$ where $D_{i}^{(3)},D_{ij}^{(3)},D_{ijk}^{(3)}~=~
D_{i},D_{ij},D_{ijk} [-p_{1},k_{1},k_{2},m_{\tilde{t_{1}}},m_{\tilde{g}},m_{\tilde{g}},
m_{\tilde{g}}]$.
The $F_{k}^{(t4 R)}~(k=1 \sim 20)$ are written explicitly as: $$\begin{array} {lll}
F_{1}^{(t4 R)} &= F_{2}^{(t4 R)} \\
& =\frac{1}{2} ((x_{1} x_{3} m_{t} (C_{11}-C_{12})\\
&+x_{2} x_{4} m_{t} C_{12}-
x_{2} x_{3} m_{\tilde{g}} C_{0})[-p_1,p_1+p_2,
m_{\tilde{g}},m_{\tilde{t}_{1}},m_{\tilde{t}_{1}}])
\end{array}
\eqno{(A.c.6)}$$ $$\begin{array} {lll}
F_{i}^{(t4 R)}~ =~ 0, (i=3,4,...20).
\end{array}$$
[s20]{}
H.E. Haber and G.L. Kane, Phys. Rep. [**117**]{}, 75(1985); J.F. Gunion and H.E. Haber, Nucl. Phys. [**B272**]{}, 1(1986).
P.C. Bhat, for the D0 collaboration, talk presented at the Wine and Cheese Seminar at Fermilab, February 1997
M. Glück, J.F. Owens and E. Reya, Phys. Rev. D17,2324(1978); B.L. Combridge, Nucl.Phys.[**B151**]{},429(1979); H. Georgi, [*et al.*]{}, Ann. Phys.(N.Y.)114,273(1978).
P. Nason, S. Dawson and R.K. Ellis, Nucl. Phys. [**B303**]{}, 607(1988); G. Altarelli, M. Diemoz, G. Martinelli and P.Nason, Nucl. Phys. [**B308**]{}, 724(1988); W. Beenakker, H. Kujif, W.L. van Neerven and J. Smith, Phys. Rev. [**D40**]{}, 54(1989).
C. Li, B. Hu, J. Yang and C. Hu, Phys. Rev. [**D52**]{},5014(1995); Z. Sullivan, Reprint: hep-ph/9611302.
H.Y. Zhou, C.S. Li, Phys. Rev. [**D55**]{},4421(1997);
A. Bartl, E. Christova and W. Majerotto, Nucl. Phys. [**B460**]{}, (1996)235.
T. Gehrmann and W.J. Stirling, Z. Phys. [**C65**]{}, (1995)461; S.J. Brodsky, M. Burkardt and I. Schmidt, Nucl. Phys. [**B441**]{}, (1995)197 and the references therein.
M.Glück, E . Reya and A. Vogt, Z. Phys. [**C48**]{},471(1990); M.Glück, E . Reya and A. Vogt, Z. Phys. [**C67**]{},433(1995).
M.Glück, E . Reya and M. Stratmann, Phys. Rev. [**D53**]{}, 4775(1996) M.Stratmann, hep-ph/9710379
S.P. Martin and M.T. Vaughn, Phys. Lett. [**B318**]{},(1993)331; W. Beenakker, R. Höpker and P.M. Zerwas, Phys. Lett. [**B378**]{}, (1996)159.
W. Beenaker,R. Höpker,T. Plehn and P.M. Zerwas, DESY.96-178, October 1996.
Bernd A. Kniehl, Phys. Rep. 240(1994)211.
G. Passarino and M. Veltman, Nucl. Phys. [**B160**]{}151(1979).
[**Figure Captions**]{}
[**Fig.1**]{} Feynman diagrams at the tree-level and one-loop level in the SUSY QCD for the $gg\rightarrow t \bar{t}$ subprocess. Fig.1 (a): Tree level diagrams. Fig.1 (b.1): Self-energy diagrams (for top-quark and gluon). Fig.1 (b.2): Vertex diagrams (including tri-gluon and gluon-top-top interactions). Fig.1 (b.3): Box diagrams (only t-channel). Dashed lines represent $\tilde{t}_1,\tilde{t}_2$ in Fig.1 (b).
[**Fig.2**]{} (a) relative corrections to polarized and unpolarized cross sections of the $t\bar{t}$ production process in pp colliders as a function of $\sqrt{s}$ with input structure functions of Brodsky et al. [@Geh](LO);\
(b) relative corrections to polarized and unpolarized cross sections of the $t\bar{t}$ production process in pp colliders as a function of $\sqrt{s}$ with input structure functions of Glück et al. [@Glk] [@GLK] [@STR](NLO), in both above figures, solid line for the MSSM QCD correction with unpolarized protons, dashed line for the MSSM QCD correction with $proton(+) proton(+)$ polarization, dotted line for the MSSM QCD correction with $proton(-)
proton(-)$ polarization and dot-dashed line for the MSSM QCD correction with $proton(+)
proton(-)$ polarization;\
(c) the CP-violating parameter $\xi_{CP}$ as a function of $\sqrt{s}$, solid line for input structure functions of Glück et al(NLO), dashed line for input structure functions of Brodsky et al.(LO)
$m_{\tilde{g}}=200~GeV$, $m_{\tilde{t}_1}=250~GeV$, $m_{\tilde{t}_2}=450~GeV$ and $\theta=\phi=45^{\circ}$.
[**Fig.3**]{} (a) relative corrections to the cross section of the $t\bar{t}$ production subprocess, $\hat{\delta}_{\pm \pm}$ as a function of $\sqrt{\hat{s}}$, solid line for the MSSM QCD correction with $gluon(+) gluon(+)$ polarization and dashed line for the MSSM QCD correction with $gluon(-) gluon(-)$ polarization. (b) the CP-violating parameter $\hat{\xi}_{CP}$ of the subprocess as a function of $\sqrt{\hat{s}}$. (c) relative corrections to the cross section of the subprocess $\hat{\delta}_{+-}$ as a function of $\sqrt{\hat{s}}$.
$m_{\tilde{g}}=200~GeV$, $m_{\tilde{t}_1}=250~GeV$, $m_{\tilde{t}_2}=450~GeV$ and $\theta=\phi=45^{\circ}$.
[**Fig.4**]{} (a) cross section of the $t\bar{t}$ production subprocess via gg fusion, $\hat{\sigma}_{\pm \pm}$ as a function of $m_{\tilde{g}}$, solid line for the MSSM QCD correction with $gluon(+) gluon(+)$ polarization and dashed line for the MSSM QCD correction with $gluon(-) gluon(-)$ polarization. (b) the CP-violating parameter $\hat{\xi}_{CP}$ of the subprocess as a function of $m_{\tilde{g}}$.
$m_{\tilde{t}_1}=100~GeV$ , $m_{\tilde{t}_2}=450~GeV$, $\sqrt{\hat{s}}= 500~GeV$ and $\theta=\phi=45^{\circ}$.
[**Fig.5**]{} (a) relative corrections to the cross section of the $t\bar{t}$ production subprocess via gg fusion, $\hat{\delta}_{\pm \pm}$ as a function of $m_{\tilde{t}_{1}}$, solid line for the MSSM QCD correction with $gluon(+) gluon(+)$ polarization and dashed line for the MSSM QCD correction with $gluon(-) gluon(-)$ polarization. (b) the CP-violating parameter $\hat{\xi}_{CP}$ of the subprocess as a function of $m_{\tilde{t}_{1}}$.
$m_{\tilde{g}}=200~GeV$, $m_{\tilde{t}_2}=450~GeV$, $\sqrt{\hat{s}}=500~GeV$ and $\theta=\phi=45^{\circ}$.
[**Fig.6**]{} (a) relative corrections to the cross section of the $t\bar{t}$ production subprocess via gg fusion, $\hat{\delta}_{\pm \pm}$ as a function of $m_{\tilde{t}_{2}}$, solid line for the MSSM QCD correction with $gluon(+) gluon(+)$ polarization and dashed line for the MSSM QCD correction with $gluon(-) gluon(-)$ polarization. (b) the CP-violating parameter $\hat{\xi}_{CP}$ of the subprocess as a function of $m_{\tilde{t}_{2}}$.
$m_{\tilde{g}}=200~GeV$, $m_{\tilde{t}_1}=100~GeV$ and $\sqrt{\hat{s}}=500~GeV$ and $\theta=\phi=45^{\circ}$.
[**Fig.7**]{} (a) relative corrections to the cross section of the $t\bar{t}$ production subprocess via gg fusion, $\hat{\delta}_{\pm \pm}$ as a function of $\phi$, solid line for the MSSM QCD correction with $gluon(+) gluon(+)$ polarization and dashed line for the MSSM QCD correction with $gluon(-) gluon(-)$ polarization. (b) the CP-violating parameter $\hat{\xi}_{CP}$ of the subprocess as a function of $\phi$.
$m_{\tilde{g}}=200 GeV$,$m_{\tilde{t}_1}=150~GeV$, $m_{\tilde{t}_2}=450~GeV$ ,$\sqrt{\hat{s}}=500~GeV$ and $\theta=45^{\circ}$.
[^1]: Supported in part by Committee of National Natural Science Foundation of China and Project IV.B.12 of scientific and technological cooperation agreement between China and Austria
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper uses two new ingredients, namely stochastic differential equations satisfied by continuous-state branching processes (CSBPs), and a topology under which the Lamperti transformation is continuous, in order to provide self-contained proofs of Lamperti’s 1967 representation of CSBPs in terms of spectrally positive Lévy processes. The first proof is a direct probabilistic proof, and the second one uses approximations by discrete processes, for which the Lamperti representation is evident.'
address:
- |
Instituto de Matemáticas\
Universidad Nacional Autonoma de México, Área de la investigación científica, Circuito Exterior Ciudad Universitaria, Coyoacán 04510, México, D.F. México
- 'UPMC Univ Paris 06, Laboratoire de Probabilités et Modèles Aléatoires, CNRS UMR 7599, Case courrier 188, 4 Place Jussieu, 75252 Paris Cedex 05, France.'
- 'Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad Nacional Autónoma de México, Mexico City, A.P. 20-726, Mexico'
author:
- 'Ma. Emilia Caballero'
- Amaury Lambert
- 'Ger[ó]{}nimo Uribe Bravo'
bibliography:
- 'treescsbp.bib'
title: 'Proof(s) of the Lamperti representation of Continuous-State Branching Processes'
---
Introduction {#intro}
============
The Lamperti representation theorem
-----------------------------------
During the 1960s and early 70s, John Lamperti provided relationships between *Lévy processes* and two other classes of Markov processes. The first class was that of *continuous-state branching processes* (CSBPs for short) in [@csbp], and the second one was that of *positive self-similar* processes in [@lamp72] (then called *positive semi-stable*; the reader might also wish to consult the recent survey [@surveyexponentialfunctionals]). Here, we are interested in the former, but both relationships have had a strong impact on recent research. From now on, we will refer to the first relationship as *the* Lamperti representation. Roughly, it provides a one-to-one correspondence, via a simple *random time change*, between CSBPs and (possibly killed) Lévy process with no negative jumps. The Lamperti representation has proved useful in the study of CSBPs (as in [@binghamCBP]), but also in that of Lévy processes (e.g. [@bertoinLocalTimesViaCSBP]) and superprocesses (see for example [@leGallSnakes]).
Lamperti [announces]{} his representation theorem in [@csbp], and assures that ‘proofs of the main results will appear elsewhere, but he never published them. Nine years have elapsed before this result was proved by I.S. Helland in [@weakConvUnderTimeChange] by discrete approximations. There is one missing step in Helland’s paper, since non-conservative cases are not included. Also, M.L. Silverstein [@silversteinTimeChange Theorem 4] gives an ‘analytic paraphrase of Lamperti’s result’, namely, he proves by analytic methods, that CSBP laws are in one-to-one correspondence with Laplace exponents of Lévy processes with no negative jumps. The Lamperti representation, as a path transformation, is not studied there.
Our goal here is to give two proofs of the Lamperti representation. One is a direct proof of the Lamperti representation (including the absence of negative jumps) using probabilistic arguments (infinite divisibility, strong Markov property, martingales, stopping theorems, stochastic differential equations). The other one is a proof by discrete approximations, in the same vein as Helland, but using a new topology on Skorohod space.\
\
The state space we will work on is $E=[0,\infty]$ with any metric $\rho$ which makes it homeomorphic to $[0,1]$. We let $*$ stand for the convolution of measures, and use the convention $z+\infty=\infty$ for any $z\in E$.
\[def : CSBP\] A [**continuous-state branching process**]{}, in short [**CSBP**]{}, is a conservative and càdlàg Markov process with values in $E$, whose transition kernels ${\ensuremath{\left( P_t\right) }}_{t\geq 0}$ satisfy the following [**branching property**]{}:
[$P_t\!{\ensuremath{\left( z_1,\cdot\right) }}$]{}\*[$P_t\!{\ensuremath{\left( z_2,\cdot\right) }}$]{}=[$P_t\!{\ensuremath{\left( z_1+z_2,\cdot\right) }}$]{}
for all $t\geq 0$ and $z_1,z_2\in E$.
We could also have defined CSBPs as stochastically continuous instead of [càdlàg]{}. In the forthcoming Proposition \[prop : prelim\], we will see that 0 and $\infty$ are *absorbing* states for a CSBP. We could give an analogous definition if the state-space were $[0,\infty)$ without the conservativity assumption; however, using $\infty$ as the cemetery point for the former we obtain an $E$-valued conservative process which will turn out to be a CSBP and have the Feller property with respect to the metric $\rho$.
We now define the *Lamperti transformation*, which acts on the Skorohod space of [càdlàg]{} trajectories with values in $E$ that are *absorbed at zero and infinity*, that will be denoted $D$. More formally, $D$ consists of functions ${\ensuremath{f:E\to E}}$ which are [càdlàg]{} (so that in particular ${\ensuremath{f\!{\ensuremath{\left( \infty-\right) }}}}:=\lim_{t\to\infty}{\ensuremath{f\!{\ensuremath{\left( t\right) }}}}$ exists in $E$), such that ${\ensuremath{f\!{\ensuremath{\left( \infty-\right) }}}}={\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}\in{\ensuremath{\left\{ 0,\infty\right\} }}$ and for which ${\ensuremath{f\!{\ensuremath{\left( t\right) }}}}=0$ (resp. $=\infty$) implies that ${\ensuremath{f\!{\ensuremath{\left( t+s\right) }}}}=0$ (resp. $=\infty$) for all $s\geq 0$.
For any $f\in D$, first introduce the additive functional $\theta$ given by
\_t:=\_0\^t f(s) ds,
and let $\kappa$ denote the right-inverse of $\theta$ on $[0,\infty]$, given by
\_t:={u0: \_u >t}
using the convention $\inf\emptyset=\infty$. Define the *Lamperti transformation* ${\ensuremath{L:D\to D}}$ by
[$L\!{\ensuremath{\left( f\right) }}$]{}=f
where one remembers that ${\ensuremath{L\!{\ensuremath{\left( f\right) }}}}(t)=f(\infty)$ if $\kappa_t=\infty$.
Notice that $0$ and $\infty$ indeed are also absorbing for $L(f)$. $L$ is a bijection of $D$. This can be checked by merely computing its inverse: setting $g=L(f)$ one rewrites $\kappa$ as
\_t:=\_0\^t1/g(s) ds.
Then $f=g\circ \theta$, where $\theta$ is the right-inverse of $\kappa$. It will always be implicit in what follows that a Lévy process is a càdlàg process with independent and homogeneous increments, sent to $\infty$ at an independent exponential time, where it is understood that an exponential distribution with parameter zero means the distribution which assigns probability 1 to the value $\infty$. A *spectrally positive* Lévy process is a Lévy process with no negative jumps. Recall (e.g. [@bertoinLevyP]) that the *Laplace exponent* of a spectrally positive Lévy process is a convex function $\Psi$ on $[0,\infty)$ satisfying
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( e^{-\lambda X_t}\right) }}$]{}=e\^[-x +t[$\Psi\!{\ensuremath{\left( \lambda\right) }}$]{}]{}t,x,0.
When $\Psi$ does not take positive values, $X$ is a.s. non-decreasing, and it is called a *subordinator*.
If $\Psi(0)=0$, it is known that $X$ has infinite lifetime. If $q:=-\Psi(0)>0$, then it is easily seen that $X$ is the Lévy process with Laplace exponent $q+\Psi$ killed at an independent exponential time with parameter $q$. Since in our setting, $X$ is set to $\infty$ after it is killed, we will consider that the killing time is the first jump with infinite size. This amounts to adding to the Lévy measure a Dirac measure at $\{+\infty\}$ with mass $q$.\
\
Let us state the Lamperti representation theorem.
\[lampRep\] The Lamperti transformation is a bijection between continuous-state branching processes and Lévy processes with no negative jumps stopped whenever reaching zero. Specifically, for any CSBP $Z$, $L(Z)$ is a Lévy process with no negative jumps stopped whenever reaching zero; for any Lévy process with no negative jumps $X$ stopped whenever reaching zero, $L^{-1}(X)$ is a continuous-state branching process.
There are two natural strategies for a proof of this theorem.\
\
The first strategy is based on generators, and consists in proving a relationship of the type $A_{Z}f(x)=xA_{X}f(x)$, where $A_{Z}$ is the local generator of $Z$ and $A_{X}$ is that of $X$. Starting either with a CSBP $Z$ or with a Lévy process $X$, one characterizes the Laplace transforms of one-dimensional distributions of the other process to obtain one part of the theorem. The other part can be obtained proving that the Lamperti transformation is onto (the Laplace exponent of any Lévy process/CSBP is attained). This method was hinted at by M.L. Silverstein [@silversteinTimeChange p.1045] in the preparatory discussion of Theorem 4, which states that Laplace transforms of CSBP are in one-to-one correspondence with Laplace exponents of spectrally positive Lévy processes (see the forthcoming Proposition \[LaplaceExponent\]). However, this discussion was not meant to be a proof, but was rather meant to guess the form of the aforementioned correspondence (which is proved by purely analytical arguments).
We wish to provide a proof of the Lamperti representation theorem in this vein, that we call ‘direct probabilistic proof’. Our goal is to emphasize the probabilistic rationale for the appearance of a spectrally positive Lévy process when applying the Lamperti transformation to CSBPs and of CSBPs when applying the inverse Lamperti transformation to spectrally positive Lévy processes. In particular, we do not wish to use analytical arguments to prove surjectivity. The study of the Lamperti transformation uses martingales (as a substitute for the delicate use of generators), and the inverse is analyzed in the spirit of [@dawsonLi] where stochastic differential equations are shown to be satisfied by affine processes, which make explicit in the special case of CSBPs.\
\
The second strategy is based on discrete approximations. In the case of Markov branching processes with integer values (discrete-state branching processes, or DSBPs), the Lamperti representation in terms of time-continuous random walks (with no negative jumps ‘larger’ than $-1$) is nearly evident (see below). After rescaling, this yields a one-to-one correspondence between rescaled DSBPs and certain compound Poisson processes which are in the domain of attraction of spectrally positive Lévy processes. The second ingredient, due to [@lampertiberkeley], is the fact that all CSBPs are limits of rescaled DSBPs. The third ingredient is a necessary and sufficient condition for a sequence $(Y_n)$ of DSBPs to converge to a certain (CSBP) process $Y$. Such a condition is given by I.S. Helland in [@weakConvUnderTimeChange] (see [@grimvall] for the case of a sequence of Bienaymé–Galton–Watson processes), and proved to be equivalent to the convergence (in finite-dimensional distributions *and* weakly in the Skorohod topology) of the sequence $X_n:=L(Y_n)$ to a spectrally positive Lévy process $X$. If the convergence of $(Y_n)$ is strong enough so as to guarantee continuity of the Lamperti transform, then Theorem \[lampRep\] follows. This difficult step is carried out in [@weakConvUnderTimeChange], even in the explosive case, where $\infty$ can be reached continuously in finite time by the CSBP (but in the exception of the non-conservative case, where $\infty$ can be reached by a jump from a finite state). More specifically, the Lamperti transformation is not continuous w.r.t. the usual Skorohod topology due to possible explosive cases. If explosive cases are excluded, one can proceed as in [@ethierKurtz Ch. VI,IX] using properties of the time-change transformation studied by Whitt in [@usefulFunctionsWhitt]. If even explosive (but conservative) cases are included, one can follow the work of Helland, introducing the (non-metrizable) Stone topology on our Skorohod space $D$. The Lamperti transformation is still not continuous under this topology, but if $(W_n)$ converges weakly to $W$ w.r.t. this topology, then under certain conditions on $W$ and the sequence $(W_n)$, there is convergence of finite-dimensional distributions of $L(W_n)$ to those of $L(W)$. This proves sufficient to achieve the proof of Theorem \[lampRep\].
We will provide a proof of the Lamperti representation theorem in the same vein, that we call ‘proof through weak convergence’, not completing the proof of Helland by allowing the non-conservative case, but rather, introducing a new topology on Skorohod space which will make the Lamperti transformation continuous on $D$.
Outline of the two proofs
-------------------------
Section \[timeChange\] is dedicated to the direct probabilistic proof of Theorem \[lampRep\], and Section \[weakConv\] to its proof through weak convergence.
Beforehand, we will recall well-known properties of CSBPs and sometimes sketch their proofs.
### Proof through martingales and stochastic calculus
Let us outline Section \[timeChange\]. First, we prove that in continuous time and continuous state-space (both conditions are needed), a branching process cannot have negative jumps. Then we show that if ${\ensuremath{e_\lambda:z\mapsto {\ensuremath{\exp\!{\ensuremath{\left( -\lambda z\right) }}}}}}$ for any $\lambda>0$, then there exists a function $F$ (the negative of the branching mechanism of Proposition \[differentiability\]) such that
M\^\_t:=[$e_\lambda\!{\ensuremath{\left( Z_t\right) }}$]{}+[$F\!{\ensuremath{\left( \lambda\right) }}$]{}\_0\^t Z\_s[$e_\lambda\!{\ensuremath{\left( Z_s\right) }}$]{} ds
is a martingale. Applying the optional stopping theorem to the time change $\kappa_t$, we get a differential equation satisfied by the Laplace transform of the marginal of the image $Y$ of $Z$ by the Lamperti transformation. Solving this differential equation yields an expression which is very close to that known for a Lévy process, when it is not stopped upon reaching 0. The conclusive step consists in proving that $Y$ indeed is a Lévy process stopped upon reaching 0. For the second part of the theorem, we can use the Lévy–Itô decomposition for the initial Lévy process. We start with any spectrally positive Lévy process $X$ with initial position $x>0$, Lévy measure $\Lambda$ and Gaussian coefficient $\sigma$. Using the Lévy–Itô decomposition, we show that the image $Z$ of the process $X$ stopped upon reaching 0, by the inverse Lamperti transformation, satisfies $$\begin{aligned}
\label{eqn : SIE}
Z_t
=x&+a\int_0^t Z_s\, ds + \sigma \int_0^t\sqrt{Z_s} \,dB_s\nonumber
\\&+\int_0^t\int_0^{Z_{s-}}\int_{[1,\infty]}rN(ds, dv, dr)+
\int_0^t\int_0^{Z_{s-}}\int_{(0,1)}r\tilde{N}(ds, dv, dr),\end{aligned}$$ for some real number $a$, where $B$ is a Brownian motion and $N$ is an independent Poisson measure ($\tilde N$ is the associated compensated measure) with intensity measure $ds\,dv\,\Lambda(dr)$. It is then immediate to deduce the branching property. This stochastic equation is similar to the ones shown to be satisfied by affine processes in [@dawsonLi (5.1)], [@bertoinLeGallFlowsII] and [@bertoinLeGallFlowsIII (9)]. The Poisson integral in equation (\[eqn : SIE\]) has the following interpretation: the role of the second coordinate of the Poisson measure $N$ is to *mark* jumps in order to have them occur only if this mark is ‘below’ the path of $Z$; thus, the jumps with size in $(r, r+dr)$ occur at a rate equal to $Z_{t}\, \Lambda(dr)$, that is, as in the discrete case discussed below, the branching process jumps at a rate which is *linear* in the population size.
### Proof through weak convergence
The second proof (Section \[weakConv\]) relies on the approximation of spectrally positive Lévy processes by compound Poisson processes and of CSBPs by (time-continuous) discrete state-space branching processes, abbreviated as DSBPs.
A [**discrete space branching process**]{} $Z=(Z_t;t\ge 0)$ is a càdlàg Markov process with values in $\overline{\ensuremath{\mathbb{N}}}={\ensuremath{\mathbb{N}}}\cup{\ensuremath{\left\{ \infty\right\} }}$ (sent to $\infty$ after possible blow-up), which jumps from state $i$ to state $i+j$, $j=-1,1,2,\ldots$, at rate $i \mu_{j+1}$, where $(\mu_k)_{k\ge 0}$ is a finite measure on $\overline{\ensuremath{\mathbb{N}}}$ with zero mass at 1.
The integer $Z_t$ can be interpreted as the size at time $t$ of a population where each individual is independently replaced at constant rate $\lambda:=\sum_k\mu_k$ by a random quantity of individuals, equal to $k$ with probability $\mu_k/\lambda$. As a consequence, it is easily seen that $Z$ satisfies the branching property.
To explain the heuristics behind the Lamperti transformation (implicit in [@csbp; @silversteinTimeChange] and also found in [@bertoinMaphysto; @kyprianouBook]), let us note that for any state $i\not\in\{0,\infty\}$, the size of the jump of $Z$ starting from $i$ does not depend on $i$. Thus, the jump chain of $Z$ is exactly that of the compound Poisson process $X$ which goes from state $i$ to state $i+j$, $j=-1,1,2,\ldots$, at rate $ \mu_{j+1}$. The only difference between those two processes lies in the waiting times between two jumps. The Lamperti transformation is a random time change that enables the paths of one process to be obtained from those of the other one by an appropriate modification of the waiting times. If $T_0=0$ and $T_1<T_2<\cdots$ are the successive jump times of $Z$, then the differences ${\ensuremath{\left( T_i-T_{i-1}\right) }}_{i\geq 1}$ are conditionally independent given the successive states ${\ensuremath{\left( Z_{T_i}\right) }}_{i\in{\ensuremath{\mathbb{N}}}}$ and conditionally on them, $T_i-T_{i-1}$ is exponential with parameter $\lambda Z_{T_{i-1}}$. The important point is to notice that defining $Y$ as the Lamperti transform of $Z$ amounts to multiplying each waiting time $T_i-T_{i-1}$ by $Z_{T_{i-1}}$; this turns the waiting time into an exponential variable with parameter $\lambda$, except when $Z_{T_{i-1}}=0$, since then the waiting time is infinite. Therefore, $Y$ is equal to the compound Poisson process $X$ with rate $\lambda$ and jump distribution $\mu_{1+\cdot}$ stopped upon reaching zero. Of course, a similar sketch of proof can be achieved for the other direction of the Lamperti transformation.
It turns out that general CSBPs can be approximated by DSBPs at the level of finite-dimensional distributions if and only if the corresponding Lamperti transforms of the latter approximate spectrally positive Lévy processes stopped whenever reaching zero (analogous to previous work of Grimvall presented in [@grimvall]). Therefore, one could hope to prove the Lamperti representation of CSBPs by weak convergence considerations. This program would be carried out in a very simple manner if the Lamperti transformation were continuous on Skorohod space but unfortunately this is not the case (Helland first reported such a phenomenon in [@weakConvUnderTimeChange]). We will therefore have to circumvent this problem by using properties of our approximations which ensure weak convergence in Skorohod space with a modified topology which makes the Lamperti transformation continuous and implies convergence in the usual Skorohod space.
Preliminary results {#subsec : prelim results}
-------------------
\[prop : prelim\] For a CSBP, both states $0$ and $\infty$ are absorbing, and for all $t,z\in(0,\infty)$,
[$P_t\!{\ensuremath{\left( z,{\ensuremath{\left\{ 0,\infty\right\} }}\right) }}$]{}<1.
In addition, there is ${\ensuremath{u_t:[0,\infty)\to [0,\infty)}}$ such that $$\label{silversteinDef}
\int_{[0,\infty]} e^{-\lambda z'}\, {\ensuremath{P_t\!{\ensuremath{\left( z,dz'\right) }}}}=e^{-z{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}}$$ for $z\in [0,\infty]$ which satisfies the composition rule
[$u_{t+s}\!{\ensuremath{\left( \lambda\right) }}$]{}=[$u_t\!{\ensuremath{\left( {\ensuremath{u_s\!{\ensuremath{\left( \lambda\right) }}}}\right) }}$]{}.
Finally, ${\ensuremath{\left( P_t\right) }}_{t\geq 0}$ is a Feller semigroup.
The absorbing character of $0$ and $\infty$ is easily handled, the composition rule follows from the Markov property, while the Feller character can be dealt as in [@lampertiberkeley Lemma 2.2]. Indeed, the [càdlàg]{} character of the trajectories implies that $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is continuous at zero, where it is equal to $\lambda$, and so the composition rule gives us continuity everywhere. The extended continuity theorem for Laplace transforms applied to implies that $P_tf$ is continuous whenever $f$ is (because the restriction of $f$ to $[0,\infty)$ would be continuous and bounded) and that it tends to $f$ pointwise as $t\to 0$.
We now provide further properties of $u_t$.
\[differentiability\] For every $\lambda>0$, the function $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is differentiable on $[0,\infty)$. Moreover,
= F([$u_t\!{\ensuremath{\left( \lambda\right) }}$]{})t,0
where
[$F\!{\ensuremath{\left( \lambda\right) }}$]{}:=.|\_[t=0]{} .
The function $\Psi:=-F$ is called the [**branching mechanism**]{} of $Z$.
This last result was proved in [@silversteinTimeChange], using a delicate analytical proof, so we prefer to provide an elementary proof resting mainly on the composition rule. An only more slightly complicated argument found in [@GikhmanSkorohod Lemma 1, Chap V.2, p.413] enables one to generalize the above proposition to stochastically continuous multi-type continuous-state branching processes.
In this proof, we exclude the trivial case where $Z$ is a.s. constant, so that ${\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}\not= \lambda$ unless $t=0$. First note that by a recursive application of the dominated convergence theorem, $\lambda\mapsto{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is infinitely differentiable in $(0,\infty)$ and strictly increasing. Next observe that the Feller property of $Z$ gives the continuity of $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$.
By the composition rule again, we may write $$\label{expInDomainForZ}
{\ensuremath{u_{t+h}\!{\ensuremath{\left( \lambda\right) }}}}-{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}
={\ensuremath{u_{t}\!{\ensuremath{\left( {\ensuremath{u_h\!{\ensuremath{\left( \lambda\right) }}}}\right) }}}}-{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}
={\ensuremath{\frac{\partial {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}}{\partial \lambda}\!{\ensuremath{\left( \lambda'\right) }}}}{\ensuremath{\left( {\ensuremath{u_h\!{\ensuremath{\left( \lambda\right) }}}}-\lambda\right) }}.$$for some $\lambda'\in[\lambda,{\ensuremath{u_h\!{\ensuremath{\left( \lambda\right) }}}}]$. Hence the increment ${\ensuremath{u_{t+h}\!{\ensuremath{\left( \lambda\right) }}}}-{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ has the same sign as ${\ensuremath{u_h\!{\ensuremath{\left( \lambda\right) }}}}-\lambda$ and so, for equally-spaced partitions ${\ensuremath{\left\{ t_i\right\} }}_{i}$ of $[0,t]$ with spacing $h$, we have:
\_[i]{}[|[$u_{t_{i+1}}\!{\ensuremath{\left( \lambda\right) }}$]{}-[$u_{t_i}\!{\ensuremath{\left( \lambda\right) }}$]{}|]{} =([$u_h\!{\ensuremath{\left( \lambda\right) }}$]{}-)\_[i]{}([$u_{t_{i+1}}\!{\ensuremath{\left( \lambda\right) }}$]{}-[$u_{t_i}\!{\ensuremath{\left( \lambda\right) }}$]{}) =[|[$u_t\!{\ensuremath{\left( \lambda\right) }}$]{}-|]{}.
We deduce that $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ has *finite variation* and hence, it is almost everywhere differentiable. Now thanks to ,
\_[h0]{} =
where the r.h.s. is nonzero, so choosing $t$ where $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is differentiable, its right-derivative exists at $0$. This, along with the last display now yields the differentiability everywhere, as well as the following equality $$\label{eqn : plein de derivees partielles}
\frac{\partial{\ensuremath{u_{t}\!{\ensuremath{\left( \lambda\right) }}}}}{\partial t}= \frac{\partial {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}}{\partial \lambda}\cdot F(\lambda),$$ where we have set
[$F\!{\ensuremath{\left( \lambda\right) }}$]{}:=.|\_[t=0]{}.
Letting $h\downarrow 0$ in ${\ensuremath{\left( u_h\circ u_t(\lambda)-u_t(\lambda)\right) }}/h$, we finally get $$\label{eqn : magic}
\frac{\partial{\ensuremath{u_{t}\!{\ensuremath{\left( \lambda\right) }}}}}{\partial t}= F({\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}),$$ which ends the proof.
\[LaplaceExponent\] The branching mechanism $\Psi$ is the Laplace exponent of a spectrally positive Lévy process.
This last proposition can be found in [@silversteinTimeChange], where it is proved by analytical methods relying on completely monotone functions. Silverstein uses this proposition to prove uniqueness of solutions to the differential equation in Proposition \[differentiability\], which we only need in the proof by weak convergence and offer a simple argument for it. He additionally proves that any Laplace exponent of a killed spectrally positive Lévy process can occur; we obtain this as a consequence of our approach to Theorem \[lampRep\]. It is also proved in [@KawazuWatanabe], where it is deduced mainly from Itô’s formula. We will rely on the convergence criteria for infinitely divisible probability measures as found in [@kallenberg Thm. 15.14, p.295].
Since for every $x\geq 0$, $\lambda\mapsto e^{-x{\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}}$ is the Laplace transform of a probability measure (on $[0,\infty]$) then $\lambda\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is the Laplace exponent of a subordinator. Recalling that the Laplace exponent of a subordinator is minus its Laplace exponent as a spectrally positive Lévy process (cf. [@bertoinLevyP]), it follows that for every ${\ensuremath{ \varepsilon}}>0$,$$\label{aproxEL}
\lambda\mapsto {\ensuremath{\left( \lambda-{\ensuremath{u_{\ensuremath{ \varepsilon}}\!{\ensuremath{\left( \lambda\right) }}}}\right) }}/{\ensuremath{ \varepsilon}}$$is the Laplace exponent of a spectrally positive Lévy process whose limit as ${\ensuremath{ \varepsilon}}\to 0+$, $\Psi$, is then the Laplace exponent of a spectrally positive Lévy process. Indeed, letting $G_{{\ensuremath{ \varepsilon}}}$ is the (infinitely divisible) law on $(-\infty,\infty]$ whose Laplace exponent is , the Helly-Bray theorem gives us a subsequence ${\ensuremath{ \varepsilon}}_k\to 0$ for which $G_{{\ensuremath{ \varepsilon}}_k}$ converges to an increasing [càdlàg]{} function $G$; we can interpret $G$ as the distribution of a probability measure $\mu$ on $[-\infty,\infty]$. To see that it doesn’t charge $-\infty$, we use Fatou’s lemma for convergence in law:
e\^[-x]{} [$G\!{\ensuremath{\left( dx\right) }}$]{}\_[k]{}e\^[-x]{} [$G_{{\ensuremath{ \varepsilon}}_k}\!{\ensuremath{\left( dx\right) }}$]{}e\^[[$\Psi\!{\ensuremath{\left( \lambda\right) }}$]{}]{}<.
Actually, $\Psi$ is the log-Laplace transform of $G$: by the convergence in the preceding display we get
\_ke\^[-x]{} [$G_{{\ensuremath{ \varepsilon}}_k}\!{\ensuremath{\left( dx\right) }}$]{}<
for all $\lambda\geq 0$. Since for $\lambda'>\lambda\geq 0$, $e^{-\lambda' y}={\ensuremath{\left( e^{-\lambda y}\right) }}
^{\lambda'/\lambda}$, the $L_p$ criterion for uniform integrability implies that
e\^[-y]{}[$G_{{\ensuremath{ \varepsilon}}_k}\!{\ensuremath{\left( dy\right) }}$]{}\_[k]{}e\^[-y]{}[$G\!{\ensuremath{\left( dy\right) }}$]{}
and so
e\^[-y]{}[$G\!{\ensuremath{\left( dy\right) }}$]{}=e\^[-[$F\!{\ensuremath{\left( \lambda\right) }}$]{}]{}.
The same argument, when applied to $\lambda\mapsto {\ensuremath{\left( \lambda-{\ensuremath{u_{t{\ensuremath{ \varepsilon}}}\!{\ensuremath{\left( \lambda\right) }}}}\right) }}/{\ensuremath{ \varepsilon}}$, tells us that $t\Psi$ is the log-Laplace of a probability measure on $(-\infty,\infty]$, so that $G$ is infinitely divisible. The fact that its Lévy measure does not charge $(-\infty,0)$ is deduced from [@kallenberg Thm. 15.14, p.295].
Direct probabilistic proof {#timeChange}
==========================
The Lamperti transform of a CSBP
--------------------------------
Let $Z$ denote a CSBP and ${\ensuremath{ {{\mathbb{P}}}} }_x$ its law when it starts at $x\in [0,\infty]$. First, we prove that $Z$ cannot have negative jumps. Fix $\delta>0$ and set $$J_\delta:=\inf\{t>0:Z_t-Z_{t-}<-\delta\}.$$ Now let $n$ be any integer such that $x/n<\delta$ and let $(Z^{(i,j)};i\ge 1, j=1,\ldots,n)$ be independent copies of $Z$ whose starting point will be defined recursively on $i$. Also set $Z^{(i)}:=\sum_{j=1}^n Z^{(i,j)}$. Let $T^{(i,j)}_\delta$ denote the first hitting time of $(\delta, +\infty]$ by $Z^{(i,j)}$ and set $\tau_\delta^{(i)}:=\inf_{1\le j\le n} T^{(i,j)}_\delta$. Now set the initial values of $Z^{(i,j)}$ as follows : $Z^{(1,j)}(0)=x/n<\delta$ for all $j$ and $$Z^{(i+1,j)}(0)=n^{-1}Z^{(i)}(\tau_\delta^{(i)})\qquad j=1,\ldots,n,\: i\ge 1,$$ so that in particular $Z^{(1)}(0)=x$ and $Z^{(i+1)}(0)=Z^{(i)}(\tau_\delta^{(i)})$. Next, define $I$ as $$I:=\min\{i\ge 1 : Z^{(i)}(\tau_\delta^{(i)}) >n\delta \}.$$ Observe that by definition of $\tau^{(i)}_\delta$, all paths $(Z^{(i,j)}_t;t<\tau^{(i)}_\delta)$ remain below $\delta$, and so all paths $(Z^{(i)}_t;t<\tau^{(i)}_\delta)$ remain below $n\delta$. Observe that each $Z^{(i)}$ has the same transition kernels as $Z$, and that $\tau_\delta^{(i)}$ is a stopping time for $(Z^{(i,j)}; j=1,\ldots,n)$, so that the concatenation, say $Z^\star$, in increasing order of $i=1,\ldots, I$, of the paths $Z^{(i)}$ all killed at $\tau^{(i)}_\delta>0$, has the same law as $Z$ killed at $T_{n\delta}$, where $$T_{n\delta}:=\inf\{t\ge 0 : Z_t >n\delta \}.$$ Now recall that for all $1\le i\le I$ and $1\le j\le n$, all paths $(Z^{(i,j)}_t;t<\tau^{(i)}_\delta)$ remain below $\delta$. Since these processes are CSBPs, they only take non-negative values, and therefore cannot have a negative jump of amplitude larger than $\delta$. Since CSBPs are Feller processes, they have no fixed time discontinuity and the independent copies $(Z^{(i,j)};j=1,\ldots,n)$ a.s. do not jump at the same time. As a consequence, $(Z^{(i)}_t;t<\tau^{(i)}_\delta)$ has no negative jump of amplitude larger than $\delta$. The same holds for $(Z^{(i)}_t;t\le\tau^{(i)}_\delta)$ because if $\tau^{(i)}_\delta$ is a jump time, it can only be the time of a positive jump. As a consequence, the process $Z^\star$ has no negative jump of amplitude larger than $\delta$, which implies $$T_{n\delta}<J_\delta.$$ Letting $n\to\infty$ and because $\delta$ is arbitrarily small, this last inequality shows that $Z$ has no negative jumps.
Now define $Y$ as the image of $Z$ by the Lamperti transformation. Specifically, let $\kappa$ be the time-change defined as the inverse of the additive functional $\theta:t\mapsto \int_0^t Z_s\, ds$ and let $Y$ be defined as $Z\circ\kappa$. Recall Proposition \[prop : prelim\] and the branching mechanism $-F$. We consider the process $M^\lambda$ defined as
M\^\_t=[$e_\lambda\!{\ensuremath{\left( Z_t\right) }}$]{}+[$F\!{\ensuremath{\left( \lambda\right) }}$]{}\_0\^t Z\_s[$e_\lambda\!{\ensuremath{\left( Z_s\right) }}$]{} ds.
We now prove that $M^\lambda$ is a martingale under ${\ensuremath{ {{\mathbb{P}}}} }$. Thanks to ,
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Z_t\right) }}}}\right) }}$]{}= -x e\^[-x[$u_t\!{\ensuremath{\left( \lambda\right) }}$]{}]{}=-x F() e\^[-x[$u_t\!{\ensuremath{\left( \lambda\right) }}$]{}]{}=F()[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Z_t\right) }}}}\right) }}$]{},
which gives as a conclusion
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Z_t\right) }}}}\right) }}$]{}=-[$F\!{\ensuremath{\left( \lambda\right) }}$]{}[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( Z_t{\ensuremath{e_\lambda\!{\ensuremath{\left( Z_t\right) }}}}\right) }}$]{}.
This last equality proves that $M^\lambda$ has constant expectation, and the fact that it is a martingale follows from the Markov property of $Z$. Now $\kappa_t$ is a stopping time, so we can use the optional stopping theorem to get that for any $s>0$,
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( M^\lambda_{\kappa_t\wedge s}\right) }}$]{}=[$e_\lambda\!{\ensuremath{\left( x\right) }}$]{}
which translates into
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Z_{\kappa_t\wedge s}\right) }}}}\right) }}$]{}=[$e_\lambda\!{\ensuremath{\left( x\right) }}$]{}-[$F\!{\ensuremath{\left( \lambda\right) }}$]{}[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( \int_0^{\kappa_t\wedge s}Z_u{\ensuremath{e_\lambda\!{\ensuremath{\left( Z_u\right) }}}}\, du\right) }}$]{}.
By the dominated convergence theorem and the monotone convergence theorem applied respectively to the l.h.s. and r.h.s. as $s\to\infty$, one obtains
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Z_{\kappa_t}\right) }}}}\right) }}$]{}=[$e_\lambda\!{\ensuremath{\left( x\right) }}$]{}-[$F\!{\ensuremath{\left( \lambda\right) }}$]{}[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( \int_0^{\kappa_t}Z_u{\ensuremath{e_\lambda\!{\ensuremath{\left( Z_u\right) }}}}\, du\right) }}$]{}
so that by using the definition of $Y$ and the fact that
\_0\^[\_t]{}Z\_u[$e_\lambda\!{\ensuremath{\left( Z_u\right) }}$]{} du =\_0\^tZ\_[\_u]{}[$e_\lambda\!{\ensuremath{\left( Z_{\kappa_u}\right) }}$]{} d\_u =\_0\^t[$e_\lambda\!{\ensuremath{\left( Z_{\kappa_u}\right) }}$]{}[[[$\bf{1}$]{}]{}\_[Z\_[\_u>0]{}]{}]{} du
(since $Z_{\kappa_u}\,d\kappa_u={{{\ensuremath{\bf{1}}}}_{Z_{\kappa_u}>0}}\, du$), we get the equality $$\label{genY}
{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}={\ensuremath{e_\lambda\!{\ensuremath{\left( x\right) }}}}-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}\int_0^t {\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_s\right) }}}}{{{\ensuremath{\bf{1}}}}_{Y_s>0}}\right) }}}}\, ds.$$We denote by $T_0$ the first hitting time of $0$ by $Y$. As a first consequence of , note that if we write ${\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}{{{\ensuremath{\bf{1}}}}_{Y_t>0}}\right) }}}}={\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}-{\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }_x\!{\ensuremath{\left( T_0\leq t\right) }}}}$, the following differential equation is satisfied $$\label{diffEq}
\frac{\partial {\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}}{\partial t}+{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}=F(\lambda){\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }_x\!{\ensuremath{\left( T_0\leq t\right) }}}}.$$ We can therefore use standard techniques of solving first order linear differential equations to deduce the following equality $$\label{explicitExponent}
{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}= e^{-\lambda x-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}t}+{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{\left( 1-e^{-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}{\ensuremath{\left( t-T_0\right) }}}\right) }}{{{\ensuremath{\bf{1}}}}_{T_0\leq t}}\right) }}}}.$$ The last step is now to deduce that $Y$ is a Lévy process stopped upon hitting 0. In the case when ${\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }_x\!{\ensuremath{\left( T_0=\infty\right) }}}}=1$ for some $x\in (0,\infty)$ the same property holds for all $x\in (0,\infty)$ and we conclude from that
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}$]{}= e\^[-x-[$F\!{\ensuremath{\left( \lambda\right) }}$]{}t]{}.
Then $Y$ is a Lévy process which remains on $(0,\infty]$ when started there. It is therefore a subordinator and, from the last display, its Laplace exponent is $-F$.
This step is more complicated when ${\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }_x\!{\ensuremath{\left( T_0=\infty\right) }}}}<1$. Because we would like to show how the Lévy process emerges without appealing to analytical properties of the function $F$, we have been able to achieve a proof which makes no use of Proposition \[LaplaceExponent\]. But since this proof is a bit long and technical, we propose hereafter a shorter one which uses Proposition \[LaplaceExponent\]. Thanks to this proposition, there is a spectrally positive Lévy process $X$ with Laplace exponent $-F$, whose law we denote by $\mathbb{Q}$.
We stick to the notation $T_0$ for both processes $X$ and $Y$. It is not difficult to arrive at the following equality $$\mathbb{Q}_x(e_\lambda(X_{t\wedge T_0}))=e^{-\lambda x-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}t}+{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{\left( 1-e^{-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}{\ensuremath{\left( t-T_0\right) }}}\right) }}{{{\ensuremath{\bf{1}}}}_{T_0\leq t}}\right) }}}}.$$ Then thanks to , the only thing we have to check is that $T_0$ has the same law under ${\ensuremath{ {{\mathbb{P}}}} }_x$ as under $\mathbb{Q}_x$. To see this, first recall that $T_0=\int_0^\infty Z_s ds$. Since the CSBP started at $x+y$ is the sum of two independent CSBPs started at $x$ and $y$ respectively, the distribution of $T_0$ under ${\ensuremath{ {{\mathbb{P}}}} }_{x+y}$ is the convolution of the laws of $T_0$ under ${\ensuremath{ {{\mathbb{P}}}} }_x$ and ${\ensuremath{ {{\mathbb{P}}}} }_y$. We can therefore conclude that the distribution of $T_0$ under ${\ensuremath{ {{\mathbb{P}}}} }_x$ is infinitely divisible on $[0,\infty]$, and that there is a nonnegative function $\phi$ on $[0,\infty)$ such that $-\phi$ is the Laplace exponent of a subordinator and $$\label{eqn : T_0 et phi}
{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( e^{-\lambda T_0}\right) }}}}=e^{-x{\ensuremath{\phi\!{\ensuremath{\left( \lambda\right) }}}}}\qquad x,\lambda\ge 0.$$ On the other hand, as is well-known [@bertoinLevyP], $${\ensuremath{\mathbb{Q}_x\!{\ensuremath{\left( e^{-\lambda T_0}\right) }}}}=e^{-x{\ensuremath{\varphi\!{\ensuremath{\left( \lambda\right) }}}}}\qquad x,\lambda\ge 0,$$ where $\varphi$ is the nonnegative function on $[0,\infty)$ characterised by $-F\circ\varphi =\mbox{Id}_{[0,\infty)}$. At this point, we have to make sure that $-F$ indeed takes positive values (i.e. $X$ is not a subordinator). On the contrary, if $F$ took only nonnegative values, then by , we would get $$\frac{\partial {\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}\right) }}}}}{\partial t}=-{\ensuremath{F\!{\ensuremath{\left( \lambda\right) }}}}{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{e_\lambda\!{\ensuremath{\left( Y_t\right) }}}}{{{\ensuremath{\bf{1}}}}_{T_0> t}}\right) }}}},$$ so that all mappings $t\mapsto {\ensuremath{{\mathbb{E}}}}_x(e_\lambda(Y_t))$ would be nonincreasing. Letting $\lambda\to\infty$, we would get that the mapping $t\mapsto {\ensuremath{ {{\mathbb{P}}}} }_x(Y_t=0)$ also is nonincreasing. But since $0$ is absorbing, this mapping is obviously nondecreasing, so that ${\ensuremath{ {{\mathbb{P}}}} }_x(Y_t=0)={\ensuremath{ {{\mathbb{P}}}} }_x(Y_0=0)=0$ for all $t\ge 0$ and $x>0$. This contradicts the assumption that $Y$ hits 0 with positive probability.
If $(K_t^\lambda;t\ge 0)$ denotes the martingale obtained by taking conditional expectations of the terminal variable $\exp(-\lambda\int_0^\infty Z_s\, ds)$, we get
K\_t\^= [$\exp\!{\ensuremath{\left( -\lambda\int_0^tZ_s\, ds-\phi(\lambda) Z_t\right) }}$]{},
so that in particular
e\^[-x()]{}=[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( {\ensuremath{\exp\!{\ensuremath{\left( -\lambda\int_0^tZ_s\, ds-\phi(\lambda) Z_t\right) }}}}\right) }}$]{}.
Informally, we evaluate the derivative w.r.t. $t$ of both sides at $t=0$ to obtain $$\label{branchingMechanismFromIntegralOfZ}
0=-\lambda x e^{-x\phi(\lambda)}-xF(\phi(\lambda))e^{-x\phi(\lambda)},$$so that $-F\circ \phi$ is the identity on $[0,\infty)$. This shows that $\phi=\varphi$, so that $T_0$ indeed has the same law under ${\ensuremath{ {{\mathbb{P}}}} }_x$ as under $\mathbb{Q}_x$. It remains to give a formal proof of . Write $K^\lambda$ as the product of the semimartingale $L_t^\lambda={\ensuremath{\exp\!{\ensuremath{\left( -{\ensuremath{\phi\!{\ensuremath{\left( \lambda\right) }}}}Z_t\right) }}}}$ and the finite variation process $N_t^\lambda={\ensuremath{\exp\!{\ensuremath{\left( -\lambda\int_0^{t} Z_s\, ds\right) }}}}$; we can write $L_t^\lambda$ as
L\^\_t=M\_t-[$F\!{\ensuremath{\left( {\ensuremath{\phi\!{\ensuremath{\left( \lambda\right) }}}}\right) }}$]{}\_0\^t Z\_se\^[-[$\phi\!{\ensuremath{\left( \lambda\right) }}$]{}Z\_s]{} ds
where $M\equiv M^{\phi(\lambda)}$ is a (formerly defined) locally bounded martingale, in particular square integrable. Integration by parts gives us
K\_t\^=e\^[-[$\phi\!{\ensuremath{\left( \lambda\right) }}$]{}x]{}+\_0\^tN\^\_[s-]{} dM\_s-\_0\^tN\^\_sL\^\_sZ\_s ds.
Since $N^\lambda$ is bounded, its stochastic integral with respect to $M^\lambda$ is a square integrable martingale. Taking expectations, the second summand vanishes, and since by stochastic continuity of $Z$, $t\mapsto{\ensuremath{{\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( L^\lambda_tN^\lambda_tZ_t\right) }}}}$ is continuous (and bounded), we get
0=.[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( K^\lambda_t\right) }}$]{}|\_[t=0]{}=-e\^[-[$\phi\!{\ensuremath{\left( \lambda\right) }}$]{}x]{}
which implies .
The inverse Lamperti transform of a spectrally positive Lévy process {#Amaury'sConverse}
--------------------------------------------------------------------
In this subsection, we consider a Lévy process $X$ with no negative jumps, started at $x\ge0$, stopped at its first hitting time $T_0$ of 0, and possibly sent to $\infty$ after an independent exponential time. Using the well-known Lévy-Itô decomposition of $X$ [@bertoinLevyP; @kyprianouBook], we can write for every $t<T_0$ $$\label{eqn : LI}
X_t= x+at+\sigma B^X_t+P^X_t+M^X_t,$$where $a$ is a real number, $\sigma$ is a nonnegative real number (the Gaussian coefficient), $B^X$ is a standard Brownian motion, $P^X$ is a compound Poisson process and $M^X$ is a square integrable martingale, all terms being independent and adapted to the same filtration. To be more specific about $P^X$ and $M^X$, we denote by $\Lambda$ the Lévy measure of $X$, which is a $\sigma$-finite measure on $(0,\infty]$ (see Introduction) such that $\int_{(0,\infty]}(1\wedge r^2)\Lambda(dr)<\infty$. Then there is a Poisson measure $N^X$ on $[0,\infty)\times(0,\infty]$ with intensity measure $dt\,\Lambda(dr)$, and associated compensated measure $\tilde{N}^X (dt,dr):=N^X(dt,dr)-dt\,\Lambda(dr)$ (defined for $r< 1$) such that
P\^X\_t:=\_0\^t\_[\[1,\]]{}rN\^X(ds,dr) M\^X\_t:=\_0\^t\_[(0,1)]{}r\^X(ds,dr),
where the second integral is the $L^2$ limit, as ${\ensuremath{ \varepsilon}}\to 0$, of
M\^[X,[$ \varepsilon$]{}]{}\_t:=\_0\^t\_[([$ \varepsilon$]{},1)]{}r\^X(ds,dr).
Notice that, at the first jump of $P^X$ of infinite size, $X$ jumps to $\infty$ and remains there. It will be implicit in the rest of the proof that equalities hold in $[0,\infty]$.
Recall that the inverse Lamperti transform $Z$ of $X$ is given as follows. Set
\_t:=\_0\^[tT\_0]{} ,
and let $\theta$ be its inverse
\_t:={u0: \_u >t},
so that $Z:=X\circ \theta$. To prove that $Z$ is a CSBP, we will use the following proposition.
\[prop : SDE\] There is a standard Brownian motion $B^Z$, and an independent Poisson measure $N^Z$ on $[0,\infty)\times(0,\infty)\times(0,\infty]$ with intensity measure $dt\,dv\,\Lambda(dr)$ such that $$\begin{aligned}
\label{eqn : SDE}
Z_t
=x&+a\int_0^t Z_s\, ds + \sigma \int_0^t\sqrt{Z_s} \,dB^Z_s\nonumber
\\&+\int_0^t\int_0^{Z_{s-}}\int_{[1,\infty]}rN^Z(ds, dv, dr)+
\int_0^t\int_0^{Z_{s-}}\int_{(0,1)}r\tilde{N}^Z(ds, dv, dr),\end{aligned}$$ where $\tilde{N}^Z$ is the compensated Poisson measure associated with $N^Z$.
Define ${\ensuremath{\mathscr{G}}}$ as the time-changed filtration, that is, ${\ensuremath{\mathscr{G}}}_t={\ensuremath{\mathscr{F}}}_{\theta_t}$. We denote by $(T_n,\Delta_n)_{n\ge 1}$ an arbitrary labelling of the pairs associating jump times and jump sizes of $Z$. By a standard enlarging procedure, we can assume we are also given an independent ${\ensuremath{\mathscr{G}}}$-Brownian motion $B$, an independent ${\ensuremath{\mathscr{G}}}$-Poisson point process $N$ on $[0,\infty)\times(0,\infty)\times(0,\infty]$ with intensity measure $dt\,dv\,\Lambda(dr)$, and an independent sequence $(U_n)_{n\ge 1}$ of random variables uniformly ditributed on $(0,1)$ such that $U_n$ is ${\ensuremath{\mathscr{G}}}_{T_n}$-measurable and independent of ${\ensuremath{\mathscr{G}}}_{T_n-}$.
As a first step, we define $B^Z$ and $N^Z$. Recall the Lévy-Itô decomposition . Notice that $Y:=B^X\circ\theta$ is a continuous local martingale w.r.t. ${\ensuremath{\mathscr{G}}}$, so we can define $B^Z$ as
B\^Z\_t:=\_0\^t dY\_s + \_0\^t [[[$\bf{1}$]{}]{}\_[Z\_s=0]{}]{} dB\_s.
Next, we define $N^Z$ as
N\^Z(dt,dv,dr):=\_[n]{}\_[{T\_n,U\_nZ\_[T\_n-]{}, \_n}]{}(dt, dv, dr)+[[[$\bf{1}$]{}]{}\_[v>Z\_[t-]{}]{}]{} N(dt, dv, dr),
where $\delta$ denotes Dirac measures.
The second step consists in proving that $B^Z$ is a ${\ensuremath{\mathscr{G}}}$-Brownian motion, and that $N^Z$ is an independent ${\ensuremath{\mathscr{G}}}$-Poisson point process with intensity $dt\,dv\,\Lambda(dr)$.
Observe that $B^Z$ is a continuous local martingale w.r.t. ${\ensuremath{\mathscr{G}}}$, and that its quadratic variation in this filtration equals
<B\^Z>\_t = \_0\^t d\_s + \_0\^t [[[$\bf{1}$]{}]{}\_[Z\_s=0]{}]{} ds=\_0\^t ([[[$\bf{1}$]{}]{}\_[Z\_s=0]{}]{}+ [[[$\bf{1}$]{}]{}\_[Z\_s=0]{}]{}) ds=t,
because $d\theta_s=Z_s\, ds$. This shows that $B^Z$ is a ${\ensuremath{\mathscr{G}}}$-Brownian motion. For $N^Z$, let $H$ be a non-negative ${\ensuremath{\mathscr{G}}}$-predictable process, let $f$ be a two-variable non-negative Borel function, and let $R^X$ be the image of $N^X$ by the mapping $(t,r)\mapsto(\theta_t,r)$. Then by predictable projection, $$\begin{aligned}
{\ensuremath{{\mathbb{E}}}}\sum_nH_{T_n}\, f(U_n Z_{T_n-},\, \Delta_n) &=& {\ensuremath{{\mathbb{E}}}}\int_0^1 du \int_{[0,\infty)}\int_{(0,\infty]}R^X(dt,dr)\,H_t\,f(uZ_{t-}, r)\\
&=& {\ensuremath{{\mathbb{E}}}}\int_0^1 du \int_{[0,\infty)}d\theta_t\int_{(0,\infty]} \Lambda(dr)\,H_t\,f(uZ_{t}, r)\\
&=&{\ensuremath{{\mathbb{E}}}}\int_0^\infty Z_{t}dt \int_{(0,\infty]}\Lambda(dr)\int_0^1 du\, H_t\,f(uZ_{t}, r) \\
&=&{\ensuremath{{\mathbb{E}}}}\int_0^\infty dt \int_{(0,\infty]}\Lambda(dr)\int_0^{Z_{t}}dv\, H_t\,f( v, r).\end{aligned}$$ Now since $$\begin{aligned}
&{\ensuremath{{\mathbb{E}}}}\int_{[0,\infty)}\int_{(0,\infty)}\int_{(0,\infty]}N(dt,dv,dr)\,{{{\ensuremath{\bf{1}}}}_{v>Z_{t-}}} \, H_t\,f(v,r)
\\&={\ensuremath{{\mathbb{E}}}}\int_0^\infty dt \int_{(0,\infty]}\Lambda(dr)\int_{Z_{t}}^\infty dv\, H_t\,f( v, r),\end{aligned}$$ we deduce
[${\mathbb{E}}$]{}\_[\[0,)]{}\_[(0,)]{}\_[(0,\]]{}N\^Z(dt,dv,dr) H\_tf(v,r)=[${\mathbb{E}}$]{} \_0\^dt\_[0]{}\^dv\_[(0,\]]{}(dr) H\_tf( v, r),
which shows that $N^Z$ is a ${\ensuremath{\mathscr{G}}}$-Poisson point process with the claimed intensity. Finally, since $B^Z$ is a ${\ensuremath{\mathscr{G}}}$-Brownian motion and $N^Z$ is a ${\ensuremath{\mathscr{G}}}$-Poisson point process, Theorem 6.3 on p.77 of [@ikedaWatanabeII] ensures that $B^Z$ and $N^Z$ are independent.
The last step is showing that $Z$ indeed solves . We will refer to the successive terms in as $A_t$ (Lebesgue integral), $\sigma W_t$ (Brownian integral), $U_t$ (Poisson integral), and $V_t$ (compensated Poisson integral). Since we want to prove that $X\circ\theta=x+A+\sigma W+U+V$, and since $a\theta_t=A_t$, it is enough to prove that $B^X\circ\theta=: Y=W$, $P^X\circ\theta=U$, and $M^X\circ\theta=V$. Denote by $T$ the absorption time of $Z$ at 0 and recall that by definition of $B^Z$,
\_0\^t dB\^Z\_s = \_0\^t dY\_s + \_0\^t [[[$\bf{1}$]{}]{}\_[Z\_s=0]{}]{}dB\_s,
where the second term vanishes. As a consequence, $W_t = Y_{t\wedge T}=Y_t$, which provides us with the first required equality. Since $P^X(\theta_t)$ is merely the sum of jumps of $X$ of size greater than 1 occurring before time $\theta_t$, it is also the sum of jumps of $Z$ of size greater than 1 occurring before time $t$. As a consequence,
P\^X(\_t)= \_[n:T\_nt]{} \_n [[[$\bf{1}$]{}]{}\_[\_n1]{}]{}=\_0\^t\_0\^[Z\_[s-]{}]{}\_[(0,\]]{}r N\^Z(ds,dv,dr)[[[$\bf{1}$]{}]{}\_[r1]{}]{},
which provides us with the second required equality. As for the third one, the same reasoning as previously yields the following, where limits are taken in $L^2$ $$\begin{aligned}
M^X(\theta_t)
=\lim_{\varepsilon\downarrow 0}&\left(\sum_{n:T_n\le t} \Delta_n {{{\ensuremath{\bf{1}}}}_{\varepsilon<\Delta_n< 1}}- \theta_t \int_{(\varepsilon,1)}\Lambda (dr)\right)
\\= \lim_{\varepsilon\downarrow 0}&\left(\int_0^t\int_0^{Z_{s-}}\int_{(0,\infty]}r N^Z(ds,dv,dr)\,{{{\ensuremath{\bf{1}}}}_{\varepsilon<r< 1}}\right.
\\&- \left.\int_0^t ds \int_0^{Z_{s-}}dv\int_{(0,\infty]} \Lambda (dr)\,{{{\ensuremath{\bf{1}}}}_{\varepsilon<r< 1}}\right),\end{aligned}$$which indeeds shows that $M^X(\theta_t)=V_t$.
Now we want to prove that $X\circ\theta$ is a CSBP. Thanks to Proposition \[prop : SDE\], we only need to check that any solution $Z$ to satisfies the branching property. Let $Z^1$ and $Z^2$ be two independent copies of $Z$, one starting from $x_1$ and the other from $x_2$. Thanks to Proposition \[prop : SDE\], we can write the sum $\zeta$ of $Z_1$ and $Z_2$ as
\_t:=Z\^1\_t+Z\^2\_t=x\_1+x\_2+A\_t+W\_t + U\_t+V\_t,
where, with obvious notation,
A\_t:=a\_0\^t (Z\^1\_s+Z\^2\_s) ds,W\_t:=\_0\^t dB\^1\_s+\_0\^t dB\^2\_s,
U\_t:=\_0\^t\_0\^[Z\^1\_[s-]{}]{}\_[\[1,\]]{}rN\^1(ds, dv, dr)+\_0\^t\_0\^[Z\^2\_[s-]{}]{}\_[\[1,\]]{}rN\^2(ds, dv, dr),
V\_t:=\_0\^t\_0\^[Z\^1\_[s-]{}]{}\_[(0,1)]{}r\^1(ds, dv, dr)+\_0\^t\_0\^[Z\^2\_[s-]{}]{}\_[(0,1)]{}r\^2(ds, dv, dr),
and $B^1, N^1, B^2, N^2$ are all independent and adapted to the same filtration, say ${\ensuremath{\mathscr{F}}} = ({\ensuremath{\mathscr{F}}}_t;t\ge 0)$. By a standard enlarging procedure, we can assume that we are also given an independent ${\ensuremath{\mathscr{F}}}$-Brownian motion $B$ and an independent ${\ensuremath{\mathscr{F}}}$-Poisson point process $N$ with intensity measure $dt\,dv\,\Lambda(dr)$.
Notice that $W$ is a continuous local martingale with quadratic variation $t\mapsto \int_0^t \zeta_s ds$. Set $$B^\zeta_t := \int_0^t \frac{{{{\ensuremath{\bf{1}}}}_{\zeta_s\not=0}}}{\sqrt{\zeta_s}}\, dW_s + \int_0^t {{{\ensuremath{\bf{1}}}}_{\zeta_s=0}}\, dB_s.$$ Then $B^\zeta$ is adapted to the filtration ${\ensuremath{\mathscr{F}}}$ and, letting $T$ denote the first hitting time of 0 by $\zeta$,
W\_t=W\_[tT]{}=\_0\^t [[[$\bf{1}$]{}]{}\_[\_s=0]{}]{}dW\_s=\_0\^t dB\^\_s-\_0\^t[[[$\bf{1}$]{}]{}\_[\_s=0]{}]{} dB\_s=\_0\^t dB\^\_s.
In addition, the quadratic variation of $B^\zeta$ in the filtration ${\ensuremath{\mathscr{F}}}$ is $$<B^\zeta>_t = \int_0^t \frac{{{{\ensuremath{\bf{1}}}}_{\zeta_s\not=0}}}{\zeta_s}\, d<W>_s + \int_0^t {{{\ensuremath{\bf{1}}}}_{\zeta_s=0}}\, ds = \int_0^t {{{\ensuremath{\bf{1}}}}_{\zeta_s\not=0}}\, ds + \int_0^t {{{\ensuremath{\bf{1}}}}_{\zeta_s=0}}\, ds = t,$$ so that $B^\zeta$ is a ${\ensuremath{\mathscr{F}}}$-Brownian motion. Now set $$N^\zeta (dt, dv, dr) = {{{\ensuremath{\bf{1}}}}_{v< Z^1_{t-}}} \,N^1(dt, dv, dr)+ {{{\ensuremath{\bf{1}}}}_{Z^1_{t-}<v<\zeta_{t-}}} \,N^2(dt, dv-Z^1_{t-}, dr)+{{{\ensuremath{\bf{1}}}}_{v>\zeta_{t-}}} \,N(dt, dv, dr)$$ Then for any non-negative ${\ensuremath{\mathscr{F}}}$-predictable process $H=(H_t;t\ge0)$ and any two-variable non-negative Borel function $f$, $$\begin{aligned}
&\int_{[0,\infty)}\int_{(0,\infty)}\int_{(0,\infty]}N^\zeta(dt,dv,dr)\,H_t \, f(v,r) \\
&= \int_{[0,\infty)}\int_{(0,\infty)}\int_{(0,\infty]}N^1(dt,dv,dr)\,H_t\,{{{\ensuremath{\bf{1}}}}_{v< Z^1_{t-}}} \, f(v,r)\\
&+ \int_{[0,\infty)}\int_{(0,\infty)}\int_{(0,\infty]}N^2(dt,dv,dr)\,H_t \,{{{\ensuremath{\bf{1}}}}_{v< Z^2_{t-}}} \,f(v+Z^1_{t-},r)\\
&+ \int_{[0,\infty)}\int_{(0,\infty)}\int_{(0,\infty]}N(dt,dv,dr)\,H_t \,{{{\ensuremath{\bf{1}}}}_{v> \zeta_{t-}}} \, f(v,r),\end{aligned}$$ so that, taking $H_t={{{\ensuremath{\bf{1}}}}_{v<\zeta_{t-}}}$ and $f(v,r)=r{{{\ensuremath{\bf{1}}}}_{r\ge 1}}$, we get
U\_t=\_0\^t\_[0]{}\^[\_[s-]{}]{}\_[\[1,\]]{}rN\^(ds, dv, dr).
In addition, by predictable projection,
[${\mathbb{E}}$]{} \_[\[0,)]{}\_[(0,)]{}\_[(0,\]]{}N\^(dt,dv,dr)H\_t f(v,r) =[${\mathbb{E}}$]{}\_[\[0,)]{}\_[(0,)]{}\_[(0,\]]{}dtdv (dr)H\_t f(v,r),
so that $N^\zeta$ is a ${\ensuremath{\mathscr{F}}}$-Poisson point process with intensity $dt\,dv\, \Lambda(dr)$. Similarly, we could get that
V\_t=\_0\^t\_[0]{}\^[\_[s-]{}]{}\_[(0,1)]{}r\^(ds, dv, dr),
concluding that $$\begin{gathered}
\zeta_t:=x_1+x_2+a\int_0^t \zeta_s\, ds +\sigma\int_0^t\sqrt{ \zeta_s}\, dB^\zeta_s\\
+\int_0^t\int_0^{\zeta_{s-}}\int_{[1,\infty]}rN^\zeta(ds, dv, dr)+
\int_0^t\int_0^{\zeta_{s-}}\int_{(0,1)}r\tilde{N}^\zeta(ds, dv, dr).\end{gathered}$$ Finally, since $B^\zeta$ is a ${\ensuremath{\mathscr{F}}}$-Brownian motion and $N^\zeta$ is a ${\ensuremath{\mathscr{F}}}$-Poisson point process, Theorem 6.3 of [@ikedaWatanabeII] ensures that $B^\zeta$ and $N^\zeta$ are independent. Pathwise uniqueness for is proved in [@dawsonLi] under the stronger integrability condition
\_[(0,\]]{}rr\^2[$\Lambda\!{\ensuremath{\left( dr\right) }}$]{}<,
which excludes jumps of infinite size. We now sketch a proof, suggested by Zenghu Li, of pathwise uniqueness for lower semi-continuous solutions to . As a consequence, we will conclude that $\zeta=Z^1+Z^2$ has the same law as the process $Z$ started at $x_1+x_2$, that is, $Z$ has the branching property.
For each integer $n$, consider the equation$$\begin{aligned}
Z_t
=x&+a\int_0^t Z_s\, ds + \sigma \int_0^t\sqrt{Z_s} \,dB^Z_s
\\&+\int_0^t\int_0^{Z_{s-}}\int_{[1,\infty]}r\wedge nN^Z(ds, dv, dr)+
\int_0^t\int_0^{Z_{s-}}\int_{(0,1)}r\tilde{N}^Z(ds, dv, dr),\end{aligned}$$Existence and pathwise uniqueness holds for this equation by Theorem 5.1 in [@dawsonLi]. Consider also two solutions $Z'$ and $Z''$ to and consider the first times $\tau'_n$ and $\tau''_n$ that they have a jump of magnitude greater than $n$. Set also $\tau_n=\tau'_n\wedge\tau''_n$. Then, $Z'$ and $Z''$ satisfy the above equation on $[0,\tau_n]$, and so they are indistinguishable on $[0,\tau_n]$. As $n\to\infty$, $\tau_n$ converges to the first instant when $Z'$ or $Z''$ have a jump of infinite size, say $\tau_{\infty}$, a jump that comes from an atom of $N^Z$ of the form ${\ensuremath{\left( \tau_\infty,v,\infty\right) }}$, so that both processes feature it. Since after this time both processes equal to $\infty$, since the integral with respect to the Poisson process diverges, then $Z'$ and $Z''$ are indistinguishable.
Proof through weak convergence {#weakConv}
==============================
Here, we provide a second proof of Theorem \[lampRep\], this time through weak convergence. We use the fact that the Lamperti representation is easy to prove on discrete state-spaces, and introduce a topology on Skorohod space for which the inverse Lamperti transformation is *continuous*. Then approximating Lévy processes by compound Poisson processes, and CSBPs by discrete-state branching processes, we will deduce the Lamperti representation on the continuous state-space.
Preliminaries
-------------
Recall that $\rho$ is any metric on $E=[0,\infty]$ that makes $E$ homeomorphic to $[0,1]$. Recall the Skorohod-type space $D$ consisting of functions ${\ensuremath{f:E\to E}}$ which are [càdlàg]{} (so that in particular $\lim_{t\to\infty}{\ensuremath{f\!{\ensuremath{\left( t\right) }}}}={\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}$), such that ${\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}\in{\ensuremath{\left\{ 0,\infty\right\} }}$ and for which ${\ensuremath{f\!{\ensuremath{\left( t\right) }}}}=0$ (resp. $=\infty$) implies that ${\ensuremath{f\!{\ensuremath{\left( t+s\right) }}}}=0$ for all $s\geq 0$ (resp. $=\infty$).
For any $t\le \infty$, we denote by $\|\cdot\|_t$ the uniform norm on $[0,t]$, and by $\rho_t^D$ the uniform distance with respect to $\rho$, that is, $$\rho_t^D(f,g):=\sup_{s\in[0,t]}\rho(f(s), g(s)).$$ Let $\Lambda_t$ be the set of increasing homeomorphisms of $[0,t]$ into itself ($[0,\infty)$ if $t=\infty$), and define the metric $d_\infty$ on $D$ as $$\begin{aligned}
d_{\infty}(f,g):=1\wedge \inf_{\lambda\in\Lambda_\infty}{\ensuremath{\rho_\infty^D\!{\ensuremath{\left( f,g\circ\lambda\right) }}}}\vee \|\lambda-\operatorname{Id}\|_\infty.\end{aligned}$$
The proofs of the two following propositions can be found in Subsection \[technicalProofsWeakConvergence\].
\[continuityInvLampTrans\] The inverse Lamperti transformation $L^{-1}$ is continuous on ${\ensuremath{\left( D,d_\infty\right) }}$.
The usual Skorohod topology on $[0,t]$ defined in [@billingsley Ch. 3.12] (resp. on $[0,\infty)$ defined in [@billingsley Ch. 3.16]) is induced by the metric $d_t$ (resp. $d$), where $$\begin{aligned}
{\ensuremath{d_t\!{\ensuremath{\left( f,g\right) }}}}=\inf_{\lambda\in\Lambda_t}\, {\ensuremath{\rho_t^D\!{\ensuremath{\left( f,g\circ\lambda\right) }}}}\vee \|\lambda-\operatorname{Id}\|_t
\quad\left(\text{resp.}\quad
{\ensuremath{d\!{\ensuremath{\left( f,g\right) }}}}=\int_0^\infty e^{- t}{\ensuremath{d_t\!{\ensuremath{\left( f,g\right) }}}}\, dt\right).\end{aligned}$$Then ${\ensuremath{d\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ as $n\to\infty$ if and only if for every continuity point $t$ of $f$, ${\ensuremath{d_t\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ (cf. Lemma 1 in [@billingsley Ch. 3.16, p. 167]), which gives a precise meaning to saying that $d$ controls the convergence of $f_n$ to $f$ only on compact subsets $[0,t]$ of $[0,\infty)$. It is easy to see, and will be repeatedly used, that $d_{\infty}(f_n,f)\to 0$ as $n\to\infty$ implies ${\ensuremath{d_t\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ for every continuity point $t$ of $f$, so that convergence with $d_\infty$ implies convergence in the usual Skorohod space. We also point out that in general,
[$d_\infty\!{\ensuremath{\left( f,g\right) }}$]{} [$\max\!{\ensuremath{\left( {\ensuremath{d_t\!{\ensuremath{\left( f,g\right) }}}} , {\ensuremath{d_\infty\!{\ensuremath{\left( f\circ s_t, g\circ s_t\right) }}}}\right) }}$]{}
where $f\circ s_t:= {\ensuremath{f\!{\ensuremath{\left( t+\cdot\right) }}}}$, since the right-hand side is obtained by taking the infimum over homeomorphisms which send $t$ to itself.
We will also need the following technical result on stopped Lévy processes, as well as its corollary.
\[convOnSkoTypeSpace\] Let $X$ and $(X^n)_n$ be spectrally positive Lévy processes with Laplace exponents $ \Psi, \Psi_n$ respectively. If for all $\lambda\geq 0$ we have
\_[n]{} \_n() = (),
then $X^n$ stopped whenever reaching zero converges weakly in $(D, d_\infty)$ to $X$ stopped whenever reaching zero. The same result holds if the processes ${\ensuremath{\left( X^n\right) }}_n$ are rescaled compound Poisson processes with jumps in ${\ensuremath{\left\{ -1\right\} }}\cup\overline{\ensuremath{\mathbb{N}}}$.
We will use the last proposition in the form of the following corollary (see e.g. Lemma 5.4 in [@limitseqbp p. 287]). For any $a,b >0$, consider the scaling operator $S^a_b$ on Skorohod space which sends $f$ to $t\mapsto {\ensuremath{f\!{\ensuremath{\left( a\cdot t\right) }}}}/b$.
\[cor : practical use\] Let $X$ be a spectrally positive Lévy process with Laplace exponent $\Psi$, started at $x\ge 0$ and stopped whenever reaching 0. There are a sequence of integers $a_n\to\infty$, and a sequence ${\ensuremath{\left( X^n\right) }}_{n}$ of compound Poisson processes started at $x_n\in {\ensuremath{\mathbb{N}}}$, stopped upon reaching 0, and whose jump distribution is concentrated on ${\ensuremath{\left\{ -1\right\} }}\cup\overline {\ensuremath{\mathbb{N}}}$, such that the Laplace exponent $\Psi^n$ of $S_{n}^{a_n}(X_n)$ converges to $\Psi$ and the sequence ${\ensuremath{\left( S_{n}^{a_n}(X_n)\right) }}_{n}$ converges weakly to $X$ in $(D, d_\infty)$.
We begin the proof of Theorem \[lampRep\] by studying the inverse Lamperti transformation.
The inverse Lamperti transform of a spectrally negative Lévy process
--------------------------------------------------------------------
Let $X$ and ${\ensuremath{\left( X^n\right) }}_{n\in{\ensuremath{\mathbb{N}}}}$ be as in Corollary \[cor : practical use\]. As we have noted in the introduction, denoting $L^{-1}$ the inverse Lamperti transformation, $L^{-1}(X^n)$ satisfies the branching property in $\overline{\ensuremath{\mathbb{N}}}$. Also it is obvious that $Z^n:=L^{-1}\circ S^{a_n}_n(X^n)$ satisfies the branching property in $n^{-1}\overline{\ensuremath{\mathbb{N}}}$ (e.g. because $L^{-1}\circ S^a_b=S^{a /b}_b\circ L^{-1}$).
Thanks to Proposition \[continuityInvLampTrans\], the sequence of branching processes $(Z^n)_n$ converges weakly in $(D,d_\infty)$ to the Markov process $Z:=L^{-1}(X)$ (time-changing a [càdlàg]{} strong Markov by the inverse of an additive functional gives another [càdlàg]{} strong Markov process, cf. [@dynkin Vol. 1, X.5]). To show that $Z$ is a CSBP, we have to check that it has inherited the branching property from the sequence ${\ensuremath{\left( Z^n\right) }}_{n\in{\ensuremath{\mathbb{N}}}}$, and thanks to the Markov property, it is sufficient to check the branching property at any fixed time. The result is due to the following two facts. First, because neither of the discrete branching processes $Z^n$ jumps at fixed times, neither does $Z$ jump at fixed times. Second, it is known that for any fixed time $t$, the mapping $D:f\mapsto f(t)$ is continuous at any $f$ which is continuous at $t$. As a conclusion, for any fixed time $t$, the mapping $D:f\mapsto f(t)$ is a.s. continuous at $Z$. This ends the proof.
Recall the usual topology on Skorohod space from the remark in the previous subsection. For this topology, the inverse Lamperti transformation $L^{-1}$ is *not* continuous, and the problem is due to explosions as seen in the example below.
\[discontinuityLampTransExample\] Consider an element $f$ of $D$ such that ${\ensuremath{f\!{\ensuremath{\left( s\right) }}}}\to \infty$,
[$\kappa_\infty\!{\ensuremath{\left( f\right) }}$]{}=\_0\^<,
and note that its inverse Lamperti transform ${\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}$ blows up at $\kappa_\infty$. If we approximate $f$ by $f_n=f{{{\ensuremath{\bf{1}}}}_{[0,n]}}$, then the inverse Lamperti transform of $f_n$ is always zero after ${\ensuremath{\kappa_n\!{\ensuremath{\left( f\right) }}}}$ so that it cannot converge to ${\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}$; it does converge to another limit however. This is illustrated in Figure \[discontinuousCharLampTrans\]. An explanation of why the problem occurs is that ${\ensuremath{\kappa\!{\ensuremath{\left( f\right) }}}}$ contracts $[0,\infty)$ into $[0,\kappa_\infty)$ and so to have convergence in Skorohod space of a sequence of functions when they approach a limit taking infinite values, we have to control the behaviour of the trajectories of the sequence on $[0,\infty)$ instead of only on its compacts subsets as with the usual metrics.
![Discontinuity of the inverse Lamperti transformation.[]{data-label="discontinuousCharLampTrans"}](exampleOfdiscontinuity.ps){width="\textwidth"}
The Lamperti transform of a CSBP {#WeakConvergenceDirectSense}
--------------------------------
Let $Z$ be a CSBP with law ${\ensuremath{ {{\mathbb{P}}}} }_x$ when it starts at $x$. As we have shown in Propositions \[prop : prelim\] and \[differentiability\], there are nonnegative real numbers $u_t(\lambda)$, $t,\lambda\ge 0$, such that
[${\ensuremath{{\mathbb{E}}}}_x\!{\ensuremath{\left( e^{-\lambda Z_t}\right) }}$]{}=e\^[-x[$u_t\!{\ensuremath{\left( \lambda\right) }}$]{}]{},
and $t\mapsto {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}$ is differentiable on $[0,\infty)$. In addition, there is a real function $\Psi$ on $[0,\infty)$ called the branching mechanism of $Z$, such that $$\label{ode}
\frac{\partial {\ensuremath{u_t\!{\ensuremath{\left( \lambda\right) }}}}}{\partial t}=-{\ensuremath{\Psi\!{\ensuremath{\left( {\ensuremath{u_t\!{\ensuremath{\left( {\lambda}\right) }}}}\right) }}}}\qquad t,\lambda\ge 0,$$and $\Psi$ is the Laplace exponent of a spectrally positive Lévy process.
Then let $X$ and ${\ensuremath{\left( X^n\right) }}_{n\in{\ensuremath{\mathbb{N}}}}$ be as in Corollary \[cor : practical use\]. Set $\tilde Z^n:=L^{-1}(S_{n}^{a_n}(X_n))$. As in the proof of the converse implication, each $\tilde Z^n$ satisfies the branching property in $n^{-1}\overline{\ensuremath{\mathbb{N}}}$, and thanks to Proposition \[continuityInvLampTrans\], the sequence of branching processes $(\tilde Z^n)_n$ converges weakly in $(D,d_\infty)$ to the Markov process $\tilde Z:=L^{-1}(X)$. (As remarked earlier, time-changing a [càdlàg]{} strong Markov by the inverse of an additive functional gives another [càdlàg]{} strong Markov process.) We end the proof showing that the finite dimensional distributions of $\tilde Z^n$ converge to those of $Z$, which will entail the equality in distribution between $Z$ and $\tilde Z$, and subsequently between $L(Z)$ and $X$, since $X=L(\tilde Z)$.
Since $\tilde Z^n$ is a branching process, there are real numbers ${\ensuremath{\tilde u^n_t\!{\ensuremath{\left( \lambda\right) }}}}$ such that
[${\ensuremath{{\mathbb{E}}}}_{x_n/n}
\!{\ensuremath{\left( e^{-\lambda \tilde Z^n_t}\right) }}$]{}= [$\exp\!{\ensuremath{\left( -(x_n/n){\ensuremath{\tilde u^n_t\!{\ensuremath{\left( \lambda\right) }}}}\right) }}$]{}t,0,
and we also have $$\label{integralEquationDSBPs}
{\ensuremath{\tilde u^n_t\!{\ensuremath{\left( \lambda\right) }}}}=\lambda-\int_0^t{\ensuremath{\Psi^n\!{\ensuremath{\left( {\ensuremath{\tilde u^n_s\!{\ensuremath{\left( \lambda\right) }}}}\right) }}}}\, ds .$$By convergence of the sequence of branching processes $(\tilde Z^n)$, ${\ensuremath{\tilde u^n_t\!{\ensuremath{\left( \lambda\right) }}}}$ converges pointwise to some nonnegative real number $\tilde u_t(\lambda)$ such that
[${\ensuremath{{\mathbb{E}}}}_x
\!{\ensuremath{\left( e^{-\lambda \tilde Z_t}\right) }}$]{}= e\^[-x[$\tilde u_t\!{\ensuremath{\left( \lambda\right) }}$]{}]{}t,0.
Since $\Psi^n$ converges to $\Psi$ pointwise and they are convex on $(0,\infty)$, convergence is uniform on compact subsets of $(0,\infty)$; by taking limits in , we obtain
[$\tilde u_t\!{\ensuremath{\left( \lambda\right) }}$]{}=-\_0\^t[$\Psi\!{\ensuremath{\left( {\ensuremath{\tilde u_s\!{\ensuremath{\left( \lambda\right) }}}}\right) }}$]{} ds.
Because of the local Lipschitz character of $\Psi$ on $(0,\infty)$ and Gronwall’s lemma $\tilde u_t=u_t$. As a consequence,
[${\ensuremath{{\mathbb{E}}}}_x
\!{\ensuremath{\left( e^{-\lambda \tilde Z_t}\right) }}$]{}= e\^[-x[$ u_t\!{\ensuremath{\left( \lambda\right) }}$]{}]{}=[${\ensuremath{{\mathbb{E}}}}_x
\!{\ensuremath{\left( e^{-\lambda Z_t}\right) }}$]{}t,0,
so that $Z$ and $\tilde Z$ have the same law.
Proof of propositions \[continuityInvLampTrans\] and \[convOnSkoTypeSpace\] {#technicalProofsWeakConvergence}
---------------------------------------------------------------------------
### Proof of Proposition \[continuityInvLampTrans\]
There are two cases to consider since every element of $D$ either tends to $0$ or to $\infty$. At the outset however, there are some simple propositions that cover both.
First of all, note that if ${\ensuremath{d\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ then ${\ensuremath{\kappa\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{\kappa\!{\ensuremath{\left( f\right) }}}}$ uniformly on compact sets of $[0,{\ensuremath{T_0\!{\ensuremath{\left( f\right) }}}})$.
Second, note that if ${\ensuremath{d^c\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ then ${\ensuremath{\theta\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{\theta\!{\ensuremath{\left( f\right) }}}}$ uniformly on compact sets of $[0,{\ensuremath{\kappa_{{\ensuremath{T_0\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}}})$. This follows from the following argument. It suffices to prove pointwise convergence on $[0,{\ensuremath{\kappa_{{\ensuremath{T_0\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}}})$; let $s<t<s'$ be three points on that interval. Then ${\ensuremath{\theta\!{\ensuremath{\left( f\right) }}}}<{\ensuremath{T_0\!{\ensuremath{\left( f\right) }}}}$ on $s,s'$ and $t$. By the preceding paragraph,
[$\kappa_{{\ensuremath{\theta_s\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f_n\right) }}$]{}[$\kappa_{{\ensuremath{\theta_s\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}$]{}=s<t
and so eventually, ${\ensuremath{\theta_s\!{\ensuremath{\left( f\right) }}}}<{\ensuremath{\theta_t\!{\ensuremath{\left( f_n\right) }}}}$. By the same argument, we see that eventually ${\ensuremath{\theta_t\!{\ensuremath{\left( f_n\right) }}}}<{\ensuremath{\theta_{s'}\!{\ensuremath{\left( f\right) }}}}$. By taking $s,s'\to t$, we see that ${\ensuremath{\theta_t\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{\theta_t\!{\ensuremath{\left( f\right) }}}}$.
Note that the preceding two facts are true even if we are working with the metric $d$. The particular nature of the metric $d_\infty$ come into play at this stage: note that if ${\ensuremath{d_\infty\!{\ensuremath{\left( f_n,f\right) }}}}\to 0$ then ${\ensuremath{f_n\!{\ensuremath{\left( \infty\right) }}}}={\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}$ from a given index onwards.
We will now consider the case when ${\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}=\infty$. Let $M>0$ be such that the $\rho$-diameter of $[M,\infty]$ is less than ${\ensuremath{ \varepsilon}}$. The quantity ${\ensuremath{L_{2M}\!{\ensuremath{\left( f\right) }}}}=\sup{\ensuremath{\left\{ t\geq 0:{\ensuremath{f\!{\ensuremath{\left( t\right) }}}}\leq M\right\} }}$ is finite and $\inf_{s\geq {\ensuremath{L_{2M}\!{\ensuremath{\left( f\right) }}}}}{\ensuremath{f\!{\ensuremath{\left( s\right) }}}}\ge M$. Also, ${\ensuremath{\kappa_{{\ensuremath{L_{2M}\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}}}<{\ensuremath{\kappa_\infty\!{\ensuremath{\left( f\right) }}}}$ (the rhs is ${\ensuremath{\kappa_{{\ensuremath{T_0\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}}}$) and so the preceding paragraphs tell us that ${\ensuremath{\theta\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{\theta\!{\ensuremath{\left( f\right) }}}}$ uniformly on ${\ensuremath{\kappa_{{\ensuremath{L_{2M}\!{\ensuremath{\left( f\right) }}}}}\!{\ensuremath{\left( f\right) }}}}$. Whitt’s result on the continuity of time-changes [@usefulFunctionsWhitt] tells us that ${\ensuremath{L^{-1}\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}$ (with respect to the Skorohod metric) on $[0,{\ensuremath{\kappa\!{\ensuremath{\left( {\ensuremath{L_{2M}\!{\ensuremath{\left( f\right) }}}}\right) }}}}{f}]$. Since
[$\rho\!{\ensuremath{\left( {\ensuremath{f\!{\ensuremath{\left( s\right) }}}},{\ensuremath{f_n\!{\ensuremath{\left( s\right) }}}}\right) }}$]{}<[$ \varepsilon$]{}
for $s>L_{2M}$, then ${\ensuremath{d_\infty\!{\ensuremath{\left( {\ensuremath{L^{-1}\!{\ensuremath{\left( f_n\right) }}}},{\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}\right) }}}}\to 0$.
The remaining case, which is handled similarly, is when ${\ensuremath{f\!{\ensuremath{\left( \infty\right) }}}}=0$. Suppose ${\ensuremath{f\!{\ensuremath{\left( 0\right) }}}}>0$, since otherwise there is nothing to prove. For ${\ensuremath{ \varepsilon}}>0$ small enough, we can introduce the (finite) quantity ${\ensuremath{L_{\ensuremath{ \varepsilon}}\!{\ensuremath{\left( f\right) }}}}=\sup{\ensuremath{\left\{ t\geq 0:{\ensuremath{f\!{\ensuremath{\left( t\right) }}}}>{\ensuremath{ \varepsilon}}\right\} }}$. Since $L_{\ensuremath{ \varepsilon}}<T_0$, by the same arguments as above, we have that ${\ensuremath{L^{-1}\!{\ensuremath{\left( f_n\right) }}}}\to{\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}$ (with respect to the Skorohod topology) on $[0,{\ensuremath{\kappa_{L_{\ensuremath{ \varepsilon}}}\!{\ensuremath{\left( f\right) }}}}]$. Since, eventually, $f_n>2{\ensuremath{ \varepsilon}}$ on $[{\ensuremath{L\!{\ensuremath{\left( {\ensuremath{ \varepsilon}}\right) }}}},\infty)$, then ${\ensuremath{d_\infty\!{\ensuremath{\left( {\ensuremath{L^{-1}\!{\ensuremath{\left( f_n\right) }}}},{\ensuremath{L^{-1}\!{\ensuremath{\left( f\right) }}}}\right) }}}}\to 0$ as $n\to\infty$.
### Proof of Proposition \[convOnSkoTypeSpace\]
Note that a given Lévy process $X$ is either killed at an independent exponential time, or drifts to $\infty$, or to $-\infty$ or has $\liminf_{t\to\infty}X_t=-\infty$ and $\liminf_{t\to\infty}=\infty$ (it oscillates). When we stop a spectrally positive Lévy process at 0 there are therefore three cases: either the stopped process jumps to $\infty$, or it drifts to $\infty$ without reaching 0 or it is stopped at 0 at a finite time. In any case, the trajectories of the stopped process belong to $D$. The convergence of the Laplace exponents of the approximating sequence $X^n$ implies the convergence of the finite-dimensional distributions and so Skorohod’s classical result implies that the convergence holds on ${\ensuremath{\left( D,d\right) }}$ (cf. [@kallenberg Thm. 15.17, p. 298]). To study the convergence of the stopped processes on ${\ensuremath{\left( D,d_\infty\right) }}$, we will use Skorohod’s representation theorem to assume that, on a given probability space, $X^n$ converges almost-surely to $X$. Let ${\ensuremath{T_{\ensuremath{ \varepsilon}}\!{\ensuremath{\left( X\right) }}}}$ denote $\inf{\ensuremath{\left\{ s\geq 0:X_t\leq {\ensuremath{ \varepsilon}}\right\} }}\in [0,\infty]$; we will add the subscript $n$ when the stopping times are defined from $X^n$. Note that on the set ${\ensuremath{T_0\!{\ensuremath{\left( X\right) }}}}<\infty$, ${\ensuremath{T_{0+}\!{\ensuremath{\left( X\right) }}}}={\ensuremath{T_0\!{\ensuremath{\left( X\right) }}}}$, by the quasi-left-continuity of Lévy processes (cf. [@bertoinLevyP Pro. I.2.7, p.21]). Stopping at the hitting time of zero is therefore a.s. continuous at $X$ (on ${\ensuremath{\left( D,d\right) }}$, as can be seen in [@pages] and Lemma VI.2.10 in [@jacodShiryaevII p. 340]) and so $X^n$ stopped at zero, denoted $\tilde X^n$, converges almost surely to $\tilde{X}$ (equal to $X$ stopped at zero). We will now divide the proof in three cases.
$X$ drifts to $-\infty$ or oscillates
: In this case, $T_0$ is finite almost surely. As we have remarked, $T^n_0\to T_0$ and so for $h>0$, $T^n_0\leq T_0+h$ from a given index onwards almost surely. Since Lévy processes do not jump at fixed times, $X$ is continuous at $T_0+h$ for $h>0$ and so ${\ensuremath{d_{T_0+h}\!{\ensuremath{\left( X,X^n\right) }}}}\to 0$ as $n\to\infty$. Note that
\_[n]{}[$d_\infty\!{\ensuremath{\left( \tilde X,\tilde X^n\right) }}$]{}\_[n]{}[$d_{T_0+h}\!{\ensuremath{\left( X,X^n\right) }}$]{}=0.
$X$ drifts to $\infty$
: We will begin by verifying that the convergence of $X^n$ to $X$ (on ${\ensuremath{\left( D,d\right) }}$) implies that we can uniformly control the overall infimum of the $X^n$. This is formally achieved in the following statement: given $\delta>0$ there exists some $M>0$ such that
[$ {{\mathbb{P}}}$ ]{}\_[2M]{}(\_[s0]{} X\_s <M) < [$ {{\mathbb{P}}}$ ]{}\_[2M]{}(\_[s0]{} X\^n\_s <M) <
from a given index onwards. For the proof, note that since $\Psi_n$ and $\Psi$ are strictly convex, we may denote their largest roots by $\Phi_n$ and $\Phi$ respectively. When $X^n$ is a spectally positive Lévy processes, the Laplace exponents $\Psi_n$ and $\Psi$ restricted to $[\Phi_n,\infty)$ and $[\Phi,\infty)$ have inverses $\phi_n$ and $\phi$. The convergence of the Laplace exponents and their convexity allow us to prove that $\Phi_n \to \Phi$ as $n\to \infty$. When $X$ drifts to infinity, then $\Phi>0$ and from [@bertoinLevyP Thm. 1, p.189] and the above, we deduce
[${\ensuremath{ {{\mathbb{P}}}} }_{2M}\!{\ensuremath{\left( \inf_{s\geq 0} X_s <M\right) }}$]{},\_[n]{} [${\ensuremath{ {{\mathbb{P}}}} }_{2M}\!{\ensuremath{\left( \inf_{s\geq 0} X^n_s <M\right) }}$]{}e\^[-M]{}.
By taking $M$ large enough, the claim follows. When the approximating sequence is constituted of rescaled left continuous compound Poisson processes, we adapt the proof of [@bertoinLevyP Thm. 1, p.189] to arrive at the same conclusion.
Since $X$ drifts to $\infty$, it reaches arbitrarily high levels, and since $X^n$ converges to $X$ on ${\ensuremath{\left( D,d\right) }}$, then $X^n$ will also reach arbitrarily high levels. Coupled with our control on the infimum, we will see that from a given (random) time onwards and from a given index, $X^n$ and $X$ are close since they are above a high enough barrier. Formally, we will now prove that $\tilde X^n$ converges to $\tilde X$ in probability (using $d_\infty$): for any ${\ensuremath{ \varepsilon}},\delta>0$ let $M>0$ be such that the $\rho$-diameter of $[M, \infty]$ is less than ${\ensuremath{ \varepsilon}}$ and ${\ensuremath{\exp\!{\ensuremath{\left( -M\Phi\right) }}}}<\delta/2$. We introduce the stopping time: $$\begin{aligned}
S_{3M}=\inf{\ensuremath{\left\{ s\geq 0 : X_s >3M\right\} }},\end{aligned}$$ as well as the corresponding hitting times times $S^n_{2M}$ of $[2M,\infty)$ for $X_n, n=1,2,\cdots $. Observe that $$\begin{aligned}
&{\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( {\ensuremath{d_{\infty}\!{\ensuremath{\left( \tilde X^n, \tilde X\right) }}}}> {\ensuremath{ \varepsilon}}\right) }}}}
\\&= {\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( {\ensuremath{d_{\infty}\!{\ensuremath{\left( \tilde X^n, \tilde X\right) }}}}> {\ensuremath{ \varepsilon}},T_0<\infty\right) }}}}+{\ensuremath{{\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( {\ensuremath{d_{\infty}\!{\ensuremath{\left( \tilde X^n, \tilde X\right) }}}}>{\ensuremath{ \varepsilon}},T_0=\infty\right) }}}}.\end{aligned}$$ The first summand of the right-hand side of the preceding inequality converges to zero by the arguments of the previous case. Consider $h>0$ and let us bound the second summand by
[${\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( C_n\right) }}$]{}+[${\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( D_n\right) }}$]{}
where
C\_n = [$\left\{ {\ensuremath{d_{S_{3M}+h}\!{\ensuremath{\left( X^n, X\right) }}}}> {\ensuremath{ \varepsilon}},T_0=\infty\right\} $]{}
and
D\_n = [$\left\{ {\ensuremath{d_{\infty}\!{\ensuremath{\left( {\ensuremath{\left( X^n_{S_{3M}+h+t}\right) }}_{t\geq 0}, {\ensuremath{\left( X_{S_{3M}+h+t}\right) }}_{t\geq 0}\right) }}}}> {\ensuremath{ \varepsilon}},T_0=\infty\right\} $]{}.
Since Lévy processes do not jump at fixed times, the Strong Markov property implies that almost surely $X$ does not jump at time $S_{3M}+h$ so that ${\ensuremath{d_{S_{3M}+h}\!{\ensuremath{\left( X^n,X\right) }}}}\to 0$ almost surely. Hence
\_[n]{} [$ {{\mathbb{P}}}$ ]{}(C\_n) = 0.
This also implies that from a given index onwards, $S^n_{2M}\leq S_{3M}+h$ so that
[${{\mathbb{P}}}\! \left( S^n_{2M}> S_{3M}+h \right)$]{}0.
Hence, it remains to bound ${\ensuremath{{{\mathbb{P}}}\! \left( D_n,S^n_{2M}\leq S_{3M}+h \right)}}$. If the $d_\infty$ distance between ${\ensuremath{\left( X_{S_{3M+h+t}}\right) }}_{t\geq 0}$ and ${\ensuremath{\left( X^n_{S_{2M+h+t}}\right) }}_{t\geq 0}$ is to be greater than ${\ensuremath{ \varepsilon}}$ while $S^n_{2M}\leq S_{3M+h}$ then either $X^n$ goes below $M$ after $S^n_{2M}$ or $X$ goes below $M$ after $S_{3M}$. The probability of both events is smaller than $\delta/2$ from a given index onwards because of our choice of $M$, so that
\_n[${\ensuremath{ {{\mathbb{P}}}} }\!{\ensuremath{\left( {\ensuremath{d_{\infty}\!{\ensuremath{\left( \tilde X^n, \tilde X\right) }}}}> {\ensuremath{ \varepsilon}}\right) }}$]{}
for every $\delta>0$. We conclude that ${\ensuremath{d_\infty\!{\ensuremath{\left( X^n,X\right) }}}}\to 0$ in probability.
$X$ jumps to $\infty$
: This case is characterized by $q:=-{\ensuremath{\Psi\!{\ensuremath{\left( 0\right) }}}}>0$. It can be reduced to the $q=0$ case by means of an independent exponential variable of rate $q$: if $X'$ is a Lévy process whose Laplace exponent is $\Psi-{\ensuremath{\Psi\!{\ensuremath{\left( 0\right) }}}}$ and $T$ is an exponential variable with mean $1$ independent of $X'$ and we define $X''$ as $X'$ sent to $\infty$ at time $T/q$, then $X''$ has the same law as $X$. If $X'^n$ is a Lévy process with Laplace exponent $\Psi_n-{\ensuremath{\Psi_n\!{\ensuremath{\left( 0\right) }}}}$ (and $q_n:=-{\ensuremath{\Psi_n\!{\ensuremath{\left( 0\right) }}}}$) then $X'^n$ converges in law to $X'$ on ${\ensuremath{\left( D,d\right) }}$; as before, we will use Skorohod’s representation theorem so that convergence holds almost surely on a given probability space. We now extend that space so as to have an additional mean $1$ exponential variable $T$ independent of $X'$ and ${\ensuremath{\left( X'^n\right) }}_{n\in{\ensuremath{\mathbb{N}}}}$ and define on that space $X''$ and $X''^n$ as above by killing $X'$ and $X'^n$ at times $T/q$ and $T/q_n$ respectively. Since $q_n\to q$ by hypothesis, and $X'$ is continuous at time $T/q+h$ almost surely (for any $h>0$), then ${\ensuremath{d_{T/q+h}\!{\ensuremath{\left( X'^n,X'\right) }}}}\to 0$ and since $T/q_n\leq T/q+h$ from a given index onwards, then ${\ensuremath{d_\infty\!{\ensuremath{\left( X''^n,X''\right) }}}}\to 0$ almost surely.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors wish to thank Zenghu Li and Steve Evans for their comments on a previous version of this paper, and Sylvie Méléard, as well as Zenghu Li again, for their help with stochastic integral equations. G.U.B.’s research was supported by CoNaCyT grant No. 174498. Logistic and financial support received from PROYECTO PAPIITT-IN120605 and Instituto de Matematicas. A.L. is very thankful to the staff and colleagues at the Instituto for their hospitality during his stay.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Jacob Abernethy\
University of Michigan\
`[email protected]`\
Chansoo Lee\
University of Michigan\
`[email protected]`\
Ambuj Tewari\
University of Michigan\
`[email protected]`\
bibliography:
- 'bandits\_hazard\_rate.bib'
title: Fighting Bandits with a New Kind of Smoothness
---
#### Acknowledgments. {#acknowledgments. .unnumbered}
J. Abernethy acknowledges the support of NSF under CAREER grant IIS-1453304. A. Tewari acknowledges the support of NSF under CAREER grant IIS-1452099.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present electro-photometric UBV and high-speed U-band flickering observations of the recurrent nova T CrB during a period when its U brightness varies by more than 2 mag. The V band is dominated by the ellipsoidal variability of the red giant, however, the variability of the hot component also causes $\sim 0.15$ mag variations in V. We define a set of parameters which characterise the flickering. The Fourier spectra of all 27 nights are similar to each other. The power spectral density of the variations has a power law component ($\propto$f$^{-1.46}$ on average). We do not detect a dependence of the Fourier spectra and autocorrelation function on the brightness of the object. Having subtracted the contribution of the red giant, we show that the flickering amplitude correlates with the average flux of the accreting component. A comparison with CH Cyg and MWC 560 indicates that the flickering of T CrB is more stable (at least during the time of our observations), than that in the jet-ejecting symbiotic stars. The data are available in electronic form from the authors.'
date: 'Accepted . Received 2004 January 30; in original form 2003 November 11'
title: 'Flickering variability of T Coronae Borealis [^1]'
---
\[firstpage\]
stars:individual: T CrB – binaries: symbiotic – binaries:novae, cataclysmic variables
Introduction
============
T CrB (HD 143454) is an interacting binary star which consists of a red giant and a white dwarf (Selvelli et al. 1992, and references therein). The star has undergone two nova eruptions (Nova CrB 1866, 1946) and is thus classified as a recurrent nova (and, due to the presence of the cool giant, plus emission lines seen at outburst, also as a symbiotic star). The red giant fills Roche lobe, and thus the accretion flow onto the white dwarf (WD) is via L$_{1}$, which is typical for cataclysmic variables. Sharing characteristics of three (partly overlapping) types of interacting binaries, T CrB is therefore an important object for our understanding of the different processes in interacting binaries.
Stochastic brightness variations (flickering), occurring on time scales of seconds and minutes with amplitudes ranging from a few millimagnitudes up to more than an entire magnitude are a phenomenon typical for cataclysmic variables, and it is rarely observed in symbiotic stars. For example, to-date it is detected in only 8 of the 220 known symbiotics (Dobrzycka et al. 1996; Belczy[ń]{}ski et al 2000, Sokoloski et al. 2001). In T CrB flickering with amplitude of $\Delta$U$\sim$0.1-0.5 mag has been observed on a time scale of minutes (Ianna 1964, Lawrence et al. 1967 Bianchini & Middleditch 1976, Walker 1977, Bruch 1980). The flickering amplitude is somewhat smaller in B and V bands (Raikova & Antov 1986, Hric et al. 1998). In addition, on some occasions such flickering disappears (Bianchini & Middleditch 1976, Oscanian 1983, Miko[ł]{}ajewski et al. 1997). In our previous investigation (Zamanov & Bruch 1998) we showed that the flickering of T CrB is indistinguishable from the flickering observed in dwarf novae, in spite of the vast difference in the geometrical size of the systems.
The exact origin of the stochastic variations is not clear, but they are considered to be a result of accretion onto the WD through a disk. The possible mechanisms include unstable mass transfer, magnetic discharges, turbulence and instability in the boundary layer (e.g. Warner 1995, Bruch 1992).
Here we present new UBV and high-speed flickering observations of T CrB, estimate the contribution of the red giant, analyse the U band variability, search for relations between the flickering quantities and the brightness of the object, and compare its behaviour with two other symbiotic stars (the “nanoquasars” CH Cyg and MWC 560).
Observations
============
JD-2400000 V B U JD-2400000 V B U JD-2400000 V B U
------------ -------- -------- -------- -- ------------ -------- -------- -------- -- ------------ -------- -------- --------
50476.655 10.309 11.570 11.568 50698.295 10.146 11.298 11.089 51008.314 10.062 11.393 11.681
50476.659 10.311 11.564 11.620 50698.301 10.142 11.269 10.979 51009.304 10.026 11.318 11.574
50477.556 10.299 11.476 11.367 50739.236 9.946 10.960 10.655 51009.308 10.072 11.403 11.720
50477.559 10.280 11.507 11.496 50739.242 9.943 11.075 11.105 51015.323 10.055 11.343 11.526
50478.580 10.276 11.490 11.363 50741.222 9.773 10.777 10.349 51015.327 10.073 11.353 11.442
50478.583 10.274 11.446 11.314 50741.227 10.036 11.950 12.671 51016.308 10.080 11.387 11.566
50479.660 10.315 11.558 11.515 50828.591 10.213 11.245 11.087 51016.313 10.086 11.395 11.643
50479.663 10.311 11.579 11.517 50828.595 10.188 11.217 11.144 51027.320 10.124 11.412 11.531
50480.651 10.322 11.693 11.780 50864.466 10.104 11.346 11.600 51027.324 10.160 11.413 11.497
50480.654 10.370 11.615 11.639 50864.470 9.852 11.226 11.508 51034.298 10.215 11.468 11.602
50504.566 10.198 11.558 11.850 50865.491 10.015 11.294 11.476 51034.302 10.231 11.470 11.583
50504.572 10.162 11.425 11.676 50865.499 9.993 11.280 11.406 51226.574 9.848 11.063 11.085
50520.518 10.025 11.456 12.039 50867.519 9.974 11.322 11.550 51226.582 9.911 11.167 11.328
50520.524 10.040 11.433 11.970 50867.523 9.958 11.313 11.484 51239.592 10.041 11.356 11.555
50628.448 9.982 11.084 10.809 50877.603 9.974 11.373 11.658 51239.597 10.012 11.303 11.387
50628.453 9.972 11.048 10.766 50877.609 10.023 11.387 11.660 51401.292 10.054 11.095 10.714
50651.337 9.911 11.184 11.319 50877.620 10.023 11.414 11.773 51401.297 10.019 11.040 10.637
50651.342 9.906 11.191 11.328 51005.484 9.841 10.828 11.412 51404.314 10.224 11.539 11.637
50652.325 9.820 10.919 10.714 51005.485 9.853 10.833 11.371 51404.318 10.274 11.649 11.787
50652.329 9.785 10.919 10.730 51007.487 10.095 11.390 11.691 51408.289 10.249 11.712 12.123
50654.307 9.844 10.948 10.766 51007.493 10.074 11.415 12.143 51408.294 10.267 11.689 12.070
50654.311 9.733 10.908 10.811 51008.310 10.090 11.440 11.715
\[tUBV\]
The observations have been performed with the 60 cm telescope of NAO Rozhen equipped with a single channel photometer. The comparison stars were HD142929 and BD$+26{^0}2761$, the check star GSC 2037.1228 and the integration time 1 or $10\ sec$. The observations with 1 sec integration time have been binned in 10 sec. APR software (Kirov, Antov & Genkov, 1991) has been used for data processing. The accuracy of the UBV photometry is better than 0.03 mag and the results are given in Table \[tUBV\].
For the flickering observations, the reduction to the standard U band is better than $\pm 0.04\,$ mag and the internal accuracy of the data (standard deviation from the average of 10 consecutive measurements) is $0.015-0.030\,$ mag. The control of the atmospheric conditions and performance of the system have been done by observing the check star, before and after T CrB. and carefully tracing of the comparison star counts (which has been observed every 20-30 minutes). In the subsequent data processing 2 nights has been rejected, because of “doubtful” behaviour of the comparison or/and check stars. Journal of flickering observations and the main characteristics of the U band variability for each run are summarised in Table \[jour\].
The error in the magnitudes (U$_{max}$, U$_{min}$, U$_{av}$) are calculated dividing every run into two parts and calculating the quantities separately for each part, in this way addressing the possible errors of the run. In Fig.\[figV\] is plotted the orbital modulation in V, in Fig.\[fl1\] the long term U band curve and the flickering observations, and in Fig.\[exam\] are given two examples of the flickering.
Contribution of the red giant
=============================
---------- --------------- --------- ----- ------------ ------------ ------------ ---------- ------ ---------- ------------- ------------
Date TJD$_{start}$ NxIT D U$_{max}$ U$_{min}$ U$_{av}$ $\sigma$ A $\gamma$ $\tau_0$ $\tau_1$
\[sec\] min \[mag\] \[mag\] \[mag\] \[mag\] \[sec\] \[sec\]
930228 49046.5117 448x10 122 12.2600.03 12.5780.01 12.3870.04 0.084 3.41 $-$1.77 789$\pm$35 65$\pm$7
940410 49453.4621 266x10 49 12.6520.02 12.9450.05 12.8200.05 0.081 3.59 $-$1.71 516$\pm$72 54$\pm$10
950613 49882.3317 242x10 45 11.7130.11 12.2220.07 11.9520.08 0.121 3.19 $-$1.32 257$\pm$49 143$\pm$21
950620 49889.3496 520x10 102 11.7980.06 12.3530.10 12.0810.05 0.099 2.14 $-$1.47 181$\pm$25 107$\pm$10
960110 50092.6671 150x10 31 11.8260.07 12.2360.03 12.0020.03 0.092 3.70 $-$1.38 138$\pm$87 98$\pm$17
960228 50141.5088 772x10 162 11.1610.02 11.7140.01 11.4060.02 0.121 2.23 $-$1.51 292$\pm$18 130$\pm$36
960229 50142.4900 9304x1 190 11.4600.01 12.0650.07 11.7700.05 0.105 1.74 $-$1.40 1253$\pm$82 131$\pm$12
960325 50167.5358 3956x1 63 12.0170.01 12.3230.02 12.1460.00 0.052 3.04 $-$1.68 132$\pm$30 105$\pm$21
961216 50433.6246 2190x1 57 11.0510.00 11.3970.03 11.2720.02 0.078 1.52 $-$1.06 152$\pm$21 132$\pm$18
961218 50435.6354 1292x1 38 11.1110.04 11.3680.03 11.2620.04 0.066 2.50 $-$1.41 66$\pm$6 78$\pm$7
970128 50476.5542 674x10 128 11.1810.05 11.6320.10 11.4560.07 0.101 2.70 $-$1.45 675$\pm$89 174$\pm$10
970129 50477.5717 522x10 98 11.3070.01 11.6070.02 11.4690.02 0.070 2.56 $-$1.47 378$\pm$47 144$\pm$22
970130 50478.5983 425x10 81 11.3330.07 11.6140.04 11.4610.05 0.068 2.98 $-$1.54 785$\pm$204 50$\pm$8
970131 50479.5537 704x10 137 11.3540.05 11.6310.04 11.4880.04 0.060 2.95 $-$1.67 416$\pm$25 185$\pm$13
970201 50480.5379 8318x1 151 11.4430.01 11.8280.02 11.6580.04 0.093 2.42 $-$1.54 626$\pm$61 169$\pm$17
970721 50651.3508 547x10 119 11.2000.02 11.4570.01 11.3260.01 0.064 2.52 $-$1.43 799$\pm$38 50$\pm$5
970722 50652.3379 96x10 21 10.5330.01 10.7620.03 10.6690.03 0.068 3.80 $-$1.73 114$\pm$13 135$\pm$16
970827 50688.3313 207x10 46 10.5610.06 10.9390.07 10.7580.06 0.086 3.57 $-$1.67 280$\pm$60 123$\pm$22
980220 50864.5092 837x10 168 11.4640.01 11.7790.02 11.6040.01 0.071 2.14 $-$1.40 571$\pm$44 124$\pm$12
980224 50868.5383 407x10 81 11.2960.06 11.7000.07 11.5050.05 0.094 3.11 $-$1.58 498$\pm$45 139$\pm$23
980713 51008.3217 256x10 52 11.5400.01 11.7740.00 11.6600.00 0.062 2.76 $-$1.28 212$\pm$25 256$\pm$46
980714 51009.3188 362x10 74 11.5040.04 11.8400.02 11.6700.05 0.078 2.94 $-$1.43 257$\pm$216 89$\pm$10
980720 51015.3342 304x10 64 11.3350.02 11.6210.00 11.4770.01 0.072 3.34 $-$1.67 191$\pm$14 125$\pm$18
980721 51016.3208 319x10 64 11.4350.03 11.7700.05 11.6130.02 0.059 2.72 $-$1.41 105$\pm$14 91$\pm$11
980802 51028.3125 344x10 69 11.4000.06 11.7590.02 11.5950.04 0.079 2.19 $-$1.09 387$\pm$118 106$\pm$14
980803 51029.3217 298x10 61 11.5210.02 11.7370.02 11.6140.01 0.042 2.81 $-$1.38 87$\pm$25 40$\pm$10
990107 51185.6225 406x10 76 11.6460.07 12.0740.05 11.8470.05 0.086 2.33 $-$1.40 495$\pm$144 54$\pm$5
\[jour\]
---------- --------------- --------- ----- ------------ ------------ ------------ ---------- ------ ---------- ------------- ------------
V band {#VgM}
------
In symbiotic stars, the mass donor is a red giant. In the case of T CrB its contribution is not negligible in the UBV bands. The V band variability of T CrB is dominated by the ellipsoidal variability of the red giant (Peel 1985, Lines et al 1988). The V band data from the long term light curve (see Stanishev et al. 2004 and references therein) are plotted in Fig.1, folded with the orbital period.
A three term truncated Fourier fit to all data gives $$\begin{aligned}
& & V = 10.056\,(0.003) \\
& & +0.007\,(0.004)\cos 2\pi\phi -0.026\,(0.004)\sin 2\pi\phi \\
& & -0.161\,(0.004)\cos 4\pi\phi -0.036\,(0.004)\sin 4\pi\phi \\
& & +0.016\,(0.004)\cos 6\pi\phi -0.037\,(0.004)\sin 6\pi\phi \end{aligned}$$ where $\phi$ is the orbital phase (hereafter the numbers in the parentheses refer to the errors). This fit is plotted in Fig.1. Typical deviation of the points from the fit line is $\pm$0.10 mag.
On Fig.1 we have plotted with different symbols the points when the object is brighter and fainter at shorter wavelengths (open circles refer to U$<$12 and filled – to U$\ge$12). It is visible that the filled circles are displaced downward slightly relative to the open ones. The U band brightness is dominated by the hot component. We can also deduce that the variability of the hot component of about 2 magnitudes in U (see also Stanishev et al. 2004) also contributes to that in V. To define this contribution we performed a simple fits (using only the main terms) to the open and filled symbols. The obtained coefficients are: $$\begin{aligned}
V=10.157\,(0.005)-0.194\,(0.007)\cos 4\pi\phi & {\rm for} & U\ge12 \\
V=10.029\,(0.003)-0.163\,(0.004)\cos 4\pi\phi & {\rm for} & U < 12 . \end{aligned}$$ The mean values of the U band magnitudes are U=12.3$\pm$0.3 and U=11.2$\pm$0.4 for the fainter and brighter points respectively. Therefore, the increase of the system U-band brightness brightness by 1.1 mag results in an increase of the V brightness by 0.128 (Eq. 3,4). We derive a relative contribution R(V)=0.205$\pm$0.035, where R(V)$=$F$_{hot}$/F$_{gM}$ is the relative contribution between the accreting object and the red giant at V=10.056. The corresponding orbital light curve of the red giant is plotted as a dashed line on Fig.\[figV\]. The calculated contribution is very similar to that obtained by Zamanov & Bruch (1998) on the basis of the average colours of the flickering source in cataclysmic variables.
Although the data in Fig.\[figV\] spread over 22 years, the typical deviation of the points from the fit line is $\pm$0.10 mag. This points to the fact that the $V$ band light curve has not changed in its main features over the last 22 years. This in turn indicates that the M giant is not variable. Indeed, we can put an upper limit on its possible variability of $\Delta V <$ 0.05 – 0.10. The stability of the red giant is better defined in IR observations (Yudin & Munari 1993, Shahbaz et al., 1997), where a upper limit of variability $\Delta J <0.02$ has been constrained.
Red giant contribution to U band flux
-------------------------------------
The latest definitions of the spectral type of the red giant in T CrB are M4III (Zhu et al. 1999) and M4.5III (M[" u]{}rset & Schmid 1999). Both are obtained on the basis of IR spectra and with typical uncertainty $\pm 1$ spectral subtype. The expected colour of a M4.5III star is (U-V)$_{M4.5III}=3.16 \pm 0.10$ (Lee 1970), (U-V)$_{M4.5III}=3.28 \pm 0.10$ (Schmidt-Kaler 1982), (U-V)$_{M4.5III}=3.25 \pm 0.10$ (calculated using the tables of Fluks et al. 1994). The New ATLAS9 model atmospheres (Pietrinferni, Cassisi, Salaris, Castelli, in preparation), for a star T$_{eff}=$3500 K and $\log (g) = 1.0$, give (U-V)=3.18 for \[Fe/H\]=0.2, (U-V)=3.26 for \[Fe/H\]=0, and (U-V)=3.42 for \[Fe/H\]=-1.5.
Using R(V)=0.222 at V=10.029 (as derived in Section \[VgM\]), the fit to V (Eq.1), E$_{B-V}=0.15$, and (U-V)$_{M4.5III} =$3.25, we can calculate the contribution of the red giant to the U band. The light curve of T CrB, our flickering observations, and the red giant contribution (as a sine wave) are plotted in Fig.2.
One way to check the calculated U brightness of the red giant is IR photometry, supposing that the red giant is the only source in J band. The J brightness of T CrB varies in the interval J=5.90 – 6.19 (Kamath and Ashok 1999). Interpolations in the tables of the colours of red giants give (U-J)$_{M4.5III}=7.75 \pm 0.25$ (from Lee 1970), (U-J)$_{M4.5III}=7.14 \pm 0.2$ combining (U-V) from Schmidt-Kaler (1982) and (V-J) from Ducati et al. (2001), (U-J)$_{M4.5III}=7.34 \pm 0.2$ (from Fluks et al. 1994), where the calculated uncertainties refer to $\pm 0.5$ spectral type. The model atmospheres give (U-J)$_{[Fe/H]=0.2}$=7.24, (U-J)$_{[Fe/H]=0}$=7.13 and (U-J)$_{[Fe/H]=-1.5}$=6.52 (Pietrinferni et al. 2004).
Shahbaz et al (1999) modeled the M giant spectrum, with enhanced abundance of lithium but normal abundance of the other metals, however they used higher gravity inconsistent with the last orbital solution.
If we suppose that the red giant is the only source in J, using the above (U-J) colours and E$_{B-V}=0.15$ we could expect value about U$\sim$ 14.4 - 13.3, which is consistent with the supposed contribution of the red giant to the U band (see Fig.2).
Flickering quantities
=====================
The U magnitudes were converted into fluxes, adopting flux for a zero magnitude star F$_{0}(U)=4.194\times10^{-11}$ ${\rm Watt\: m}^{-2}{\rm nm}^{-1}$ (Bessel 1979). In addition to our data, we used those from Bruch (1992) and the data of Oskanian (1983). Bruch’s data are reddened with $E_{B-V}=0.12$ and they were corrected for the difference in the adopted zero-point of the flux scale. For Oskanian’s data, we assumed that the instrumental difference $\Delta u=0$ corresponds to magnitude U=11.83. We have used only positive detections of the flickering. After the observed flux during a given night was corrected for the contribution of the red giant the following quantities were calculated:
F$_{av}$ – the average flux of the hot component;
F$_{max}$ – the maximum flux of the hot component;
F$_{min}$ – the minimum flux of the hot component;
F$_{fl}$ – the average flux of the flickering, F$_{fl}=$F$_{av} -$F$_{min}$.
In Fig.\[fl2\] are plotted the flickering quantities versus the average flux of the hot component. It is obvious they have to be connected, however it is not clear a priori how the different quantities will depend on each other. The least squares fits to data in Fig.4 give: $$\begin{aligned}
F_{max}= -0.22\,(0.07) + 1.252\,(0.006)\,\,F_{av} \\
F_{min}= -0.32\,(0.56) + 0.858\,(0.054)\,\,F_{av} \\
F_{fl} = +0.19\,(0.42) + 0.156\,(0.042)\,\,F_{av} \end{aligned}$$
Power spectra
-------------
For each run we also calculated the power spectrum and the autocorrelation function. Two examples are shown in Fig.\[exam\]. Over a wide range of frequencies the power spectra of T CrB light curves follow power law $P(f)\propto f^\gamma$, where $P$ is the power and $f$ is the frequency. Such a power-law shape is commonly observed in the light curves of cataclysmic variables and is attributed the flickering. The power-law index $\gamma$ was determined in the frequency interval from 3 to 160 cycles/hour. Practically, we fitted the power spectra over this interval in log-log scale with least-squares linear fit: $\log(P)=A + \gamma \log(f)$. $\gamma$ and A are given in Table 2. The typical errors are $\Delta$A$=\pm 0.8$ and $\Delta \gamma=\pm 0.25$. The visual comparison/inspection shows that all power spectra are very similar. This is confirmed from the fits. They give an average value of $\overline{\gamma}= -1.46 \pm 0.17$. There are two runs where $\gamma$ is $\approx -$1.1. In both cases two stronger flashes are visible in the variability with amplitude about $\sim$ 0.2 mag.
The power-law spectrum of the type observed in T CrB is expected in the model of flickering proposed by Yonehara, Mineshige, & Welsh(1997). They proposed as origin of flickering self-organised critical state of the disk in which seemingly chaotic fluctuations can be produced. Such a model implies $\gamma \simeq -1$ – $-2$. Our observations of T CrB do not contradict to this model.
The correlation analysis shows no correlation of A and $\gamma$ with U$_{av}$ or F$_{av}$, (r$_P < 0.2$), indicating that the flickering is “stable”, i.e. without considerable changes of the nature of the power spectrum in spite of the variability.
This is however not the situation in the symbiotic star CH Cyg, where the power spectrum changes dramatically. There are even moments when CH Cyg’s power spectrum cannot be fitted with a simple power-law as a result of an unstable disk and disruption of the inner disk during the jet ejection (Sokoloski & Kenyon, 2003b). In T CrB we did not observe instabilities like those observed in CH Cyg.
Autocorrelation function (ACF)
------------------------------
Another objective way to investigate flickering behaviour is the ACF (see also Bruch 2000). The ACFs were calculated according to Edelson & Krolik (1988) for unevenly spaced data.
The typical time scale of the flickering may be defined as the time shift at which the ACF first reaches the value 1/e. Thus determined correlation times are influenced by the presence of periodic brightness variations or some trends in the data. As Robinson & Nather (1971) and Panek (1980) note, these correlation times can be additionally biased by the presence of weakly correlated noise and the process of trend removal (if applied).
The e-folding time is given in Table 2. The errors are determined from the errors of the autocorrelation coefficients. We calculated e-folding times in two different ways: (i) the e-folding time of the ACF of the original data in each run ($\tau_{0})\;$ and (ii) after subtraction of a spline fit $(\tau_{1})$. Operation (ii) has been done in order to obtain the typical time of the flickering on shorter time scales. A tension spline interpolation was undertaken. We subtracted this spline fit through the mean points in non-overlapping bins of length about $\sim$20 minutes in a way that is identical to that applied to TT Ari by Kraicheva et al. (1999).
The e-folding time of the ACF varies over a wide interval. The average values, standard deviation of the average, and median value of the e-folding time are $\overline{\tau_{0}}= 394 \pm 287$, $<\tau_{0}>=292$, $\overline{\tau_{1}}= 115 \pm 48$, $<\tau_{1}>=123$, where all values are in seconds. The values of $\tau_1$ are more or less similar to the values of TT Ari as defined in Kraicheva et al. (1999).
The correlation analysis showed that there are no correlations between the so-defined e-folding times and the brightness of the object, or the flickering quantities (F$_{av}$, F$_{fl}$, or the size of the boundary layer). Linear Pearson correlation coefficient and Spearman’s Rank correlation give values between 0 and 0.2, indicating that there is no dependence of the time scale of variability on the luminosity of the hot component, i.e. the the characteristic time scales of the flickering are not connected with the brightness of the object.
Boundary layer {#BL}
--------------
The origin of flickering is still not clear although this phenomenon is observed for many stars. Bruch (1992) and Bruch & Duschl (1993) identify the boundary layer between the accretion disk and the white dwarf as the most probable place for the origin of flickering. Bruch & Duschl (1993) consider that the ratio $\frac{F_{fl}}{F_{min}}$ is connected with the size of the boundary layer between the white dwarf and the accretion disk.
In T CrB F$_{fl}$ is well correlated with F$_{min}$ (see Fig.\[bl\]) . The Pearson’s Correlation Coefficient is r$_P$=0.72 and Spearman’s Rank Correlation r$_S$=0.56. Assuming that the deviation of the points from the fit lines (Fig 2. and Eqs. 3,4,5) is due only to the errors of the measurements, the fits (Eqs. 3,4,5) indicate that in spite of variations in $F_{av}$ (which we suppose is related to the mass accretion rate), the ratios $ F_{fl}/F_{min}$ and ${F_{max}}/{F_{min}}$ do not change markedly. In terms of Bruch & Duschl (1993) this means that the size of the boundary layer remains almost constant independently of changes in the mass accretion rate. Here, adding more data (see Fig.\[bl\]) we confirm the conclusion of Zamanov & Bruch (1998) that F$_{fl}$ increases linearly with the increase of F$_{min}$. This, within the limits of Bruch & Duschl’s model, means that the size of the boundary layer in T CrB remains almost constant independently of the changes in the mass accretion rate.
Flickering amplitude
====================
Flickering amplitude ($\Delta F=F_{max}-F_{min}$) is also measured in the extensive observations of CH Cyg (Miko[ł]{}ajewski et al., 1990) and MWC 560 (Tomov et al. 1996). The flickering amplitude versus the average flux of the hot component, after subtraction of the red giant contribution is presented for all three stars in Fig.\[3stars\].
T CrB
-----
The data in Fig.\[3stars\] show that the flickering amplitude depends on the average hot component flux. The correlation is well defined with $r_P=0.72$ and $r_S=0.56$. Searching for a dependence of the type $\Delta F \propto (F_h)^k$, we obtain a best fit for T CrB $k=1.09\pm0.11$. The fit and the corresponding error have been calculated in two ways: (1) Using the errors into corresponding quantities as given in Table 2. (2) Bootstrapping simulations (e.g. Efron & Tibshirani, 1993) over the points plotted in Fig.5, i.e. taking $\sim$20 subsamples from our points. Using only our own and Bruch’s points (i.e. data well calibrated in U) we obtain $k=1.03\pm0.09$. Using different subsamples of the whole sample we obtained values $0.93\le k \le 1.22$. An error in the subtraction of the M giant contribution of 25% will cause error in k of about 0.05.
CH Cyg {#amplCH}
------
Dependence of the flickering amplitude on the brightness has been reported for CH Cyg by Miko[ł]{}ajewski et al. (1990). Their results show that the flickering amplitude in CH Cyg is a power law function of the hot component luminosity, i.e. $\Delta F \propto (F_h)^k$, where $k=1.40-1.45$. Here, using their data, we subtracted the contribution of the red giant and the resulting points are plotted on Fig.\[3stars\]. To subtract the contribution of the red giant we assume that at the minimum of V flux all the light is due only to the red giant, and it has a colour corresponding to (U-V)$_{M6III}$= 2.43 - 2.70 (Lee, 1980; Schmidt-Kaler 1982; Fluks 1994). The minimum brightness of CH Cyg is V=10.0 (Miko[ł]{}ajewski et al. 1996), and we adopt a contribution of the red giant of U=12.45. We will not go into details as to whether the system is triple (Hinkle et al. 1993; Skopal et al. 1998) or binary (Munari et al. 1996). Here we only note that an error of $\pm$0.7 mag in the subtraction of the red star(s) flux would influence the obtained slope to less than $\pm$ 0.02.
The analysis gives a very high correlation between F$_{av}$ and $\Delta$F. r$_P$ and r$_S$ are always about 0.88-0.94 using (1) all points, (2) the propeller state observations, and (3) the accretor state (for further discussion of propeller and accretor states in CH Cyg, see Miko[ł]{}ajewski et al 1990).
After the subtraction of the red giant contribution we obtain $k=1.08\pm0.05$ if we use all points, $k=1.48\pm0.05$ for propeller state points only, and $k=1.41\pm0.05$ for the accretor state only. It has to be noted that the fit $\Delta F \propto (F)^{1.08}$ obtained on the basis of all points is very similar to the value for T CrB.
MWC 560
-------
For MWC 560 the minimum brightness is V=10.2 (Tomov et al. 1996) and the red giant is classified as M5.5 III (Schmid et al 2001). Using (U-V)$_{M5.5III}$=2.80 - 2.95 (Lee, 1980; Schmidt-Kaler 1982; Fluks 1994), we adopt a contribution of the red giant equivalent to U$\approx 13.0$. It deserves noting that the calculated red giant contribution in MWC 560 and CH Cyg is considerably smaller than that in T CrB, which is in accordance with the fact that in these two objects the V band variability is dominated by the hot component (not by the red giant as in T CrB).
The flickering amplitude versus F$_{av}$ for MWC 560 is plotted in Fig.\[3stars\] (lower panel). The correlation is not very significant (r$_P=0.25$ r$_S=0.27$ – although, if we delete 3 points with $\log \Delta F <0.6$ we can obtain a moderate correlation up to r$_P \sim 0.4$).
During the time of these observations MWC 560 is in the process of jet ejection, and the jet is even precessing (Iijima 2002). The lack of significant correlation between the flickering amplitude and hot component flux probably is a result of the outflow, i.e. the jet ejection destroys the innermost parts of the accretion disk, where the flickering is formed.
The connection between the jet and flickering is not investigated in MWC 560, however it is visible in the 1997 jet launch in CH Cyg (Sokoloski & Kenyon 2003a). The disks and jets are also connected in quasars and microquasars (see Livio, Pringle & King 2003 and references therein). In this sense the fact that the correlation between $\Delta$ F and F$_{av}$ is loose in MWC 560, is probably due to the outflow and its connection with the accretion disk.
Discussion of flickering amplitude
==================================
In all three symbiotic stars with available data the flickering amplitude shows the same tendency to increase with increases of the hot component flux. If we accept that there are no different states in CH Cyg and all points lay on the same line it means that the obtained value of the slope is very similar to T CrB, which in terms of the Bruch & Dushl (1993) model in turn means that in both stars the size of the boundary layer remains constant (see Sect.\[BL\]). In MWC 560 the flickering is (probably) influenced by the outflow and $\Delta$F depends weakly on F$_{av}$, and the correlation is not well defined.
The other possibility that the flickering amplitude of CH Cyg lays on two parallel lines would give different values of k in the relationship $\Delta F \propto (F)^k$. One of the reasons for this difference could be the magnetic field of the WD. Here we want to point out that the flickering amplitude could be connected with the magnetic field. The most probable place for the origin of the flickering is the inner parts of the accretion disk. If the flickering is a result of the turbulence in the inner parts of the disk then the energy available in the turbulence will be proportional to the density where the flickering forms (Bruch, 1992). If the white dwarf is magnetic, the inner parts of the accretion disk will be destroyed by the magnetic field. Different instabilities can appear then at the inner edge of the disk. These instabilities permit the accreting material to be absorbed from the magnetosphere as blobs. The energy releasing will be unsteady and we suppose that the amplitude of the flickering will be proportional to the typical mass of the blobs, and the mass of the blobs will in turn be proportional to the density at the inner edge of the disk. The density in the disk can be estimated as (Lipunov 1992):
$$\rho_{in} \approx \alpha^{-1} \left( \frac{R}{H} \right)
\frac {\dot {M_a}} { 4 \pi R^{3/2} \sqrt {2GM} },
\label{eq2}$$
where $(R/H)$ is the ratio between the radius and the vertical size of the disk, ($R/H$) is usually adopted to be a constant of order 0.01–0.1. $M$ is the mass of the white dwarf.
If the white dwarf is non-magnetic, the inner radius of the disk will be approximately equal to the white dwarf radius (for a thin boundary layer) and consequently the density at the inner edge is given by $\rho_{in} \propto \dot M_a$. The same relationship will be fulfilled if the boundary layer is not thin and its size does not change. If the white dwarf is magnetic the radius $R_0$ of the inner disk edge may be expressed as (Lamb, Pethick & Pines 1973): $$R_0=N (GM)^{-1/7} \mu ^{4/7} \dot {M_a}^{-2/7} ,
\label{eq3}$$ where $N$ is a constant of order 1, and $\mu$ is the white dwarf magnetic moment. In this case (from Eqn. \[eq2\] and \[eq3\]) $$\rho_{in} \propto \dot M_a^{k},\ \ k = \frac {10}{7}
\label{eq4}$$ where $k = \frac {10}{7}=1.43 $ is in agreement with the behaviour of CH Cyg (Miko[ł]{}ajewski et al., 1990, see also Sect.\[amplCH\] ), if the suppositions (1) $\Delta F \propto \rho_{in}$, and (2) accretor-propeller states in CH Cyg (Miko[ł]{}ajewski et al. 1990) are correct.
The data for T CrB are consistent with $k=1$ as expected for a low or non-magnetic white dwarf, i.e. the position of the inner edge of the accretion disk does not depend on the mass accretion rate.
The presence of a magnetic WD in CH Cyg is not a certain fact. Sokoloski & Kenyon (2002a), Crocker et al. (2001), Ezuka et al. (1998) threw doubts about the presence of such a magnetic WD in CH Cyg. However, the magnetic propeller model of (Miko[ł]{}ajewski & Miko[ł]{}ajewska, 1988) still remains the most promising for the variability of this object.
If the differences in the behaviour of the flickering in T CrB, CH Cyg, and MWC 560 are not connected with the magnetic field and jet ejection, other possible reasons may be the changes of the energy distribution, or different mechanisms generating the flickering in these objects.
Conclusions
===========
We have analysed the U band variability of the recurrent nova and symbiotic star T CrB, and compared its behaviour with two other symbiotic stars CH Cyg and MWC 560. During the period of our observations T CrB brightness varies in between U$=$13 and U$=$10 mag. The analyses we performed show that: (1) the V brightness during the last 22 years is dominated by the ellipsoidal variability of the red giant, however the hot component variability with $\Delta U {\lower.5ex\hbox{{$\; \buildrel > \over \sim \;$}}}$2 mag, introduces a shift in V with about 0.15 mag. No signs of variability of the red giant has been detected. (2) the power spectrum of the flickering does not change during our observations, remaining always with slope $\gamma \approx -1.5$ in spite of the changes in U. We do not detect changes in the power spectrum like those observed in CH Cyg; (3) The calculated e-folding time of the ACF also does not show dependence on the changes in U; (4) The flickering amplitude is strongly correlated with the average flux of the hot component. (5) The differences in the dependence of the flickering amplitude between T CrB, CH Cyg and MWC 560 could be connected with jet ejection and the possible presence of a magnetic white dwarf in the last two. In general, in T CrB, we have observed flickering, which does not change considerably its characteristics (at least during the time of our observations). In the future it would be very interesting to determine the behaviour of the flickering amplitude, ACF, power spectra, etc. of other symbiotic stars with flickering (in particular RS Oph, RT Cru, $o$ Ceti) as well as the flickering of MWC 560 during phases without outflow, as well the connection of flickering with jet precession. Simultaneous spectral and photometric observations over a wide spectral range from UV to IR could be very useful to investigate in detail the flickering behaviour and its connection with accretion disk instabilities and jet ejections.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research has made use of SIMBAD, IRAF, and Starlink. RZ is supported by a PPARC Research Assistantship and MFB is a PPARC Senior Fellow.
[99]{} Belczy[ń]{}ski, K., Miko[ł]{}ajewska, J., Munari, U., Ivison, R. J., & Friedjung, M., 2000, A&AS, 146, 407 Bessell, M. S. 1979, PASP, 91, 589
Bianchini, A., & Middleditch, J., 1976, IBVS 1151 Bruch, A., 1980, IBVS 1805 Bruch, A. 2000, A&A , 359, 998 Bruch, A. & Duschl, W. J. 1993, A&A, 275, 219 Bruch, A. 1992, A&A, 266, 237 Crocker, M. M., Davis, R. J., Eyres, S. P. S., Bode, M. F., Taylor, A. R., Skopal, A., & Kenny, H. T., 2001, MNRAS, 326, 781 Dobrzycka, D., Kenyon, S. J., & Milone, A. A. E. 1996, AJ, 111, 414
Ducati, J. R., Bevilacqua, C. M., Rembold, S. B., & Ribeiro, D., 2001, ApJ, 558, 309
Efron B., Tibshirani R.J., 1993, An introduction to the Bootstrap, Monographs on Statistics and applied Probability vol.57, Chapman & Hall, New York Edelson, R. A. & Krolik, J. H., 1988, ApJ, 333, 646 Ezuka, H., Ishida, M., & Makino, F., 1998, ApJ, 499, 388 Fluks, M. A., Plez, B., The, P. S., de Winter, D., Westerlund, B. E., & Steenman, H. C. 1994, A&AS, 105, 311
Glass I. S., Schultheis M., 2002, MNRAS, 337, 519 Ghosh, P. & Lamb, F. K. 1979, ApJ, 232, 259 Hinkle, K. H., Fekel, F. C., Johnson, D. S., & Scharlach, W. W. G., 1993, AJ, 105, 1074 Hric, L., Petrik, K., Urban, Z., Niarchos, P., & Anupama, G. C. 1998, A&A, 339, 449 Ianna, P.A., 1964, ApJ, 139, 780 Iijima, T. 2002, A&A 391, 617 Kamath, U. S. & Ashok, N. M. 1999, A&AS, 135, 199 Kirov N.K., Antov A.P., Genkov V.V., 1991, C.R. Acad.Bulg.Sci. 44, 5 Kraicheva, Z., Stanishev, V., Genkov, V., & Iliev, L., 1999, A&A, 351, 607
Lamb, F. K., Pethick, C. J., & Pines, D. 1973, ApJ, 184, 271
Lawrence, G. M., Ostriker, J. P., & Hesser, J. E. 1967, ApJ 148, L161
Lee, T. A. 1970, ApJ, 162, 217
Lipunov, V. M. 1992, Astrophysics of Neutron Stars, Berlin, New York : Springer-Verlag, QB843, 5713 Lines, H. C., Lines, R. D., & McFaul, T. G., 1988, AJ, 95, 1505 Livio, M., Pringle, J. E., & King, A. R. 2003, ApJ 593, 184 Miko[ł]{}ajewski, M. & Miko[ł]{}ajewska, J. 1988, ASSL Vol. 145: IAU Colloq. 103: The Symbiotic Phenomenon, 233 Miko[ł]{}ajewski, M., Tomov, T. V., Kolev, D., & Leedjarv, L., 1996, IBVS 4368
Miko[ł]{}ajewski, M., Tomov, T., & Kolev, D. 1997, Informational Bulletin on Variable Stars, 4428
Miko[ł]{}ajewski, M., Miko[ł]{}ajewska, J., Tomov, T., Kulesza, B., & Szczerba, R. 1990, Acta Astronomica, 40, 129
Munari, U., Yudin, B. F., Kolotilov, E. A., & Tomov, T. V. 1996, A&A, 311, 484
M[" u]{}rset, U. & Schmid, H. M. 1999, A&AS, 137, 473
Oskanian, A. V. 1983, IBVS 2349
Panek, R. J., 1980, ApJ, 241, 1077 Peel, M., 1985, JAAVSO, 14, 8
Pietrinferni, A., Cassisi, S., Salaris, M., Castelli, F., 2004, in preparation
Raikova, D. & Antov, A., 1986, IBVS 2960 Robinson, E. L. & Nather, R. E., 1979, ApJS, 39, 461
Schmidt-Kaler, T. H. 1982, in Landolt-Börnstein, New Series, Group VI, Vol. 2b, Stars and Star Clusters, ed. K. Schaifers & H. H. Voigt (New York: Springer)
Schmid, H. M., Kaufer, A., Camenzind, M., Rivinius, T., Stahl, O., Szeifert, T., Tubbesing, S., & Wolf, B. 2001, A&A, 377, 206
Selvelli, P. L., Cassatella, A., & Gilmozzi, R., 1992, ApJ, 393, 289 Shahbaz T., Somers M., Yudin B., Naylor T., 1997, MNRAS 288, 1027 Skopal, A., Bode, M. F., Lloyd, H. M., & Drechsel, H., 1998, A&A, 331, 224
Sokoloski, J. L. & Kenyon, S. J. 2003a, ApJ 584, 1021 Sokoloski, J. L. & Kenyon, S. J. 2003b, ApJ 584, 1027 Sokoloski, J. L., Bildsten, L., & Ho, W. C. G. 2001, MNRAS, 326, 553 Stanishev, V., Zamanov R., Tomov N., Marziani P., 2004, A&A, in press (astro-ph/0311309) Tomov, T., Kolev D., Ivanov M., Antov A., Jones A., Miko[ł]{}ajewski M., Lepardo A., et al. 1996, A&AS 116, 1
Walker, A. R., 1977, MNRAS 179, 587
Warner, B. 1995, Cataclysmic variable stars, Cambridge Astrophysics Series, Cambridge, New York: Cambridge University Press
Yonehara, A., Mineshige, S., & Welsh, W. F., 1997, ApJ, 486, 388
Yudin, B. & Munari, U. 1993, A&A, 270, 165
Zamanov, R. K. & Bruch, A. 1998, A&A, 338, 988 Zhu, Z. X., Friedjung, M., Zhao, G., Hang, H. R., & Huang, C. C. 1999, A&AS, 140, 69
[^1]: based on observations obtained in NAO Rozhen, Bulgaria
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define persistent homology groups over any set of spaces which have inclusions defined so that the underlying graph between the spaces is directed and acyclic. This method simultaneously generalizes standard persistent homology, zigzag persistence and multidimensional persistence to arbitrary directed acyclic graphs, and it also allows the study of arbitrary families of topological spaces or point-cloud data. We give an algorithm to compute the persistent homology groups simultaneously for all subgraphs which contain a single source and a single sink in $O(n^4)$ time, as well as an algorithm to compute persistence for any arbitrary subgraph in the same running time. We then demonstrate as an application of these tools a method to overlay two distinct filtrations of the same underlying space, which allows us to calculate significant barcodes using considerably fewer points than standard persistence.'
author:
- 'Erin W. Chambers [^1]'
- David Letscher
bibliography:
- 'dag-persistence.bib'
title: Persistent Homology Over Directed Acyclic Graphs
---
Introduction {#sec:intro}
============
Since its introduction in [@persistence], the concept of topological persistence has become one of the most utilized tools in computational geometry. It has found numerous applications in diverse areas such as surface reconstruction, sensor networks, bioinformatics, and cosmology.
In this paper, we give a generalization of persistence to spaces where the underlying inclusions form a directed acyclic graph. This simultaneously generalizes both zigzag and multidimensional persistence, which can be viewed as special cases of these underlying graphs on the maps between the spaces.
We then give algorithms to compute the persistent homology for DAGS in various settings. Our algorithms are analyzed in terms of $e$ and $v$, which are the number of edges and vertices in the directed acyclic graph $G$, as well as $l$, the longest directed path in $G$. (We will in general let $n = v + e$ in our algorithm analysis for simplicity.) In Section \[sec:single\], we give an $O(n^4)$ algorithm for computing persistent homology over a finite field when the underlying DAG has only a single source. In Section \[sec:algorithm\], we give an algorithm to calculate the homology of a fixed graph in $O(n^4)$ time; this procedure is based on Gram-Schmidt orthonormalization, and hence only works over rational coefficients. In addition, we give an algorithm in Section \[sec:finitefield\] for computing the persistent homology of a DAG over a finite field in $O(n^5)$ time.
Potential applications of this are extensive, including any spaces where inclusions are more general than previous settings. We present two such applications in Section \[sec:applications\]. The first uses multiple samples of the same space to accurately find significant topological features with far fewer sample points than other methods require. The second application uses DAG persistence to measure the similarity between two spaces.
Background and related work
---------------------------
For completeness, we briefly survey some results from persistent homology with an emphasis on tools and techniques used in this paper, although a full coverage is beyond the scope of this paper. See any of the recent books or surveys on topological persistence for full coverage of this broad topic and its applications [@eh-phs-08; @eh-cti-10; @ghristoverview; @z-tc-05].
The original algorithm for computing persistent homology [@persistence] was quickly followed by more work on efficiently computing these groups in several different settings [@persistencealgorithm] as well as showing desirable properties such as stability under certain assumptions [@stability] and accurate reconstruction in the presence of noise [@noisypersistence].
In the course of adding elements to the space, homology classes are “born" and “die" at a certain time. This data can be summarized in a *barcode*, which shows the intervals over which the homology classes endure [@module]. An equivalent representation is a persistence diagram that plots the birth and death times on a coordinate plane [@stability]. The use of barcodes is of particular interest to us, as it relates to our application in Section \[sec:applications\]. Through computation of barcodes, one can capture a representation of a topological feature by an interval that represents how long this interval survives in our filtration; at a high level, more important topological features will persist longer in a filtration and will therefore have longer intervals.
Several extensions to topological persistence have appeared, most notably zigzag persistence [@zigzag] and multidimensional persistence [@multid-persistence]. In each case, these generalize the simple inclusions from the main algorithm to more general setups.
Zigzag persistence considers spaces with maps of the form ${X}_1 \leftrightarrow {X}_2
\leftrightarrow \ldots \leftrightarrow {X}_n$, where the maps can go in either direction. These maps between the spaces induce maps between chain complexes which pass to homology as homomorphisms $H({X}_1) \leftrightarrow H({X}_2) \leftrightarrow \ldots \leftrightarrow H({X}_n)$; this is known as a zigzag module. Recent work in this setting includes an algorithm which examines the order of the necessary matrix multiplications quite carefully and is able to get a running time for sequence of $n$ simplex deletions or additions which is dominated by the time to multiply two $n \times n$ matrices [@zigzag-mm].
Multidimensional persistence extends standard persistence to work not just along a single dimension, but rather on maps between spaces which are parameterized with respect to multiple dimensions [@multid-persistence]. While no analog to the barcode exists in this setting to capture all topological information, the rank invariant is a natural extension of the barcode which captures relevant topological information in many settings.
Definition {#sec:def}
==========
We recall some relevant definitions and background before presenting our definition of persistent homology over directed acyclic graphs. For a full presentation of homology groups see any introductory text in algebraic topology [@hatcher; @munkres].
For a simplicial complex $X$ and an abelian group $A$, the $k$-chains $C_k(X)$ is the space of formal linear combinations of the $k$-simplices of $X$ where the coefficients are in $A$. $\partial_k: C_k(X) \to C_{k-1}$ is a linear map that calculates the boundary of a chain. The cycle group $Z_k(X) = \{ c \in C_k(X) \ | \ \partial_k(c) = 0 \} = ker(\partial_k)$ and the boundary group $B_k(X) = \{ c \in C_k(X) | \ \exists d \in C_{k+1}(X)$ with $\partial_{k+1} d = c \} = im (\partial_{k+1})$. The homology group is defined as $H_k(X) = Z_k(X)/B_k(X)$. Note that if $A$ is a field then $C_k(X), Z_k(X), B_k(X)$ and $H_k(X)$ are all vector spaces.
Given a filtration $X_0 \subset X_1 \subset \cdots \subset X_n$, the persistent homology group $H_k^p(X_i)$ can defined in multiple ways. Traditionally, it is defined as $Z_k(X_i) / B_k(X_{i+1}) \cap Z_k(X_{i+p})$ and can be viewed as a quotient group of $H_k(X_{i+p})$. An equivalent definition is $H_k^p(X_i) = im(i_*)$, where $i_*: H_k(X_i) \to H_k(X_{i+p})$ is the map induced by the inclusion $i:X_i \to X_{i+p}$. So $H_k^p(X_i)$ can also be thought of as a subgroup of $H(X_{i+p})$.
When viewed from the perspective of zigzag persistence and the use of the zigzag module, the persistence group for some interval can be thought as the subgroups which are common to all of the the homology groups. This motivates the definition of persistence for arbitrary directed graphs. The main restriction on this graph is that it must be acyclic, which is a natural for any graph which represents a set of inclusions.
For a directed acyclic graph $G = (V,E)$, a *graph filtration* $\mathcal{X}_G$ of a topological space $X$ is a pair $( \{ X_v \}_{v \in V}, \{f_e\}_{e \in E} )$ such that
1. $X_v \subset X$ for all $v \in V$
2. If $e = (v_1, v_2) \in E$ then $f_e: X_{v_1} \to X_{v_2}$ is continuous embedding (or inclusion) of $X_{v_1}$ into $X_{v_2}$.
Given a graph filtration there is a corresponding directed graphs of homology groups, where each edge corresponds to the map induced by inclusion $(f_e)_*: H_k(X_{v_1}) \to H_k(X_{v_2})$, see Figure \[fig:filtration\].
$\xymatrix@R=1.4pc @C=1.4pc{
& X_1 \ar[rr] \ar[dr] & & X_5 \\
& & X_3 \ar[r] \ar[dr] & X_6 \\
& X_2 \ar[ur] \ar[r] & X_4 \ar[ur] \ar[r] & X_k \\
}$ $\xymatrix@R=1.4pc @C=1.4pc{
& H_k(X_1) \ar[rr] \ar[dr] & & H_k(X_5) \\
& & H_k(X_3) \ar[r] \ar[dr] & H_k(X_6) \\
& H_k(X_2) \ar[ur] \ar[r] & H_k(X_4) \ar[ur] \ar[r] & H_k(X_7) \\
}$
We will define persistent homology groups for any connected subgraph of $G$ as the “largest” group that injects into all of the homology groups at the vertices such that the image group is preserved by all of the edges, see figure \[fig:definition\]. To make this definition precise, we will assume that the coefficients for our homology groups are from some field $\mathbb{F}$. Throughout this paper, we will assume that $H_k(X)$ means $H_k(X; \mathbb{F})$ (unless explicitly stated otherwise, as in Section \[sec:algorithm\]); this ensures that all of the homology groups are vector spaces over $\mathbb{F}$. In this setting, we can consider one to be “larger” than another if it has larger rank or dimension. See Figure \[fig:definition\] for an example diagram.
Given a graph filtration $\mathcal{X}_G$ and a connected subgraph $G' \subset G$, the $G'$-persistent homology group, $H_k^{G'}(\mathcal{X}_G)$, is the largest rank vector space $\mathcal{P}$ equipped with linear injections $i_v: \mathcal{P}
\to H_k(X_v)$ for all $v \in G'$ such that $f_e \circ i_u = i_v$ for all edges $e = (u,v) \in G'$. In other words, the graph of homology groups along with the inclusions of $\mathcal{P}$ form a commutative diagram.
Note that $H_k^{G'}
(\mathcal{X}_G)$ always exist if the homology groups are finitely generated and are unique up to isomorphism. The maps, however, are not necessarily unique. In Section \[subsec:relations\], we connect this definition to persistence to known models and prove that it generalizes standard persistence, zigzag persistence and multidimensional persistence.
$$\xymatrix@R=1.4pc @C=1.4pc{
& & H_k(X_1) \ar[rr] \ar[dr] & & H_k(X_5) \\
P \ar@{.>}[urr] \ar@{.>}[urrrr]\ar@{.>}[rrr]\ar@{.>}[drr] \ar@{.>}@/_1.2pc/[rrrr] \ar@{.>}[drrr] \ar@{.>}[drrrr]
& & & H_k(X_3) \ar[r] \ar[dr] & H_k(X_6) \\
& & H_k(X_2) \ar[ur] \ar[r] & H_k(X_4) \ar[ur] \ar[r] & H_k(X_7) \\
}$$
Flow graph view {#sec:flow}
---------------
There is also an equivalent and quite interesting view of the structure of the persistent homology group over $G'$, where $G'$ is again a subgraph of the directed acyclic graph $G$. In the definition of $H^{G'}_k(\mathcal{X}_G)$, we assumed that there are maps $i_v$ for every vertex of $G$. However, since the diagram commutes, we only need to define these injections from $P$ to the source vertices of $G'$, and we can then use the edge homomorphisms from $G'$ to construct the remainder of the homomorphisms from $P$ to any vertex in $G'$.
We note that in the setting, all of the composite maps will also be injective. Also, since $\mathcal{P}$ injects in each of the sink vertices, we can define maps from the homology groups at these sinks to $\mathcal{P}$, such that the composite map $\mathcal{P} \to H_k(X_v) \to \mathcal{P}$ is the identity.
In essence, this means that for any subgraph $G'$, the relevant information can be captured by collapsing $G'$ to a directed bipartite graph, where one independent set consists of the sources in $G'$ and the other is the sinks. The choice of edge homomorphisms from a source to a sink in $G'$ is irrelevant due to the commutativity of the diagram, and therefore we can replace $G'$ with a much simpler conceptual picture. (See Figure \[fig:bipartite\] for an example.)
This allows an alternate view of this computation of $P$. Consider collapsing an arbitrary path from each source to each sink of $G'$ into a single edge which is the composition of that path; this results in a bipartite graph $H$ between the sources and sinks of $G'$. Now in order to compute $P$ in this reduced graph $H$, we can view this as a flow computation: we place $P$ as a “source” and a “sink” and turn $H$ into a bipartite flow network where the goal is to maximize the rank of $P$ such that the induced homomorphism from each edge in $H$ is a injection and the composite math from $P$ to $P$ through *any* edge is the identity.
-------------------------------------------------- -------------------------------------------------------------
[@R=1.5pc @C=1.5pc[ & H\_k(X\_1) & & H\_k(X\_5)\ [@R=1.5pc @C=1.5pc[ & H\_k(X\_1) @/\_/\[ddr\] & H\_k(X\_5)\
V & & H\_k(X\_3) & H\_k(X\_6) & W\ V & & H\_k(X\_6) & W\
& H\_k(X\_2) & H\_k(X\_4) & H\_k(X\_7)\ & H\_k(X\_2) & H\_k(X\_7) ]{}]{}
]{}]{}
-------------------------------------------------- -------------------------------------------------------------
Persistence module
------------------
A graph of vector spaces of a fixed directed acyclic graph $G$ form a module structure. For standard persistence this is the same as the persistence module [@module] and also coincides with the zigzag module [@zigzag].
For a directed acyclic graph $G = (V,E)$, a *commutative $G$-module* is the pair $( \{ W_v \}_{v \in V}, \{ f_e \}_{e \in E}$) where $V_v$ is a vector space and for any edge $e = (v,w)$, $f_e: W_v \to W_w$ is a linear map with the condition that the resulting diagram is commutative.
This definition provides the framework for discussing the persistence module for a graph that extends the definition for the zigzag persistence module.
The persistence module for $\mathcal{X}_G$, $\mathcal{PH}_k(\mathcal{X}_G)$, is the commutative $G$-module formed from the graph of homology groups.
The theory for $G$-modules is very similar to the zigzag persistence module. Given a $G$-module $\mathcal{W} = (\{ W_v \}, \{ f_e \}$, a $G$-module $\mathcal{V} = ( \{ V_v \}, \{ g_e \} )$ is a *submodule* if $V_v \subset W_v$ for all $v$ and $f_e|_{V_v} = g_e$. Similarly, given two commutative $G$-modules $\mathcal{V} = ( \{V_v\}, \{f_e\})$ and $\mathcal{W} = ( \{W_v\}, \{g_e\})$, we can define their connected sum $\mathcal{V} \oplus \mathcal{W}$ as $( \{V_v \oplus W_v\}, \{f_e \oplus g_e\})$. A commutative $G$-module is said to be *indecomposable* if it cannot be written as a non-trivial connected sum. Any commutative $G$-module, $\mathcal{V}$, can be written as $\mathcal{V} = \mathcal{V}_1 \oplus \cdots
\oplus \mathcal{V}_n$, where each $\mathcal{V}$ is indecomposable. For a connected subgraph $G'$ of $G$, we will define the commutative $G$-module $\mathbb{F}_{G'}$ as the module with a copy of $\mathbb{F}$ at each vertex of $G'$; we will put the identity map on each edge of $G'$ and make every other map trivial. We will call this module *elementary*.
In representation theory, a directed graph is known as a quiver [@quiver] and a representation of that quiver is an assignment of vector spaces to each vertex and a linear map for each edge. The study of quivers provides the underlying theory for the zigzag persistence module. If we add the conditions that the quiver representation must be a commutative diagram, then we get a quiver with relations. Every finite dimensional algebra occurs as a quiver with relations. The following theorem allows us to decompose our persistence modules.
The decomposition of a commutative $G$-module is unique up to isomorphism and permutation of the summands.
In the case of zigzag persistence the relevant indecomposable modules are always elementary. This is implied by Gabriel’s theorem [@quiver] which provides an enumeration of indecomposable modules for particular graph types. Figure \[fig:indecompossible\] gives an example of an indecomposable module that is not elementary. The example consists of a sphere with four punctures, and the inclusion of each of the four boundary components. Unfortunately, the existence of such examples tells us that there is no simple “barcode” representation for DAG persistence. In Section \[subsec:barcodes\], we will generalize barcodes for our context.
In some special cases, irreducible submodules can be shown to be elementary. In particular, certain submodules carried by subgraphs with a single source and a single sink are elementary. And in practice, many other relevant submodules are also elementary.
If $M$ is an irreducible module of $\mathcal{PH}_k(\mathcal{X}_G)$ carried by a single-source single-sink subgraph $G$ with $H_k^{G'}(\mathcal{X}_G) \neq 0$ then $M \cong \mathbb{F}_{G}$.
Consider $x \neq 0$ in $H_k^{G'}(\mathcal{X}_G)$. We can think of $x$ as an element of $W_v$ for every vertex $v$ of $G$. Consider an edge $e$ from vertex $u$ to $v$, and choose bases $y_1, \ldots, y_k$ and $z_1, \ldots, z_l$ for the vector spaces $W_u$ and $W_v$, respectively. Without loss of generality, assume that $y_1$ and $z_1$ both represent the element $x$. Let $g: W_v \to W_t$ be the composition of the maps from $W_v$ to the sink vertex $t$. Note that $g(y_i) = \alpha_i x
+ b_i$ where each $b_i$ is in the span of the basis vectors of $W_t$ other than $x$. Define a new basis for $W_v$ as $x, y_2 - \alpha_2 x,
\ldots, y_k - \alpha_k x$. Observe that in this new basis the only non-zero entry in the first row and column of $f_e$ is a $1$ in the upper left hand corner. This process can be repeated for each vertex. This allows the decomposition of the module unless there is only a single bases element for each of the $W_v$. Since $M$ is irreducible, this implies that all of the vector spaces are generated by $x$ and all of the maps are the identity. Thus $M \cong \mathbb{F}_G$.
Annotated barcodes {#subsec:barcodes}
------------------
The annotated barcode for an elementary submodule is just the subgraph carrying that submodule. More generally, we will define the *barcode* for an irreducible submodule to be the subgraph where the submodule is non-trivial annotated by the dimension of the submodule. The *dimension* of a module is defined as $$\sum_{v \in V} \dim W_v - \sum_{e \in E} \text{rank} f_e$$
In the case of standard or zigzag persistence, this is equivalent to the traditional definition of a barcode. However, for general DAGs this is not a complete invariant. For a non-elementary submodule, the subgraph and dimension do not completely determine the submodule. For example, in Figure \[fig:indecompossible\] the annotated barcode would be the entire graph annotated with the dimension of the submodule which is 3.
Relationship with other models of persistence {#subsec:relations}
---------------------------------------------
We saw in Section \[sec:def\] the definition for standard persistence. Multidimensional persistence can also be defined in terms of the image of a map. Consider a multifiltration, or $d$-dimensional grid of spaces equipped with a partial ordering on vertices, where $u = (u_1,\ldots, u_d) \le v = (v_1, \ldots, v_d)$ if and only if $u_i \le v_i$ for all $i$. $$\xymatrix@R=1pc @C=1pc{
X_{x} \ar[r] & \cdots\ar[r] & X_{v} \\
\vdots \ar[u] & & \vdots \ar[u] \\
X_{u} \ar[r]\ar[u] & \cdots\ar[r] & X_{y} \ar[u] \\
}$$
The *rank invariant*, $\rho_{X,k}(u,v)$, is defined as the dimension of the image of the induced map $H_k(X_u) \to H_k(X_v)$ [@zigzag]. Since the diagram commutes, this map can be found following any path from $u$ to $v$ in the graph. Zigzag persistence is defined in terms of the zigzag module. The definition of a commutative $G$-module and a $\tau$-module for zigzag persistence are identical for a zigzag graph. The following proposition specifies how DAG persistence generalizes these three notions of persistence.
\[prop:comparison\] Suppose $\mathcal{X}_G$ is a graph filtration of $X$. Then:
1. (Standard persistence) If $G$ is a the graph corresponding to the filtration $X_0 \to X_1 \to \cdots \to X_n$ and $I_{i,p}$ is the subgraph consisting of vertices $\{X_i, \ldots, X_{i+p}\}$ then $H_k^{I_{i,p}}(\mathcal{X}_G) \cong H_k^p(X_i)$. Furthermore, $\mathcal{PH}_k(\mathcal{X}_G)$ coincides with the persistence module.
2. (Zigzag persistence) If $G$ is the graph for zigzag persistence $\mathbb{X} = X_0 \leftrightarrow X_1 \leftrightarrow \cdots \leftrightarrow X_n$ where each arrow could go in either direction then $H_k(\mathbb{X}) \cong \mathcal{PH}_k(\mathcal{X}_G)$.
3. (Multidimensional persistence) If $\mathcal{X} = \{ X_v \}_{v \in \{0,\ldots,m\}^d}$ is a multifiltration with underlying graph $G$. If $G_{u,v}$ is the subgraph with vertices $\{ w \in G \ | \ u \le w \le v \}$ then the rank invariant $\rho_{X,k}(u,v) = \dim H_k^{G_{u,v}}(\mathcal{X}_G)$.
First consider standard persistence and the relevant portion of the filtration of the homology groups: $H_k(X_i) \to \cdots \to H_k(X_{i+p})$ and let $i_*: H_k(X_i) \to H_k(X_{i+p})$ be the composition of the maps. Since $H_k^p(X_i) = im(i_*)$ is a subgroup of $H_k(X_{i+p})$ there is a natural inclusion $g: H_k^p(X_i) \to H_k(X_{i+p})$. Since $i_*$ surjects on the persistence group there exists an injection. $f: H_k^p(X_i) \to H_k(X_i)$ such that $i_* \circ f = g$. This shows us that $H_k^{I_{i,p}}(\mathcal{X}_G)$ is at least as large as $H_k^p(X_i)$. And it cannot be larger since any map in $H_k(X_i)$ that realizes persistence must have its image in $im(i_*)$ which is equal to $H_k^p(X_i)$.
The proof for multidimensional persistence is almost identical as the one for standard persistence since the rank invariant is defined as the rank of the image of any path from $H_k(X_u)$ to $H_k(X_v)$. Note that for zigzag persistence on a graph, the definition of a commutative $G$-module coincides with the definition of the zigzag module, so the statement follows immediately.
An example {#subsec:ex}
----------
In Figure \[fig:example\] we see an example of a set of spaces with inclusions that form a directed acyclic graph. At the top level is a genus two surface; the directed arrows indicate the inclusion maps in our directed acyclic graph, down to our two source vertices in the graph which include one space with two disjoint annuli and one space that is a disk with three boundaries. The graph forms a poset that demonstrates the non-trivial intersections and unions of the three surfaces. In Figure \[fig:example\], middle, we see the persistence module for the entire space, and on the left are diagrams showing the indecomposable summands of the module. Notice that each of these submodules is elementary yielding barcodes without annotation. From these indecomposables it is possible to read off the persistence for any subgraph $G'$ by counting how many of the elementary modules have $G'$ as a subgraph.
---------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------
$\vcenter{\hbox{\includegraphics[width=2in]{figures/example-curves.pdf}}}$ $\vcenter{\hbox{
{\xymatrix@R=2pc @C=2pc{
& \mathbb{F}^4 & \\
\mathbb{F}^3 \ar[ur] & & \mathbb{F}^3 \ar[ul] \\
\mathbb{F} \ar[u] & \mathbb{F}^3 \ar[ul]\ar[ur] & \mathbb{F}^2 \ar[u] \\
\mathbb{F}^2 \ar[u]\ar[ur] & & \mathbb{F}^2 \ar[ul]\ar[u]
}}}}$
$\vcenter{\hbox{\begin{tabular}{cc} $\vcenter{\hbox{\includegraphics[width=1.6in]{figures/example-barcodes.pdf}}}$
{\tiny\xymatrix@R=.75pc @C=.75pc{
& \mathbb{F} & \\
\mathbb{F} \ar[ur] & & \mathbb{F} \ar[ul] \\
\mathbb{F} \ar[u] & \mathbb{F} \ar[ul]\ar[ur] & {\color{gray} 0} \ar@{.}\ar@{.}[u] \\
\mathbb{F} \ar[u]\ar[ur] & & {\color{gray} 0} \ar@{.}[ul]\ar@{.}[u]
}} &
{\tiny\xymatrix@R=.75pc @C=.75pc{
& \mathbb{F} & \\
\mathbb{F} \ar[ur] & & \mathbb{F} \ar[ul] \\
{\color{gray} 0} \ar@{.}[u] & \mathbb{F} \ar[ul]\ar[ur] & \mathbb{F} \ar[u] \\
{\color{gray} 0} \ar@{.}[u]\ar@{.}[ur] & & \mathbb{F} \ar[ul]\ar[u]
}} \\
{\tiny\xymatrix@R=.75pc @C=.75pc{
& \mathbb{F} & \\
\mathbb{F} \ar[ur] & & {\color{gray} 0} \ar@{.}[ul] \\
{\color{gray} 0} \ar@{.}[u] & {\color{gray} 0} \ar@{.}[ul]\ar@{.}[ur] & {\color{gray} 0} \ar@{.}[u] \\
{\color{gray} 0} \ar@{.}[u]\ar@{.}[ur] & & {\color{gray} 0} \ar@{.}[ul]\ar@{.}[u]
}} &
{\tiny\xymatrix@R=.75pc @C=.75pc{
& \mathbb{F} & \\
{\color{gray} 0} \ar@{.}[ur] & & \mathbb{F} \ar[ul] \\
{\color{gray} 0} \ar@{.}[u] & {\color{gray} 0} \ar@{.}[ul]\ar@{.}[ur] & \mathbb{F} \ar[u] \\
{\color{gray} 0} \ar@{.}[u]\ar@{.}[ur] & & {\color{gray} 0} \ar@{.}[ul]\ar@{.}[u]
}} \\
{\tiny\xymatrix@R=.75pc @C=.75pc{
& {\color{gray} 0} & \\
{\color{gray} 0} \ar@{.}[ur] & & {\color{gray} 0} \ar@{.}[ul] \\
{\color{gray} 0} \ar@{.}[u] & \mathbb{F} \ar@{.}[ul]\ar@{.}[ur] & {\color{gray} 0} \ar@{.}[u] \\
\mathbb{F} \ar@{.}[u]\ar[ur] & & \mathbb{F} \ar[ul]\ar@{.}[u]
}}
\end{tabular}}}$
---------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------
Single Source-Single Sink Subgraphs {#sec:single}
===================================
In this section, we consider $\mathcal{X}_G$ where $G$ is a directed acyclic graph with a single source vertex $s$. To limit the number of cells which $f_e$ can introduce (where $e = (u,v) \in G$), we assume that each inclusion $f_e$ adds a single cell to the underlying space. As a result, we will also assume that the space $X_s$ will consist of the empty set. This assumption is both standard in most persistence algorithms [@zigzag-mm; @persistencealgorithm] and quite natural given that we can decompose any inclusion map into a series of inclusions of one simplex at a time.
First, note if we consider a single source-single sink subgraph, it suffices to take any directed path between them in $G$ and compose the maps, since the persistent homology diagram commutes. Therefore, a straightforward application of the $O(n^3)$ standard persistence algorithm would yield a $O(n^5)$ algorithm. Using tools from recent work to compute zigzag persistence in matrix multiple time [@zigzag-mm] would give a running time of $O(n^2 (M(l))+l^2 \log^2 l))$, where $l$ is the length of the longest path between any source and sink and $M(l)$ is the time to multiply two $l \times l$ matrices.
\[thm:SSSS\] If $G$ is a directed acyclic graph whose longest path has length $l$, then the persistent homology groups with coefficients from a finite field for all single source-single sink subgraphs of $G$ can be calculated in $O(v^2 l^2)$ time.
We note that if we remove the assumption that coefficients are from a finite field, our theorem still accurately counts the number of arithmetic operations, although running times for each operation might take longer. In 3 dimensions, however, no information is lost when persistent homology is taken with respect to a finite field [@persistenthomotopy], so this is not an overly restrictive assumption.
This algorithm is actually a relatively straightforward extension of a standard persistence algorithm over finite fields [@persistencealgorithm]. In this paper they initially represent the homology basis as a matrix, but they are able to eliminate the need for row operations. This results in a simpler data structure, namely an array where each simplex is annotated and marked appropriately (as basis elements are created and destroyed) during the course of the algorithm. The running time is equivalent to Gaussian elimination, or $O(m^3)$ where $m$ is the number of simplices in the filtration.
We will rely on their matrix representation (rather than on their improved data structure) for our algorithm. We first fix a single source vertex $s$, and compute a (directed) tree $T$ which reaches all possible sinks in $G$. If our underlying tree is a path, then we can directly apply their algorithm in $O(l^3)$ time. In the more general case where it is not a path, however, we must consider the matrix representation more carefully.
In general, the homology basis at dimension $k$ will be a matrix with each $(k+1)$-simplex represented as a column and each $k$-simplex represented as a row. While in their algorithm this resulted in a running time of $O(n^2)$ for each operation (due to the Gaussian elimination step), we note that in our algorithm, each such operation is bounded by $O(l^2)$, since the fact that $G'$ only has a path of length $l$ means that our matrix has size at most $l \times l$.
We can also store a representation of the current matrix at each branch point in $T$, which means we will not need to repeat our calculation starting at the source each time. Since we have at most $n$ edges in $T$ and adding each edge triggers an $O(l^2)$ computation, we take overall $O(n l^2)$ time to compute each path from a single source to all sinks. Repeating this for each source vertex yields the running time in Theorem \[thm:SSSS\]. We do note that all of the matrices involved are sparse, as in [@persistencealgorithm], so we expect the running time in this setting to be faster in practice.
We also note that this algorithm can be adapted to multidimensional persistence to give an improvement over the known polynomial time algorithm [@computingmdp]:
\[prop:lattice\] All of the rank invariants for $d$-dimension lattice with $n$ nodes can be calculated in $O(n^{4-1/d})$ time.
The improvement in the run-time for lattices is that we do not need to run the tree based algorithm for every vertex of the graph. Instead we can choose trees with roots at the vertices $X_{(0,v_2,\ldots,v_d)}$ that include all of the paths of the form $X_{(0,v_2,\ldots,v_d)} \to X_{(1,v_2,\ldots,v_d)} \to \cdots \to X_{(m,v_2,\ldots,v_d)}$. Every pair of vertices that can be connected in the lattice by a directed path can also be connected in one of these trees. There are $n^\frac{d-1}{d} = n^{1-1/d}$ such trees, yielding the improved running time.
General Subgraphs with Rational Coefficients {#sec:algorithm}
============================================
There are exponentially many subgraphs of a graph $G$, so it is unreasonable to expect to be able to calculate persistent homology for all subgraphs in an efficient manner. However, in this section we will given an algorithm calculate $H_k^{G'}(\mathcal{X}_G)$ for fixed $G'$ in $O(n^4)$ time.
Our algorithm is based on the Gram-Schmidt orthonormalization procedure [@gs]. Gram-Schmidt does not work if $\mathbb{F}$ is a finite field, since we do not have an inner product space in finite fields. Instead, we will work with rational coefficients for our homology groups. If we are working with datasets in $\mathbb{R}^3$, then all homology groups are torsion-free and the ranks of all of the persistent homology groups are the same if we have rational coefficients or coefficients in a finite field [@persistenthomotopy]. So there is no information lost when rational coefficients are used in $3$-dimensions. (In the next section we will present a slower algorithm that will allow calculation with coefficients in a finite field.)
At a high level, our algorithm is simply an inductive one that adds each simplex to the subgraph one at a time. At each stage, we will maintain the persistent homology for the current subgraph; however, this requires that we store a matrix representing the boundary, chain, and cycle groups at each vertex so that we can track changes to the persistent homology group across the entire graph as each edge is added. The end result is the following:
\[thm:generalalg\] $H_k^{G'}(\mathcal{X_G}; \mathbb{Q})$ can be found using $O(e l^3) = O(n^4)$ arithmetic operations.
The running time for this algorithm could actually be much worse since we are not working in a finite field, so there is the potential of an exponential growth in the numbers involved in these calculations. Ideally, we would work in a finite field to avoid this issue, but as previously noted the Gram-Schmidt process will not work in finite fields.
Let $G_0, \ldots, G_m$ be a sequence of subgraphs of $G'$ where $G_0$ consists only of the vertices of $G'$, $G_m = G'$ and $G_{a+1}$ is $G_a$ with a single additional edge. We will calculate the persistent homology of $G_a$ inductively.
For each vertex $i$ of $G'$, we start with an $s \times s$ matrix $M_i$, where $s$ is the number of $k$ simplices of $X_i$. The columns of $M_i$ will be partitioned into three subset: $M_i = \left( B_i \ | \ Z_i \ | \ C_i \right)$. The rows of $M_i$ correspond to the $k$-simplices of $X_i$. Initially, the columns in $B_i$ span the boundaries $B_k(X_i)$ and the the cycle space $Z_k(X_i)$ is spanned by the combination of the columns of $Z_i$ and $B_i$. The entries of $C_i$ fill out a basis of $\mathbb{Q}^s$. Let $b_i$, $z_i$ and $c_i$ be the number of columns in each portion of $M_i$. These entries can be found by calculating the homology groups $H_k(X_i)$ in $O(M(s))$ time [@mm-homology]. We will abuse notation slightly and use $B_i$ and $Z_i$ to represent both the specified columns of the matrix and the subspace spanned by those columns. Similarly, we use $Z_i+B_i$ to denote the subspace spanned by the union of $Z_i$ and $B_i$.
The first step of the algorithm will be to use Gram-Schmidt to turn $M_i$ into an orthonormal matrix, which is a matrix with the norm of each column equal to one and the dot product between distinct columns equal to zero. Let $(M_i)_c$ denote the $c$-th column of $M_i$. The orthonormalization procedure proceeds as follows:
1. Replace $(M_i)_1$ with $\frac{(M_i)_1}{||(M_i)_1||}$
2. For $j=2..s$, let $v = (M_i)_j - \sum_{l=1}^{j-1} ((M_i)_j \cdot (M_i)_l) \ (M_i)_l$ and replace $(M_i)_j$ with $\frac{v}{||v||}$.
This process does not change the span of the first $j$ columns. So the first $b_i$ columns still span $B_k(X_i)$, and the first $b_i + z_i$ columns span $Z_k(X_i)$.
The invariants we will maintain through this inductive algorithm are:
- For every component $C$ of $G_a$ and vertex $i \in C$, $H_k^C(\mathcal{X}_G) \cong (Z_i+B_i)/B_i$.
- For a component $C$ of $G_a$, the possible maps $\{ g_i : H_k^C(\mathcal{X}_G) \to C_k(X_i) \}$ that realize persistence for the graph $C$ are precisely the ones with $im(g_i) \subset Z_i+B_i$ and $im(g_i) \cap B_i = 0$ and satisfy the commutativity assumptions $f_e \circ g_i = g_j$ for every edge $e=(i,j)$ of $C$.
- For any edge $e = (i,j)$ of $G_a$, we have $f_e(Z_i+B_i) = Z_j + B_j$ and $f_e^{-1}(B_j) = B_j$.
- $M_i$ is orthonormal.
As edges are added to the subgraphs, the groups $Z_i+B_i$ will shrink and the group $B_i$ will grow, but these invariants will be maintained.
Consider adding an edge $e$ connecting vertices $i$ and $j$, and consider injections that realize persistence at both vertices and yield a commutative diagram with the following conditions:
---------------------------------------- --------------------------- --
[$\vcenter{\hbox{\xymatrix{ $\begin{array}{l}
& V \ar[dl]_{g_i} \ar[dr]^{g_j} & \\ g_i(V) \subset B_i+Z_i \\
C_k(X_i) \ar[rr]^{f_e} & & C_k(X_j) \\ g_i(V) \cap Z_i = 0 \\
}}}$]{} g_j(V) \subset B_j+Z_j \\
g_j(V) \cap Z_j = 0 \\
\end{array}$
---------------------------------------- --------------------------- --
To realize persistence for the components of the graph $G_a$, the above conditions must be met. Any realization of persistence for the larger graph, $G_{a+1}$, must also satisfy a few additional properties:
1. For the diagram to commute, $g_j(V)$ must be completely contained in the image $f_e(Z_i+B_j)$ which can be guaranteed if we replace $Z_j+B_j$ with $(Z_j+B_j) \cap f_e(Z_i+B_i)$.
2. Also, $g_i(V)$ must be contained in $f_e^{-1}(Z_j+B_j)$. To ensure this we will replace $Z_i+B_i$ with $(Z_i+B_i) \cap f_e^{-1}(Z_j+B_j)$.
3. If $g_i(V)$ intersects $f_e^{-1}(B_j)$ non-trivially then the composition $g_j = f_e \circ g_i$ intersections $B_j$ non-trivially. Thus we must ensure that $g_i(V) \cap f_e^{-1}(B_j) = 0$ so we will replace $B_i$ with $B_i \cup f_e^{-1}(B_j)$.
4. Finally, if $g_i(V) \cap B_i = 0$ then $g_j(V) \cap f_e(B_i) = 0$, so $g_j(V)$ must meet $f_e(B_i)$ trivially. So we will replace $B_j$ with $B_j \cup f_e(B_i)$.
Notice that after these changes $f_e(Z_i+B_i) = Z_j+B_j$ and $f_e^{-1}(B_j) = B_i$. However, this property is no longer guaranteed at all of the other edges. To fix this, we will traverse the graph identifying deficiencies. At any such edge $e=(i,j)$ we will make the following replacements:
1. Replace $Z_j+B_j$ with $(Z_j+B_j) \cap f_e(Z_i+B_i)$.
2. Replace $B_i$ with $B_i \cup f_e^{-1}(B_j)$.
This process must terminate since any time changes are made either the dimension of some $Z_i+B_i$ or $Z_i$ decreases; we bound the exact number of iterations later when discussing the specifics of the algorithm.
The following lemma states that the invariants are satisfied through the process.
\[lem:invariants\] The invariants are satisfied for the graph $G_{a+1}$ if they are met for $G_a$.
Before making changes to the matrices $M_i$ and $M_j$ we have a full characterization of all maps realizing persistence for the components of $G_a$. The steps we made to change these two matrices yield the overlap in the options. This yields the largest group that injects into both. So, after these changes the new matrices $M_i$ and $M_j$ completely characterize the maps realizing persistence for the components of $G_{a+1}$. The propagation process ensures that any choice of realization at one vertex can be extended to the adjacent vertex. Thus, after the propagation is complete, the invariants are satisfied at all of the vertices.
To implement the algorithm, we must perform matrix operations on each $M_i$. Our basic operations are calculating intersections and the union of subspaces; we also note that unions are the same as doing intersections of perpendicular subspaces.
As a map on chain complexes, $f_e$ is either the identity map (when there is a death of a $k$-cycle) or its image has codimension one (when there is a birth of a cycle). In the second case, the coordinate corresponding to the new simplex is always zero. Because of this, it is natural to think of chains in $C_k(X_i)$ also as chains in $C_k(X_j)$. Furthermore, chains in $C_k(X_j)$ that are zero in the extra coordinate belong to $C_k(X_i)$. We will denote the coordinate vector for this extra coordinate by $u_{ij}$. With this perspective, we can summarize our operations as follows:
--------------------------------------------------------- ------------------------------------------------------
$Z_j+B_j \Rightarrow (Z_j+B_j) \cap f_e(Z_i+B_i)$) Add $u_{ij}$ and each column of $C_i$ to $C_j$
and re-orthonormalize
$Z_i+B_i \Rightarrow (Z_i+B_i) \cap f_e^{-1}(Z_j+B_j)$) Add each column of $C_j$ projected to $u_{ij}^\perp$
to $C_i$ and re-orthonormalize
$B_i \Rightarrow B_i \cup f_e^{-1}(B_j)$ Add each column of $B_j$ to $B_i$ and
re-orthonormalize
$B_j \Rightarrow B_j \cup f_e(B_i)$ Add each column of $B_i$ to $B_j$ and
re-orthonormalize
--------------------------------------------------------- ------------------------------------------------------
Notice instead of shrinking $Z_i+B_i$, we perform the equivalent operation of enlarging its perpendicular subspace $(Z_i+B_i)^\perp = C_i$. The process of adding a column $v$ to $C_i$ proceeds like a step of Gram-Schmidt. First, project $v$ to a vector, $\bar{v}$, that is perpendicular to the rest of $C_i$. If $\bar{v}$ is non-zero, it is added to $C_i$. At this point $M_i$ is no longer square. However, the vectors in $B_i$ and then $Z_i$ can be reduced to be perpendicular to $\bar{v}$. At some point, one of these vectors will be zero and can be removed from $M_i$. Adding a vector to $B_i$ works in the same manner. Since the matrices involved are at most $l \times l$, this single update can be performed in $O(l^2)$ arithmetic operations.
Each update either increases $c_i$ or increases $z_i$, and we will avoid adding the same column twice. So, at most $O(l)$ updates can occur for each matrix. Each update that occurs for a matrix triggers potential updates for all of its neighbors. If $d_i$ is the degree of that vertex then it takes $O(d_i l^3)$ operations to do all of the updates to vertex $i$ and perform checks on each neighbor to see if they must be updated. So the total number of arithmetic operations is $\sum O(d_i l^3) = O(e l^3)$, as claimed in Theorem \[thm:generalalg\].
Persistence with Coefficients in a Finite Field {#sec:finitefield}
===============================================
As previously mentioned, the techniques used in Section \[sec:algorithm\] will not work when the coefficients are from a finite field. However, our algorithm can be modified to work in more general situations. The basic steps of the algorithm involve taking unions and intersections of subspaces; the use of Gram-Schmidt orthonormalization allows these operations to be done efficiently with rational coefficients. However, it is possible to do such intersections and unions in a finite field at the cost of a slower runtime, yielding the $O(n^5)$ algorithm that is our main result in this section.
Calculating the union of two subspaces is relatively straightforward. Assume that each subspace is represented by a matrix in row echelon form. The concatentation of the two matrices can be put in row echelon form using Gaussian elimination in $O(m^2 n)$ time where the original matrices have $m$ rows and $n$ columns.
To calculate the intersection two subspaces spanned by the rows of the matricies $M$ and $N$ respectively, consider solving the following linear system: $$\left( \alpha \beta \right) \left( \begin{array}{c} M \\ N \end{array} \right) = 0$$ The intersection of the two subspaces is spanned by the vectors $\alpha_1 M, \ldots, \alpha_k M$. Putting these vectors in a matrix and then reducing it to row echelon form yields a representation of the intersection. This can be done in cubic time in the size of the matrices.
As in the previous section, there are a quadratic number of updates to the subspaces, yielding the following theorem:
For any finite field $\mathbb{F}$, $H_k^{G'}(\mathcal{X_G}; \mathbb{F})$ can be found in $O(n^5)$ time.
Applications {#sec:applications}
============
Estimating Persistence Using Multiple Subsamples
------------------------------------------------
The first application is estimating persistence for a point sample using a pair of much smaller subsamples. Let $X_0 \to X_1 \to \cdots \to X_n$ and $Y_0 \to Y_1 \to \cdots \to Y_n$ be filtrations for union of balls of various radii for two subsamples of a common point set. Moreover, we assume that each of the $X_i$ is contained in some $Y_j$ and vice-versa. This yields a directed graph of the form shown in Figure \[fig:parallel\_graph\].
In Figure \[fig:parallel\], we show a compelling example of the utility of our algorithm. Each subfigure is a persistence diagram, where each cycle is considered as a pair of birth and death times. The persistence of each cycle is the difference in these times which is equal to the distance to the diagonal. The space considered is genus two surface sampled with 5000 points, at which level the persistent features, which are the 4 generators of homology, are clearly seen as significant. Note that these four points blur together in pairs in the figure and are difficult to distinguish from each other. In the remaining pictures, we calculate persistent homology for a simple directed acyclic graph that consists of two different 200 point subsamples including into a larger 400 point sample. We note that individually, each sample’s persistent homology is quite noisy and does not indicate the 4 generators at all. However, the persistent homology for the directed acyclic graph (shown on the top right) clearly separates the 4 main generators from the noise, at a far lower level of sampling than is possible with standard persistent homology. The table shows the persistence values for the four generating cycles for the surface and the next largest cycle. The values have been renormalized to the length of the cycle with largest persistence in the original sample.
Additional possible applications of this are numerous. For example, we could use directed acyclic graphs of various witness complexes and seek improved results using a DAG over a small set of witness complexes; recent work using zigzag persistence seems relevant in this setting [@zigzagapplications]. Also, the example in Section \[subsec:ex\] demonstrates how homology classes from pieces of a space can be “aligned” similar to the bootstrapping method of [@zigzagapplications].
----------------------------------------------------------------------------------------------------------------
[@R=.75pc @C=.75pc[ X\_0 & X\_1 @[.>]{}\[rr\] & & X\_i @[.>]{}\[rr\] & & X\_j @[.>]{}\[rr\] & & X\_n\
\
Y\_0 & Y\_1 @[.>]{}\[rr\] & & Y\_i @[.>]{}\[rr\] & & Y\_j @[.>]{}\[rr\] & & Y\_n ]{}]{}
----------------------------------------------------------------------------------------------------------------
[ccc]{} Full (5000 points) & First half (200 points) & Parallel\
$\vcenter{\hbox{\includegraphics[width=1.35in]{figures/big.pdf}}}$ & $\vcenter{\hbox{\includegraphics[width=1.35in]{figures/half1.pdf}}}$ & $\vcenter{\hbox{\includegraphics[width=1.35in]{figures/parallel.pdf}}}$\
& &\
Subsample (400 points) & Second half (200 points) & Persistence values\
$\vcenter{\hbox{\includegraphics[width=1.35in]{figures/full.pdf}}}$ & $\vcenter{\hbox{\includegraphics[width=1.35in]{figures/half2.pdf}}}$ &
Top 4 Fifth
----------- ------------ -------
Full 80.4–100.0 12.8
Subsample 78.8–88.7 72.1
1st half 72.8–78.9 68.2
2nd half 75.2–78.5 72.1
Parallel 72.8–77.6 37.1
------------------------------------------------------
[@R=.75pc @C=.75pc[ & & & &\
& & & &\
& & X\_2 Y\_2 @[.>]{}\[uu\] &\
X\_2 @[.>]{}\[uuu\] & & & Y\_2 @[.>]{}\[uuu\]\
& X\_2 Y\_2 @[.>]{}\[uuuu\] & &\
& & & &\
& & X\_1Y\_1 &\
X\_1 & & & Y\_1\
& X\_1 Y\_1 & &\
& & & &\
& & X\_0Y\_0 &\
X\_0 & & & Y\_0\
& X\_0 Y\_0 & &\
]{}]{}
------------------------------------------------------
Shape comparison
----------------
Given filtrations of two overlapping shapes, a natural question is to measure how similar they are. DAG persistence can provide a method for comparison. Consider the graph in Figure \[fig:comparison\]; it includes fitrations for two shapes $X$ and $Y$ as well as their intersection and union. If $X$ and $Y$ are very similar and well aligned then this would be detected in the annotated barcodes.
This comparison is made by building persistence diagrams for each filtration $\{ X_i \}$ and $\{ Y_i \}$ using standard persistence. A persistence diagram is built for the module over $G$ by first calculating its annotated barcodes. Each barcode is converted to a point $(i,j)$ in the persistence diagram if $i$ is the smallest index and $j$ is the largest index of the vertices of the subgraph carrying the barcode. The multiplicity of this point is equal to the annotation of the barcode.
This comparison was performed on two 1000 point subsamples from the dataset in the previous example (see Figure \[fig:parallel\]). The bottleneck distance distance between the persistence diagrams of each of the two samples and the persistence diagram for the comparison graph were both under 5% of the lifespan of the 4 significant topology features of the shapes. This demonstrates that the two shapes being compared have nearly identical topological features except on a very small scale.
Future Work {#sec:future}
===========
A practical algorithm for decomposition the persistence module into indecomposibles would be useful in finding an annotated barcode representation of the persistence module. Such an algorithm could generalize the right filtrations used in [@zigzag; @zigzag-mm].
Recent work has focused on developing parallel frameworks for persistence [@DBLP:journals/corr/BauerKR13]. A second practical application of our formulation would be using the type of splitting shown is Section \[subsec:ex\] as a basis for a divide and conquer algorithm that allows parallel computation of standard persistence. For example, in Figure \[fig:example\], we can calculate the persistence of the entire space by combining the computation for the subspaces $X,Y,$ and $Z$, which can be done in parallel.
There is the potential to improve our algorithm for all single source-single sink subgraphs using tools from recent work that computes zigzag homology in matrix multiply time [@zigzag-mm]. Consider a directed tree $T$ in the DAG for a single sink $s$; this tree is composed of paths which are zigzag homology complexes, so we can use the recent zigzag algorithm for each path. If this tree has common subpaths from previous calls to the zigzag algorithm, we could potentially speed up our algorithm by using this information. On the other hand, if the tree has no such common subpaths, then intuitively we should be able to balance the fact that the disjoint paths sum to $n$ (so that we either have many very short paths or the tree is a single path). Balancing this recursion based on their analysis, however, give no better running time than the naive $O(n^2 M(l))$ algorithm we briefly outlined in Section \[sec:single\], since the time to combine the information from a previous call with the new recursive call will take longer than the call itself. A better algorithm that uses this information successfully is an interesting direction to consider.
Another potential for algorithm improvement follows for our flow characterization in Section \[sec:flow\]. It is unclear how to adapt standard flow algorithm to the computation of this group $P$; nonetheless, the picture is quite compelling, and the literature on maximum flow computations has many algorithms and techniques. It will be interesting to examine this connection more carefully to see if max flow techniques could provide additional insight or improvements.
Acknowledgements
================
The authors would like to thank Afra Zomorodian for suggesting this problem during a visit, as well as for additional comments and suggestions along the way. We would also like to thank Greg Marks and Michael May for helpful conversions during the course of this work.
[^1]: Department of Mathematics and Computer Science, Saint Louis University, {echambe5,letscher}@slu.edu. Research supported in part by the National Science Foundation under Grant No. CCF 1054779 and IIS-1319573.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Max Pettini
- 'Sara L. Ellison'
- Jacqueline Bergeron
- Patrick Petitjean
date: 'Received / Accepted'
title: |
The Abundances of Nitrogen and Oxygen in\
Damped [Lyman $\alpha$]{} Systems
---
Introduction
============
The nucleosynthetic origin of nitrogen continues to be a subject of considerable interest and discussion. There is general agreement that the main pathway is a six step process in the CN branch of the CNO cycle which takes place in the stellar H burning layer, with the net result that $^{14}$N is synthesised from $^{12}$C and $^{16}$O. The continuing debate, however, centres on which range of stellar masses is responsible for the bulk of the nitrogen production. A comprehensive reappraisal of the problem was presented by Henry, Edmunds, & Köppen (2000) who compiled an extensive set of abundance measurements and computed chemical evolution models using published yields. Briefly, nitrogen has both a primary and a secondary component, depending on whether the seed carbon and oxygen are those manufactured by the star during helium burning, or were already present when the star first condensed out of the interstellar medium (ISM).
[lcccccccc]{} QSO & V & $z_{\rm em}$ & $z_{\rm abs}$ & $N$(H I)$^{a}$ & Exp. Time & Resolution & Wavelength Range & S/N$^{b}$\
& (mag) & & & (cm$^{-2}$) & (s) & (km s$^{-1}$) & (Å) &\
\
Q1409$+$095 & 18.6 & 2.856 & 2.45620 & $3.5 \times 10^{20}$ & 16200 & 7 & 3793–4989, 6726–10255$^{c}$ & 15–25\
Q1444$+$014 & 18.5 & 2.206 & 2.08681 & $1.6 \times 10^{20}$ & 18000 & 7 & 3295–6650$^{d}$ & 10–20\
Q2206$-$199 & 17.33 & 2.559 & 2.07623 & $2.7 \times 10^{20}$ & 24300 & 7 & 3295–10255$^{d}$ & 20–25\
\
Determined by profile fitting to the damping wings of the [Lyman $\alpha$]{} absorption line (see section 2 and Figure 1). Typical error in $N$(H I) is $\pm 10\%$.
The signal-to-noise ratio varies along each spectrum. These values (per pixel) refer to the QSO continuum in the regions of the N I $\lambda\lambda 1200$ and O I $\lambda 1302.2$ lines.
With a wavelength gap between 8520 and 8667 Å.
With some wavelength gaps.
Nitrogen and Oxygen in H II regions
-----------------------------------
Observational evidence for this dual nature of nitrogen is provided mainly from measurements of the N and O abundances in H II regions. (For simplicity in this paper we use parentheses to indicate logarithmic ratios of number densities; adopting the recent reappraisal of solar photospheric abundances by Holweger (2001), we have (N/H)$_{\odot} = -4.07$; (O/H)$_{\odot} = -3.26$; and (N/O)$_{\odot} = -0.81$). In H II regions of nearby galaxies, (N/O) exhibits a strong dependence on (O/H) when the latter is greater than $\sim 2/5$ solar; this is generally interpreted as the regime where secondary N becomes dominant.[^1] At low metallicities on the other hand, when (O/H)$\,\simlt -4.0$ (that is, $\simlt 1/5$ solar), N is mostly primary and tracks O; this results in a plateau at (N/O) $\simeq -1.5$.
The principal sources of primary N are thought to be intermediate mass stars ($4 \simlt M/M_{\odot} \simlt 7$) during the asymptotic giant branch (AGB) phase. Henry et al. (2000) showed that, integrating along the IMF the N yields from intermediate mass stars by van den Hoek & Groenewegen (1997) and the O yields from massive stars by Maeder (1992) at metallicity $Z = 1/20 Z_{\odot}$, one obtains (N/O)$ = -1.41$ in good agreement with the observed plateau.
A corollary of the hypothesis that intermediate mass stars are the main producers of primary nitrogen is that its release in the ISM should lag behind that of O, since the latter is widely believed to be produced by massive stars which explode as Type II supernovae (SN) soon after an episode of star formation. Henry et al. (2000) calculated this time delay to be approximately 250Myr; at low metallicities the (N/O) ratio could then perhaps be used as a clock with which to measure the past rate of star formation, as proposed by Edmunds & Pagel (1978). Specifically, in metal-poor galaxies which have only recently experienced a burst of star formation one may expect to find values of (N/O) [*below*]{} the primary plateau at (N/O) $\simeq -1.5$.
This scenario has been discussed at length over the last few years with claims of both conflicting and supporting evidence from the data. Izotov & Thuan (1999) and Izotov et al. (2001) were struck by the constant values of all the element ratios, including (N/O), which they measured in blue compact dwarf (BCD) galaxies with (O/H)$ \simlt -4.4$ (corresponding to an oxygen abundance less than $\sim$1/15 of solar). This finding led them to the radical proposal that (a) these are galaxies undergoing their first burst of star formation, and (b) all the elements they observed have a primary origin in massive stars, so that they are released at the same time as O, thereby disposing altogether of the notion of a time delay. On the other hand, a variety of studies using high resolution [*Hubble Space Telescope*]{} ([*HST*]{}) images have shown that many of these BCDs host old stellar populations, with ages greater than 1Gyr (e.g. Schulte-Ladbeck et al. 2001 and references therein; Crone et al. 2002). Generally, these low mass galaxies tend to have low rates of star formation (even though they are observed during a bursting phase). If the lag between O and N production is only 250 Myr, as proposed by Henry et al. (2000), it is perhaps not surprising that most of them lie near the primary value of the (N/O) ratio expected from intermediate mass stars (Pilyugin 1999).
In contrast, a recent survey by Contini et al. (2002) found that UV-selected galaxies at intermediate redshifts (0$<$[*z*]{}$<$0.4) exhibit a wide range of (N/O) values at a given (O/H), in many cases well below the primary level. While Contini et al. interpret their results as evidence in favour of a delayed production of primary N, one may well wonder at the high fraction of galaxies apparently caught within 250Myr since the last major episode of oxygen enrichment. Possibly the UV-continuum selection technique strongly favours galaxies with high rates of star formation in the recent past.
Nitrogen and Oxygen in QSO Absorption Line Systems
--------------------------------------------------
As pointed out by Pettini, Lipman, & Hunstead (1995), clues to the nucleosynthetic origin of nitrogen can also be provided by QSO absorption line systems, particularly the high column density ($N$(H I) $\geq 2 \times 10^{20}~$cm$^{-2}$) damped [Lyman $\alpha$]{} absorbers (DLAs). These are thought to represent an early stage in the evolution of galaxies, when most of their baryonic mass was in the interstellar medium. Apart from the obvious interest in measuring element abundances in the distant past, when galaxies were young, one of the advantages of DLAs is that they are generally of low metallicity, approximately between $1/10$ and $1/100$ solar (Pettini et al. 1999; Prochaska & Wolfe 2002). Thus, they probe a regime where local H II region abundance measurements are sparse or non-existent and where the effect of a delayed production of primary nitrogen should be most pronounced.
There are practical difficulties, however, in measuring (N/O) in DLAs. The resonance lines most easily accessible, the N I triplet $\lambda\lambda 1199.5, 1200.2, 1200.7$ and O I $\lambda 1302.2$, have very different optical depths. While the latter is normally saturated in DLAs, the former is weak and can be difficult to detect, particularly as it falls within the [Lyman $\alpha$]{} forest. Pettini et al. (1995), and subsequently Lu, Sargent, & Barlow (1998) and Centurión et al. (1998), attempted to circumvent the first problem by using the abundances of other $\alpha$-capture elements, such as Si and S, as proxies for (O/H). These studies found that the (N/$\alpha$) ratio in DLAs spans a wide range of values, broadly within the range bracketed by the predictions for primary and secondary production of nitrogen. However, these results have been criticized by Izotov & Thuan (1999) and more recently Izotov, Schaerer, & Charbonnel (2001). Unlike O and N, which are mostly neutral in H I gas, Si and S are singly ionised. Since in principle Si II and S II can also occur in H II regions, Izotov and collaborators have proposed that unaccounted ionisation corrections, rather than real abundance variations, are the cause of the observed scatter of the (N/$\alpha$) ratios in DLAs. Their model, however, makes the radical assumption that essentially all the metals are located in (presumably self-enriched) H II regions, and that the metallicity of the neutral phase is negligible.
Motivations for the Present Study
---------------------------------
In an attempt to clarify this confusing state of affairs, we have begun a new programme of observations with the Ultraviolet-Visual Echelle Spectrograph (UVES) on the VLT (Kueyen) telescope (Dekker et al. 2000). In order to minimize the practical difficulties outlined above, we targeted DLAs which (a) have relatively low values of hydrogen column density ($N$(H I)$\simlt 4 \times 10^{20}$cm$^{-2}$), thus increasing the probability that the O I $\lambda 1302.2$ line may not be strongly saturated), and (b) are at relatively low redshifts ($z_{\rm abs} \simlt 2.7$), where the [Lyman $\alpha$]{} forest begins to thin out. Despite this observing strategy, which capitalises on the superior performance of UVES at ultraviolet wavelengths (D’Odorico et al. 2000), our endeavours have been only partially successful, as we shall see. In addition, we bring together a number of relevant abundance measurements which have been published since the surveys by Lu et al. (1998) and Centurión et al. (1998) for a comprehensive reappraisal of the abundances of oxygen and nitrogen in DLAs.
Observations and Data Reduction
===============================
Details of the DLAs observed are collected in Table 1. The data were secured over the three nights of 28–30 May 2000. By using different cross-disperser gratings and dichroic filters we recorded the spectra of three QSOs over the wavelength ranges listed in the penultimate column of Table 1. With a 1 arcsec wide entrance slit and $2 \times 2$ binning on the CCDs, the resulting spectral resolution was between 6.7 and 7.1km s$^{-1}$ FWHM, measured from the widths of emission lines of the Th-Ar hollow-cathode lamp used for wavelength calibration. The observations consisted of a number of one hour long exposures; the QSOs were moved along the slit between exposures so as to use different portions of the detectors to record the spectrum of each object. The slit was aligned at the parallactic angle throughout the observations.
The two-dimensional images were processed with the UVES standard pipeline software; see for example Ellison, Ryan, & Prochaska (2001) for a description of the main steps. The individual spectra extracted for each QSO were mapped onto a common, linear, vacuum heliocentric, wavelength scale; the interval between successive wavelength bins was set to the average value of the original CCD pixels, typically $\delta \lambda = 0.03\,$Åand 0.04Å in the blue and red wavelength regions respectively. The individual extractions were then co-added, with weights proportional to the S/N ratios, to produce final blue and red spectra for each QSO. With several independent exposures at our disposal, we were able to exclude from the sum wavelength bins which deviated from the mean by several standard deviations, because of residuals from the subtraction of cosmic-ray events, strong sky emission lines, and other invalid pixels.
Figure 1 shows portions of the final spectra in the regions near the damped [Lyman $\alpha$]{} lines at $z_{\rm abs}=2.45620$ in Q1409$+$095, $z_{\rm abs}=2.08681$ in Q1444$+$014, and $z_{\rm abs}=2.07623$ in Q2206$-$199, after division by the underlying QSO continuum. Superposed on the data are theoretical damped absorption profiles corresponding to the values of $N$(H I) listed in Table 1; the typical error in the determination of $N$(H I) is about 10%. Our observations revealed the existence of a second high column density absorption system in Q1409$+$095, at $z_{\rm abs} = 2.66820$. For completeness, we have included it in Figure 1; the profile fit shown is for $N$(H I)$= 5 \times 10^{19}$cm$^{-2}$.
Metal Absorption Lines in the Four DLAs
=======================================
Rationale
---------
As is the case with modern echelle spectra, our observations of the three QSOs in Table 1 provide a wealth of information on several absorption systems, as well as on the [Lyman $\alpha$]{} forest. In the present study we are primarily interested in the abundances of nitrogen and oxygen, detected respectively via the N I $\lambda \lambda 1199.5, 1200.2, 1200.7$ triplet and the O I $\lambda 1302.2$ line. Also of relevance to our analysis are the abundances of silicon (measured via Si II $\lambda 1304.4$) and iron (several absorption lines of Fe II throughout the spectral regions covered). Silicon is an $\alpha$-capture element which can provide an additional check on the oxygen abundance in cases where the O I $\lambda 1302$ line is saturated and the corresponding value of $N$(O I) is uncertain. Sulphur would also be useful in this context, but unfortunately the relevant absorption lines are either blended or undetected in most of the cases considered here. Iron is of interest because it is thought to be produced primarily by Type Ia supernovae; like nitrogen, its release into the interstellar medium should be delayed relative to that of oxygen. The four elements in question, N, O, Si, and Fe, respond differently to depletion onto dust grains. In the local ISM, N and O are normally present in the gas phase with near-solar abundances; Si and Fe, on the other hand, are susceptible to dust depletion, the latter normally more than the former (e.g. Savage & Sembach 1996). Thus, by considering all four elements together, it may be possible to decouple nucleosynthesis and depletion effects.
Profile Fitting and Abundance Determinations
--------------------------------------------
[lccccccc]{} QSO & Component & $z_{\rm abs}$ & $b$ & log $N$(O I) & log $N$(Si II) & log $N$(Fe II) & log $N$(N I)\
& Number & & (km s$^{-1}$) & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$)\
\
Q1409$+$095 & 1 & 2.45597 & 12.9 & $14.76 \pm 0.1$ & $13.82 \pm 0.03$ & $13.46 \pm 0.03$ & $\leq 12.80$\
& 2 & 2.45630 & 5.8 & $14.27 \pm 0.03$ & $13.08 \pm 0.03$ & $12.81 \pm 0.03$ & $\leq 12.50$\
& 3 & 2.45648 & 7.3 & $14.82 \pm 0.1$ & $13.63 \pm 0.03$ & $13.29 \pm 0.03$ & $\leq 12.77$\
\
Q1409$+$095 & 1 & 2.66820 & 7.5 & $15.27 \pm 0.1$ & $14.02 \pm 0.03$ & $13.85 \pm 0.03$ & $\leq 13.45$\
\
Q2206$-$199 & 1 & 2.07623 & 6.1 & $15.24 \pm 0.09$ & $13.65 \pm 0.02$ & $13.37 \pm 0.02$ & $\leq 12.88$\
\
Normalised profiles of the absorption lines mentioned above were fitted with Voigt profiles using the VPFIT package.[^2] VPFIT deconvolves the composite absorption profiles into the minimum number of discrete components and returns for each the most likely values of redshift $z$, Doppler width $b$ (km s$^{-1}$), and column density $N$ (cm$^{-2}$) by minimizing the difference between observed and computed profiles. The profile decomposition takes into account the instrumental point spread function of UVES. As we shall see shortly, the velocity structure of three out of the four DLAs studied here is very simple. In each case the same set of values of $z$ and $b$ was found to provide a satisfactory fit to the absorption lines of O I, Si II, and Fe II. Oscillator strengths and rest wavelengths were adopted from the up to date compilation of atomic parameters maintained by J. X. Prochaska at http://kingpin.ucsd.edu/\~hiresdla/atomic.dat (see also Prochaska et al. 2001). The solutions returned by VPFIT are collected in Table 2.
The error estimates $\sigma$(log$N$) returned by VPFIT were typically $< 0.05$dex for the optically thin Si II and Fe II lines and $< 0.1$dex for the saturated O I lines. To test whether these errors are realistic we performed Monte Carlo type simulations similar to those described by Bowen, Blades, & Pettini (1995) and Bowen, Pettini, & Blades (2002). Briefly, for each DLA we used as the starting point the theoretical spectrum produced by VPFIT with the best fit parameters given in Table 2. From this template we generated 300 spectra with the same S/N as the original data; these were then refitted to deduce values of $b$ and $N$. From gaussian fits to the distributions of 300 values of $b$ and $N$ we determined values of $\sigma$($b$) and $\sigma$(log$N$). The former are typically $\pm 0.2$[km s$^{-1}$]{}; the latter, which are comparable to the values returned by VPFIT, are listed in Table 2. Despite the good agreement between the two methods, it must be remembered that these errors reflect primarily the S/N of the data and do not include the possibility that the velocity dispersion parameter $b$ may be different between the different ions considered. The consequences of letting $b$ vary from ion to ion are discussed below for each DLA individually.
The species observed are the major ionisation stages of their respective elements in H I regions. Thus, to obtain their abundances relative to hydrogen we simply divide the column densities returned by VPFIT by $N$(H I) (after adding together the values for individual velocity components within the same DLA, when more than one component is present). Finally these values are referred to the solar abundance scale (Holweger 2001). The results are collected in Table 3, where the solar reference values are listed in the footnotes.
This procedure would overestimate the abundances of Si and Fe if some of the Si II and Fe II absorption we detect were due to ionised gas associated with the neutral DLA. Conversely, the abundances of O and N would be underestimated if these elements were partly ionised in the H I region. In the cases under consideration, however, there is little evidence that these reservations are justified. First, all four species exhibit very similar velocity structure; we see no indication of components present in the first ions and missing in the neutrals, or vice versa. Second, the recent reappraisal of this problem by Vladilo et al. (2001) shows that at the values of $N$(H I) of the DLAs studied here, the ionisation corrections to the abundances determined as above are smaller than the typical error of $\pm 0.15$ dex of our abundance estimates. We now briefly comment on each DLA in turn.
[lcccccccccc]{} QSO & $z_{\rm abs}$ & log $N$(H I)$^{a}$ & log $N$(O I) & log $N$(Si II) & log $N$(Fe II) & log $N$(N I) & \[O/H\]$^{b}$ & \[Si/H\]$^{b}$ & \[Fe/H\]$^{b}$ & \[N/H\]$^{b}$\
& & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & & & &\
\
Q1409$+$095 & 2.45620 & 20.54 & $15.15 \pm 0.06$ & $14.08 \pm 0.02$ & $13.74 \pm 0.02$ & $\leq 13.19$ & $-2.13$ & $-2.00$ & $-2.25$ & $\leq -3.28$\
\
Q1409$+$095 & 2.66820 & 19.70 & $15.27 \pm 0.1$ & $14.02 \pm 0.03$ & $13.85 \pm 0.03$ & $\leq 13.45$ & $-1.17$ & $-1.22$ & $-1.30$ & $\leq -2.18$\
\
Q2206$-$199 & 2.07623 & 20.44 & $15.24 \pm 0.09$ & $13.65 \pm 0.02$ & $13.37 \pm 0.02$ & $\leq 12.88$ & $-1.94$ & $-2.33$ & $-2.52$ & $\leq -3.49$\
\
Typical error in log $N$(H I) is $\pm 0.04$.
In the usual notation: \[X/H\] = log(X/H) $-$ log(X/H)$_{\odot}$. Solar abundances are from the recent updates by Holweger (2001): log(O/H)$_{\odot} = -3.264$; log(Si/H)$_{\odot} = -4.464$; log(Fe/H)$_{\odot} = -4.552$; and log(N/H)$_{\odot} = -4.069$.
The $z_{\rm abs} = 2.45620$ DLA in Q1409$+$095
----------------------------------------------
The Si II $\lambda 1304.4$ and Fe II $\lambda 2344.2$ lines in this DLA are unsaturated and are well reproduced with three absorption components spanning a velocity range $\Delta v = 44$[km s$^{-1}$]{} (see Figure 2 and Table 2). This is a very metal-poor system. Adding together the column densities for the three components we deduce that the abundance of Si is 1/100 of solar, and that of Fe only little more than 1/200 of solar (see Table 3).
Even at these low abundances, and with a hydrogen column density of $3.5 \times 10^{20}$[cm$^{-2}$]{} (near the lower limit of $2 \times 10^{20}$[cm$^{-2}$]{} commonly adopted as the definition of a DLA), the O I $\lambda 1302.2$ line is saturated (see Figure 2). If we adopt the same set of absorption parameters as for Si II and Fe II, we deduce an oxygen abundance of 1/135 of solar, in reasonable agreement with that of Si. The good model fit to O I $\lambda 1302.2$ makes it unlikely that we have overestimated $N$(O I) (the line would be broader than observed). The oxygen abundance could be higher if our model underestimates the saturation of the O I $\lambda 1302.2$ line; this possibility, however, does not alter the interpretation of our results below because it would tend to emphasize the ‘anomalously’ low (N/O) ratio (see §5).
The N I $\lambda 1200$ triplet at $z_{\rm abs} = 2.45620$ falls within the [Lyman $\alpha$]{} forest of this $z_{\rm em} = 2.856$ QSO. Figure 2 shows the best fit to this spectral region obtained with the model parameters of the other three lines, but it is difficult to be confident that the weak absorption we see, particularly in the $\lambda 1200.2$ transition, is actually N I and not a [Lyman $\alpha$]{} forest line. Consequently, we prefer to adopt the corresponding $N$(N I) as an upper limit to the column density of N I, from which we deduce that the abundance of N is less than 1/2000 of solar.
The $z_{\rm abs} = 2.66820$ sub-DLA in Q1409$+$095
--------------------------------------------------
There is a second high column density system in Q1409$+$095; although its $N$(H I)$ = 5.0 \times 10^{19}$[cm$^{-2}$]{}is below the limit normally adopted for DLAs (such systems are sometimes referred to as sub-DLAs), the damping wings of the [Lyman $\alpha$]{} line are easily recognisable in the UVES spectrum (see Figure 1) and $N$(H I) is well determined.
The O I, Si II, and Fe II absorption lines in this system are well reproduced with a single absorption component with a narrow velocity dispersion, $b = 7.5$[km s$^{-1}$]{}(see Figure 3); although there may be further substructure within this component, it is difficult to resolve it unambiguously in our data. All three lines give consistent abundances of between 1/15 and 1/20 of solar (see Table 3). Again, the N I triplet is blended with [Lyman $\alpha$]{} forest absorption. While there are features which match the expected positions of two of the N I lines, we take the fit shown in Figure 3 as indicative of the maximum amount of N I which could be present. In this case, N is less abundant than 1/150 of solar.
The $z_{\rm abs} = 2.08681$ DLA in Q1444$+$014
----------------------------------------------
Unlike the other DLAs in this study, this system exhibits a very complex velocity structure. The fit to the Si II $\lambda 1304.4$ line requires at least 13 components spread over the velocity interval $\Delta v \simeq 350$[km s$^{-1}$]{}, from $-170$ to $+176$[km s$^{-1}$]{} relative to $z_{\rm abs} = 2.08681$ (see Figure 4). Unfortunately, many of the components are strongly saturated in O I; while the weaker ones are well matched by the model parameters returned by VPFIT for Si II, we cannot determine a reliable value of the total column density of O I as it is dominated by the heavily saturated absorption between $v \simeq 0$ and $-150$[km s$^{-1}$]{}. In the bottom panel we show the partial fit, by the weaker components only, to the region encompassing the N I triplet. The combination of 13 absorption components which partially overlap in the closely separated triplet lines, together with blending with [Lyman $\alpha$]{} forest lines, makes it impossible to estimate the abundance of N. As we are unable to determine (N/O) in this system, we do not consider it further in this paper.
The $z_{\rm abs} = 2.07623$ DLA in Q2206$-$199
----------------------------------------------
This is a well-known DLA, one of the first to be studied at echelle spectral resolution thanks to the brightness of the QSO (Pettini et al. 1990; Pettini & Hunstead 1990). It has a remarkably simple velocity structure—the absorption apparently arises in a single component with $b \simeq 6$[km s$^{-1}$]{}(Figure 5 and Table 2). This property makes it one of the few DLAs suitable for measuring the abundance of deuterium (Pettini & Bowen 2001). It also exhibits a very low metallicity; O, Si and Fe are below solar by factors of 90, 210, and 330 respectively (Table 3). For the elements in common (Si and Fe), our results are in good agreement (within 0.1 dex) with the Keck data analysed by Prochaska & Wolfe (1997). As can be seen from Figure 5, we have a tentative detection of N I from which we deduce an upper limit to the abundance of N of only $3 \times 10^{-4}$ of solar.
It is interesting that in this system we deduce \[O/Si\]$ = +0.4 \pm 0.1$ (Table 3). Possibly, this is a hint of the trend of increasing \[O/Si\] with decreasing metallicity seen in Galactic halo stars with \[Fe/H\]$\simlt -2$ (Edvardsson et al. 1993; Ryan, Norris, & Beers 1996; Israelian et al. 2001). Less interesting explanations are that Si is depleted onto dust (although a depletion by a factor of 0.4 in the log seems somewhat high, given the low metallicity of this DLA—see Pettini et al. 1997) or that we have overestimated $N$(O I) when fitting the saturated $\lambda 1302$ absorption line. The abundance of S, an undepleted element which tracks O more closely than Si (Israelian & Rebolo 2001; Takada-Hidai et al. 2001—see discussion in §4) would help discriminate between these three possibilities, but unfortunately all three transitions of the S II$\lambda1256$ multiplet in this system are blended with [Lyman $\alpha$]{} forest lines.
The DLA Sample
==============
In the last few years there have been several new measurements of the abundance of N, generally reported as part of abundance studies of individual DLAs. We have collected these data in Table 4 where references to the original papers can also be found. All of the measurements in the table are from spectra of high signal-to-noise ratio and high (echelle) resolution obtained with 8-10m class telescopes. For simplicity we limit ourselves to absorption systems with $N$(H I)$ \geq 2 \times 10^{20}$cm$^{-2}$; inclusion of similar data available for a few sub-DLAs would not change any of our conclusions.
[llclccclcc]{} QSO & $z_{\rm abs}$ & log $N$(H I) & log $N$(N I) & log $N$(O I) & log $N$(S II) & (O/H)+12$^{a}$ & (N/O)$^{a}$ & (Fe/O)$^{a}$ & Ref.$^{b}$\
& & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & (cm$^{-2}$) & & & &\
\
Q0000$-$2620 & 3.3901 & 21.41 & 14.73 & 16.45 & …& 7.04 & $-1.72$ & $-1.58$ & 1\
Q0100$+$1300 & 2.3090 & 21.32 & 15.29 & …& 15.12 & 7.34 & $-1.37$ & $-1.56$ & 2, 3\
Q0201$+$1120 & 3.38639 & 21.26 & 15.33 & …& 15.21 & 7.49 & $-1.42$ & $-1.40$ & 4\
J0307$-$4945 & 4.46658 & 20.67 & 13.57 & 15.91 & …& 7.24 & $-2.34$ & $-1.70$ & 5\
Q0347$-$3819 & 3.02486 & 20.63 & 14.89 & 16.64 & …& 8.01 & $-1.75$ & $-2.21$ & 6\
Q1409$+$095 & 2.45620 & 20.54 & $\leq 13.19$ & 15.15 & …& 6.61 & $\leq -1.96$ & $-1.41$ & 7\
GB1759$+$7539 & 2.62528 & 20.76 & 14.99 & …& $15.21^c$ & 7.99 & $-1.76^c$ & $-1.81$ & 8,9\
Q2206$-$199 & 2.07623 & 20.44 & $\leq12.88$ & 15.24 & …& 6.80 & $\leq -2.36$ & $-1.87$ & 7\
Q2343$+$1230 & 2.4313 & 20.34 & 14.67 & …& 14.71 & 7.91 & $-1.58$ & …& 2\
Q2348$-$1444 & 2.27936 & 20.56 & $\leq 13.22$ & …& 13.73 & 6.71 & $\leq -2.05$ & $-1.48$ & 3\
\
When the oxygen abundance is not available, S has been used as a proxy for O by assuming the solar ratio (O/S)$_{\odot} = +1.54$ (Grevesse & Sauval 1998; Holweger 2001); see text (§4) for justification.
References—1: Molaro et al. (2000); 2: Lu, Sargent, & Barlow (1998); 3: Prochaska & Wolfe (1999); 4: Ellison et al. (2001); 5: Dessauges-Zavadsky et al. (2001); 6: Levshakov et al. (2002); 7: This work; 8: Outram, Chaffee, & Carswell (1999); 9: Prochaska et al. (2002).
Prochaska et al. (2002) found that only about half of the metals in this DLA arise in a neutral component, while the reminder are in partially ionised gas. It is unclear what fraction of S in the ionised gas is S II; in the extreme case that half of the observed S II is in an ionised region where O I and N I are absent, the value of (N/O) for this DLA should be increased by +0.3dex.
The full sample consists of 10 DLAs in which the abundance of N has been measured (seven detections and three upper limits). In one half of the cases the abundance of O is available directly, either because the O I $\lambda 1302.2$ line is not strongly saturated, or from weaker, unsaturated, transitions to higher energy levels. In the other five cases we take $N$(S II) as a proxy for $N$(O I), after correcting for the solar abundance ratio (O/S)$_{\odot} = +1.54$ (Grevesse & Sauval 1998; Holweger 2001). The justification for this approach is as follows.
First, in Galactic stars the (O/S) ratio appears to be approximately solar over a wide range of metallicities down to \[O/H\]$ \simlt -2$ (Israelian et al. 2001; Israelian & Rebolo 2001; Takada-Hidai et al. 2001; Nissen et al. in preparation). Second, in the diffuse Galactic interstellar medium both elements show little affinity for dust (Savage & Sembach 1996); dust corrections for O and S are almost certainly unimportant for our purposes at the reduced levels of dust depletions typical of metal-poor DLAs (Pettini et al. 1997; Vladilo 1998). Finally, differential ionisation corrections between O I and S II are likely to be small. Theoretically, the photoionisation calculations by Vladilo et al. (2001) show that, for all the models they considered, if log (O/S)=log \[$N$(O I)/$N$(S II)$]+{\cal C}$, then $|{\cal C}| \simlt 0.2$ ($\simlt 0.1$) when log $N$(H I)$\geq 20.2$ ($\geq 20.6$). This condition is met by 5 (4) out of the 5 DLAs in Table 4 for which we use S as a proxy for O. For all the DLAs considered here $|{\cal C}|$\[N/O\]$\leq 0.002$ so that log (N/O)=log \[$N$(N I)/$N$(O I)\].
Some authors (e.g. Lu et al. 1998) have also used Si to deduce the abundance of O, by assuming that in DLAs $N$(O I)/$N$(Si II)=(O/Si)$_{\odot}$. This assumption, however, is less secure than the equivalent one for S and has been criticised, for example, by Matteucci, Molaro, & Vladilo (1997). First, Si can be depleted onto dust. Second, it is unclear whether the abundances of O and Si do track each other at low metallicities (Edvardsson et al. 1993; Ryan, Norris, & Beers 1996; Israelian et al. 2001); theoretically they are not expected to, if some of the Si is produced by type Ia supernovae (Matteucci et al. 1997). Neither of these complications applies to S. In the past, the use of Si as an indicator of the O abundance had been motivated by the paucity of measurements of \[O/H\] and \[S/H\] in DLAs; with the larger body of data now available this approach is no longer necessary. For this reason, we restrict ourselves to O and S in the following analysis.
Nitrogen and Oxygen in DLAs
---------------------------
We are now ready to compare the abundances of N and O in high redshift DLAs with those measured in H II regions in the local universe (Figure 6). For the latter, we use the compilation assembled by Henry et al. (2000—references to the original surveys are given in the figure caption), augmented by some additional work (Ferguson, Gallagher, & Wyse 1998; van Zee 2000). The dashed lines in the figure are approximate boundaries of the region in the (N/O) vs. (O/H) plot where we may expect DLAs to fall. The line labelled ‘Secondary’ is simply an extrapolation to low (O/H) values of the (N/O) vs. (O/H) trend at high metallicities, where secondary N presumably dominates. [*If*]{} the extrapolation to low metallicities is valid (we do not know that it is), this is the minimum amount of N (relative to O) which one may expect to find, irrespectively of the previous star formation history of the system under consideration. The ‘Primary’ level is just the plateau at (N/O)$ = -1.5$ indicated by existing measurements in low metallicity H II regions. We stress that these are just approximate empirical limits which are not derived from a full chemical evolution treatment of the available data (e.g. Larsen, Sommer-Larsen, & Pagel 2001)
As can be seen from Figure 6, the DLAs in our sample do indeed populate the region of the diagram delimited by the broken lines. Six of the DLAs considered exhibit levels of N enrichment which are similar to those of the most metal-poor H II regions. On the other hand, in four cases N appears to be below the ‘Primary’ line, by up to nearly one order of magnitude.
{width="40pc"}
Discussion
==========
The results presented here, in Figure 6 and Table 4, confirm the earlier conclusion by Lu et al. (1998) that the (N/O) ratio in DLAs exhibits a larger range of values than that of any other pair of heavy elements which are not depleted onto dust (Prochaska & Wolfe 2002). We now consider possible causes of this scatter.
The simplest explanation may be just measurement error. Although the errors on $N$(N I) and $N$(O I) quoted by the sources referenced in Table 4 are normally less than 0.1dex (and always less than 0.2dex), it must be remembered that the N I lines (a) are generally weaker than most other metal transitions used for abundance work in DLAs, and (b) occur within the [Lyman $\alpha$]{} forest where blending can complicate their analysis. It is thus conceivable that the true uncertainties, particularly in the N abundance, may have been underestimated. On the other hand, if measurement error were the sole cause of the dispersion of the DLA points in Figure 6, it would be a remarkable coincidence that all 10 values in our sample should fall within the boundaries delimited by the Primary and Secondary lines in the figure. In principle one would expect measurement errors to scatter points above the Primary line and, at the larger values of (O/H), below the Secondary boundary. For many of the low points in Figure 6, it would be very hard to understand how an absorption line $\sim 3-5$ times stronger than observed could have been missed. We conclude, therefore, that measurement errors are not the principle cause of the dispersion observed.
We do not subscribe to the view proposed by Izotov et al. (2001) that unrecognised ionisation corrections are responsible for the low (N/O) values. As explained above, in about half of the cases (N/O)$_{\rm DLA}$ is comparable to the value measured in nearby metal-poor H II regions; there is no evidence that these DLAs are different from the rest of the sample in their level of ionisation. Similarly, there is no separation in Figure 6 between DLAs where (O/H) has been measured directly and those where it has been deduced by reference to S.
Some may be concerned about possible systematic differences between abundances measured via absorption lines in cool interstellar H I clouds and those deduced from the analysis of nebular emission lines from H II gas. However, there are no precedents for such systematic differences in the solar vicinity (Meyer, Cardelli, & Sofia 1997; Meyer, Jura, & Cardelli 1998) and, in any case, here we are dealing with a scatter rather than a systematic offset.
Qualitatively, our findings are consistent with, and indeed provide empirical evidence in favour of, an origin of primary nitrogen in intermediate mass stars. The observed range in the (N/O) ratio at low (O/H) is, in this picture, a natural consequence of the delayed release of N into the ISM relative to the O produced by Type II supernovae. However, there are puzzling aspects in this scenario too, in that the proportion of DLAs with values of (N/O) below the primary level seems at first sight to be surprisingly high. Recall that Henry et al. (2000) estimated the time delay between N and O release into the ISM to be $\sim 250$Myr, if the main source are stars in the mass range $7 - 4\,M_{\odot}$. The median redshift of the DLA sample in Table 4 is $\langle z \rangle = 2.5$; in a $\Omega_{\rm M} = 0.3$, $\Omega_{\Lambda} = 0.7$, $h = 0.65$ cosmology this corresponds to a look-back time of $11.7$Gyr. [^3] There is now evidence that star formation in the universe was already at high levels well before this epoch, certainly by $z \simgt 4.5$ (Steidel et al. 1999), and possibly up to $z \simeq 6$ (Shapley et al. 2001; Becker et al. 2001; Dawson et al. 2001). In this cosmology, $z = 6$ corresponds to a look-back time of $\sim 13.5$Gyr. Thus, the time interval between $z = 6$ and our median $\langle z_{\rm DLA}\rangle = 2.5$ is seven times higher than the estimated time delay for the release of primary nitrogen. If the galaxies giving rise to DLAs formed continuously between $z = 6$ and 2.5, naively one may expect to find only one DLA out of seven in the interim period when primary nitrogen has not yet caught up with the oxygen released by the same episode of star formation. Instead we find at least four such cases out of ten (see Figure 6).
Perhaps we are just seeing the effects of small number statistics and a larger sample of (N/O) measurements will in future reveal that the present 40% is an overestimate of the number of DLAs with a N deficiency. However, if such a high fraction is confirmed by future observations, we can think of two explanations for this apparent puzzle. The four DLAs with values of (N/O) clearly below the primary level in Figure 6 have abundances (O/H) less than $\sim 1/30$ of solar. Thus, they have presumably experienced little star formation up to the time when we observe them; in a closed box model, they would have turned less than 1/30 of their gas into stars. It therefore seems at least plausible that these DLAs arise preferentially in young galaxies, which have only recently condensed out of the intergalactic medium. In other words, the very fact that the difference between primary and secondary nitrogen is most pronounced at low metallicities introduces a bias in our sample such that it may include a relatively high proportion of galaxies caught in the initial stages of their chemical evolution (e.g. Cen et al. 2002).
A second possibility is that the time delay for the release of primary nitrogen may be longer when metal abundances are lower. This dependence may arise if at lower metallicities stars of progressively lower masses can synthesize and release nitrogen, because their interiors are hotter (Lattanzio & Forestini 1999) and/or through the increasing importance of rotation (Meynet & Maeder 2002a,b; see also Marigo 2001). The slope of the initial mass function would then shift the peak of the N production to lower mass stars, with longer evolutionary times. While these effects remain to be fully quantified, there have been suggestions that at the metallicities of DLAs the full release of primary N may lag behind that of O by up to 800Myr (Lattanzio et al., in preparation). If confirmed, this proposal would go a long way towards explaining the high proportion of DLAs which are underabundant in N at $\langle z \rangle = 2.5$.
Clues from the Abundance of Iron?
---------------------------------
It should be possible to test the hypothesis that the dispersion of (N/O) values in DLAs is a consequence of the delayed release of primary nitrogen by considering the abundances of other elements synthesized by low mass stars. Observationally, iron is one of the more easily accessible through several rest frame ultraviolet transitions. Measurements of (Fe/H) are available for 9 out of the 10 DLAs in our sample[^4]; the corresponding values of (Fe/O) are listed in Table 4.
It is generally thought that approximately 2/3 of the iron is produced in Type Ia supernovae. While it is still unclear which types of binary system give rise to such events, there seems to be a consensus that the evolutionary timescales of Type Ia SN progenitors are of the same order as, or longer than, those of the $7-4\,M_{\odot}$ stars which are the source of primary nitrogen (e.g. Matteucci & Recchi 2001). From this it follows that DLAs deficient in N (in the sense of having less N than the primary level) should also exhibit sub-solar (Fe/O) ratios. We examine this point in Figure 7, by plotting the quantity \[(N/O) + 1.5\] (that is, the difference between the measured (N/O) and the primary plateau at (N/O)$ = -1.5$) vs. \[(Fe/O) + 1.29\] (the difference between observed and solar (Fe/O) adopting (Fe/O)$_{\odot} = -1.29$ from Holweger 2001).
There are two complications which affect the interpretation of the data in Figure 7. First, if massive stars which explode as Type II SN also produce some Fe, the effect of time delay will be less pronounced for Fe than for N. Indeed, Galactic stars with metallicities similar to those of the DLAs considered here, between $\sim 1/10$ and $\sim 1/100$ of solar, are deficient in Fe only by factors of 2–3. In the units of Figure 7, \[Fe/O\]$\equiv$(Fe/O)$-$(Fe/O)$_{\odot} = -0.3$ to $-0.5$ (e.g. Israelian et al. 2001). This is larger, but not by much, than the typical accuracy of 0.1–0.2 dex with which this ratio is measured in DLAs. A more serious problem is that, unlike O, Fe can be depleted onto dust and this effect will mimic the intrinsic underabundance which we are testing for. The combination of the subtlety of the effect with the uncertainty in correcting for dust depletions may well explain, at least in part, why the $\alpha$-element enhancement well established in Galactic metal-poor stars has proved so difficult to pin down in DLAs (Vladilo 1998; Pettini et al. 2000; Prochaska & Wolfe 2002, Ledoux, Bergeron, & Petitjean 2002). Accounting for the fraction of Fe in solid form requires taking into consideration the full complement of abundance measures in a DLA (e.g. Vladilo 2002), and is beyond the scope of this work. The values of \[Fe/O\] plotted in Figure 7 should therefore be considered as indicative of the maximum intrinsic deficiency of Fe relative to O in the DLAs in our sample.
These complications make it difficult to draw firm conclusions from the results in Figure 7. Some of the DLAs appear to meet our expectations, by exhibiting low values of (N/O) matched by corresponding Fe underabundances. Similarly, DLAs which lie closer to the Primary line at (N/O)$= -1.5$ show a range of \[Fe/O\] values, as expected if the time delay for the release of Fe is longer than that of N. But there are also cases which evidently do not fit this pattern. The DLA in Q0347$-$3819 may be a special case, in that it is the [*only*]{} DLA known to date with such a strong $\alpha$-element enhancement (Levshakov et al. 2002). More problematic, however, are the two DLAs with low (N/O) (both are upper limits) and only a minor (if any) Fe deficiency. Possibly we have underestimated the N abundance in these cases because of blending with [Lyman $\alpha$]{} forest lines.[^5]
We conclude that the results of Figure 7 do not provide clear support for the scenario where both N and Fe lag behind the release of O into the ISM. Possibly, as more data become available, a trend between (N/O) and (Fe/O) will emerge, if one exists. But Figure 7 may also be telling us that there is a limit to the degree to which our ideas on the chemical evolution of the Milky Way can be applied to the high redshift DLA galaxies.
Summary and Conclusions
=======================
We have presented new measurements of the abundances of N, O, Si, and Fe in two metal-poor damped [Lyman $\alpha$]{} systems (plus one sub-DLA) obtained with UVES on the VLT. We have combined these data with others from the literature to form a sample of 10 DLAs for which reliable measurements of the abundance of N (or useful upper limits) are available. The sample consists exclusively of high resolution, high S/N measurements obtained with 8-10m class telescopes. Our aim was to test current ideas on the nucleosynthesis of N, by extending measurements of the abundances of N and O from H II regions in the local universe to the younger and less enriched galaxies at high redshift ($\langle z \rangle = 2.5$) that give rise to the DLAs. Our principal conclusions are as follows.
1. The (N/O) ratio remains a difficult quantity to measure, because the absorption lines available generally have widely different optical depths, so that when the N I $\lambda 1200$ triplet is detected O I $\lambda 1302.2$ is saturated. Furthermore, the N I lines are often blended within the [Lyman $\alpha$]{} forest. We have not identified an optimum strategy for dealing with these difficulties. DLAs at higher redshifts than those shown here make accessible weaker transitions of O I at far-UV wavelengths, but the chances of confusion of N I with [Lyman $\alpha$]{} forest lines are then increased.
2. We have argued that S, when available, can be used as a proxy for O, because they are both undepleted $\alpha$-elements and the differential ionisation corrections between O I and S II should not be important when the column density of neutral hydrogen is as high as in DLAs. However, it would obviously be reassuring to confirm empirically that (O/S)$\equiv N$(O I)/$N$(S II) in at least a few cases, with specially targeted observations.
3. We confirm earlier reports that the (N/O) ratio shows a larger dispersion of values than other ratios of heavy elements in DLAs. All 10 measurements in our sample fall within the region in the (N/O) vs. (O/H) plot bounded by the primary and secondary levels of N production.
4. We have considered several possible explanations for this scatter and conclude that the most plausible one is that we are seeing the effects of a time delay, by about 250Myr, in the release of primary N from $7 - 4\,M_{\odot}$ stars relative to the massive stars which release O when they explode as Type II supernovae. Thus, our results provide empirical evidence in support of currently favoured ideas for the nucleosynthesis of N. The uniform value (N/O)$\simeq -1.5$ (on a log scale) seen in nearby metal-poor star-forming galaxies can be understood in this scenario if these galaxies are not young, but contain older stellar populations, as indicated by a number of imaging studies with [*HST*]{}.
5. A surprisingly high proportion (40%) of DLAs in our sample have apparently not yet attained the full primary level of N enrichment at (N/O)$\simeq -1.5$. Possibly, the low metallicity regime—where the difference between secondary and primary nitrogen enrichment is most pronounced—preferentially selects young galaxies which have only recently condensed out of the intergalactic medium and begun forming stars. A more speculative alternative, which needs to be explored computationally, is that at low metallicities stars with masses lower than $4 M_{\odot}$ may make a significant contribution to the overall N yield. The release of primary N may, under these circumstances, continue for longer than 250Myr, perhaps for a substantial fraction of the Hubble time at the median $\langle z \rangle = 2.5$ of our sample.
6. In this scenario, the abundance of Fe—two thirds of which is presumed to originate from Type Ia supernovae—should also lag behind that of O, by at least as long as that of N. However, conflicting clues are provided by a plot (N/O) vs. (Fe/O) for the present data; while some DLAs do match the predictions of our working hypothesis, there also appear to be some deviant cases. The interpretation is complicated by the fact that some fraction of the Fe is also produced by massive stars and that depletion onto dust can be an issue. A larger sample of measurements is required to assess whether the relative abundances of N, O and Fe in DLAs are consistent with current ideas of chemical evolution in galaxies.
7. Finally, we point out that the large dispersion in the abundance of N at these low metallicities essentially precludes the use of N V $\lambda\lambda 1238, 1242$ doublet in constraining models of the ionisation of the intergalactic medium from the ratios of different ions in the [Lyman $\alpha$]{} forest (e.g. Boksenberg et al., in preparation).
It is a pleasure to acknowledge the competent assistance with the observations by the ESO support staff at Paranal. We are very grateful to Dick Henry for kindly providing us with electronic tables of most of the data (on local H II region abundances) plotted in Figure 6, and to Jason X. Prochaska for making widely available his comprehensive compilation of DLA measurements and related atomic parameters. Special thanks are due to David Bowen for running his Monte Carlo simulations on our behalf. Mike Edmunds, Paolo Molaro, Bernard Pagel and Sam Rix made valuable comments on early versions of the manuscript. This work was supported in part by the European RTN programme “The Intergalactic Medium”. Max Pettini thanks the Instituto de Astrofisica de Canarias and the Instituto de Astronomía, UNAM for their hospitality during visits when this work was completed.
Becker, R.H., Fan, X., White, R.L., et al. 2001, AJ, 122, 2850
Bowen, D.V., Blades, J.C., & Pettini, M. 1995, ApJ, 448, 634
Bowen, D.V., Pettini, M., & Blades, J.C. 2002, ApJ, in press
Cen, R., Ostriker, J.P., Prochaska, J.X., & Wolfe, A.M. 2002, ApJ, submitted (astro-ph/0203524)
Centurión, M., Bonifacio, P., Molaro, P., & Vladilo, G. 1998, ApJ, 509, 620
Contini, T., Treyer, M.A., Sullivan, M., & Ellis, R.S. 2002, MNRAS, 330, 75
Crone, M.M., Schulte-Ladbeck, R.E., Greggio, L., & Hopp, U. 2002, ApJ, in press (astro-ph/0111248)
Dawson, S., Stern, D., Bunker, A.J., Spinrad, H., & Dey, A. 2001, AJ, 122, 598
Dekker, H., D’Odorico, S., Kaufer, A., Delabre, B., & Kotzlowski, H. 2000, Proc. SPIE Conf., 4008, 534
Dessauges-Zavadsky, M., D’Odorico, S., McMahon, R.G., Molaro, P., Ledoux, C., Péroux, C., & Storrie-Lombardi, L.J. 2001, A & A, 370, 426
D’Odorico, S., Cristiani, S., Dekker, H., et al. 2000, Proc. SPIE Conf., 4005, 121
Edmunds, M.G., & Pagel, B.E.J. 1978, MNRAS, 185, 77P
Edvardsson, B., Andersen, J., Gustafsson, B., Lambert, D.L., Nissen, P.E., & Tomkin, J. 1993m A&A, 275, 101
Ellison, S.L., Pettini, M., Steidel, C.C., & Shapley, A.E. 2001, ApJ, 549, 770
Ellison, S.L., Ryan, S.G., & Prochaska, J.X. 2001, MNRAS, 326, 628
Ferguson, A.M.N., Gallagher, J.S., & Wyse, R.F.G. 1998, AJ, 116, 673
Grevesse, N., & Sauval, A.J. 1998, Space Sci Rev, 85, 161
Henry, R.B.C., Edmunds, M.G., & Köppen, J. 2000, ApJ, 541, 660
Holweger, H. 2001, in Solar and Galactic Composition, ed. R.F. Wimmer-Schweingruber, (Berlin: Springer), 23
Israelian, G., & Rebolo, R. 2001, ApJ, 557, L43
Israelian, G., Rebolo, R., García López, R.J., Bonifacio, P., Molaro, P., Basri, G., & Shchukina, N. 2001, ApJ, 551, 833
Izotov, Y. et al. 2001, in Gaseous Matter in Galaxies and Intergalactic Space, eds. R. Ferlet, M. Lemoine, J.M. Desert, & B. Raban (Frontier Group), 225
Izotov, Y., Schaerer, D., & Charbonnel, C. 2001, ApJ, 549, 878
Izotov, Y.I., & Thuan, T.X. 1999, ApJ, 511, 639
Kobulnicky, H.A., & Skillman, E.D. 1996, ApJ, 471, 211
Larsen, T.I., Sommer-Larsen, J., & Pagel, B.E.J. 2001, MNRAS, 323, 555
Lattanzio, J.C. & Forestini, M.F. 1999, in Asymptotic Giant Branch Stars, eds. T. Le Bertre, A. Lébre, & C. Waelkens, IAU Symp. 191, 31.
Ledoux, C., Bergeron, J., & Petitjean, P. 2002, A&A, in press (astro-ph/0202134)
Levshakov, S.A., Dessauges-Zavadsky, M., D’Odorico, S., & Molaro, P. 2002, ApJ, 565, 696
Lu, L., Sargent, W.L.W., & Barlow, T.A. 1998, AJ, 115, 55
Maeder, A. 1992, A&A, 264, 105
Marigo, P. 2001, A&A, 370, 194
Matteucci, F., Molaro, P., & Vladilo, G. 1997, A&A, 321, 45
Matteucci, F., & Recchi, S. 2001, ApJ, 558, 351
Meyer, D.M., Cardelli, J.A., & Sofia, U.J. 1997, ApJ, 490, L103
Meyer, D.M., Jura, M., & Cardelli, J.A. 1998, ApJ, 493, 222
Meynet, G. & Maeder, A. 2002a, A&A, 381, L25
Meynet, G. & Maeder, A. 2002b, A&A, submitted
Molaro, P., Bonifacio, P., Centurión, M., D’Odorico, S., Vladilo, G., Santin, P., & Di Marcantonio, P. 2000, ApJ, 541, 54
Outram, P.J., Chaffee, F.H., & Carswell, R.F. 1999, MNRAS, 310, 289
Pettini, M. & Bowen, D.V. 2001, ApJ, 560, 41
Pettini, M., Ellison, S. L., Steidel, C. C., & Bowen, D. V. 1999, ApJ, 510, 576
Pettini, M., Ellison, S. L., Steidel, C. C., Shapley, A.E., & Bowen, D. V. 2000, ApJ, 532, 65
Pettini, M. & Hunstead, R.W. 1990, Aust. J. Phys., 43, 227
Pettini, M., Hunstead, R.W., Smith, L.J., & Mar, D.P. 1990, MNRAS, 246, 545
Pettini, M., King, D.L., Smith, L.J., & Hunstead, R.W. 1997, ApJ, 478, 536
Pettini, M., Lipman, K., & Hunstead, R.W. 1995, ApJ, 451, 100
Pilyugin, L.S. 1999, A&A, 346, 428
Prochaska, J.X., et al. 2001, ApJS, 137, 21
Prochaska, J.X., Howk, J.C., O’Meara, J.M., Tytler, D., Wolfe, A.M., Kirkman, D., Lubin, D., & Suzuki, N. 2002, ApJ, in press (astro-ph/0202140)
Prochaska, J.X., & Wolfe, A.M. 1997, ApJ, 474, 140
Prochaska, J.X., & Wolfe, A.M. 1999, ApJS, 121, 369
Prochaska, J.X., & Wolfe, A.M. 2002, ApJ, 566, 68
Ryan, S.G., Norris, J.E., & Beers, T.C. 1996, ApJ, 471, 254
Savage, B.D., & Sembach, K.R. 1996, ARA&A, 34, 279
Schulte-Ladbeck, R.E., Hopp, U., Greggio, L., Crone, M.M., Drozdovsky, I.O., & Hopkins, A.M. 2001, in Dwarf Galaxies and Their Environment, ed. K.S. de Boer, R.J. Dettmar, & U. Klein, (Aachen: Shaker), 83
Shapley, A.E., Steidel, C.C., Adelberger, K.L., Dickinson, M., Giavalisco, M., & Pettini, M. 2001, ApJ, 562, 95
Steidel, C.C., Adelberger, K.L., Giavalisco, M., Dickinson, M., & Pettini, M. 1999, ApJ, 519, 1
Takada-Hidai, M., et al. 2001, ApJ, submitted (astro-ph/0103481)
Thurston, T.R., Edmunds, M.G., & Henry, R.B.C. 1996, MNRAS, 283, 990
van den Hoek, L.B., & Groenewegen, M.A.T. 1997, A&AS, 123, 305
van Zee, L., 2000, ApJ, 543, L31
van Zee, L., Salzer, J.J., Haynes, M.P., O’Donoghue, A.A., & Balonek, T.J. 1998, AJ, 116, 2805
Vladilo, G. 1998, ApJ, 493, 583
Vladilo, G. 2002, ApJ, in press (astro-ph/0112338)
Vladilo, G., Centurión, M., Bonifacio, P., & Howk, J.C. 2001, ApJ, 557, 1007
[^1]: As an aside, Henry et al. (2000) pointed out that the rise in (N/O) with (O/H) is steeper than would be normally expected for a purely secondary element and proposed that the effect is augmented by a decreasing O yield with increasing metallicity (Maeder 1992).
[^2]: VPFIT is available at http://www.ast.cam.ac.uk/\~rfc/vpfit.html
[^3]: The age of the universe in this cosmology is $14.5$Gyr.
[^4]: The same is not true for other Fe-peak elements, especially Zn whose abundance may be easier to interpret because, unlike Fe, Zn is not depleted onto dust.
[^5]: A similar apparent inconsistency was pointed out by Centurión et al. (1998) for the DLA in Q0100$+$1300. The HIRES observations of this QSO by Lu et al. (1998) and Prochaska & Wolfe (1999), used here, give (N/O)+1.5$ = +0.13$ and \[Fe/O\]$\,=-0.27$, which are consistent with our working hypothesis. Centurión et al. (1998), on the other hand, derived a much lower column density of N—by a factor of four—from observations of the weaker N I $\lambda\lambda1134$ triplet with the 4m William Herschel telescope on la Palma. Observations of the N I $\lambda\lambda1134$ triplet with an 8-10m telescope are required to resolve this discrepancy.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Davide Modolo and Vittorio Ferrari
bibliography:
- '../../bibtex/shortstrings.bib'
- '../../bibtex/vggroup.bib'
- '../../bibtex/calvin.bib'
title: |
Learning Semantic Part-Based Models\
from Google Images
---
-based models have gained significant attention in the last few years. The key advantages of exploiting part representations is that parts have lower intra-class variability than whole objects, they deal better with pose variation and their configuration provides useful information about the aspect of the object. Parts localization has therefore been addressed in the context of several vision tasks, such as object recognition [@chen14cvpr; @endres13cvpr; @felzenszwalb10pami], object segmentation [@arbelaez12cvpr; @wang15cvpr], fine-grained classification [@liu12eccv; @parkhi12cvpr; @zhang14eccv], human pose-estimation [@liu14eccv; @sun11iccv_art; @ukita12cvpr], attribute prediction [@bourdev09iccv; @zhang13iccv], action classification [@gkioxari15iccv] and scene classification [@juneja13cvpr], achieving state-of-the-art results in many of them.
Part-based methods can be grouped into two sets. The first set of works define an object part as any patch that is discriminative for the object class [@endres13cvpr; @felzenszwalb10pami; @arbelaez12cvpr; @wang15cvpr; @juneja13cvpr]. These works typically discover parts in the training images automatically, without human supervision. However, their resulting parts do not have a meaning for humans (e.g. a patch straddling between the wheel and the chassis of a car [@felzenszwalb10pami]). The second set of works define parts semantically (e.g. ‘wheel’) [@chen14cvpr; @wang15cvpr; @liu12eccv; @parkhi12cvpr; @liu14eccv; @sun11iccv_art; @ukita12cvpr; @zhang13iccv]. These are more interpretable for a human and are necessary to obtain fine descriptions of objects and their interactions. For example, “the headlights of the bus are turned on” and “the cat is touching the TV with its tail”. Moreover, part localization is necessary for a robot to correctly grasp an object (e.g. grasp a mug by the handle). However, existing works on semantic part detection require part location annotations in the training images, which are very expensive to obtain.
In this paper we try to get the best of both worlds by proposing a novel method to train semantic part models of object classes without manual location annotations. We train these models on images automatically collected from Google Images. We represent an object class as a mixture over multiple viewpoints. We learn a collection of semantic part appearance models, and models of their spatial arrangement on the object, specific to each viewpoint. Moreover, we also train models capable of predicting the viewpoint of the object, which we then use to select an appropriate location model to guide part localization on novel instances. Learning from the web has been addressed before [@fergus05iccv; @vijayanarasimhan:cvpr08; @li10ijcv; @schroff11pami; @liq13cvpr; @chen13iccv; @divvala14lcvpr; @chen15cvpr; @novotny16eccvworkshop], but mostly at the level of object classes. Instead, here we learn complex semantic part-based models from the web.
We learn these rich models fully automatically, entirely from Google Images, by collecting training instances for both *parts* and *objects* (fig. \[fig:front\]), and automatically connecting the two levels. Our technique incrementally learns from easy examples first, and then gradually adapts to harder examples. This adaptation is done within Google Images, where part images offer easy examples (fig. \[fig:front\] left), and then harder examples are mined from object instances (fig. \[fig:front\] right). The move from part images to object images also enables us to learn the spatial arrangement of parts on the object (location models). In a final step, we further adapt our models on an external, non-Google image domain to adapt to even harder examples, e.g. on PASCAL VOC [@everingham10ijcv].
{width="\textwidth"}
\[fig:front\]
We demonstrate the effectiveness of our incremental learning algorithm on the PASCAL-Part dataset [@chen14cvpr]. Interestingly, the performance of our part models increases at every step of the learning, with the final models more than doubling the initial performance (from 12.9 to 27.2 AP). Moreover, we compare to two other webly-supervised works (LEVAN [@divvala14lcvpr] and NEIL [@chen13iccv]) and show that our part models perform better. Finally, we also show that our part models can help object detection performance by enriching the R-CNN detector [@girshick14cvpr] with parts.
Related work {#sec:rel_work}
============
Many works learn non-semantic part models, where the parts are arbitrary patches that are discriminative for an object class [@endres13cvpr; @felzenszwalb10pami; @arbelaez12cvpr; @wang15cvpr; @juneja13cvpr]. Our work is more related to semantic part-based models, and to techniques for learning object classes from image search engines.\
[**Semantic part-based models.**]{} There is a considerable amount of work on using semantic parts to help recognition tasks. The largest part of this has focused on the fine-grained recognition problem in several animal domains, such as birds [@liu14eccv; @zhang13iccv; @zhang14eccv; @lin15cvpr] and pets [@liu12eccv; @parkhi12cvpr]. In these works, an object is treated as a collection of parts that models its shape and appearance. Semantic parts help capturing subtle object appearance differences that could not be captured by a monolithic object model. These differences are crucial to discriminate between animal breeds. Other applications where semantic part-based models have been used are object detection [@chen14cvpr], articulated human and animal pose estimation [@liu14eccv; @sun11iccv_art; @ukita12cvpr] and attribute prediction [@zhang13iccv; @gkioxari15iccv]. In object detection, parts help dealing with deformed, occluded and low resolution objects. In articulated pose estimation, parts help identifying objects in special configuration (e.g. jumping and sitting) as opposed to canonical ones. Finally, in attribute prediction, attributes are predicted best by the part containing direct evidence about them.
All the above mentioned methods require accurate part location annotations for training (either in terms of keypoints [@liu12eccv; @liu14eccv] or bounding boxes [@chen14cvpr; @parkhi12cvpr; @sun11iccv_art; @ukita12cvpr; @zhang13iccv; @zhang14eccv; @lin15cvpr; @gkioxari15iccv]). Our framework instead does not require annotations of part positions nor extent, and automatically learns from Google Images instead.\
[**Learning from image search engines.**]{} Several works have tried to learn visual models from training samples collected automatically from image search engines [@fergus05iccv; @vijayanarasimhan:cvpr08; @li10ijcv; @schroff11pami; @liq13cvpr; @chen13iccv; @divvala14lcvpr; @chen15cvpr; @novotny16eccvworkshop]. Most of them tackle image classification [@fergus05iccv; @vijayanarasimhan:cvpr08; @li10ijcv; @schroff11pami; @liq13cvpr] and develop algorithms to find good training samples and learn iteratively.
Some works try to learn object class detectors from the web [@chen13iccv; @divvala14lcvpr; @chen15cvpr]. Chen *et al.* [@chen15cvpr] considers only objects and not parts. LEVAN [@divvala14lcvpr] leverage Google Books Ngrams to discover all appearance variations of an object class, then trains an object detector with a separate component per variation. While some of the components happen to represent parts (e.g. ‘horse head’), these are treated just like other independent components (at the same level as ‘jumping horse’ and ‘racing horse’). NEIL [@chen13iccv] mines web images to discover common sense relationships between object classes (e.g. ‘car is found in raceway’), including also some part-of relations (e.g. ‘wheel is part of car’). Importantly, both LEVAN and NEIL learn simple object class detectors, consisting of ‘root filters’ only. Instead we learn more complex, structured models of object classes, which include semantic part appearance models and their spatial arrangements within the object, conditioned on object viewpoint. Note how [@chen13iccv; @divvala14lcvpr; @chen15cvpr] do not report quantitative localization results for part detection. Finally, we believe our work is complementary to [@chen13iccv; @divvala14lcvpr; @chen15cvpr]. The frameworks of [@chen13iccv; @divvala14lcvpr] could provide a list of which parts belong to which object class, which could be passed on to our technique to learn more complex models. Moreover, our part models could be used in combination with the strong R-CNN root filters learned from the web by [@chen15cvpr] (analog to sec. \[sec:det\]). The concurrent work [@novotny16eccvworkshop] is the only other one to learn semantic part models from the web. It learns the structure of objects in an embedding space where geometric relationships are implicitly conveyed by non-semantic mid-level parts. Instead, we learn explicit relationships on the semantic parts themselves, based on object and parts instances automatically mined from web images.
{width="\textwidth"}
\[fig:schema\]
Overview of our approach {#sec:overview}
========================
We present here an overview of our framework for automatically learning compositional semantic part models. We learn these models for each object class separately. For each class, we use Google Images to collect images of the object under several pre-defined viewpoints, and images of its parts. We use the part samples to train initial part appearance models, which are later used to learn the connection between the parts and the whole object. Learning this association is the key to our compositional part models.
For each class, we learn part appearance models $\mathcal{A}$, part location models $\mathcal{L}$, and object viewpoint classifiers $\mathcal{V}$. Our framework operates in four stages: $\mathcal{T}_0 - \mathcal{T}_3$ (fig. \[fig:schema\]). In the first stage $\mathcal{T}_0$, we collect training samples from Google Images (objects and parts, fig. \[fig:front\]). Then, we iteratively learn the components of our part models, each time learning from harder examples: ($\mathcal{T}_1$) images containing only parts from Google, ($\mathcal{T}_2$) part examples mined from object images from Google, and ($\mathcal{T}_3$) part examples mined from object images from the PASCAL VOC 2010 dataset [@everingham10ijcv]. Every stage is fully automatic and does not require human intervention.
For each object part, we learn one appearance model and $V$ location models, one for each viewpoint in our predefined set. A single location model is not sufficient to capture the position of a part with respect to the object, as this is strongly affected by viewpoint changes. For example, the front view of a bicycle has one wheel on the bottom-center of the bicycle, while a bicycle from the side has two wheels on the bottom left and right (fig. \[fig:front\], right).
For simplicity, in the rest of the paper we use superscripts to indicate the stage a model component is trained at. For example, our part appearance model $\mathcal{A}^2$ is trained at stage $\mathcal{T}^2$, while $\mathcal{A}^3$ at stage $\mathcal{T}^3$.\
[**$\mathbf{\mathcal{T}^0}$: Collecting data.**]{} Our framework queries Google Images for images of an object under canonical viewpoints and images of each of its parts (sec. \[sec:data\]). These images are biased towards simple representations, in a uniform background and they reliably contain the wanted object (or part). However, one image may contain multiple instances or objects not appearing nicely in the centre (fig \[fig:segm\]). It would be better if each object/part instance would be enclosed in a tight bounding-box. Bounding-boxes around parts help learning accurate appearance models as they exclude background pixels, and around objects they help learning accurate part location models as they provide a stable coordinate frame common to all instances. We therefore devise a simple, yet effective algorithm to fit a tight bounding-box around each part/object instance (sec. \[sec:data\]). Finally, we consider each bounding-box as a separate image, obtaining our initial training set. We denote with $\mathcal{O}_j \in \mathcal{O}$ the set of images of the object under viewpoint $v_j$ and with $\mathcal{P}_i \in \mathcal{P}$ the set of images of part $p_i$.\
[**$\mathbf{\mathcal{T}^1}$: Learning from Google’s easy examples.**]{} For each part $p_i$, our framework learns an appearance model $\mathcal{A}^1_i$ on the part images $\mathcal{P}_i$ (sec. \[sec:app\]). These are the easiest samples, as in these images the part often appears isolated from the object and against a clean background (fig. \[fig:front\] left).\
[**$\mathbf{\mathcal{T}^2}$: Learning from Google’s harder examples.**]{} In this stage our framework moves on to object images $\mathcal{O}$. It learns part location models $\mathcal{L}^2$ and updates all part appearance models by using additional samples from $\mathcal{O}$ (sec. \[sec:app\]). Moreover, it trains an object viewpoint classifier $\mathcal{V}^2$ on $\mathcal{O}$ (sec. \[sec:view\]).
For each viewpoint $v_j$ and part $p_i$, it learns $\mathcal{L}^2_{i, j}$. The key idea is to run $\mathcal{A}^1_i$ on the object images $\mathcal{O}_j$. The top-scored part detections are likely to be correct and, importantly, they are now [*localized within an object image*]{}. Therefore, they provide valuable training samples for the location of the part within the object (sec. \[sec:loc\]). The intuition here is that objects captured under the same viewpoint have parts in similar spatial arrangements. For example, all horses from the side have the head on the left side of the image, mostly on top (fig. \[fig:schema\]). Since all objects in $\mathcal{O}_j$ are in the same viewpoint $v_j$, correct detections of a part will be consistently found at similar locations across different object instances. Subsequently, the framework mines part samples automatically from each $\mathcal{O}_j \in \mathcal{O}$ using $\mathcal{A}^1_i$ and the corresponding $\mathcal{L}^2_{i,j}$ (sec. \[sec:mining\]). The process looks for detections that have a high score according to $\mathcal{A}^1_i$ and are at the right location according to $\mathcal{L}^2_{i,j}$. By combining these two source of information, we consistently discover correct part samples. Finally, the framework uses these samples to update part appearance models to $\mathcal{A}^2_i$. Note how these new samples are more difficult than the ones in $\mathcal{T}^1$, since come from images showing whole objects and against natural backgrounds. Lastly, the framework trains an object viewpoint classifier $\mathcal{V}^2$ on $\mathcal{O}$, by using each set of images $\mathcal{O}_j \in \mathcal{O}$ as training set for viewpoint $v_j$ (sec. \[sec:view\]). Finally, note how at this stage the framework has trained a complete, rich model (part appearance $\mathcal{A}^2$, part location $\mathcal{L}^2$, object viewpoint $\mathcal{V}^2$) entirely and automatically from Google Images (fig. \[fig:schema\], top).\
[**$\mathbf{\mathcal{T}^3}$: Learning from PASCAL VOC.**]{} In this final stage the framework refines all $\mathcal{A}^2$ and $\mathcal{L}^2$ using even more difficult training samples from another domain (sec. \[sec:app\], \[sec:loc\]). These samples are mined automatically from the PASCAL VOC dataset, which contains photographs depicting challenging objects in natural scenes, often occluded or truncated (fig. \[fig:schema\], bottom). These are much harder than the ones in stage $\mathcal{T}^2$, where each image had a single whole object. The framework mines positives as in step $\mathcal{T}^2$, but using the updated $\mathcal{A}^2$ instead of the initial $\mathcal{A}^1$ (sec. \[sec:mining\]). Similarly to other works [@parkhi12cvpr; @lin15cvpr], we only search for parts inside the ground-truth bounding-boxes of the object class (which are provided with PASCAL VOC). We call this set $\mathcal{O}^{voc}$. Furthermore, we call the set of mined positives $\mathcal{P}^{voc}$. In order to use our viewpoint-specific location models, we need to determine the viewpoint of the images in $\mathcal{O}^{voc}$. We automatically predict $v_j$ for each image in $\mathcal{O}^{voc}$ using the object viewpoint classifier $\mathcal{V}^2$. Finally, after mining new positives, the framework finally trains final location models $\mathcal{L}^3_{i, j}$ and part appearance models $\mathcal{A}^3_i$ (sec. \[sec:mining\]).
The components of our approach {#sec:tech}
==============================
We detail below the components of our approach. In sec. \[sec:data\] we describe our data collection mechanism. In sec. \[sec:app\], \[sec:loc\] and \[sec:view\] we describe how to train $\mathcal{A}$, $\mathcal{L}$ and $\mathcal{V}$, respectively. In sec. \[sec:mining\] we then present our procedure to automatically mine new part instances from objects.
Data collection and preprocessing {#sec:data}
---------------------------------
This section describes how we download part and object images from Google and how we fit a tight bounding-box around each part/object instance in them.\
[**Querying Google Images.**]{} We collect images of an object under multiple viewpoints and of its parts (fig. \[fig:front\]). We keep the top 100 retrieved images for each object viewpoint and the top 25 for each object part. We observed these numbers to produce good, clean images. Collecting more than 25 part images sometimes delivers spurious images without the part, which would introduce noise in the learning process.
For each object class, we use the names of its parts as listed in the PASCAL-Part Dataset [@chen14cvpr] and the viewpoint names specified by PASCAL VOC 2010 [@everingham10ijcv] (*front, back, left, right*). As [*left*]{} and [*right*]{} is not a level of granularity satisfied by Google Images yet, we query for a generic [*side*]{} viewpoint (fig. \[fig:front\] right-top) and then automatically split the retrieved images into left and right subsets. In order to do this, we first augment the image set by mirror flipping all images horizontally, and then we cluster them by minimizing the intra-cluster HOG compactness, similarly to [@felzenszwalb10pami].
[**Fitting bounding-boxes.**]{} As mentioned in sec. \[sec:overview\], we want to fit a tight bounding-box around each object/part instance. These bounding-boxes help learning accurate appearance and location models for the parts. Fortunately, Google Images results are biased towards whole objects in a uniform background (fig. \[fig:segm\]a) and unoccluded. These are easy to localize. We formulate this task as a pixel labelling problem, where each pixel $\phi_i$ can take a label $l_i \in \{0, 1\}$ (background or foreground). We aim at finding the best labelling $\psi^* = \argmin_L E(\mathcal{\psi})$. Similar to other segmentation works [@Rother04-tdfixed; @kuettel12cvpr; @RosenfeldICCV11], we define an energy function: $$\small
\vspace{-1mm}
E(\mathcal{\psi}) = \sum_i M_i(l_i) + \sum_i G_i(l_i) + \alpha \sum_{i, j} V(l_i, l_j),
\label{eq:Seg}$$ where, the pairwise potential $V$ encourages smoothness by penalising neighbouring pixels taking different labels and the unary potential $G_i$ evaluates how likely a pixel $i$ is to take label $l_i$ according to an appearance model which consists of two GMMs [@Rother04-tdfixed] (foreground and background). Inspired by [@kuettel12cvpr], we produce an initial rough estimate $M$ of which pixels lie on the object, and use it both to estimate the appearance models $G$, and as a unary potential of its own. We do this in an unsupervised manner, based purely on the spatial distribution of object proposals [@uijlings13ijcv] in the image (fig. \[fig:segm\]b). We define the likelihood $M_i(1)$ of a pixel $i$ to be foreground as the number of proposals that contain it, divided by the total number of proposals in the image ($M_i(0) = 1-M_i(1)$). The idea is that if a pixel is contained in many proposals, then it is likely to belong to the object. As in GrabCut [@Rother04-tdfixed], we iteratively alternate between minimizing the energy (eq. \[eq:Seg\]) to obtain a segmentation, and updating the appearance models based on this segmentation. After a few iterations this process converges and we fit a tight bounding-box around each connected component in the final segmentation (fig. \[fig:segm\]d). We apply this procedure to all images collected from Google, obtaining the initial training set of object images $\mathcal{O}$ and part images $\mathcal{P}$ (treating each bounding-box as a separate image).\
[**Part proposals.**]{} We generate class-independent part proposals inside each image in $\mathcal{O}$ using [@uijlings13ijcv]. As observed by Zhang *et al.* [@zhang14eccv], these proposals achieve low recall on small semantic parts. In order to overcome this difficulty, we changed the standard settings of [@uijlings13ijcv] to return smaller proposals and increase part recall. This results in about 2000 proposals per object image in $\mathcal{O}$, likely to cover all parts. In the rest of the paper we use $\mathcal{W}$ to refer to the set all part proposals over all object images.
![*Examples of the steps of our procedure to fit bounding-boxes to object/part instances in web images (sec. \[sec:data\]). (a) is the input image; (b) is the initial rough foreground estimate $M$; (c) is the output of the segmentation process; and (d) are the bounding-boxes fit to connected components in the segmentation. Note how the two images ‘horse leg’ and ‘bicycle front’ have multiple part/object instances and our method is able to fit a separate bounding-box around each of them.*](figures/segmentation_old.pdf){width="\textwidth"}
\[fig:segm\]
Training part appearance models $\mathcal{A}$ {#sec:app}
---------------------------------------------
Each stage of our learning framework updates the part appearance models of the object class. We describe here how these models are trained at each stage.
[**Stage $\mathcal{T}^1$.**]{} We train $\mathcal{A}^1$ on the image set $\mathcal{P}$, containing simple part images. As appearance model we use a convolutional neural network (CNN) and train it to distinguish between the $P$ parts. More specifically, we start from AlexNet pre-trained on the ImageNet 2012 classification challenge [@krizhevsky12nips] and replace its original 1000-way $\mathtt{fc8}$ classification layer with a $P$-way $\mathtt{fc8}$ layer. We then finetune the whole network for part classification on the images in $\mathcal{P}$. Note how $\mathcal{P}$ only contains 25 samples per part. In order to avoid overfit we use a learning rate of $10^{-4}$ and apply early stopping ($1000$ iterations, 5 epochs). Higher learning rates cause the parameters to vary abruptly over iterations, whereas $10^{-4}$ results in a smooth learning curve. At test time we use the softmax at layer $\mathtt{fc8}$ to predict how likely a proposal is to contain each of the parts. At all times we use the publicly available CNN implementation [@jia13caffe]. [**Stage $\mathcal{T}^2$.**]{} At this stage we learn $\mathcal{A}^2$ on a larger training set of examples from both part images and part samples automatically mined from object images using the appearance model from stage $\mathcal{T}^1$ and the location model from stage $\mathcal{T}^2$ (sec. \[sec:mining\], fig. \[fig:schema\]). As appearance model we train a similar CNN to the one of stage $\mathcal{T}^1$, but with a difference: we use a richer $(P+1)$-way $\mathtt{fc8}$ layer, where the additional output is used to classify background patches. Note how by mining for positive part instances in object images (sec. \[sec:mining\]) we indirectly discover negative proposals (those with intersection-over-union ($IoU$) [@everingham10ijcv] $\le 0.3$ with mined positives). [**Stage $\mathcal{T}^3$.**]{} In the last stage we train $\mathcal{A}^3$ on the harder image set $\mathcal{P}^{voc}$, using as training samples parts automatically mined from $\mathcal{O}^{voc}$ using the part detector from stage $\mathcal{T}^2$ (sec. \[sec:mining\]). As appearance model we train a CNN as in $\mathcal{T}^2$.
{width="100.00000%"}
\[fig:loc\]
Learning part location models $\mathcal{L}$ {#sec:loc}
-------------------------------------------
The appearance model $\mathcal{A}$ scores part proposals in an image based on their appearance only. We build location models to capture complementary knowledge about likely positions and scales of the object parts within the the coordinate frame of the object. In stage $\mathcal{T}^2$ we learn the location models purely from Google Images and in stage $\mathcal{T}^3$ we adapt them to a different domain. In this subsection we use $\mathcal{W}_j$ to refer to the set of all part proposals in object images $\mathcal{O}_j$.\
[**Part training samples.**]{} For each viewpoint $v_j$ and part $p_i$, we learn a separate location model $\mathcal{L}_{i,j}$ from a set of training proposals $\mathcal{W}_{j,i} \in \mathcal{W}_j$ likely to contain part $p_i$. We describe here how we acquire these part samples $\mathcal{W}_{i, j}$ at stage $\mathcal{T}^2$, i.e. from object images $\mathcal{O}_j$. The key idea is to run the part detector $\mathcal{A}^1_i$ on these images and retain the top-scored part detections. As these detections are localized within an object image, they provide examples of the location of the part within the object. More precisely, for each image we score all part proposals with the appearance model $\mathcal{A}^1_i$, perform non-maximum suppression, and pick up to 3 detections per image (the top scored ones, if they score above a minimum confidence threshold). These detections form $\mathcal{W}_{i, j}$. This way of picking detections strikes a good trade-off between keeping all correct locations, but without including too many false-positives. At $\mathcal{T}^3$, we enrich the sample set $\mathcal{W}_{j, i}$ with the top detections produced by running the appearance model $\mathcal{A}^3_i$ on $\mathcal{O}^{voc}$. More specifically, each of these detections gets assigned to the $\mathcal{W}_{j, i}$ of viewpoint predicted by $\mathcal{V}^2$. These new samples are used to train the refined location models $\mathcal{L}^2_{i,j}$.\
[**Training a location model.**]{} The location model $\mathcal{L}_{i,j}$ scores on an input part proposal $w'$ by the density of the training set $\mathcal{W}_{j,i}$ at $w'$ $$\small
\vspace{-1mm}
\mathcal{L}_{i,j}(w') = \frac{1}{|\mathcal{W}_{j,i}| \cdot h}\sum_{w \in \mathcal{W}_{j,i}} K \left(\frac{D(w',w)}{h}\right)
\label{eq:lm}$$ where $D(w',w)$ is distance between two proposals: $D(w',w) = 1 - \frac{w' \cap w}{w' \cup w}$ and $K(u)$ is the uniform density function $K(u) = \frac{1}{2} \mathds{1}(|u| \le 1)$. In this formulation $\mathcal{L}_{i,j}(w')$ has an intuitive interpretation as the percentage of proposals in $\mathcal{W}_{j,i}$ which are close to $w'$ ($IoU < h$). If many training proposals are near $w'$, then $\mathcal{L}_{i,j}(w')$ is large, indicating that $w'$ is likely to contain the part. Conversely, if only a few proposal are near $w'$, then $\mathcal{L}_{i,j}(w')$ is small, indicating it is more likely to cover a background patch. The bandwidth $h$ controls the degree of smoothing and in our experiments we set it to $h=0.5$.
Note how $D(w',w)$ compares part proposals across different images. For this to be meaningful, $D$ operates in a coordinate frame common to all images in $\mathcal{O}_j$ (by normalizing it by the average width and height of all images). This normalization is specific to a viewpoint $v_j$, so it preserves its aspect-ratio.\
[**Model behaviour.**]{} Fig. \[fig:loc\] shows examples of some location models learned at stage $\mathcal{T}^2$. Thanks to the way we build them, our location models are robust to errors in the training set: correct training samples tend to cluster around the right locations of a part, whereas incorrect ones tend to scatter across the whole object. This results in strong peaks at the correct locations in the model, with only lower values everywhere else (e.g. headlight for Bus Front). Moreover, note how our location model is suitable for a variety of cases. Unique parts of rigid objects form unimodal distributions (Bicycle saddle, Car license plate), while Bicycle wheel and Horse leg form bimodal ones. Even in the hard case of highly movable parts of deformable object classes (e.g. Cat tail), the model learns that they can appear over broader regions and spreads the density accordingly.
Training the viewpoint classifier $\mathcal{V}^2$ {#sec:view}
-------------------------------------------------
During stage $\mathcal{T}^2$ we train classifiers $\mathcal{V}^2$ on $\mathcal{O}$ for predicting the viewpoint of the object in an image. We train a CNN to distinguish between the four viewpoints (*front, back, left, right*) for which we collected object images from Google in sec. \[sec:data\] (fig. \[fig:front\] right). We used these images to train $\mathcal{V}^2$ and, as for $\mathcal{A}$ (sec. \[sec:app\]), we took the CNN pre-trained on the ImageNet classification challenge and replaced its original 1000-way $\mathtt{fc8}$ classification layer with a 4-way $\mathtt{fc8}$. During stage $\mathcal{T}^3$, the viewpoint classifier is useful to select an appropriate location model for object images $\mathcal{O}^{voc}$. Given an input image, we select the viewpoint with the highest probability. Note that the PASCAL VOC 2010 dataset has manual viewpoint annotations for some objects ($\sim60\%$ for the classes we consider). We use these annotations in sec. \[sec:res\_view\] to evaluate how well our viewpoint classifier $\mathcal{V}^2$ works.
Mining for new part instances {#sec:mining}
-----------------------------
In stages $\mathcal{T}^2$ and $\mathcal{T}^3$ we mine for new part instances in $\mathcal{O}$ and $\mathcal{O}^{voc}$, respectively. For simplicity, we describe the process to mine from $\mathcal{O}$. Given each set of images $\mathcal{O}_j \in \mathcal{O}$ showing viewpoint $v_j$, we mine positives for part $p_i$ using the appearance model $\mathcal{A}^1_i$ and the location model $\mathcal{L}^2_{i, j}$. For each image, we score all its part proposals with $\mathcal{A}^1_i$ and $\mathcal{L}^2_{i, j}$, perform non-maximum suppression and keep only the proposals with high score. We repeat this for all $\mathcal{O}_j \in \mathcal{O}$ and obtain our final set of new samples. Importantly, we mine for new part instances within object bounding boxes only. Even though the initial appearance models $\mathcal{A}^1$ were trained on 25 samples only, they still manage to localize new part instances, as the search space is very limited. Moreover, this process is able to mine new instances that look significantly different than those in the initial set of easy examples $\mathcal{P}$ (e.g. a frontoparallel wheel against a white background vs a out-of-plane rotated wheel on an actual car, fig. \[fig:res\]), as new part instances can be selected if at the right location according to $\mathcal{L}$ , even when $\mathcal{A}$ is not confident about them.
Mining from $\mathcal{O}^{voc}$ is analogous, but requires an extra step, where we use the viewpoint classifier $\mathcal{V}^2$ (sec. \[sec:view\]) to predict the otherwise unknown viewpoints of objects $\mathcal{O}^{voc}$.
Experiments and conclusions {#sec:exp}
===========================
Datasets
--------
We evaluate our framework and all its intermediate stages on the recent PASCAL-Part dataset [@chen14cvpr], which augments PASCAL VOC 2010 [@everingham10ijcv] with pixelwise semantic part annotations. For evaluation we fit a bounding-box to each part segmentation mask. Finally, the dataset contains a `train` and a `validation` subsets. We mine new part instances from `train` in stage $\mathcal{T}^3$, and measure the performance of our framework on `validation`. We verified by using a near-duplicate detector that none of the images we collected from Google Images are in Pascal Parts. We evaluate on nine diverse object classes ([*bicycle, bird, bus, car, cat, cow, dog, horse, sheep*]{}), three parts each (table \[table:res\]). We treat each leg as a separate instance, rather than grouping them into a ‘super-part’ (as done by [@chen14cvpr]). Note how previous works evaluating on PASCAL-Parts consider fewer classes/parts [@chen14cvpr; @wang15cvpr; @hariharan15cvpr] and operate in a fully supervised scenario (training from manual part location annotations).
Viewpoint prediction {#sec:res_view}
--------------------
We evaluate here our viewpoint classifier $\mathcal{V}^2$ trained purely from Google Images (sec. \[sec:view\]). We compare it against a viewpoint classifier $\mathcal{V}^{FS}$ trained using manual annotations from PASCAL VOC 2010 `train`. In both cases we use the same CNN model and training procedure (sec. \[sec:view\]). We evaluate both classifiers in terms of accuracy on `validation` (table \[table:view\]). Results show that our viewpoint classifier $\mathcal{V}^2$ considerably outperforms the fully supervised classifier $\mathcal{V}^{FS}$. Results are not surprising, as objects in the PASCAL VOC dataset appear often truncated or occluded and sometimes labelled with the wrong viewpoint. Instead, the images from Google have clean objects with well defined viewpoints (fig. \[fig:front\]). Moreover, objects appear as a whole, leading to better prediction performance.
\[table:res\]
Part localization {#sec:part_loc}
-----------------
In this section we evaluate how good our part models are at localizing parts in novel images. We evaluate part localization in terms of average precision (AP) on the PASCAL-Part `validation` set (which was never seen by our learning procedure). As in [@chen14cvpr], a part is considered correctly localized if is has an $IoU \geq 0.4$ with a ground-truth bounding-box.
Our location models are conditioned on the object viewpoint, which is however unknown for the objects in `validation`. We apply our viewpoint classifier $\mathcal{V}^2$ on all objects in `validation` to select what location model to use. When detecting parts we use a linear combination of the score given by the appearance and location models ($\mathcal{A}+\mathcal{L})$. We evaluate each component of our system at each stage of the learning, from $\mathcal{T}^0$ to $\mathcal{T}^3$. For comparison, we trained additional part appearance models $\mathcal{A}^0$ directly on images retrieved by Google, *before* fitting a bounding-box around each training instance (sec. \[sec:data\]).
{width="\textwidth"}
\[fig:res\]
Results are presented in table \[table:res\] and fig. \[fig:res\]. Naively using images as returned by Google Images ($\mathcal{A}^0$) leads to an AP of only 12.9. This reveals how challenging is the task of localizing object parts on a dataset like PASCAL VOC. Our refined models $\mathcal{A}^1$ perform already better and improve $\mathcal{A}^0$ by +2.8, showing that our polishing process is useful and provides cleaner examples that lead to better performance. The really interesting leap however is achieved in stage $\mathcal{T}^2$ when our framework associates the object to its parts and learns the connection. More precisely, learning the location of the parts under the different viewpoints increases AP to 18.3 ($\mathcal{A}^1 +\mathcal{L}^2$). Using this information to mine for more part instances and update the appearance model improves performance even further to 22.0 ($\mathcal{A}^2$). Ultimately, the combination of these two models ($\mathcal{A}^2 +\mathcal{L}^2$) brings the performance to 23.5. This is almost double the initial AP of naively training detectors directly from part images ($\mathcal{A}^0$). Importantly, at this point we have a complete class model (appearance, location, viewpoint) trained entirely and automatically from Google Images. Finally, if we additionally migrate to the PASCAL VOC domain ($\mathcal{T}^3$) and adapt appearance and location models to it, the performance further improves to a final AP of 27.2 ($\mathcal{A}^3+\mathcal{L}^3$). The steady improvement exhibited from stage $\mathcal{T}^0$ to $\mathcal{T}^3$ by our incremental learning framework demonstrates its potential in learning complex part models automatically.
For reference, we train two fully supervised models: $\mathcal{A}^{FS1}$ and $\mathcal{A}^{FS2}$. The former uses manual part bounding-boxes from PASCAL-Part `train`, while the latter also uses the part instances from $\mathcal{T}^0$ collected from the web. Similarly to sec. \[sec:app\], for each class we took AlexNet CNN and replaced its last layer with a $(P+1)$-way $\mathtt{fc8}$ ($P$ parts an one background class). These models provide an upper-bound on what can be achieved by any weakly supervised procedure on this dataset. Our final part detector achieves 27.2 AP, which is 62% of the performance of $\mathcal{A}^{FS1}$. This is very encouraging, given that we train without part location annotations, whereas $\mathcal{A}^{FS1}$ trains from 15K part bounding-boxes on PASCAL-Part `` (covering our 9 classes with 3 parts each). These take a lot of time as the parts are small and difficult to annotate. Interestingly, $\mathcal{A}^{FS2}$ performs a little worse than $\mathcal{A}^{FS1}$, despite being trained from more data. We attribute this to the difference between the type of images in PASCAL-Part and on the web.
[**Comparison to LEVAN [@divvala14lcvpr] and NEIL [@chen13iccv].**]{} LEVAN and NEIL learn detectors from the web and their original papers do not present quantitative evaluation on part detection. Nonetheless, a few of their models represent semantic parts. We evaluate them in this section, using their DPM models [@felzenszwalb10pami] they released online [@divvala14lcvpr; @chen13iccv]. LEVAN learns multi-component object class detectors. The components within each object model are labelled with a name, like ’horse jumping’ or ’horse head’. We downloaded the detectors for our nine object classes and selected all components matching our parts. For example, to detect *car licence plate* we run the models labelled as ’plate\_car\_super3’ and ’plate\_car\_super6’. NEIL, instead, learns a collection of separate models, some representing object classes and others part classes, as well as part-of relation between them. We downloaded all NEIL’s models and selected those in a part-of relation with any of the object classes we consider. This only matches one part *wheel*. Only one generic wheel model is available, not associated to a specific object class.
We run all these part models on PASCAL-Part `validation` and show results in table \[table:res\] (rightmost two columns). Note how most of the parts we consider are missing from the components learned by NEIL and LEVAN. On the few parts that they learned, our part detectors outperform LEVAN and NEIL by a large margin. The main reason is that their models are trained from part instances downloaded from the web with no (or minimal) refinement: LEVAN uses instances similar to our $\mathcal{T}^0$, and NEIL uses something in between our $\mathcal{T}^1$ and $\mathcal{T}^2$. A second reason is that LEVAN and NEIL’s components are based on simple HOG features, which are weaker than CNNs.
Object detection {#sec:det}
----------------
In this section we augment the R-CNN object class detector [@girshick14cvpr] with our part models. The standard R-CNN detector scores each object proposal $w$ in an image with a root filter $\mathcal{R}$ covering the whole object. Inspired by [@felzenszwalb10pami] we add a collection of parts arranged in a deformable configuration: $$\small
\vspace{-2mm}
\text{score}(w) = \mathcal{R}(w) + \sum_{i}^{N} \max_{w' \in \Upsilon} (\alpha_{i} \cdot \mathcal{A}_i(w') + \beta_{i} \cdot \mathcal{L}_{i, \mathcal{V}(w)}(w'))
\label{eq:obj_det}$$ where $\Upsilon$ is the set of part proposals inside $w$. For each part $i$, the $\max$ operation looks for the best fitting proposal $w'\in \Upsilon$ according to the part appearance model ($\mathcal{A}_i$) and location model ($\mathcal{L}_i$), measuring how likely part $i$ is to appear at the location $w'$. As we have a separate location model per viewpoint, we use our classifier $\mathcal{V}$ to select which one to use on $w$. We use the same set of proposals for both objects and parts (sec. \[sec:data\]). We set the weights $\alpha$ and $\beta$ by cross-validation on `train`. This overall object class model is similar to [@felzenszwalb10pami], but instead of using Gaussian part location models, we have a full probability distributions given by kernel density estimators (eq. \[eq:lm\]).
\[table:objLoc\]
We train the root filer on PASCAL-Parts `train` as in [@girshick14cvpr]. The other elements of the model are learned from the web and PASCAL-Parts using our technique (sec. \[sec:tech\]), i.e. $\mathcal{A}^3$ as part filters, $\mathcal{L}^3$ as location models and $\mathcal{V}^2$ as viewpoint classifiers. No manual part location annotations is used for training. We report object detection results on `validation` of PASCAL VOC 2010 and 2012, in terms of AP (table \[table:objLoc\]). Compared to using the R-CNN root filter alone (R-CNN$_1$), adding parts increases its performance by 2-3% on all classes, and on both test sets (R-CNN$_1+$ parts). This shows that our part models can help object class detection, even when added to an already strong fully supervised detector like R-CNN. This is an interesting result, especially considering that our parts are designed to be [*semantic*]{}, as opposed to discriminative arbitrary patches [@endres13cvpr; @felzenszwalb10pami].
For a fully fair comparison, we also train another R-CNN model on object instances from both PASCAL-Part [*and*]{} the images we downloaded from the web ($\mathcal{T}^0$). Interestingly, training using this additional data decreases performance by 0.8% (R-CNN$_2$). Again, we attribute this to the difference between the type of images in PASCAL-Part and on the web.
Conclusions {#sec:conl}
===========
We presented a technique for learning part-based models from the web. It operates by collecting object and part instances and by automatically connecting them in an incremental learning procedure. Our models encompass the appearance of parts and their spatial arrangement on the object, specific to each viewpoint. We reported results on the challenging PASCAL-Parts which show that our technique is able to learn good part detectors from the web. Finally, we demonstrated the value of our part models by enriching the R-CNN object detector with parts.
[**Acknowledgments.**]{} Support by ERC Starting Grant VisCul.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Overestimation of the maximum action-value is a well-known problem that hinders Q-Learning performance, leading to suboptimal policies and unstable learning. Among several Q-Learning variants proposed to address this issue, Weighted Q-Learning (WQL) effectively reduces the bias and shows remarkable results in stochastic environments. WQL uses a weighted sum of the estimated action-values, where the weights correspond to the probability of each action-value being the maximum; however, the computation of these probabilities is only practical in the tabular setting. In this work, we provide the methodological advances to benefit from the WQL properties in Deep Reinforcement Learning (DRL), by using neural networks with Dropout Variational Inference as an effective approximation of deep Gaussian processes. In particular, we adopt the Concrete Dropout variant to obtain calibrated estimates of epistemic uncertainty in DRL. We show that model uncertainty in DRL can be useful not only for action selection, but also action evaluation. We analyze how our novel *Deep Weighted Q-Learning* algorithm reduces the bias w.r.t. relevant baselines and provide empirical evidence of its advantages on several representative benchmarks.'
bibliography:
- 'wdqn.bib'
---
Introduction {#s:intro}
============
Reinforcement Learning (RL) aims at learning how to take optimal decisions in unknown environments by solving credit assignment problems that extend in time. In order to be sample efficient learners, agents are required to constantly update their own beliefs about the world, about which actions are good and which are not. Temporal difference (TD) [@sutton1998reinforcement] and off-policy learning are the constitutional elements of this kind of behavior. TD allows agents to bootstrap their current knowledge to learn from a new observation as soon as it is available. Off-policy learning gives the means for exploration and enables experience replay [@lin1991er]. Q-Learning [@qlearning] implements both paradigms.
Algorithms based on Q-learning are, in fact, driving Deep Reinforcement Learning (DRL) research towards solving complex problems and achieving super-human performance on many of them [@mnih2015human; @hessel2018rainbow]. Nonetheless, Q-Learning is known to be positively biased [@hasselt2010double] since it learns by using the maximum over the - noisy - bootstrapped TD estimates. This overoptimism can be particularly harmful in stochastic environments and when using function approximation [@thrun1993issues], notably also in the case where the approximators are deep neural networks [@hasselt2016ddqn]. Systematic overestimation of the action-values coupled with the inherently high variance of DRL methods can lead to incrementally accumulate errors, causing the learning algorithm to diverge.
Among the possible solutions, the Double Q-Learning algorithm [@hasselt2010double] and its DRL variant Double DQN [@hasselt2016ddqn] tackle the overestimation problem by disentangling the choice of the target action and its evaluation. The resulting estimator, while achieving superior performance in many problems, is negatively biased [@hasselt2013estimating]. Underestimation, in fact, can lead in some environments to lower performance and slower convergence rates compared to standard Q-Learning [@deramo2016wql; @lan2020maxmin]. Overoptimism, in general, is not uniform over the state space and may induce to overestimate the value of arbitrary bad actions, throwing the agent completely off. The same holds true, symmetrically, for overly pessimistic estimates that might undervalue a good course of action. Ideally, we would like DRL agents to be aware of their own uncertainty about the optimality of each action, and be able to exploit it to make more informed estimations of the expected return. This is exactly what we achieve in this work.
We exploit recent developments in Bayesian Deep Learning to model the uncertainty of DRL agents using neural networks trained with dropout variational inference . We combine, in a novel way, the dropout uncertainty estimates with the Weighted Q-Learning algorithm [@deramo2016wql], extending it to the DRL settings. The proposed *Deep Weighted Q-Learning* algorithm, or *Weighted DQN* (WDQN), leverages an approximated posterior distribution on Q-networks to reduce the bias of deep Q-learning. WDQN bias is neither always positive, neither negative, but depends on the state and the problem at hand. WDQN only requires minor modifications to the baseline algorithm and its computational overhead is negligible on specialized hardware.
The paper is organized as follows. In Section \[s:bg\] we define the problem settings, introducing key aspects of value-based RL. In Section \[s:overest\] we analyze in depth the problem of estimation biases in Q-Learning and sequential decision making problems. Then, in Section \[s:wdqn\], we first discuss how neural networks trained with dropout can be used for Bayesian inference in RL and, from that, we derive the WDQN algorithm. In Section \[s:exp\] we empirically evaluate the proposed method against relevant baselines on several benchmarks. Finally, we provide an overview of related works in Section \[s:related\], and we draw our conclusions and discuss future works in Section \[s:end\].
Preliminaries {#s:bg}
=============
A Markov Decision Process (MDP) is a tuple $\langle \mathcal{S}, \mathcal{A}, \mathcal{P}, \mathcal{R}, \gamma \rangle$ where $\mathcal{S}$ is a state space, $\mathcal{A}$ is an action space, $\mathcal{P}: \mathcal{S} \times \mathcal{A} \rightarrow \mathcal{S}$ is a Markovian transition function, $\mathcal{R}: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ is a reward function, and $\gamma \in [0,1]$ is a discount factor. A sequential decision maker ought to estimate, for each state $s$, the optimal value $Q^*(s,a)$ of each action $a$, i.e., the expected return obtained by taking action $a$ in $s$ and following the optimal policy $\pi^*$ afterwards. We can write $Q^*$ using the *Bellman optimality equation* [@bellman1954theory] $$\begin{gathered}
Q^*(s, a) = {}\\
\mathbb{E}\left[ r_{t+1} + \gamma \max_{a^\prime} Q^*(s_{t + 1}, a^\prime) \vert s_t = s, a_t = a\right].\end{gathered}$$
#### (Deep) Q-Learning
A classical approach to solve finite MDPs is the Q-Learning algorithm [@qlearning], an off-policy value-based RL algorithm, based on TD. A Q-Learning agent learns the optimal value function using the following update rule: $$\begin{gathered}
\label{eq:q-update}
Q(s_t, a_t) \gets Q(s_t, a_t) + \alpha \left( y_t^{QL} - Q(s_t, a_t) \right),\end{gathered}$$ where $\alpha$ is the learning rate and, following the notation introduced by [@hasselt2016ddqn], $$\label{eq:ql}
y_t^{QL} = r_{t} + \gamma \max_{a} Q(s_{t+1}, a).$$ The popular Deep Q-Network algorithm (DQN) [@mnih2015human] is a variant of Q-Learning designed to stabilize off-policy learning with deep neural networks in highly dimensional state spaces. The two most relevant architectural changes to standard Q-Learning introduced by DQN are the adoption of a replay memory, to learn offline from past experience, and the use of a target network, to reduce correlation between the current model estimate and the bootstrapped target value.
In practice, DQN learns the Q-values online, using a neural network with parameters $\bm{\theta}$, sampling the replay memory, and with a target network whose parameters $\bm{\theta}^{-}$ are updated to match those of the online model every $C$ steps. The model is trained to minimize the loss $$L(\bm{\theta}) = \underset{{\langle s_i,a_i,r_i,s_i^\prime \rangle \sim m}}{\mathbb{E}}\left[\left(y_i^{DQN} - Q(s_i, a_i; \bm{\theta})\right)^2\right],$$ where $m$ is a uniform distribution over the transitions stored in the replay buffer and $y^{DQN}$ is defined as $$\label{eq:dqn}
y^{DQN}_i = r_i + \gamma \max_{a} Q(s_i^\prime, a; \bm{\theta}^{-}).$$
#### Double DQN
Among the many studied improvements and extensions of the baseline DQN algorithm [@wang2016dueling; @schaul2016prioritized; @distributional2017bellemare; @hessel2018rainbow], Double DQN (DDQN) [@hasselt2016ddqn] reduces the overestimation bias of DQN with a simple modification of the update rule. In particular, DDQN uses the target network to decouple action selection and evaluation, and estimates the target value as $$y^{DDQN}_i = r_i + \gamma Q(s_i^\prime, \operatorname*{argmax}_{a}Q\left(s_i^\prime, a; \bm{\theta}); \bm{\theta}^{-}\right).$$ DDQN improves on DQN converging to a more accurate approximation of the value function, while maintaining the same model complexity and adding a minimal computational overhead.
Estimation biases in Q-Learning {#s:overest}
===============================
Choosing a target value for the Q-Learning update rule can be seen as an instance of the Maximum Expected Value (MEV) estimation problem for a set of random variables, here the action-values $Q(s_{t+1},{}\cdot{})$. Q-Learning uses the Maximum Estimator (ME) [^1] to estimate the maximum expected return and exploits it for policy improvement. It is well-known that ME is a positively biased estimator of MEV [@smith2006optimizer]. The divergent behaviour that may occur in Q-Learning, then, may be explained by the amplification over time effect on the action-value estimates caused by the overestimation bias, which introduces a positive error at each update [@hasselt2010double]. Double Q-Learning [@hasselt2010double], on the other hand, learns two value functions in parallel and uses an update scheme based on the Double Estimator (DE). It is shown that DE is a negatively biased estimator of MEV, which helps to avoid catastrophic overestimates of the Q-values. The DDQN algorithm, introduced in Section \[s:bg\], is the extension of Double Q-Learning to the DRL settings. In DDQN, the target network is used as a proxy of the second value function of Double Q-Learning to preserve sample and model complexity.
In practice, as shown by @deramo2016wql and @lan2020maxmin, the overestimation bias of Q-Learning is not always harmful, and may also be convenient when the action-values are significantly different among each other (e.g., deterministic environments with a short time horizon, or small action spaces). Conversely, the underestimation of Double Q-Learning is effective when all the action-values are very similar (e.g., highly stochastic environments with a long or infinite time horizon, or large action spaces). In fact, depending on the problem, both algorithms have properties that can be detrimental for learning. Unfortunately, a prior knowledge about the environment is not always available and, whenever it is, the problem may be too complex to decide which estimator should be preferred. Given the above, it is desirable to use a methodology which can robustly deal with heterogeneous problems.
Weighted Q-Learning
-------------------
@deramo2016wql proposes the Weighted Q-Learning (WQL) algorithm, a variant of Q-Learning based on the therein introduced Weighted Estimator (WE). The WE estimates MEV as the weighted sum of the random variables sample means, weighted according to their probability of corresponding to the maximum. Intuitively, the amount of uncertainty, i.e., the entropy of the WE weights, will depend on the nature of the problem, the number of samples and the variance of the mean estimator (critical when using function approximation). WE bias is bounded by the biases of ME and DE [@deramo2016wql].
The target value of WQL can be computed as $$\label{eq:wql}
y_t^{WQL} = r_t + \gamma \sum_{a \in \mathcal{A}} w_a^{s_{t+1}}Q(s_{t+1},a),$$ where $w_a^{s_{t+1}}$ are the weights of the WE and correspond to the probability of each action-value being the maximum: $$\label{eq:wqlweights}
w^{s}_a = P\left( a = \operatorname*{argmax}_{a^\prime} Q(s, a^\prime)\right).$$ The update rule of WQL can be obtained replacing $y_t^{QL}$ with $y_t^{WQL}$ in Equation \[eq:q-update\]. The weights of WQL are estimated in the tabular setting assuming the sample means to be normally distributed.
WE has been studied also in the Batch RL settings, with continuous state and action spaces, by using Gaussian Process regression [@d2017estimating].
Deep Weighted Q-Learning {#s:wdqn}
========================
Q-network parameters $\bm{\theta}$, dropout rates $p_1, \dots, p_L$, a policy $\pi$, replay memory $\mathcal{D}$ $\bm{\theta}^- \gets \bm{\theta}$ Initialize memory $\mathcal{D}$. Select an action $a_i$ according to some policy $\pi$ given the distribution over action-value functions $Q(s, {}\cdot{}; \bm{\theta}, \bm{\omega})$ Execute $a_t$ and add $\langle s_t, a_t, r_t, s_{t+i} \rangle$ to $\mathcal{D}$ Sample a mini-batch of transitions $\{\langle s_i,a_i,r_i,s_i^\prime \rangle, i= 1,\dots, M\}$ from $\mathcal{D}$ Take $K$ samples from $Q(s_i,{}\cdot{}; \bm{\theta}^-, \bm{\omega})$ by performing $K$ stochastic forward passes Use the samples to compute the WDQN weights (Eq.\[eq:weights\]) and targets (Eq.\[eq:wdqn\]) Perform a SGD step on the WDQN loss (Eq. \[eq:wdqnloss\]) Eventually update $\bm{\theta}^-$
A natural way to extend the WQL algorithm to the DRL settings is to consider the uncertainty over the model parameters using a Bayesian approach. Among the possible solutions to estimate uncertainty, bootstrapping has been the most successful in RL problems, with BootstrappedDQN (BDQN) [@osband2016boot; @osband2018deeprandomized] achieving impressive results in environments where exploration is critical. On the other hand, using bootstrapping necessitates significant modifications to the baseline DQN architecture and requires to train a model for each sample of the approximate posterior distribution. This limits the number of samples available considerably and is a major drawback in using BDQN to approximate the WE weights. Using dropout, conversely, does not impact model complexity and allows to compute the weights of the WE by using infinitely many samples.
In the following we first introduce how neural networks trained with dropout can be used for approximated Bayesian inference and discuss how this approach has been used with success in RL problems. Then, we propose a novel approach to exploit the uncertainty over the model parameters for action evaluation, adapting the WE to the DRL settings. Finally we analyze a possible shortcoming of the proposed method and identify a solution from the literature to address it.
Bayesian inference with dropout
-------------------------------
Dropout [@srivastava2014dropout] is regularization technique used to train large neural networks by randomly dropping units during learning. In recent years, dropout has been analyzed from a Bayesian perspective [@kingma2015vardrop; @gal2016dropout], and interpreted as a variational approximation of a posterior distribution over the neural network parameters. In particular, @gal2016dropout show how a neural network trained with dropout and weight decay can bee seen as an approximation of a deep Gaussian process [@damianou2013deepgauss]. The result is a theoretically grounded interpretation of dropout and a class of Bayesian neural networks (BNNs) that are cheap to train and can be queried to obtain uncertainty estimates. In fact, a single stochastic forward pass through the BNN can be interpreted as taking a sample from the model’s predictive distribution, while the predictive mean can be computed as the average of multiple samples. This inference technique is known as *Monte Carlo* (MC) *dropout* and can be efficiently parallelized in modern GPUs.
A straightforward application of Bayesian models to RL is Thompson Sampling (TS) [@thompson1933likelihood]. TS is an exploration technique that aims at improving the sample complexity of RL algorithms by selecting actions according to their probability of being optimal given the current agent’s beliefs. A practical way to use TS in Deep Reinforcement Learning is to take a single sample from a Q-network trained with dropout and select the action that corresponds to the maximum sampled action-value [@gal2016dropout]. TS based on dropout achieves superior data efficiency compared against naïve exploration strategies, such as $\varepsilon$-greedy, both in sequential decision making problems [@gal2016dropout; @stadie2015exploration] and contextual bandits [@riquelme2018deepbandits; @collier2018deepcontextual]. Furthermore, dropout has been successfully used in model-based RL, to estimate the agent’s uncertainty over the environment dynamics [@gal2015bayespilco; @kahn2017uncertaintyaware; @malik2019calibrated].
Here we focus on the problem of action evaluation. We show how to use approximate Bayesian inference to evaluate the WE estimator by introducing a novel approach to exploit uncertainty estimates in DRL. Our method empirically reduces Q-Learning bias, is grounded in theory and simple to implement.
Weighted DQN
------------
Let $Q({}\cdot{}, {}\cdot{}; \bm{\theta}, \bm{\omega})$ be a BNN with weights $\bm{\theta}$ trained with a Gaussian prior and Dropout Variational Inference to learn the optimal action-value function of a certain MDP. We indicate with $\bm{\omega}$ the set of random variables that represents the dropout masks, with $\bm{\omega}_i$ the $i$-th realization of the random variables and with $\bm{\Omega}$ their joint distribution: $$\begin{gathered}
\label{eq:dropmasks}
\bm{\omega} = \{ \omega_{lk} : l= 1, \dots, L , k=1,\dots,K_{l}\},\\
\begin{aligned}
&\omega_{lk} \sim \text{Bernoulli}(p_l), &\bm{\omega}_i \sim \bm{\Omega}(p_1, \dots, p_L),
\end{aligned}\end{gathered}$$ where $L$ is the number of weight layers of the network and $K_l$ is the number of units in layer $l$.
Consider a sample $q(s, a)$ the MDP return, obtained taking action $a$ in $s$ and following the optimal policy afterwards. Following the GP interpretation of dropout of @gal2016dropout, we can approximate the likelihood of this *observation* as a Gaussian such that $$\label{eq:gaussianapprox}
q(s,a) \sim \mathcal{N}\left( Q(s, a; \bm{\theta}, \bm{\omega}); \tau^{-1}\right),$$ where $\tau$ is the model precision. We can approximate the predictive mean of the process, and the expectation over the posterior distribution of the Q-value estimates, as the average of $T$ stochastic forward passes through the network: $$\begin{gathered}
\mathbb{E}\left[Q(s, a; \bm{\theta}, \bm{\omega})\right] \approx {}\\
\hat{Q}_T(s, a; \bm{\theta}) = \frac{1}{T}\sum_{t=1}^T Q(s, a; \bm{\theta}, \bm{\omega}_t).\end{gathered}$$ $\hat{Q}_T(s, a; \bm{\theta})$ is the BNN prediction of the action-values associated to state $s$ and action $a$. A similar, more computationally efficient, approximation can be obtained through *weight averaging*, which consists in scaling the output of the neurons in layer $l$ by $1-p_l$ during training and leaving them unchanged at inference time. We indicate this estimate as $\tilde{Q}(s, a; \bm{\theta})$ and we use it for action selection during training.
The model *epistemic* uncertainty, i.e., the model uncertainty over its parameters, can be measured similarly as the sample variance across $T$ realizations of the dropout random variables: $$\begin{gathered}
\label{eq:qvar}
{Var}\left[Q(s, a; \bm{\theta}, \bm{\omega})\right] \approx {}\\
\frac{1}{T}\sum_{t=1}^T Q(s, a; \bm{\theta}, \bm{\omega}_t)^2 - \left(\frac{1}{T}\sum_{t=1}^T Q(s, a; \bm{\theta}, \bm{\omega}_t)\right)^2.\end{gathered}$$
As shown in @gal2016dropout the predictive variance can be approximated with the variance of the estimator in Eq. \[eq:qvar\] plus the model inverse precision $\tau^{-1}$.
We can estimate the probability required to calculate the WE in a similar way. Given an action $a$, the probability that $a$ corresponds the maximum expected action-value can be approximated as the number of times in which, given $T$ samples, the sampled action-value of $a$ is the maximum over the number of samples $$\begin{aligned}
\label{eq:weights}
w^{s}_a(\bm{\theta}) & = P\left( a = \operatorname*{argmax}_{a^\prime} Q(s, a^\prime; \bm{\theta}, \bm{\omega})\right) \nonumber\\
& \approx \frac{1}{T}\sum_{t=1}^T \left\llbracket a = \operatorname*{argmax}_{a^\prime} Q(s, a^\prime; \bm{\theta}, \bm{\omega}_t) \right\rrbracket,\end{aligned}$$ where $\llbracket \dots \rrbracket$ are the Iverson brackets ($\llbracket P \rrbracket$ is $1$ if $P$ is true, $0$ otherwise). The weights can be efficiently inferred in parallel with no impact in computational time.
We can define the WE given the Bayesian target Q-network estimates using the obtained weights as: $$\label{eq:wdqn}
y^{WDQN}_i = r_i + \gamma \sum_{a \in \mathcal{A}} w^{s^\prime_i}_a(\bm{\theta}^-) \hat{Q}_T(s_i^\prime, a; \bm{\theta}^-).$$
Finally we report for completeness the loss minimized by WDQN, where the parameter updates are backpropagated using the dropout masks: $$\begin{aligned}
\label{eq:wdqnloss}
L(\bm{\theta}) & = {} \nonumber\\
&\quad \underset{{{ \vcenter{ \Let@ \restore@math@cr \default@tag
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\ialign{\hfil$\m@th\scriptstyle##$&$\m@th\scriptstyle{}##$\hfil\crcr
\langle s_i,a_i,r_i,s_i^\prime \rangle & \sim m \\ \bm{\omega}_i & \sim \bm{\Omega}\crcr
} }}}}{\mathbb{E}}\left[\left(y_i^{WDQN} - Q(s_i, a_i; \bm{\theta}, \bm{\omega_i})\right)^2\right]\nonumber\\
& \quad + \lambda \sum_{l=1}^L \left\| \bm{\theta}_l \right\|^2_2,\end{aligned}$$ where $ \bm{\theta}_l$ are the weights of layer $l$ and $\lambda$ the weight decay coefficient. Using weight decay is necessary for the the variational approximation being valid. The complete WDQN algorithm is reported in Algorithm \[alg:wdqn\].
Concrete Dropout
----------------
The dropout probabilities are variational parameters and influence the quality of the approximation. Ideally, they should be tuned to maximize the log-likelihood of the observations using a validation method. This is clearly not possible in RL where the available samples, and the underlying distribution generating them, change as the policy improves. In fact, using dropout with a fixed probability might lead to a poor uncertainty estimates [@osband2016boot; @osband2018deeprandomized; @gal2017concrete]. Concrete Dropout [@gal2017concrete] mitigates this problem by using a differentiable continuous relaxation of the Bernoulli distribution and learning the dropout rate from data.
In practice, this means that the distribution of the dropout random variables $\omega_{lk}$ becomes $$\label{eq:concretemasks}
\displaystyle \omega_{lk} = \sigma \left( {\beta} \left( \log p_l - \log (1-p_l) + \log u - \log (1-p_l) \right) \right),
$$ where $\beta$ is a temperature parameter (fixed at $\beta = 10$), $u$ is a uniform random variable $u \sim \mathcal{U}(0, 1)$ and $\sigma({}\cdot{})$ is the *sigmoid* function. With this formulation the sampling procedure becomes differentiable and the loss in Eq. \[eq:wdqnloss\] can be rewritten as: $$\begin{aligned}
\label{eq:concreteloss}
L(\bm{\theta}, \bm{p}) & = {} \nonumber\\
&\quad \underset{{{ \vcenter{ \Let@ \restore@math@cr \default@tag
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\ialign{\hfil$\m@th\scriptstyle##$&$\m@th\scriptstyle{}##$\hfil\crcr
\langle s_i,a_i,r_i,s_i^\prime \rangle & \sim m \\ \bm{\omega}_i & \sim \bm{\Omega}\crcr
} }}}}{\mathbb{E}}\left[\left(y_i^{WDQN} - Q(s_i, a_i; \bm{\theta}, \bm{\omega_i})\right)^2\right]\nonumber\\
& \quad + \sum_{l=1}^L \left(\lambda \left\| \bm{\theta}_l \right\|^2_2 - \zeta K_l \mathcal{H}(p_l)\right),\end{aligned}$$ where $\zeta$ is a dropout regularization coefficient, $\mathcal{H}(p)$ is the entropy of a Bernoulli distribution with parameter $p$ and $K_l$ the number of neurons in layer $l$.
Experiments {#s:exp}
===========
{width="0.95\linewidth"}
In this section we compare WDQN against the standard DQN algorithm and its DDQN variant to measure the effect of the different estimators on the quality of the learned value functions and policies. First we run a proof of concept experiment on the Lunar Lander environment [@brockman2016gym]. Then we perform an in depth analysis on Asterix, one of the Atari games in the Arcade Learning Environment (ALE) [@bellemare13arcade] where DQN is known to overestimate the action-values [@hasselt2016ddqn]. Finally we test WDQN in three environments of the MinAtar benchmark [@young19minatar]. WDQN achieves a more accurate estimation of the expected return which results, on average, in better policies.
All the agents are evaluated after each epoch of training using the greedy policy, for Asterix we follow the evaluation protocol of @mnih2015human. For each experiment we report both the average cumulative reward at each evaluation step and prediction of the expected return, compared against the real discounted return obtained by the agents. Even if TS is the natural choice for WDQN, all the algorithms, unless explicitly stated, use an $\varepsilon$-greedy policy to guarantee the fairness of the comparison. For WDQN the greedy action is selected during training as the action corresponding the maximum Q-value estimated with weight averaging, while during evaluation we take the action with the highest probability of being optimal computed as in Eq. \[eq:weights\]. We found WDQN to be robust to the number of dropout samples used to compute the WE (e.g., $T \geq 30$). We used low values for the weight decay term, as in [@farebrother2018generalization], and tuned the dropout regularization coefficient for the problem at hand. The algorithms and the experimental setup have been developed using the open-source RL libraries MushroomRL [@deramo2020mushroomrl] and OpenAI Gym [@brockman2016gym]. Full description of the complete experimental setup and further results for each experiment can be found in the appendix.
Lunar Lander
------------
![Learning curves on **Lunar Lander**. On the **left** the evaluation scores of each epoch ($20$k steps). On the **right** the maximum action-value estimated by each agent at the initial state of the MDP; the dashed lines are the real obtained discounted return. The results shown are the average of 20 independent runs, shaded areas are 95% confidence intervals. The curves are smoothed using a moving average of 10 epochs.[]{data-label="fig:lunar"}](imgs/lunar){width="0.99\linewidth"}
Lunar Lander is an MDP from the Gym collection. In order to solve the environment, the agent has to control the thrusters of a spacecraft to safely land at a specific location. In order to make prediction and control more challenging, we increased the stochasticity of the environment adding a 10% probability of repeating the last executed action, instead of the one selected by the agent. The three agents use the same exact network architecture and hyperparameters, the only difference is that WDQN uses Concrete Dropout in each hidden layer. Figure \[fig:lunar\] shows the learning curves of the three agents, with WDQN achieving a significantly higher average reward and prediction accuracy.
![Comparison of **WDQN** against **WDQN with TS** and vanilla **DQN with Concrete Dropout** (DropDQN). On the **left** the evaluation scores. On the **right** the maximum estimated action-value at the starting screen of the game; dashed lines are the real obtained discounted return. The results shown are the average across 3 different random seeds, the shaded areas are 95% confidence intervals. The curves are smoothed using a moving average of 10 epochs.[]{data-label="fig:ablation"}](imgs/atari_var){width=".99\linewidth"}
Asterix
-------
{width="0.95\linewidth"}
We use the same neural network and hyperparameters of [@mnih2015human], except for the optimizer that we replace with Adam [@kingma2014method], following more recent best practices [@hessel2018rainbow; @dabney2018qrdqn]. For WDQN we use Concrete Dropout only in the fully connected layer after the convolutional block. We use sticky-actions, with a probability of repeating last action of $25\%$, as proposed by [@machado18arcade]. Figure \[fig:asterix\] shows the result of the comparison in terms of average reward and prediction accuracy. Differently from what observed by [@hasselt2016ddqn], DQN does not diverge in our setting. In fact, while widely overestimating the discounted return, DQN achieves better sample-complexity and higher cumulative reward w.r.t. DDQN.
WDQN, on the other hand, is clearly the less biased option and despite the slow start - that can be explained by the regularization introduced by dropout - reaches the highest average reward. After an initial phase, the average entropy of the WE weights remains almost constant with two possible interpretations: 1) in many states there is not a single action that dominates the others, 2) the agent is not able to completely resolve its uncertainty over the action-value functions.
In order to gather more insights on WDQN and on how each component (namely the policy and the MEV estimator) influence the observed behavior, in Figure \[fig:ablation\] we compare it against a baseline DQN agent trained with Concrete Dropout (DropDQN in the figure) and a version of WDQN that uses a TS policy as in [@gal2016dropout]. In this setting TS is not beneficial and underperforms. DropDQN is clearly the worst performer of the tested algorithms, indicating that the use of WE is indeed beneficial and suggesting that WDQN may improve with ad hoc hyperparameter tuning.
MinAtar
-------
MinAtar [@young19minatar] is a RL testbed with environments mimicking the dynamics of games from ALE, but with a simplified state representation. MinAtar implements also sticky actions and difficulty ramping [@machado18arcade]. The results of the experiment are shown in Figure \[fig:minatar\]. We use the same convolutional neural network and hyperparameters of [@young19minatar], but we replace RMSProp with the Adam optimizer [^2] with learning rate $1\mathrm{e}{-4}$. For WDQN we choose the dropout regularization coefficient with random search and use the same value in the three games.
WDQN achieves higher average reward in Breakout, performs on par with the other two algorithms in Seaquest and is more sample efficient in Freeway. For what concerns the estimation of the expected return, WDQN shows a much lower bias in the first epochs of the learning procedure and converge to more accurate estimates in Breakout and Freeway, while performing similarly to the other algorithms in Seaquest. Since the number of actions and the regularization coefficient are the same among the three environments, it is clear that the entropy of the WE weights heavily depends on the dynamics and input space of the problem.
Related works {#s:related}
=============
The Weighted Q-Learning algorithm [@deramo2016wql] joins previous works in the attempt to address the bias of Q-Learning. Double Q-Learning [@hasselt2010double] firstly showed how the use of DE, to underestimate the action-values, can be helpful to stabilize learning in highly stochastic environments. Bias-corrected Q-Learning [@lee2013bias] improves the learning stability of Q-Learning by subtracting a quantity to the target value that depends on the standard deviation of the reward. The Double Weighted Q-Learning algorithm [@zhang2017doubleweighted] (which, despite the name, is not related to WQL) uses instead a combination of ME and DE to balance between the overestimation and underestimation of the two estimators. DE and WE have also been introduced in Batch RL as alternatives to the ME used in Fitted Q-Iteration [@ernst2005tree], respectively in the Double Fitted Q-Iteration and Weighted Fitted Q-Iteration algorithms [@d2017estimating]. Among the three, Weighted Fitted Q-Iteration is the only algorithm able to handle continuous action spaces.
Overestimation of the action-values can be even more problematic in the DQN algorithm [@mnih2015human], due to the high variance typical of DRL approaches. The Double DQN algorithm [@hasselt2016ddqn] introduces the use of DE in DQN, and shows better estimate of action-values and superior performance w.r.t. vanilla DQN. Other variants of DQN addressing the overestimation problem are based on exploiting multiple estimates of the action-values. For instance, Averaged DQN [@anschel2017averaged] controls the variance of the estimation by averaging the target of the update over an arbitrary number of previous checkpoints of the target network. Then, the recent Maxmin DQN [@lan2020maxmin] reduces the bias keeping several estimates of the action-values in parallel, and computing the target values using the maximum of the minimums of each action-value estimate. Overestimation is also detrimental in actor-critic DRL algorithms, e.g. Deep Deterministic Policy Gradient (DDPG) [@lillicrap2015continuous]. Notably, the Twin Delayed DDPG (TD3) algorithm [@fujimoto2018addressing] is a variant of DDPG exploiting several tricks, e.g. a clipped version of Double Q-Learning update, that significantly improves the approximation of the value function and the overall performance.
Conclusion and future works {#s:end}
===========================
We present WDQN, a new value-based Deep Reinforcement Learning algorithm, that extends the Weighted Q-Learning algorithm to work in environments with an highly dimensional state representation. WDQN is a principled and robust method to exploit uncertainty in DRL to accurately estimate the maximum action-value in a given state. We empirically support our claims by showing that WDQN consistently reduces the bias of DQN across different tasks. Our results corroborate the findings of previous works [@gal2016dropout; @gal2017concrete], confirming that dropout can be used successfully for approximate Bayesian inference in DRL. Future works may explore the combination of WDQN with the other, orthogonal, DQN extensions and may attempt to adapt the WDQN approach to other techniques modelling uncertainty in deep neural networks.
Appendix {#appendix .unnumbered}
========
0.1in The appendix is organized as follows:
- in Section \[a:mev\] we further discuss the problem of estimating the maximum expected value (MEV) of a set of random variables;
- in Section \[a:exp\] we provide, together with some additional empirical results, details on the hyperparameters and experimental setup used to evaluate the agents.
Estimating the Maximum Expected Value {#a:mev}
=====================================
Let $X = \lbrace X_1, \dots, X_M \rbrace$ be a set of $M \geq 2$ independent random variables with unknown means $\mu_i = \mathbb{E}[X_i]$, and $S = \bigcup_{i=1}^M S_i$ be a set of noisy samples, where $S_i$ corresponds to the subset of samples drawn from the unknown distribution of the random variable $X_i$. The problem of estimating the MEV consists in using the sample means $\hat{\mu}_i(S)$ to approximate the largest true mean $\max_i \mu_i = \mu_*$. Unfortunately, for many distributions (e.g. Gaussian), no unbiased estimator of the MEV exists [@blumenthal1968estimation; @hasselt2013estimating].
#### Maximum Estimator
The Maximum Estimator (ME) can be defined as $$\label{eq:me}
\hat{\mu}_*^{\text{ME}}(S) = \max_i \hat{\mu}_i \approx \mu_*.$$ The ME is the most straightforward way of estimating the MEV and is positively biased [@smith2006optimizer].
#### Double Estimator
The Double Estimator (DE) is a negatively biased estimator for the MEV [@stone1974cross; @hasselt2013estimating]. DE splits the set of samples $S$ into two independent sets $S^A$ and $S^B$, with sample means $\hat{\mu}_i(S^A)$ and $\hat{\mu}_i(S^B)$. Using an cross-validation approach, the DE is computed by selecting the random variable with the highest sample mean in one set, and computing its sample mean from samples in the other one. The resulting values are $\hat{\mu}_{b^*}(S^A)=\hat{\mu}_{\operatorname*{argmax}_i \hat{\mu}_i(S^B)}(S^A)$ and $\hat{\mu}_{a^*}(S^B)=\hat{\mu}_{\operatorname*{argmax}_i \hat{\mu}_i(S^A)}(S^B)$. Finally, the DE can be obtained as $$\label{eq:de}
\hat{\mu}_*^{\text{DE}}(S) = \frac{\hat{\mu}_{b^*}(S^A) + \hat{\mu}_{a^*}(S^B)}{2}.$$
#### Weighted Estimator
@deramo2016wql introduced the Weighted Estimator (WE), defined as $$\label{eq:we}
\hat{\mu}_*^{\text{WE}}(S) = \sum_{i=1}^M w_i^S \hat{\mu}_i(S),$$ where $w_i^S$ is the probability of $\hat{\mu}_i$ being the maximum sample mean. Since the distribution of the sample mean of a random variable with an unknown distribution is itself unknown, @deramo2016wql approximates the sample means as normally distributed random variables by the central limit theorem. WE can be both positively and negatively biased, and its bias is bounded by the one of ME and DE: $$\label{e:bias}
- \frac{1}{2}\left(\sqrt{\sum_{i=1}^M\frac{\sigma_i^2}{S_i^A}} + \sqrt{\sum_{i=1}^M\frac{\sigma_i^2}{S_i^B}}\right) \leq
\text{Bias}(\hat{\mu}_*^{\text{DE}}(S)) \leq \text{Bias}(\hat{\mu}_*^{\text{WE}}(S)) \leq \text{Bias}(\hat{\mu}_*^{\text{ME}}(S)) \leq
\sqrt{\frac{M-1}{M} \sum_{i=1}^M\frac{\sigma_i^2}{S_i}},$$ where $\sigma_i$ is the variance of the $i$-th random variable.
Furthermore, WE has often a variance that is empirically smaller to the the variance of the other two estimators [@deramo2016wql]. A variant of WE that extends to infinite random variables has been proposed in @d2017estimating.
Experiments details {#a:exp}
===================
In this section we provide details on the experimental setup used for the empirical evaluation of the proposed method.
For the implementation of the algorithms and the simulation environments we rely on the following open-source libraries:
- MushroomRL [@deramo2020mushroomrl];
- Gym [@brockman2016gym];
- ALE [@bellemare13arcade];
- MinAtar [@young19minatar];
- PyTorch [@paszke2019pytorch].
Asterix
-------
For the experiment on the Asterix Atari game we use the settings of @mnih2015human, but we add sticky actions [@machado18arcade], i.e., a probability of $p_\text{repeat} = 25\%$ of repeating the action executed at the previous frame instead of the one selected by the agent. For the agents we use the same neural network of @mnih2015human with, as already mentioned, the Adam [@kingma2014method] optimizer. The additional hyperparameters introduced by WDQN are tuned with a - small - random search. Concrete Dropout is used only on the neurons of the last hidden layer. Agents are evaluated during training every $1$M frames, with the 30 no-op starting condition. During evaluation the episode length is capped at 30 minutes. Table \[tab:atari\] reports a list of additional relevant hyperparameters.
**Hyperparameter** **Value**
--------------------------------------------------------- -------------------
Optimizer Adam
Learning rate $5\mathrm{e}{-5}$
Adam epsilon ($\epsilon_\text{Adam}$) $0.01/32$
Batch size $32$
Loss function Huber
Training frequency $4$
Target network update frequency $10000$
Min. memory size $50000$
Max. memory size $1\text{M}$
Discount factor ($\gamma$) $0.99$
Initial exploration rate ($\varepsilon_{start}$) $1.0$
Final exploration rate ($\varepsilon_{end}$) $0.1$
Exploration steps $1\text{M}$
Evaluation exploration rate ($\varepsilon_{test}$) $0.001$
$\text{MC dropout samples}^*$ $100$
$\text{Weight decay coefficient}$ $(\lambda)^*$ $1\mathrm{e}{-6}$
$\text{Dropout regularization coefficient}$ $(\zeta)^*$ $5\mathrm{e}{-4}$
$\text{Initial dropout rate}$ $(p)^*$ $0.5$
: DQN hyperparameters on Asterix. Hyperparameters marked with a $*$ are used only for WDQN.
\[tab:atari\]
Lunar Lander
------------
**Hyperparameter** **Value**
--------------------------------------------------------- ---------------------
Units per layer $[100,100]$
Activation $relu$
Optimizer Adam
Learning rate $3\mathrm{e}{-4}$
Batch size $32$
Loss function MSE
Training frequency $1$
Target network update frequency $300$
Min. memory size $250$
Max. memory size $10000$
Discount factor ($\gamma$) $0.99$
Initial exploration rate ($\varepsilon_{start}$) $1.0$
Final exploration rate ($\varepsilon_{end}$) $0.01$
Exploration steps $1000$
Evaluation exploration rate ($\varepsilon_{test}$) $0.0$
$\text{MC dropout samples}^*$ $50$
$\text{Weight decay coefficient}$ $(\lambda)^*$ $1\mathrm{e}{-6}$
$\text{Dropout regularization coefficient}$ $(\zeta)^*$ $2.5\mathrm{e}{-3}$
$\text{Initial dropout rate}$ $(p)^*$ $0.2$
: DQN hyperparameters on Lunar Lander. Hyperparameters marked with a $*$ are used only for WDQN.
\[tab:lunar\]
For Lunar Lander [@brockman2016gym] we limit the length of an episode (during both evaluation and training) at 1000 steps. To increase the complexity of the problem, we make the environment stochastic using sticky actions (see previous subsection) with a $p_\text{repeat} = 10\%$. We use a small neural network with only two hidden layers. For WDQN, we use Concrete Dropout on each hidden layer. Table \[tab:lunar\] reports the hyperparameters used to train the agents.
MinAtar
-------
MinAtar [@young19minatar] offers a collection of environments resembling games from the Atari Learning Environment [@bellemare13arcade]. The state representation of MinAtar environments is a matrix with multiple channels, where each channel gives specific information about some aspects of the environment (e.g., position and speed of a moving object). MinAtar implements the ALE modifications suggested by @machado18arcade, i.e., sticky actions and difficulty ramping.
We use the same hyperparameters and convolutional neural network of @young19minatar, but we use Adam for training. The WDQN dropout regularization coefficient is tuned with random search: we select the value providing qualitatively better learning curves and keep it fixed across the three games. Table \[tab:minatar\] shows the relevant hyperparameters.
**Hyperparameter** **Value**
--------------------------------------------------------- -------------------
Optimizer Adam
Learning rate $1\mathrm{e}{-4}$
Batch size $32$
Loss function Huber
Training frequency $1$
Target network update frequency $1000$
Min. memory size $5000$
Max. memory size $100000$
Discount factor ($\gamma$) $0.99$
Initial exploration rate ($\varepsilon_{start}$) $1.0$
Final exploration rate ($\varepsilon_{end}$) $0.1$
Exploration steps $100000$
Evaluation exploration rate ($\varepsilon_{test}$) $0.0$
$\text{MC dropout samples}^*$ $100$
$\text{Weight decay coefficient}$ $(\lambda)^*$ $1\mathrm{e}{-6}$
$\text{Dropout regularization coefficient}$ $(\zeta)^*$ $1\mathrm{e}{-4}$
$\text{Initial dropout rate}$ $(p)^*$ $0.1$
: DQN hyperparameters on Minatar environments. Hyperparameters marked with a $*$ are used only for WDQN.
\[tab:minatar\]
For WDQN, we run an additional experiment to asses the impact of the dropout regularization coefficient. We set the initial dropout rate at $p=0.5$ (which corresponds to the maximum entropy) and test the agents using regularization coefficients of different magnitude. The results, reported in Figure \[fig:reg\], show how higher levels of regularization generally correspond to higher entropy of the weights used to compute the WE.
{width="0.95\linewidth"}
[^1]: Details about the estimators considered in this section are provided in the appendix.
[^2]: We found Adam to provide significantly more stable learning across all the agents.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With the development of range sensors such as LIDAR and time-of-flight cameras, 3D point cloud scans have become ubiquitous in computer vision applications, the most prominent ones being gesture recognition and autonomous driving. Parsimony-based algorithms have shown great success on images and videos where data points are sampled on a regular Cartesian grid. We propose an adaptation of these techniques to irregularly sampled signals by using continuous dictionaries. We present an example application in the form of point cloud denoising.'
author:
-
- |
\
Tel-Aviv University\
-
bibliography:
- 'pc\_references.bib'
title: |
Cloud Dictionary:\
Coding and Modeling for Point Clouds
---
Introduction
============
Sparse coding for point clouds
==============================
Dictionary learning
===================
Experiments
===========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An array is [*row-Latin*]{} if no symbol is repeated within any row. An array is [*Latin*]{} if it and its transpose are both row-Latin. A [*transversal*]{} in an $n\times n$ array is a selection of $n$ different symbols from different rows and different columns. We prove that every $n \times n$ Latin array containing at least $(2-\sqrt{2}) n^2$ distinct symbols has a transversal. Also, every $n \times n$ row-Latin array containing at least $\frac14(5-\sqrt{5})n^2$ distinct symbols has a transversal. Finally, we show by computation that every Latin array of order $7$ has a transversal, and we describe all smaller Latin arrays that have no transversal.'
author:
- 'Darcy Best[^1]'
- Kevin Hendrey
- 'Ian M. Wanless[^2]'
- 'Tim E. Wilson'
- 'David R. Wood[^3]'
title: Transversals in Latin arrays with many distinct symbols
---
Introduction
============
This paper deals with square arrays of symbols. By an [*entry*]{} of such an array $A$, we mean a triple $(i,j,A_{ij})$ where $A_{ij}$ is the symbol in cell $(i,j)$ of $A$. A [*partial transversal of length $k$*]{} in an array is a selection of $k$ entries no pair of which agree in any of their three coordinates. A [*transversal*]{} of an $n\times n$ array is a partial transversal of length $n$ and a [*near transversal*]{} is a partial transversal of length $n-1$. An array is [*Latin*]{} if no symbol appears more than once in any row or column. Thus, an $n\times n$ Latin array may contain anywhere from $n$ to $n^2$ distinct symbols. If it has just $n$ distinct symbols, then it is a [*Latin square*]{}. Transversals of Latin squares were first studied to construct mutually orthogonal Latin squares. Since then they have garnered a lot of interest in their own right and lead to several famous long-standing conjectures (see [@Wan11] for a survey).
For even orders $n$ there are at least $n^{n^{3/2}(1/2-o(1))}$ Latin squares that do not have transversals [@CW17]. However, for $n\times n$ Latin arrays, as the number of distinct symbols increases, there must come a point beyond which it becomes impossible to avoid transversals. This paper is motivated by the question of when this threshold occurs. Let $\ell(n)$ be the least positive integer such that $\ell(n){\geqslant}n$ and every Latin array of order $n$ with at least $\ell(n)$ distinct symbols contains a transversal. This function was introduced by Akbari and Alipour [@AkbariAlipour04], who calculated $\ell(n)$ for $n{\leqslant}4$ and showed that $\ell(5){\geqslant}7$ and $\ell(2^k-2)>2^k$ for every integer $k>2$. Counter-intuitively, every Latin square of order 5 contains a transversal, but there is a Latin array of order 5 with six symbols and no transversal. Hence, it is not always true that increasing the number of symbols increases the number of transversals. Nevertheless, $\ell(n)$ is well defined since an $n\times n$ Latin array with $n^2$ different symbols certainly has a transversal. Akbari and Alipour put forward the following conjectures:
\[conj:ln\] For every integer $n{\geqslant}3$, we have $\ell(n){\leqslant}n^2/2$.
\[conj:noconst\] For every integer $c$, there exists a positive integer $n$ such that $\ell(n)>n+c$.
Up until this point, it was unknown whether there is some constant $c<1$ such that $\ell(n){\leqslant}cn^2$ for every integer $n>1$. In Sections \[sec:probabilistic\] and \[sec:Latin-arrays\], we provide two independent proofs of such a result. The proof in Section \[sec:Latin-arrays\] gives a better bound, but the other is of independent interest since it demonstrates an entirely different (probabilistic) approach. In Section \[sec:small-values\], we determine $\ell(n)$ exactly for $n{\leqslant}7$.
On first glance, Conjecture \[conj:ln\] seems very generous and that maybe $\ell(n)$ even has a linear upper bound. However, the problem is deceptively hard, and the following observation gives some hint as to why.
\[prop:ln-implies-brualdi\] Let $k$ be a non-negative integer. If $\ell(n) {\leqslant}2kn+n-k^2-k$ for all $n$, then every Latin square of order $n$ has a partial transversal of length $n-k$.
Let $L$ be any Latin square of order $n$. Let $M$ be a Latin array of order $n+k$, which has $L$ as the top-left $n \times n$ subarray and all remaining entries are new distinct symbols. The number of symbols in $M$ is $n+2nk+k^2{\geqslant}\ell(n+k)$, so there must be a transversal in $M$. This transversal hits at most $2k$ cells in the last $k$ rows or columns of $M$, so it must intersect the copy of $L$ in at least $n-k$ cells, each of which contains a different symbol.
Putting $k=1$, we see that if $\ell(n){\leqslant}3n-2$ for all $n$, then every Latin square has a near transversal. This would prove a famous conjecture attributed to Brualdi (see [@Wan11]). Indeed, any linear upper bound on $\ell(n)$ would imply the existence of a constant $c$ such that every Latin square of order $n$ has a partial transversal of length $n-c$. The best result to date [@HS08] is that every Latin square has a partial transversal of length $n-O(\log^2n)$.
There is a broader setting in which quadratically many symbols is known to be best possible, namely row-Latin arrays. An array is [*row-Latin*]{} if no symbol appears more than once in any row. For every positive integer $n$, let $\ell_r(n)$ be the least positive integer such that $\ell_r(n){\geqslant}n$ and every $n\times n$ row-Latin array with at least $\ell_r(n)$ distinct symbols contains a transversal. Bar[á]{}t and Wanless [@BaratWanless14] showed that $\ell_r(n)>\frac12
n^2-O(n)$. In Section \[sec:Latin-arrays\], we prove that $\ell_r(n){\leqslant}\big\lceil\frac{1}{4}(5-\sqrt{5})n^2\big\rceil$ for every integer $n>1$.
Probabilistic Approach {#sec:probabilistic}
======================
In this section we use probabilistic methods to prove a bound on $\ell(n)$. Let $\mathcal{B} = \{B_1,\ldots,B_t\}$ be a set of events in a probability space. Usually the events $\mathcal{B}$ are called the *bad events* because the aim is for them to not occur. Define ${\overline{B_i}}$ to be the complement of the event $B_i$. A graph $G$ with vertex set $\mathcal{B}$ is a *lopsidependency graph* if for all $B_i \in \mathcal{B}$ and for every subset $S$ of the complement of the closed neighbourhood of $B_i$ in $G$, $$\begin{aligned}
\label{eq:lopsi-def}
\mathbb{P}\Big( B_i \Big| \bigcap_{j \in S} {\overline{B_j}} \Big)
{\leqslant}\mathbb{P}(B_i).\end{aligned}$$ Lopsidependency graphs were introduced by Erdős and Spencer [@erdosspencer1991] and are useful because they have fewer edges than a naively defined dependency graph. Intuitively, a lopsidependency graph says that the probability of an event does not increase when conditioned on an arbitrary set of non-adjacent events not occurring.
The Clique Lovász Local Lemma by Kolipaka, Szegedy and Xu [@kolipaka2012sharper] gives a condition under which none of the bad events occur. Specialising their formulation, we get:
\[l:CLL\] Let $\mathcal{B} = \{B_1,\ldots,B_t\}$ be a set of events with lopsidependency graph $G$. Let $\{K_1,\ldots,K_s\}$ be a set of cliques in $G$ covering all the edges, and assume $\kappa {\geqslant}\max_i|K_i|$. Suppose that no event $B_i$ is in more than $\mu$ of the cliques $K_1,\dots,K_s$. If there exist $x \in (0, 1/\kappa)$ such that $ \mathbb{P}(B_i) {\leqslant}x \left( 1 - \kappa x \right)^{\mu-1}
$for all $1 {\leqslant}i {\leqslant}t$, then $$\mathbb{P}\Big( \bigcap_{i=1}^t {\overline{B_i}} \Big) > 0.$$
We use this lemma to prove:
Let $L$ be a Latin array of order $n$. If $L$ has at least $({229}n^2+{27}n)/256 \approx 0.8945n^2$ distinct symbols, then $L$ has a transversal.
Suppose $L$ has at least $n^2-cn^2-dn$ distinct symbols. Let $\sigma$ be a permutation picked uniformly at random from the symmetric group on $\{1,2,\dots,n\}$. Think of $\sigma$ as choosing the cells $(i,\sigma(i))$ for $1{\leqslant}i{\leqslant}n$, which might correspond to a transversal. Define the bad events, $$\begin{aligned}
\mathcal{B} =
\{(i,j,i',j'):1{\leqslant}i < i'{\leqslant}n, \sigma(i)=j, \sigma(i')=j',
L_{ij}=L_{i'j'}\}.\end{aligned}$$ These events correspond to $\sigma$ choosing a pairs of cells in $L$ that contain the same symbol. To prove that a transversal exists we just need to prove that, with positive probability, none of the bad events occur.
The next task is to define the lopsidependency graph which will be used in applying Lemma \[l:CLL\]. Let $G$ be a graph with vertex set $\mathcal{B}$. An edge $\{(a,b,x,y),(a',b',x',y')\}$ is in $G$ if and only if at least two of the cells $(a,b)$, $(x,y)$, $(a',b')$ and $(x',y')$ share a row or column. This occurs only if at least one of $x=x'$, $x=a'$, $a=x'$, $a=a'$, $y=y'$, $y=b'$, $b=y'$ or $b=b'$. Erdős and Spencer [@erdosspencer1991] showed that $G$ is a lopsidependency graph.
Let $\mathcal{K} = \{K_1,K_2,\ldots,K_{2n}\}$ be a set of cliques of $G$ defined as follows. Each clique corresponds to a row or column of $L$. An event $(a,b,x,y)$ is in a clique $K_i$ if $(a,b)$ or $(x,y)$ is in the row or column corresponding to $K_i$. Note that $K_i \in \mathcal{K}$ is a clique because the events in $K_i$ share a row or column (the one corresponding to $K_i$) and so they are adjacent in $G$. These cliques cover every edge of $G$ because two events are adjacent only if they share a row or column. Each event in $\mathcal{B}$ corresponds to two cells in distinct rows and columns, so each event is within exactly four cliques. Thus we take $\mu=4$. To find the bound $\kappa$, consider a clique $K\in\mathcal{K}$ which, without loss of generality, corresponds to the first row. Each event in $K$ corresponds to two cells of $L$, one in the first row and one not in that row. Let $D$ be the set of cells outside the first row that are included in some event in $K$. Each cell in $D$ shares a symbol with exactly one cell in the first row. Hence $|K|=|D|$ and the cells not in $D$ contain as many distinct symbols as $L$ does. Hence $n^2-|K|{\geqslant}n^2-cn^2-dn$, which means that we may take $\kappa=cn^2+dn$.
Taking $x=1/(4\kappa)$, we find that to apply Lemma \[l:CLL\] we need $$\frac{1}{n(n-1)} = \mathbb{P}(B_i) {\leqslant}x(1-\kappa x)^3=\frac{27}{256\kappa},$$ which is satisfied when $c={27}/{256}$ and $d=-{27}/{256}$.
A Better Bound {#sec:Latin-arrays}
==============
In this section, we prove a better bound on $\ell(n)$ using non-probabilistic methods. We start by proving results about general square arrays, then later use these results to give bounds on the number of symbols in transversal-free row-Latin arrays and transversal-free Latin arrays.
We call a symbol in an array $A$ a [*singleton*]{} if it occurs exactly once in $A$ and a [*clone*]{} otherwise. We define $R_i(A)$ and $C_j(A)$ to be the set of symbols occurring in row $i$ and column $j$ of $A$, respectively. Let $A(i\mid j)$ denote the array formed from $A$ by deleting row $i$ and column $j$ and let $\Psi_{ij}(A)$ be the set of symbols that appear in $A$ and not in $A(i\mid j)$.
\[lem:upper-bound-from-larges\] Let $A$ be a transversal-free array of order $n$. If $A(n\mid n)$ has a transversal and if $|R_n(A) \cup C_n(A)| {\geqslant}(k+1)n-1$, then $A$ has at most $$\frac12(k^2 - 2k + 2)n^2 + \frac12(3k-2)n$$ distinct symbols.
Assume that $T$ is a near transversal of $A$ that does not meet the last row or column and minimises the number of symbols that it has from $R_n(A) \cup C_n(A)$.
We call a symbol *large* if it appears in both $T$ and $R_n(A)\cup C_n(A)$ and *small* otherwise. Let $\lambda$ be the number of large symbols. Permute the first $n-1$ rows and columns of $A$ so that $T$ is located along the main diagonal and all of the large symbols of $T$ appear in the top $\lambda$ rows. For $1{\leqslant}i<n$, note that $A_{in}$ and $A_{ni}$ cannot be two different small symbols. Otherwise, $\left(T \setminus \{(i,i,A_{ii})\}\right) \cup
\{(i,n,A_{in}),(n,i,A_{ni})\}$ would be a transversal of $A$. So there are at most $n-1$ distinct small symbols in the last row and column. Thus, $$\label{eq:bound-on-large-syms}
\lambda {\geqslant}|R_n(A) \cup C_n(A)|-(n-1) {\geqslant}(k+1)n-1-(n-1) = kn.$$
We now define a subset $\Gamma$ of the entries of $A$ in which each symbol in $A$ is represented exactly once. We populate $\Gamma$ in three steps. First, $T \subseteq \Gamma$. Second, for every small symbol $s$ that occurs in the last row or column, select one such entry containing $s$ and add it to $\Gamma$. Finally, for every symbol $s'$ in $A$ that does not appear in $T$ or in the last row or column, select one entry with the symbol $s'$ and add it to $\Gamma$.
We claim that if $(i,j)$ is in the top $\lambda$ rows of $A$ with $i<j<n$, then at most one of $(i,j,A_{ij})$ and $(j,i,A_{j,i})$ can be in $\Gamma$. Suppose otherwise, and consider $$\label{e:wasP}
\big(T \setminus \big\{(i,i,A_{ii}),(j,j,A_{jj})\big\}\big) \cup
\big\{(i,j,A_{ij}),(j,i,A_{ji})\big\}.$$ Note that the symbol $A_{ii}$ is contained in the last row or column of $A$. By the definition of $\Gamma$, we know that $(i,j)$ and $(j,i)$ do not have the same symbol and neither one shares a symbol with any entry in $T$ or in the last row or column. So (\[e:wasP\]) is a near transversal that contains fewer symbols in $R_n(A) \cup
C_n(A)$ than $T$, contradicting the choice of $T$. This implies that within the first $\lambda$ rows and columns of $A(n\mid n)$, there are at least $$(n-2)+(n-3)+\dots+(n-\lambda-1) = \lambda n - \frac{\lambda(\lambda+3)}{2}$$ entries not contained in $\Gamma$. Within the last row and column of $A$, there are at most $n-1$ entries in $\Gamma$ (all containing small symbols), so at least $n$ entries are not in $\Gamma$. Thus, the number of distinct symbols in $A$ is $$\label{eq:gamma-upper-1}
|\Gamma| {\leqslant}n^2 - \left(\lambda n - \frac{\lambda(\lambda+3)}{2}\right) - n
= \frac12\lambda^2-\left(n-\frac32\right)\lambda+n(n-1).
$$ This quadratic in $\lambda$ decreases weakly on the integer points in the interval $kn{\leqslant}\lambda{\leqslant}n-1$. Given (\[eq:bound-on-large-syms\]), we may substitute $\lambda=kn$ into (\[eq:gamma-upper-1\]) to get the desired result.
\[lem:good-col\] Let $A$ be an $n\times n$ array with $\beta n^2$ distinct symbols. If there are $d {\geqslant}1$ clones in row $i$, then there is some clone $A_{ij}$ such that $$|R_i(A) \cup C_j(A)| {\geqslant}|R_i(A)|+\frac{\beta n^2-(n-d)(n-1)-|R_i(A)|}{d}.$$
We will endeavour to find a column $j$ such that $|C_j(A) \setminus R_i(A)|$ is large. Without loss of generality, assume that the rightmost $d$ columns of row $i$ contain clones. First, remove all occurrences of the symbols in $R_i(A)$ from the array. Now, arbitrarily select a representative entry for each of the remaining symbols in the array. Note that there are no representatives in row $i$ and so there are at most $(n-d)(n-1)$ representatives in the first $n-d$ columns. Of the original $\beta n^2$ symbols, at least $\beta n^2-(n-d)(n-1)-|R_i(A)|$ must have their representative in the last $d$ columns. By the pigeon-hole principle, the desired clone $A_{ij}$ occurs in one of the last $d$ columns.
Let $\mathcal{A}$ be some class of square arrays of symbols that has the following two properties: (i) if any row and column of an array in $\mathcal{A}$ is deleted, the resulting array is in $\mathcal{A}$ and (ii) if in one entry of the array, the symbol is changed to a new symbol that appears nowhere else in the array, then the resulting array is in $\mathcal{A}$. Note that $\mathcal{L}$, the set of all Latin arrays, and $\mathcal{R}$, the set of all row-Latin arrays, both satisfy the requirements listed.
Let $\frac12{\leqslant}\alpha{\leqslant}1$. Define $\mathcal{M}_{\mathcal{A}}(\alpha)$ to be the set of transversal-free arrays in $\mathcal{A}$ whose ratio of number of distinct symbols to cells is at least $\alpha$. Suppose that $\mathcal{M}_{\mathcal{A}}(\alpha)$ is non-empty. Define $\mathcal{M}^*_{\mathcal{A}}(\alpha) \subseteq
\mathcal{M}_{\mathcal{A}}(\alpha)$ by the rule that if $A \in
\mathcal{M}^*_{\mathcal{A}}(\alpha)$, then no array in $\mathcal{M}_{\mathcal{A}}(\alpha)$ has an order smaller than $A$ and no array in $\mathcal{M}_{\mathcal{A}}(\alpha)$ of the same order as $A$ contains more distinct symbols than $A$. For example, both $\mathcal{M}^*_{\mathcal{L}}(1/2)$ and $\mathcal{M}^*_{\mathcal{R}}(1/2)$ consist solely of the Latin squares of order $2$. For the remainder of the section, we will bound the number of symbols in arrays by examining properties of the arrays in $\mathcal{M}^*_{\mathcal{A}}(\alpha)$.
\[lem:uniquemin\] Let $A \in \mathcal{M}^*_{\mathcal{A}}(\alpha)$ be an array of order $n$. If $A_{ij}$ is a singleton, then $|\Psi_{ij}(A)|>\alpha(2n-1)$ and $R_i(A)$ (resp., $C_j(A)$) contains more than $(2\alpha-1)n$ symbols that appear only in row $i$ (resp., column $j$) of $A$.
Any array of order $1$ has a transversal, so $n {\geqslant}2$. There is no transversal $T$ of $A(i\mid j)$, or else $T\cup\{(i,j,A_{ij})\}$ would be a transversal of $A$. As $A\in\mathcal{M}^*_{\mathcal{A}}(\alpha)$, we have that $A(i\mid j) \not\in\mathcal{M}_{\mathcal{A}}(\alpha)$, so the number of distinct symbols in $A(i\mid j)$ is strictly less than $\alpha(n-1)^2$. Thus, $$|\Psi_{ij}(A)|>\alpha n^2-\alpha(n-1)^2= \alpha(2n-1).$$ At most $n-1$ of the symbols in $\Psi_{ij}(A) \setminus \{A_{ij}\}$ appear in $C_j(A)$, so at least $$|\Psi_{ij}(A)|-(n-1)>\alpha(2n-1)-(n-1) {\geqslant}(2\alpha-1)n$$ symbols appear in row $i$ and nowhere else in $A$. A similar argument applies to $C_j(A)$.
\[lem:clone-near-trans-implies-few-syms\] Let $A \in \mathcal{M}^*_{\mathcal{A}}(\alpha)$ be an array of order $n$. If $A_{ij}$ is a clone and $|R_i(A) \cup C_j(A)|{\geqslant}(k+1)n-1$, then $A$ has at most $$\frac12(k^2 - 2k + 2)n^2 + \frac12(3k-2)n$$ distinct symbols.
Without loss of generality, $i=j=n$. Create $A'$ by changing the symbol in the $(n,n)$ cell of $A$ to a symbol that did not previously appear in $A$. Since $A_{ij}$ is a clone in $A$, we know that $A'$ contains strictly more symbols than $A$. Since $A \in
\mathcal{M}^*_{\mathcal{A}}(\alpha)$, we conclude that $A'$ has a transversal, although $A$ does not. Hence there is a near transversal of $A$ that does not meet row $n$ or column $n$. By applying Lemma \[lem:upper-bound-from-larges\], the result follows.
In the best case, Lemma \[lem:clone-near-trans-implies-few-syms\] falls just short of proving Conjecture \[conj:ln\].
\[cor:best-case\] Let $A \in \mathcal{M}^*_{\mathcal{A}}(\alpha)$ be an array of order $n$. If $A_{ij}$ is a clone and $|R_i(A) \cup C_j(A)| = 2n-1$, then $A$ has at most $(n^2+n)/2$ distinct symbols.
Lemmas \[lem:good-col\], \[lem:uniquemin\] and \[lem:clone-near-trans-implies-few-syms\] form the main framework needed to bound the number of symbols. We will utilise Lemmas \[lem:good-col\] and \[lem:uniquemin\] in different ways to find an entry $(i,j,k)$ where $k$ is a clone and row $i$ and column $j$ contain many different symbols. We then apply Lemma \[lem:clone-near-trans-implies-few-syms\] to bound the number of symbols overall. The following subsections concentrate on specific classes for $\mathcal{A}$.
Row-Latin Arrays
----------------
In this subsection, we consider $\mathcal{A}=\mathcal{R}$, the set of row-Latin arrays.
\[lem:rowLatin-clone\] Let $M \in \mathcal{M}^*_{\mathcal{R}}(\alpha)$ be a row-Latin array of order $n$. There exists a clone $M_{ij}$ for which $|R_i(M)
\cup C_j(M)| {\geqslant}2\alpha n-1$.
First suppose that there is a clone $M_{ij}$ that appears in the same column as a singleton. By Lemma \[lem:uniquemin\], $C_j(M)$ contains at least $(2\alpha - 1)n$ symbols that appear only in $C_j(M)$. One of these symbols may be $M_{ij}$, but $$|R_i(M) \cup C_j(M)|
= |R_i(M)| + |C_j(M) \setminus R_i(M)| {\geqslant}n + (2\alpha - 1)n-1
= 2\alpha n-1,$$ as required.
Hence we may assume that no column contains a singleton and a clone. Let $d$ be the number of columns that contain clones.
If $d {\leqslant}n/2$, then we can find a transversal in the following way. Let $R$ be the $n \times d$ subarray of $M$ that contains the clones of $M$. A result of Drisko [@Dri98] implies that $M$ has a partial transversal of length $d$ that is wholly inside $R$. Since this partial transversal covers all columns that contain clones, it can trivially be extended to a transversal using singletons.
So we may assume that $d > n/2$. Since each row contains $d$ clones, we may use Lemma \[lem:good-col\] with $\beta{\geqslant}\alpha$ to find some clone $M_{ij}$ such that $$|R_i(M) \cup C_j(M)| {\geqslant}\frac{\alpha-1}{d}n^2 + 2n - 1
> 2(\alpha-1)n+2n-1=2\alpha n-1. \qedhere$$
We now show one of our main results, that row-Latin arrays with many symbols must have a transversal.
\[th:rowLatin-bound\] Let $L$ be a row-Latin array of order $n$. If $L$ has at least $\frac{1}{4}(5-\sqrt{5})n^2 \approx 0.6910n^2$ distinct symbols, then $L$ has a transversal.
Aiming for a contradiction, suppose that $L\in\mathcal{M}_\mathcal{R}(\alpha)$ for $\alpha=(5-\sqrt{5})/4$. Then there exists $M \in \mathcal{M}^*_{\mathcal{R}}(\alpha)$. Let $M$ have order $m$. By Lemma \[lem:rowLatin-clone\], there is a clone $M_{ij}$ such that $|R_i(M) \cup C_j(M)| {\geqslant}2\alpha m - 1$. By Lemma \[lem:clone-near-trans-implies-few-syms\], the number of distinct symbols in $M$ is at most $$\label{eq:rowLatin-poly}
\frac{1}{2}\left((2\alpha-1)^2-2(2\alpha-1)+2\right)m^2+
\frac{1}{2}\left(3(2\alpha-1)-2\right)m
=\alpha m^2-\frac14(3\sqrt{5}-5)m.$$ This contradicts the fact that $M$ has at least $\alpha m^2$ distinct symbols, and we are done.
Latin Arrays
------------
In this subsection, we consider $\mathcal{A}=\mathcal{L}$, the set of Latin arrays.
We call a Latin array $L$ of order $n$ *focused* if every singleton in $L$ occurs in a row or a column that contains only singletons and $|\Psi_{ij}(L)|=2n-1$ for some $(i,j)$ (that is, row $i$ and column $j$ contain only singletons). We deal with focused and unfocused arrays separately.
For focused arrays we use the following simple adaptation of a result of Woolbright [@Woolbright78]. The original proof was for Latin squares, but it works without change for Latin arrays (in fact for row-Latin arrays, but we do not need that).
\[th:partial\] Let $L$ be an $n \times n$ Latin array and $0 {\leqslant}t < n$. If $(n-t)^2 > t$, then $L$ has a partial transversal of length $t+1$.
In the following result, recall that we assume $\alpha{\geqslant}1/2$.
\[lem:focused\] Let $M \in \mathcal{M}^*_{\mathcal{L}}(\alpha)$ be a Latin array of order $n$. If $M$ is focused, then $M$ contains at most $\frac{1}{8}({6-\sqrt{2}})n^2 \approx 0.5732n^2$ distinct symbols.
Let $\delta = \left\lceil(2\alpha-1)n\right\rceil$. Suppose $M$ has $r$ rows and $c$ columns that contain singletons. Permute the rows and columns of $M$ so that these singletons occur in the top $r$ rows and leftmost $c$ columns. Since $M$ is focused, $\min(r,c) {\geqslant}1$ and the bottom-right $(n-r)\times(n-c)$ subarray does not contain any singletons. Thus, if we consider any singleton in the last row or last column, we get $\min(r,c) {\geqslant}\delta$ by Lemma \[lem:uniquemin\].
If $\alpha {\geqslant}3/4$, then $\min(r,c) {\geqslant}n/2$ and so $\left\{(i,n-i+1) : 1 {\leqslant}i {\leqslant}n\right\}$ is a set of cells containing only singletons, contradicting the fact that $M$ has no transversal. So $\alpha < 3/4$.
Let $M'$ be the subarray formed by the last $n-\delta$ rows and columns of $M$. Suppose that $M$ has a partial transversal of length $n-2\delta$ wholly inside $M'$. Then this partial transversal can easily be extended to a transversal by selecting singletons in the first $\delta$ rows and $\delta$ columns of $M$. By assumption $M$ has no transversal, so applying Theorem \[th:partial\] to $M'$ we find that $(\delta+1)^2{\leqslant}n-2\delta-1$. Hence $$\label{e:quad}
0{\geqslant}\delta^2+4\delta+2-n{\geqslant}(2\alpha-1)^2n^2+(8\alpha-5)n+2.$$ From the discriminant of this quadratic we learn that $32\alpha^2-48\alpha+17{\geqslant}0$. Since $\alpha<3/4$ we have $\alpha{\leqslant}({6-\sqrt{2}})/{8}$.
For any $\alpha>1/2$, it is worth noting that (\[e:quad\]) fails for all large $n$. So we get an asymptotic version of Conjecture \[conj:ln\] holding for focused Latin arrays. We are not able to reach such a strong conclusion for the unfocused case.
\[lem:entrychoice\] Let $M \in \mathcal{M}^*_{\mathcal{L}}(\alpha)$ be a Latin array of order $n$. If $M$ is unfocused, then there exists some clone $M_{ij}$ such that $\left|R_i(M) \cup C_j(M) \right| {\geqslant}(\alpha+1)n-1.$
Firstly, we consider the case that $M$ has some row or column that contains only clones. Without loss of generality, row $i$ contains only clones. By Lemma \[lem:good-col\], there is some clone $M_{ij}$ such that $|R_i(M) \cup C_j(M)| {\geqslant}n+(\alpha n^2 - n)/{n}
=(\alpha+1)n-1$.
Secondly, we consider the case that every row and column of $M$ contains a singleton. Since $M$ is unfocused, there is some singleton $M_{ik}$ such that there is a clone in both row $i$ and column $k$. By Lemma \[lem:uniquemin\], we have $|\Psi_{ik}(M)|>\alpha(2n-1)$. Each symbol in $\Psi_{ik}(M)$ appears in either $R_i(M)$ or $C_k(M)$. Also, $M_{ik}$ appears in both $R_i(M)$ and $C_k(M)$, so without loss of generality, $R_i(M)$ contains at least $\left(|\Psi_{ik}(M)|+1\right)/2 >
\alpha(n-1/2)+1/2{\geqslant}\alpha n$ symbols that are in $\Psi_{ik}(M)$. Let $M_{ij}$ be a clone in the same row as $M_{ik}$. Except possibly for $M_{ij}$, none of the $n$ symbols in $C_j(M)$ are in $\Psi_{ik}(M)$. Hence, $|R_i(M) \cup C_j(M)|{\geqslant}\alpha n+n-1$ as required.
We now show a stronger result than Theorem \[th:rowLatin-bound\] holds for Latin arrays.
\[thm:main-result\] Let $L$ be a Latin array of order $n$. If $L$ has at least $\left(2-\sqrt{2}\right)n^2 \approx 0.5858n^2$ distinct symbols, then $L$ has a transversal.
Aiming for a contradiction, suppose that $L\in\mathcal{M}_\mathcal{L}(\alpha)$ for $\alpha=2-\sqrt{2}$. Then there exists $M \in \mathcal{M}^*_{\mathcal{R}}(\alpha)$. Let $M$ have order $m$. Note that $M$ cannot be focused, by Lemma \[lem:focused\]. So, by Lemma \[lem:entrychoice\], there is a clone $M_{ij}$ such that $\left|R_i(M) \cup C_j(M) \right| {\geqslant}(\alpha+1)m-1$. By Lemma \[lem:clone-near-trans-implies-few-syms\], the number of distinct symbols in $M$ is at most $$\label{eq:Latin-poly}
\frac{1}{2}\left(\alpha^2-2\alpha+2\right)m^2+
\frac{1}{2}\left(3\alpha-2\right)m
=\alpha m^2-\frac12(3\sqrt{2}-4)m.
$$ This contradicts the fact that $M$ has at least $\alpha m^2$ distinct symbols, and we are done.
Small Values {#sec:small-values}
============
We now shift our attention to small values of $n$ where we can compute $\ell(n)$ exactly. Akbari and Alipour [@AkbariAlipour04] determined $\ell(n)$ for $n {\leqslant}4$. We extend this search to $n{\leqslant}7$ and catalogue all Latin arrays of small orders with no transversals. For $n{\geqslant}8$, computing $\ell(n)$ seems challenging. We will mention a couple of unsuccessful attempts to find examples that would provide some insight.
Following [@EW16], we say that two Latin arrays are [ *trisotopic*]{} if one can be changed into the other by permuting rows, permuting columns, permuting symbols and/or transposing. The set of all Latin arrays trisotopic to a given array is a *trisotopy class*. The number of transversals is a trisotopy class invariant, so to find all transversal-free Latin arrays of a given order it suffices to consider trisotopy class representatives. However, for orders $n>5$ it becomes difficult to construct a representative of every trisotopy class. The following method allows us to push our results a couple of orders further.
Let $L$ be a transversal-free Latin array. In the first two rows of $L$, select two entries that do not share a column or symbol (this can always be done for $n {\geqslant}3$). Without loss of generality, we may assume that these two entries are $(1,1,x)$ and $(2,2,y)$. Let $L'$ be the bottom-right $(n-2) \times (n-2)$ subarray of $L$ where all occurrences of $x$ and $y$ are replaced with a hole (that is, a cell with no symbol; we forbid holes from being chosen in a transversal or partial transversal). There cannot be a partial transversal of length $n-2$ in $L'$, otherwise the corresponding entries in $L$, together with $(1,1,x)$ and $(2,2,y)$, would form a transversal of $L$.
Thus, to search for transversal-free Latin arrays of order $n$, we first build a catalogue $\mathcal{C}_{n-2}$ of trisotopy class representatives of transversal-free partial Latin arrays of order $n-2$ with at most two holes in each row and each column. Starting with this catalogue, we can reverse the argument above. At least one representative of each trisotopy class of transversal-free Latin array of order $n$ can be obtained by taking an element of $\mathcal{C}_{n-2}$, filling its holes with $x$ and $y$, then extending it to a Latin array of order $n$.
By the above technique we are able to give a complete catalogue of the transversal-free trisotopy classes for orders $n {\leqslant}7$. Table \[tab:l\_n\] gives the value of $\ell(n)$ and the number of trisotopy classes with a specific number of symbols.
----- ----------- ------------- --------------- --------------- ------- --
$n$ $\ell(n)$ $n$ symbols $n+1$ symbols $n+2$ symbols Total
2 3 1 - - 1
3 3 - - - 0
4 6 1 1 - 2
5 7 - 2 - 2
6 9 8 19 1 28
7 7 - - - 0
----- ----------- ------------- --------------- --------------- ------- --
: \[tab:l\_n\]Values of $\ell(n)$ and the number of trisotopy classes of transversal-free Latin arrays.
Representatives of the trisotopy classes of transversal-free Latin arrays of orders $4$ and $5$ are: $$\left(\begin{array}{cccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d\\
{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}a\\
{\texttt}c&{\texttt}d&{\texttt}a&{\texttt}b\\
{\texttt}d&{\texttt}a&{\texttt}b&{\texttt}c\\
\end{array}\right),\quad
\left(\begin{array}{cccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d\\
{\texttt}b&{\texttt}c&{\texttt}a&{\texttt}e\\
{\texttt}c&{\texttt}a&{\texttt}d&{\texttt}b\\
{\texttt}e&{\texttt}d&{\texttt}b&{\texttt}a\\
\end{array}\right),\quad
\left(\begin{array}{ccccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e\\
{\texttt}b&{\texttt}c&{\texttt}a&{\texttt}e&{\texttt}f\\
{\texttt}c&{\texttt}a&{\texttt}b&{\texttt}f&{\texttt}d\\
{\texttt}e&{\texttt}d&{\texttt}f&{\texttt}c&{\texttt}a\\
{\texttt}d&{\texttt}f&{\texttt}e&{\texttt}a&{\texttt}b\\
\end{array}\right),\quad
\left(\begin{array}{ccccc}
{\texttt}f&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e\\
{\texttt}b&{\texttt}c&{\texttt}a&{\texttt}e&{\texttt}f\\
{\texttt}c&{\texttt}a&{\texttt}b&{\texttt}f&{\texttt}d\\
{\texttt}e&{\texttt}d&{\texttt}f&{\texttt}c&{\texttt}a\\
{\texttt}d&{\texttt}f&{\texttt}e&{\texttt}a&{\texttt}b\\
\end{array}\right).$$ Note that our two representatives of order 5 differ only in their first entry. Both can be completed to Latin squares of order 6; in the first case this Latin square has no transversals, but in the second case it has eight transversals.
Many of the transversal-free Latin arrays for order $6$ also turn out to be quite similar to one another. There are exactly $28$ trisotopy classes for $n=6$. Previously, nine of these classes were known: eight Latin squares and the array constructed by Akbari and Alipour [@AkbariAlipour04] by removing two rows and columns from the elementary abelian Cayley table of order $8$. We will now describe the $19$ transversal-free trisotopy classes of order $6$ with seven symbols. We will denote their representative arrays by $L_1,L_2,\dots,L_{19}$. Let $$L_1 =
\left(\begin{array}{cccccc}
{\cellcolor[gray]{.8}{\texttt}}{a}&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e&{\texttt}f\\
{\texttt}b&{\cellcolor[gray]{.8}{\texttt}}{c}&{\texttt}a&{\texttt}f&{\texttt}d&{\texttt}e\\
{\texttt}c&{\texttt}a&{\cellcolor[gray]{.8}{\texttt}}{b}&{\texttt}e&{\texttt}f&{\texttt}d\\
{\texttt}d&{\texttt}e&{\texttt}f&{\texttt}g&{\texttt}b&{\texttt}c\\
{\texttt}f&{\texttt}d&{\texttt}e&{\texttt}b&{\texttt}g&{\texttt}a\\
{\texttt}e&{\texttt}f&{\texttt}d&{\texttt}c&{\texttt}a&{\texttt}g
\end{array}\right)
\mbox{ \ and \ }
L' =
\left(\begin{array}{cccccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e&{\cellcolor[gray]{.8}{\texttt}}{f}\\
{\texttt}c&{\texttt}f&{\texttt}b&{\cellcolor[gray]{.8}{\texttt}}{e}&{\texttt}d&{\texttt}a\\
{\texttt}b&{\texttt}c&{\texttt}e&{\texttt}f&{\cellcolor[gray]{.8}{\texttt}}{a}&{\texttt}d\\
{\texttt}d&{\texttt}e&{\cellcolor[gray]{.8}{\texttt}}{f}&{\texttt}a&{\texttt}b&{\texttt}c\\
{\cellcolor[gray]{.8}{\texttt}}{e}&{\texttt}d&{\texttt}a&{\texttt}c&{\texttt}f&{\texttt}b\\
{\texttt}f&{\cellcolor[gray]{.8}{\texttt}}{a}&{\texttt}d&{\texttt}b&{\texttt}c&{\texttt}e
\end{array}\right).$$
From $L'$, we define $L_{2},\dots,L_{8}$ by changing some entries on the main diagonal to a new symbol, ${\texttt}g$, in the following way. Let $$R' \in \big\{ \{1,2,3,4,5,6\}, \{1,2,4,5,6\},
\{1,3,4,5\}, \{1,3,6\}, \{1,4\}, \{2,3,5,6\}, \{3,4,5,6\} \big\}.$$ For all $r \in R'$, change the symbol on the main diagonal in row $r$ of $L'$ to ${\texttt}g$. It turns out that changing the shaded entries in $L_1$ to ${\texttt}g$ results in an array that is trisotopic to $L_{2}$. Next, $L_{9}$ is obtained by changing the symbol of the shaded entries in $L'$ to a new symbol, ${\texttt}g$. Let $$L_{10} =
\left(\begin{array}{cccccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e&{\texttt}f\\
{\texttt}b&{\texttt}c&{\texttt}g&{\texttt}a&{\texttt}f&{\texttt}e\\
{\texttt}c&{\texttt}f&{\texttt}d&{\texttt}g&{\texttt}a&{\texttt}b\\
{\texttt}d&{\texttt}a&{\texttt}f&{\texttt}e&{\texttt}g&{\texttt}c\\
{\texttt}e&{\texttt}g&{\texttt}a&{\texttt}f&{\texttt}c&{\texttt}d\\
{\texttt}g&{\texttt}e&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}a
\end{array}\right)
\mbox{ \ and \ }
L'' =
\left(\begin{array}{cccccc}
{\texttt}a&{\texttt}b&{\texttt}c&{\texttt}d&{\texttt}e&{\texttt}f\\
{\texttt}b&{\texttt}c&{\texttt}a&{\texttt}e&{\texttt}f&{\texttt}d\\
{\texttt}c&{\texttt}a&{\texttt}b&{\texttt}f&{\texttt}d&{\texttt}e\\
{\texttt}d&{\texttt}e&{\texttt}f&{\texttt}a&{\texttt}c&{\texttt}b\\
{\texttt}e&{\texttt}f&{\texttt}d&{\texttt}c&{\texttt}b&{\texttt}a\\
{\texttt}f&{\texttt}d&{\texttt}e&{\texttt}b&{\texttt}a&{\texttt}c
\end{array}\right).$$ From $L_{10}$, we can either change the symbol in the $(3,3)$ cell to ${\texttt}e$, giving $L_{11}$, or change the symbol in the $(4,4)$ cell to ${\texttt}b$, giving $L_{12}$.
The array $L_{13}$ is obtained by changing the ${\texttt}d$ in rows 2 and 3 of $L''$ to ${\texttt}g$, as well as changing the ${\texttt}f$ in row 2 to ${\texttt}d$. Next, $L_{14}$ is obtained by changing the ${\texttt}e$ in row 3 of $L_{13}$ to ${\texttt}d$. From the Latin square $L''$, any subset of entries that contain ${\texttt}d$ may be changed to a new symbol, ${\texttt}g$. This gives rise to 5 trisotopy classes. In particular, we define $L_{15},\dots,L_{19}$ by changing some occurrences of ${\texttt}d$ to a new symbol, ${\texttt}g$, in the following way. Let $$R'' \in\big\{\{1\},\{1,2\},\{1,2,3\},\{1,3,5\},\{1,4\}\big\}.$$ For all $r\in R''$, change the ${\texttt}d$ in row $r$ of $L''$ to ${\texttt}g$.
One can check that $L_{15},\dots,L_{19}$ are transversal-free by exhaustive computation, but next we give a reason why they have no transversals. The argument is in the style of the highly successful $\Delta$-lemma (see [@Wan11]). Let $L$ be any Latin array obtained by replacing any subset of the occurrences of ${\texttt}d$ in $L''$ by ${\texttt}g$. Define functions $\rho,\nu$ to ${\mathbb{Z}}_3$ by: $$\begin{aligned}
&\rho(1)=\rho(4)=0,\ \rho(2)=\rho(5)=1,\ \rho(3)=\rho(6)=2,\\
&\nu({{\texttt}a})=\nu({{\texttt}d})=\nu({{\texttt}g})=0,\ \nu({{\texttt}b})=\nu({{\texttt}e})=1,\ \nu({{\texttt}c})=\nu({{\texttt}f})=2.\end{aligned}$$ Define a function $\Delta$ from the entries of $L$ to ${\mathbb{Z}}_3$ by $\Delta(r,c,s)=\rho(r)+\rho(c)-\nu(s)$. Let $D$ denote the bottom-right $3\times 3$ subsquare of $L$. Suppose that $T$ is a transversal of $L$ and that $\bar s$ is the only symbol in $\{{{\texttt}a,b,\dots,g}\}$ that does not appear in $T$. Then $$\label{e:deltalem}
\sum_{(r,c,s)\in T}\Delta(r,c,s)=
2\sum_{i=1}^6\rho(i)-\sum_{(r,c,s)\in T}\nu(s)=\nu(\bar s).$$ Also, if $T$ includes $x$ entries in $D$ then overall it has $2x$ entries with symbols in $\{{{\texttt}a,b,c}\}$, which means that $x=1$ and $\bar s\in\{{{\texttt}a,b,c}\}$. However, $\Delta(r,c,s)=0$ for all entries of $L$, except those in $D$, where $\Delta(r,c,s)=\nu(s)$. Hence to satisfy (\[e:deltalem\]), the symbol in the only entry of $T$ in $D$ has to be $\bar s$, contradicting the fact that this symbol does not appear in $T$.
The argument we have just presented is specific to order $n=6$ and does not seem to easily generalise to arrays of larger orders.
When performing the search for transversal-free Latin arrays of order $n = 7$, we found $15\,611\,437$ trisotopy classes of transversal-free partial Latin arrays of order 5 and at most two holes in each row and column. Table \[tab:5x5-partial\] provides counts of the trisotopy classes based on number of holes and number of symbols. Since none of these arrays can extend to a Latin array of order $7$ with no transversals, we have the following result.
\[th:7x7-no-trans\] Every Latin array of order $7$ has a transversal.
-- ---- -- --- ----- -------- --------- --------- -------- ------- ------ ---- ---- ---- -- --
3 4 5 6 7 8 9 10 11 12 13
0 - - - 2 - - - - - - -
1 - - 1 17 - - - - - - -
2 - - 9 271 13 - - - - - -
3 - - 137 4893 1179 61 5 - - - -
4 - - 1484 54911 31342 5539 1906 462 62 4 -
5 - 3 10686 341251 319750 58257 9823 1175 86 4 -
6 - 19 48436 1155690 1420192 299951 33366 1953 56 - -
7 - 151 124275 2045859 2754143 670137 63480 2676 30 - -
8 - 632 159295 1720463 2198260 549316 43912 1710 78 8 1
9 - 916 80609 557285 603320 134056 7120 148 7 1 -
10 3 320 9420 40418 34218 6014 159 1 - - -
-- ---- -- --- ----- -------- --------- --------- -------- ------- ------ ---- ---- ---- -- --
: \[tab:5x5-partial\]Counts of trisotopy classes of transversal-free $5 \times 5$ partial Latin arrays, categorised by number of symbols and number of holes.
The approach that we used to prove Theorem \[th:7x7-no-trans\] is infeasible for $n {\geqslant}8$, although we did examine certain interesting sets of Latin arrays of order 8. There are 68 different transversal-free Latin squares of order 8, up to trisotopy. We also considered all Latin arrays which are obtained by removing one row and one column from a Latin square of order 9. We could immediately eliminate any square of order 9 that contains a transversal through every entry. Latin squares that do not contain a transversal through every entry are called *confirmed bachelor squares*. The confirmed bachelor squares of order 9 were generated for [@EW12], providing us with a set of trisotopy class representatives. None of these squares has an order 8 transversal-free subarray. Lastly, we searched all Latin arrays of order 8 with exactly 9 symbols where one of the symbols appears at most 4 times. None of these were transversal-free. The arrays that we have checked are a tiny subset of all Latin arrays of order 8. Without theoretical insight, it seems hopeless to check them all. So all that we can conclude at this stage is that $\ell(8){\geqslant}9$.
It is known that all Latin squares of order $9$ have transversals (see, e.g. [@EW12]). We tried, unsuccessfully, to build a transversal-free Latin array of order $9$. We did this by removing a row and column from Latin squares of order $10$. The squares that we used were representatives of all trisotopy classes for which the autoparatopy group has order $3$ or higher, as generated for [@MMM07].
The results of our investigations lead us to be skeptical that Conjecture \[conj:noconst\] is true. However, proving that it is false is likely to be extremely hard, for the reasons explained after Proposition \[prop:ln-implies-brualdi\]. Yet, it also seems hard to prove a subquadratic bound on $\ell(n)$, or even to prove Conjecture \[conj:ln\]. For $\ell_r(n)$ we know more. Thanks to [@BaratWanless14] and Theorem \[th:rowLatin-bound\], we know that $\frac12 n^2-O(n)<\ell_r(n){\leqslant}\big\lceil\frac{1}{4}({5-\sqrt{5}})
n^2\big\rceil$.
Acknowledgement {#acknowledgement .unnumbered}
---------------
The authors are grateful to Gweneäl Joret and János Barát for interesting discussions.
[10]{}
Saieed Akbari and Alireza Alipour. Transversals and multicolored matchings. , 12(5):325–332, 2004.
J[á]{}nos Bar[á]{}t and Ian M. Wanless. Rainbow matchings and transversals. , 59:211–217, 2014.
Nicholas Cavenagh and Ian M. Wanless. Latin squares with no transversals. arXiv:1609.03001 \[math.CO\].
Arthur A. Drisko. Transversals in row-[L]{}atin rectangles. , 84(2):181–195, 1998.
Judith Egan and Ian M. Wanless. atin squares with restricted transversals. , 20(7):344–361, 2012.
Judith Egan and Ian M. Wanless. Enumeration of [MOLS]{} of small order. , 85(298):799–824, 2016.
Paul Erd[ő]{}s and Joel Spencer. Lopsided [L]{}ovász local lemma and [L]{}atin transversals. , 30(2-3):151–154, 1991.
Pooya Hatami and Peter W. Shor. A lower bound for the length of a partial transversal in a [L]{}atin square. , 115(7):1103–1113, 2008.
Kashyap Kolipaka, Mario Szegedy, and Yixin Xu. A sharper local lemma with improved applications. In [*Approximation, randomization, and combinatorial optimization*]{}, volume 7408 of [*Lecture Notes in Comput. Sci.*]{}, pages 603–614. Springer, Heidelberg, 2012.
Brendan D. McKay, Alison Meynert, and Wendy Myrvold. Small [L]{}atin squares, quasigroups, and loops. , 15(2):98–119, 2007.
Ian M. Wanless. Transversals in [L]{}atin squares: a survey. In [*Surveys in combinatorics 2011*]{}, volume 392 of [*London Math. Soc. Lecture Note Ser.*]{}, pages 403–437. Cambridge Univ. Press, Cambridge, 2011.
David E. Woolbright. An [$n\times n$]{} [L]{}atin square has a transversal with at least [$n-\surd n$]{} distinct symbols. , 24(2):235–237, 1978.
[^1]: Research supported by Endeavour Postgraduate Scholarship and the NSERC PGS-D.
[^2]: Research supported by ARC grant DP150100506.
[^3]: Research supported by ARC grant FT1310464.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a formula for the values of the sign representations of the complex reflection groups $G(r,p,n)$ in terms of its image under a generalized Robinson-Schensted algorithm.'
address:
- 'Department of Mathematics, Bowdoin College'
- 'Department of Mathematics, Bowdoin College'
- 'Department of Computer Science, Bowdoin College, and Cognex Corporation'
author:
- Aba Mbirika
- Thomas Pietraho
- William Silver
title: 'On the sign representations for the complex reflection groups $G(r,p,n)$'
---
[^1]
Introduction
============
The classical Robinson-Schensted algorithm establishes a bijection between permutations $w \in \operatorname{\mathit{S_n}}$ and ordered pairs of same-shape standard Young tableaux of size $n$. This map has proven particularly well-suited to certain questions in the representation theory of both $\operatorname{\mathit{S_n}}$ and the semisimple Lie groups of type $A$. For instance, Kazhdan-Lusztig cells as well as the primitive spectra of semisimple Lie algebras can be readily described in terms of images of this correspondence.
Other sometimes more elementary representation-theoretic information requires more work to extract from standard Young tableaux. For instance, in independent work, A. Reifegerste [@reifegerste] and J. Sjöstrand [@sjostrand] developed a method for reading the value of the sign representation of a permutation $w \in \operatorname{\mathit{S_n}}$ based on two tableaux statistics. Let $w \in \operatorname{\mathit{S_n}}$ and write $RS(w) = (P,Q)$ for its image under the classical Robinson-Schensted map. If we write $e$ for the number of squares in the even-indexed rows of $P$, let $\operatorname{\mathit{sign}}(T)$ be the sign of a tableau $T$ derived from its inversion number, and let $\operatorname{\mathit{sgn}}$ be the usual sign representation on $\operatorname{\mathit{S_n}}$, then $$\operatorname{\mathit{sgn}}(w) = (-1)^{e} \cdot \operatorname{\mathit{sign}}(P) \cdot \operatorname{\mathit{sign}}(Q).$$
The focus of our paper is to extend this result to the complex reflection groups $G(r,p,n).$ Its two main ingredients generalize readily to this setting. First, the Robinson-Schensted algorithm admits a straightforward extension mapping each $w \in G(r,p,n)$ to a same-shape pair of $r$-multitableaux, see R. Stanley [@stanley:some] and L. Iancu [@iancu]. At the same time, the sign of a permutation in $\operatorname{\mathit{S_n}}$ extends to a family of $r$ one-dimensional representations of $G(r,p,n)$. After defining new spin and sign statistics on $r$-multitableaux, we prove the following extension of the result of A. Reifegerste and J. Sjöstrand:
Let $w \in G(r,p,n)$ and write $\mathbf{RS}(w) = (\mathbf{P},\mathbf{Q})$ for its image under the generalized Robinson-Schensted map. Given a primitive $r^{th}$ root of unity $\zeta$ and the associated family $\{\operatorname{\mathit{sgn}}_i\}_{i=0}^{r-1}$ of representations of $G(r,p,n)$, we have $$\operatorname{\mathit{sgn}}_i(w) = (-1)^{e({\mathbf{P}})} \cdot (\zeta^i)^{\operatorname{\mathit{spin}}(\mathbf{P})+\operatorname{\mathit{spin}}(\mathbf{Q})} \cdot \operatorname{\mathit{sign}}(\mathbf{P}) \cdot \operatorname{\mathit{sign}}(\mathbf{Q}),$$ where $e({\mathbf{P}})$ is the total sum of the lengths of the even-indexed rows of the component tableaux of ${\mathbf{P}}$.
To prove this, we define a particular set of equivalence classes on $G(r,1,n)$ and show that the formula either holds or fails for all members of a given class. By construction, each class contains a representative for which the theorem is easy to verify by a direct appeal to the original symmetric group formula, thereby completing the proof.
For the classical Weyl groups, all of which appear among the above complex reflection groups, the Möbius function of the Bruhat order can be expressed in terms of the values of the sign representation [@verma:mobius]. Its use is ubiquitous in Kazhdan-Lusztig theory, and the above formulas allow the values of the Möbius function to be read off from the images of the appropriate Robinson-Schensted map. This is especially relevant in light of the tableaux classification of Kazhdan-Lusztig cells. In type $A$, left Kazhdan-Lusztig cells consist of those permutations whose recording tableaux agree, see A. Joseph [@joseph1] or S. Ariki [@ariki]. A similar result holds for the so-called asymptotic left cells in the Iwahori-Hecke algebras in type $B$, this time in terms of $2$-multitableaux [@bonnafe:iancu].
Preliminaries
=============
After defining the family of complex reflection groups $G(r,p,n)$ and describing their one-dimensional representations, we define multipartitions, a generalization of the Robinson-Schensted algorithm, and tableaux statistics that we will use to describe these representations.
Sign representations
--------------------
Consider positive integers $r$, $p$, and $n$ with $p$ dividing $r$ and let $\zeta= \text{exp}(2 \pi \sqrt{-1})/ r)$. We define the complex reflection groups $G(r,p,n)$ as subgroups of $GL_n(\mathbb{C})$ consisting of matrices such that
- the entries are either $0$ or powers of $\zeta$,
- there is exactly one non-zero entry in each row and column,
- the $(r/p)$-th power of the product of all non-zero entries is $1$.
Together with thirty-four exceptional groups, the groups $G(r,p,n)$ account for all finite groups generated by complex reflections [@shephard-todd], and include among them all the classical Weyl groups. In our work the parameter $r$ will generally be fixed allowing us to write simply $W_n$ for the group $G(r,1,n).$ In order to establish succinct notation, we will write $$[\zeta^{a_1} \sigma_1, \zeta^{a_2} \sigma_2, \ldots, \zeta^{a_n} \sigma_n]$$ for the matrix whose non-zero entry in the $i$th column is $\zeta^{a_i}$ and appears in row $\sigma_i$. Utilizing this notation, define the set $S=\{s_0, \ldots, s_{n-1}\}$ where $$\begin{aligned}
s_0 & = [\zeta \cdot 1, 2, 3, \ldots, n] \text{, and } \\
s_i & =[1,2, \ldots, i-1, i+1, i, i+2, \ldots, n].\end{aligned}$$ Further, let $S'= \{s_0^p, s_0 s_1 s_0, s_i \; | \; 1 \leq i \leq n-1\}.$ The set $S$ generates $W_n$ with presentation given as $$W_n = \langle s_i \, | \, s_0^r, s_i^2, (s_j s_k)^2, (s_0 s_1)^4, (s_l s_{l+1})^3, \text{ $i \geq 1$, $|j-k|>1$, $l \in [1, n-2]$} \rangle.$$ Subject to similar relations, $S'$ generates a subgroup $G(r,p,n)$ of $W_n$ of index $p$, see S. Ariki [@ariki95].
There are exactly $2r$ one-dimensional representations of $W_n$; they divide naturally into two families.
\[def:one-diml-representations\] For each integer $i$ between $0$ and $r-1$, we define representations $\sigma_i$ and $\operatorname{\mathit{sgn}}_i$ of $W_n$ by specifying their values on the generating set $S$. Let
$$\tau_i^\epsilon(s_j) = \left\{
\begin{array}{ll}
\zeta^{i} & \text{ if $j=0$, and} \\
(-1)^\epsilon & \text{ if $j=1, \ldots, n-1$}
\end{array}
\right.$$ and define $\sigma_i = \tau_i^0$ and $\operatorname{\mathit{sgn}}_i = \tau_i^1$. Each becomes a representation of the subgroup $G(r,p,n)$ by restriction.
Multitableaux {#subsection:multitableaux}
-------------
We write a partition $\lambda$ of an integer $m$ as a nonincreasing sequence of positive integers $(\lambda_1, \lambda_2, \ldots, \lambda_k)$ and define its rank as $|\lambda|=m$. A [*Young diagram*]{} $[\lambda]$ of $\lambda$ is a left-justified array of boxes containing $\lambda_i$ boxes in its $i$th row. The shape of a Young diagram will refer to its underlying partition. With the integer $r$ fixed, a [*multipartition of rank $n$*]{} is an $r$-tuple $$\boldsymbol{\lambda} = (\lambda^0, \lambda^1, \ldots, \lambda^{r-1})$$ of partitions the sum of whose individual ranks equals $n$. The [*Young diagram*]{} $[\boldsymbol{\lambda}]$ of $\boldsymbol{\lambda}$ is the $r$-tuple $([\lambda^0], \ldots, [\lambda^{r-1}])$. We refer to $\boldsymbol{\lambda}$ as the [*shape*]{} of the diagram $[\boldsymbol{\lambda}]$ and define $|\boldsymbol{\lambda}|=n$. We will follow a convention of denoting objects derived from multipartitions in boldface while writing those derived from single partitions using a normal weight font.
A [*standard Young tableaux of shape*]{} $\boldsymbol{\lambda}$ is the Young diagram $[\boldsymbol{\lambda}]$ of rank $n$ together with a labeling of each of its boxes of with the elements of $\mathbb{N}_n$ in such a way that each number is used exactly once, and the labels of the boxes within each component Young diagram $[\lambda^i]$ increase along its rows and down its columns. Remembering that $r$ is fixed, we will write $\mathbf{SYT}_n$ for the set of all standard Young tableaux of rank $n$ whose shape is a multipartition with $r$ components.
A tableau $\mathbf{T} = (T_1, T_2, \ldots, T_r) \in \mathbf{SYT}_n$ will be called [*ascending*]{} if for each $i<r$, the elements of the set of labels of the component tableau $T_i$ are pairwise smaller than the elements of the set of labels of $T_{i+1}$.
Take $r=3$. The following standard Young tableau $\mathbf{T}$ has shape $\boldsymbol{\lambda}=((2,1), (1,1), (3,3))$, is of rank $11$, and is ascending: $$\raisebox{.07in}{$\mathbf{T}=\Big($}
\hspace{.05in}
\begin{tiny}
\bgroup\catcode`\:=13 \catcode`\.=13
\catcode`\;=13 \catcode`\>=13 \catcode`\^=13
\setlength{\tabheight}{3ex}\setlength{\tabwidth}{3ex} \def\b##1##2##3{\gentabbox{##1}{##2}{1.2pt}{\vbox{##3}}} \def\n##1##2##3{\gentabbox{##1}{##2}{0.4pt}{\vbox{##3}}} \vbox\bgroup\offinterlineskip
:.{1}.{3}\\
:.{2}\\
\egroup\egroup
,\hspace{.1in}
\bgroup\catcode`\:=13 \catcode`\.=13
\catcode`\;=13 \catcode`\>=13 \catcode`\^=13
\setlength{\tabheight}{3ex}\setlength{\tabwidth}{3ex} \def\b##1##2##3{\gentabbox{##1}{##2}{1.2pt}{\vbox{##3}}} \def\n##1##2##3{\gentabbox{##1}{##2}{0.4pt}{\vbox{##3}}} \vbox\bgroup\offinterlineskip
:.{4}\\
:.{5}\\
\egroup\egroup
\hspace{.02in},\hspace{.1in}
\bgroup\catcode`\:=13 \catcode`\.=13
\catcode`\;=13 \catcode`\>=13 \catcode`\^=13
\setlength{\tabheight}{3ex}\setlength{\tabwidth}{3ex} \def\b##1##2##3{\gentabbox{##1}{##2}{1.2pt}{\vbox{##3}}} \def\n##1##2##3{\gentabbox{##1}{##2}{0.4pt}{\vbox{##3}}} \vbox\bgroup\offinterlineskip
:.{6}.{7}.{10}\\
:.{8}.{9}.{11}\\
\egroup\egroup
\end{tiny}
\hspace{.05in}
\raisebox{.07in}{\Big)}.$$
Following [@stanley:some] and [@iancu], we define a map from $W_n$ to same-shape pairs of $r$-tuples of standard Young tableaux. Consider $w = [\zeta^{a_1} \sigma_1, \zeta^{a_2} \sigma_2, \ldots, \zeta^{a_n} \sigma_n] \in W_n$ and define ordered sets $w^{(k)}= (\sigma_i \, | \, a_i = k)$ for $0 \leq k < r$. Let $RS(w^{(k)}) = (P_k, Q_k)$ be the image of the sequence $w^{(k)}$ under the usual Robinson-Schensted map, labeling squares of $Q_k$ according to the relative positions of $i \in w^{(k)}$ within $w$, and define $$\mathbf{P}(w) = (P_0, P_1, \ldots, P_{r-1}) \text{ \phantom{XX} and \phantom{XX}} \mathbf{Q}(w) = (Q_0, Q_1, \ldots, Q_{r-1}).$$ The multitableaux Robinson-Schensted map is defined by $\mathbf{RS}(w) = (\mathbf{P}(w), \mathbf{Q}(w)).$ It maps $W_n$ onto the set of same-shape pairs of elements of $\mathbf{SYT}_n$ and is in fact a bijection.
Tableaux and multitableaux statistics
-------------------------------------
Our goal is to describe values of the sign representations on $W_n$ under the above generalization of the Robinson-Schensted map. To do so, we rely on a few statistics that can be readily computed from multitableaux.
An [*inversion*]{} in a Young tableau $T$ is a pair $(i,j)$ with $j>i$ for which the box labeled by $i$ is contained in a row strictly below the box labeled $j$. Let $\operatorname{\mathit{Inv}}(T)$ be the set of inversions in $T$, and write $\operatorname{\mathit{inv}}(T)$ for its cardinality. If $\mathbf{T}= (T_0, T_1, \ldots T_{r-1})$ is a multitableau, we extend this notion and define: $$\operatorname{\mathit{Inv}}(\mathbf{T}) = \bigsqcup_k \operatorname{\mathit{Inv}}(T_k) \sqcup \bigsqcup_{k<l} \operatorname{\mathit{Inv}}(T_k,T_l)$$ where $\operatorname{\mathit{Inv}}(T_k,T_l) = \{(j,i) \, | \, j>i, j \text{ is a label in $T_k$, $i$ is a label in $T_l$}\}.$ We will be mainly interested in the parity of the size of this set and define $$\operatorname{\mathit{sign}}(\mathbf{T}) = (-1)^{\operatorname{\mathit{inv}}(\mathbf{T})}.$$
For a Young tableau $T$, write $e(T)$ for the total number of boxes in its rows of even index. For a multitableau $\mathbf{T}= (T_0, T_1, \ldots T_{r-1})$, we write $\operatorname{sh}(T_k)$ for the shape of the Young diagram underlying $T_k$ and define the statistics $e$ and $\operatorname{\mathit{spin}}$ as follows: $$e(\mathbf{T}) = \sum_{k=0}^{r-1} e(T_k) \; \text{ and } \; \operatorname{\mathit{spin}}(\mathbf{T}) = \frac{1}{2} \sum_{k=0}^{r-1} k \cdot |\operatorname{sh}(T_k)|.$$
The $\operatorname{\mathit{spin}}$ statistic provides a simple description of the image of the subgroup $G(r,p,n)$ under the $r$-multitableaux Robinson-Schensted map.
$(\mathbf{P}, \mathbf{Q}) \in \mathbf{RS}(G(r,p,n))$ if and only if $2 \operatorname{\mathit{spin}}(\mathbf{P}) \equiv 0 \pmod{p}.$
Write $w= [\zeta^{a_1} \sigma_1, \zeta^{a_2} \sigma_2, \ldots, \zeta^{a_n} \sigma_n ] \in G(r,p,n)$ and for each $0 \leq k \leq r-1$, set $\mathcal{A}_k := \{ i \, | \, a_i = k \}$. If $\mathbf{RS}(w)=(\mathbf{P},\mathbf{Q})$, then $|\operatorname{sh}(P_k)| = |\mathcal{A}_k|$, and hence it follows that $k \cdot |\operatorname{sh}(P_k)|$ equals $\sum_{i \in \mathcal{A}_k} a_i$. Since $\mathbb{N}_n = \bigsqcup_{k=0}^{r-1} \mathcal{A}_k$, we conclude that $$\sum_{k=1}^{n} a_k = \sum_{k=0}^{r-1} k \cdot |\operatorname{sh}(P_k)| = 2 \operatorname{\mathit{spin}}(\mathbf{P}).$$ Finally, since $w \in G(r,p,n)$ if and only if $(\zeta^{a_1}\zeta^{a_2}\cdots\zeta^{a_n})^{\frac{r}{p}} = 1$, it follows that $w \in G(r,p,n)$ if and only if $2 \operatorname{\mathit{spin}}(\mathbf{P}) \equiv 0 \pmod{p},$ as claimed.
A set of functions and an example
---------------------------------
We define a sequence of functions on $W_n$. In the next section we will show that they coincide with the sign representations on $W_n$. Again, for $w \in W_n$, let $\mathbf{RS}(w)=(\mathbf{P},\mathbf{Q})$. For $0 \leq i < r$, we will write $$\pi_i(w) = (-1)^{e({\mathbf{P}})} \cdot (\zeta^{i})^{\operatorname{\mathit{spin}}(\mathbf{P}) + \operatorname{\mathit{spin}}(\mathbf{Q})} \cdot \operatorname{\mathit{sign}}(\mathbf{P}) \cdot \operatorname{\mathit{sign}}(\mathbf{Q}).$$
\[running\_example\] Consider $w = [ \zeta^1 \, 5, 1, \zeta^2 \, 3, 6, \zeta^2 \, 7, \zeta^1 \, 4, 2, 8]$ in $G(4,1,8).$ Recalling the notation in Section \[subsection:multitableaux\], we have $w^{(0)} = (1, 6, 2, 8)$, $w^{(1)} = (5,4)$, $w^{(2)} = (3,7)$, and $w^{(3)} = \emptyset$. Further, $$\begin{aligned}
& RS(w^{(0)}) = \begin{tiny} \left(\; \young(128,6)\hspace{.05in} ,\hspace{.05in} \young(248,7) \;\right) \end{tiny}
& RS(w^{(1)}) & = \begin{tiny} \left(\; \young(4,5)\hspace{.05in} , \hspace{.05in} \young(1,6) \;\right)\end{tiny} \\
& RS(w^{(2)}) = \begin{tiny}\left(\; \young(37)\hspace{.05in} , \hspace{.05in}\young(35) \;\right)\end{tiny}
& RS(w^{(3)}) & = \begin{tiny} \left(\; \emptyset, \emptyset \;\right)\end{tiny}.\end{aligned}$$ From these we construct the Robinson-Schensted image of $w$: $${\mathbf{RS}}(w)=({\mathbf{P}}, {\mathbf{Q}}) = \begin{tiny} \left( \; \left( \, \young(128,6) \, , \, \young(4,5) \, , \, \young(37) \, , \, \emptyset \, \right) \;, \;\left( \, \young(248,7) \, , \, \young(1,6) \, , \, \young(35) \, , \, \emptyset \, \right) \; \right).\end{tiny}$$ We read off $\operatorname{\mathit{inv}}({\mathbf{P}}) = 10$, $\operatorname{\mathit{inv}}({\mathbf{Q}}) = 12$, $e({\mathbf{P}}) = 2$, and $\operatorname{\mathit{spin}}({\mathbf{P}}) = \operatorname{\mathit{spin}}({\mathbf{Q}}) = 3$. Hence $\pi_i(w) =(\zeta^i)^2$ which coincides with $\operatorname{\mathit{sgn}}_i(w)$.
Sign under the Robinson-Schensted map
=====================================
The aim of this section is to verify the formulas for the one-dimensional representations of $W_n$ given above. We reduce the problem to the setting of the classical Robinson-Schensted map by considering a particular equivalence relation on the elements of $W_n$.
Ascending elements and admissible transformations
-------------------------------------------------
Recall the notation from Section \[subsection:multitableaux\] and write $w \in W_n$ in one-line notation as $$w=[\zeta^{a_1} \sigma_1, \zeta^{a_2} \sigma_2, \ldots, \zeta^{a_n} \sigma_n ].$$ We will say that $w$ is [*ascending*]{} if the exponents $a_k$ increase weakly and for every integer $0 \leq i<r-1$ all elements of $w^{(i)}$ are smaller than the elements of $w^{(i+1)}$. The following fact is immediate:
Suppose that $w \in W_n$ is ascending. Then its left and right Robinson-Schensted tableaux $\mathbf{P}{(w)}$ and $\mathbf{Q}{(w)}$ are both ascending.
We proceed to define equivalence classes on $W_n$, each with an ascending representative. Adopting the notation from the definition above, consider $i>0$ and define $L_i(w)=s_i \cdot w$ whenever $a_i \neq a_{i+1}$ and $R_i(w)=w \cdot s_i$ whenever $\sigma_l =i$, $\sigma_k=i+1$, and $a_l \neq a_{k}$. We will call operators of this form [*left*]{} and [*right admissible*]{}, respectively. To define an equivalence relation, we let $L_i(w) \sim w$ and $R_i(w) \sim w$ for every $i$ and take its transitive closure.
A right admissible operator permutes adjacent entries of $w$, while each left admissible operator exchanges some $\sigma_k$ and $\sigma_l$ which have different exponents in $w$ and satisfy $|\sigma_k-\sigma_l|=1$ while preserving the exponents in each position. Thus by successive application of right admissible operators, the entries of $w$ can be arranged to have weakly increasing exponents. Furthermore, by successive application of left admissible operators, the $\sigma_k$ can be permuted so that for every integer $0 \leq i<r-1$ all entries with coefficient $\zeta^i$ are smaller than those with coefficient $\zeta^{i+1}$. We have shown the following:
Each $w \in W_n$ has an ascending representative $\tilde{w}$ in the equivalence class defined by left and right admissible operators.
Admissible operators behave well with respect to the multitableaux Robinson-Schensted map:
For $w \in W_n$, $\mathbf{P}(w)=\mathbf{P}(R_i(w))$ and $\mathbf{Q}(w)=\mathbf{Q}(L_i(w)).$
Applying a right admissible transformation $R_i$ to $w$ does not change the relative order of the entries sharing the coefficient $\zeta^{a_k}$. Thus as ordered sets, we have $w^{(k)} = (R_i(w))^{(k)}$ for all $k$ and the first part of the proposition follows. Suppose that $\sigma_k$ and $\sigma_l$ have different exponents in $w$ and satisfy $|\sigma_k-\sigma_l|=1$. Exchanging the two while preserving the exponents in each position preserves the order new boxes are appended in constructing the component tableaux of $\mathbf{P}(w)$, thus a left admissible operation fixes $\mathbf{Q}(w)$, as claimed.
The multitableau $\mathbf{Q}(R_i(w))$ is obtained from $\mathbf{Q}(w)$ by exchanging boxes with labels $i$ and $i+1$. The result is standard as the boxes labeled $i$ and $i+1$ lie in different component tableaux of $\mathbf{Q}(w)$. Furthermore, $\operatorname{\mathit{Inv}}(\mathbf{Q}(R_i(w)))$ equals $\operatorname{\mathit{Inv}}(\mathbf{Q}(w))$ with $i$ and $i+1$ interchanged and the inversion $(i+1,i)$ either added or removed. In particular, a right admissible transformation changes $\operatorname{\mathit{inv}}(\mathbf{Q}(w))$ by one while fixing $\operatorname{\mathit{inv}}(Q_k(w))$ for each $k$. Similar reasoning can be applied to $\mathbf{P}(L_i(w))$, thereby proving:
\[proposition:inversions\] A right admissible transformation changes $\operatorname{\mathit{inv}}(\mathbf{Q}(w))$ by one while fixing $\operatorname{\mathit{inv}}(Q_k(w))$ for each $k$, while a left admissible transformation changes $\operatorname{\mathit{inv}}(\mathbf{P}(w))$ by one while fixing $\operatorname{\mathit{inv}}(P_k(w))$ for each $k$.
We are ready to prove the main result of this section which shows that it is only necessary to verify our formula for the $\operatorname{\mathit{sgn}}_i$ representations in the setting of ascending elements of $W_n$.
\[lemma:ascending\] Consider $w \in W_n$ and let $\tilde{w}$ be an ascending representative of $w$. Then for $0 \leq i < r$, $$\begin{aligned}
\pi_i(w) = \operatorname{\mathit{sgn}}_i(w) & \; \mbox{ if and only if } \; \pi_i(\tilde{w}) = \operatorname{\mathit{sgn}}_i(\tilde{w}).\end{aligned}$$
First note that $\tilde{w} = a \cdot w \cdot b$ for $a,b \in S_n$ and consequently for each $i$, $$\operatorname{\mathit{sgn}}_i(\tilde{w})= (-1)^{\ell(a)+ \ell(b)} \cdot \operatorname{\mathit{sgn}}_i(w).$$ Since $e$ and $\operatorname{\mathit{spin}}$ are shape-based statistics and admissible transformations preserve the shapes of the underlying tableaux, the only possible difference between $\pi_i(w)$ and $\pi_i(\tilde{w})$ lies in the signs of their left and right tableaux. This is addressed explicitly in Proposition \[proposition:inversions\]. We obtain $\pi_i(\tilde{w})= (-1)^{\ell(a)+ \ell(b)} \cdot \pi_i(w)$ and the lemma follows.
The general sign formula
------------------------
\[main\_theorem\] Let $w \in W_n$ and write ${\mathbf{RS}}(w) = ({\mathbf{P}},{\mathbf{Q}})$ for its image under the generalized Robinson-Schensted map. Given a primitive $r^{th}$ root of unity $\zeta$ and the associated family $\{\operatorname{\mathit{sgn}}_i \}_{i=0}^{r-1}$ of representations of $W_n$, we have $$\operatorname{\mathit{sgn}}_i(w) = (-1)^{e({\mathbf{P}})} \cdot (\zeta^i)^{\operatorname{\mathit{spin}}({\mathbf{P}}) + \operatorname{\mathit{spin}}({\mathbf{Q}})} \cdot \operatorname{\mathit{sign}}({\mathbf{P}}) \cdot \operatorname{\mathit{sign}}({\mathbf{Q}}).$$
By Lemma \[lemma:ascending\], it is enough to verify the above formula for ascending elements of $W_n$. Consequently, consider an ascending $w \in W_n$, writing it in one-line notation as $w=[\zeta^{a_1} \sigma_1, \zeta^{a_2} \sigma_2, \ldots, \zeta^{a_n} \sigma_n ].$ Let $\{P_k\}_{k=0}^{r-1}$ and $\{Q_k\}_{k=0}^{r-1}$ be the component tableaux of $\mathbf{P}$ and $\mathbf{Q}$. Since the latter tableaux are both ascending and of the same shape, for each $k$, $P_k$ and $Q_k$ are labeled by the same set of integers, say $\mathbf{n}_k$. Let $S_{\mathbf{n}_k}$ be the group of permutations of $\mathbf{n}_k$ and define $u_k = {RS}^{-1}(P_k, Q_k) \in S_{\mathbf{n}_k}$ for the natural extension of the classical Robinson-Schensted map to the ordered set $\mathbf{n}_k$. Let $u_k= 1 \in W_n$ if $\mathbf{n}_k =\emptyset$. Also let $u_r = [ \zeta^{a_1} 1, \zeta^{a_2} 2, \ldots, \zeta^{a_n} n ] \in W_n$. Note that we can view $S_{\mathbf{n}_k}$ as a subgroup of $S_n$ for each $k$. From this perspective, since $\mathbf{P}$ and $\mathbf{Q}$ are ascending, we have $w = u_0 \cdot u_1 \cdots u_r$ and consequently $$\begin{aligned}
\operatorname{\mathit{sgn}}_i(w) & = \operatorname{\mathit{sgn}}_i(u_0) \cdot \operatorname{\mathit{sgn}}_i(u_1) \cdots \operatorname{\mathit{sgn}}_i(u_r)\\
& = \operatorname{\mathit{sgn}}(u_0) \cdot \operatorname{\mathit{sgn}}(u_1) \cdots \operatorname{\mathit{sgn}}({u_{r-1}}) \cdot \operatorname{\mathit{sgn}}_i(u_r)\end{aligned}$$ where $\operatorname{\mathit{sgn}}$ is the usual sign of a permutation.
We would like to reduce this expression to the formula proposed by the theorem. The first step invokes the original result of A. Reifegerste [@reifegerste] and J. Sjöstrand [@sjostrand]. For $k<r$, we have $\operatorname{\mathit{sgn}}(u_k) = (-1)^{e_k} \cdot \operatorname{\mathit{sign}}(P_k) \cdot \operatorname{\mathit{sign}}(Q_k),$ where $e_k = e(P_k)$. Furthermore, $\operatorname{\mathit{sgn}}_i(u_r)$ is easily computed to be $(\zeta^i)^{\sum_{k=1}^n a_k}$ and consequently we have: $$\operatorname{\mathit{sgn}}_i(w)=(-1)^{\sum_{k=0}^{r-1} e_k} \cdot (\zeta^i)^{\sum_{k=1}^n a_k} \cdot (-1)^{\sum_{k=0}^{r-1} \operatorname{\mathit{inv}}(P_k) + \operatorname{\mathit{inv}}(Q_k)}.$$ To further simplify the above expression we note that $\sum_{k=0}^{r-1} e_k = e(\mathbf{P})$, by definition. Furthermore, if $n_k=|w^{(k)}|$ is the cardinality of the set of entries of $w$ with exponent equal to $k$, then $\sum_{k=1}^n a_k = \sum_{k=0}^{r-1} k \cdot n_k = 2 \operatorname{\mathit{spin}}(\mathbf{P}) = \operatorname{\mathit{spin}}(\mathbf{P}) + \operatorname{\mathit{spin}}(\mathbf{Q}).$
Finally, since both $\mathbf{P}$ and $\mathbf{Q}$ are ascending tableaux, the sets of inversions $\operatorname{\mathit{Inv}}(P_k, P_l)$ and $\operatorname{\mathit{Inv}}(Q_k, Q_l)$ are both empty for $k < l$. Consequently, $\operatorname{\mathit{inv}}(\mathbf{P})$ and $\operatorname{\mathit{inv}}(\mathbf{Q})$ are just the sums of the numbers of inversions in their component tableaux. The theorem now holds by direct substitution.
Continuing with $w$ from Example \[running\_example\], we have the ascending representative $\tilde{w} = [1, 3, 2, 4, \zeta^1 \, 6, \zeta^1 \, 5, \zeta^2 \, 7, \zeta^2 \, 8]$. Its Robinson-Schensted image is: $${\mathbf{RS}}(\tilde{w})=({\mathbf{P}}, {\mathbf{Q}}) = \begin{tiny}\left( \; \left( \, \young(124,3) \, , \, \young(5,6) \, , \, \young(78) \, , \, \emptyset \, \right) \;, \;\left( \, \young(124,3) \, , \, \young(5,6) \, , \, \young(78) \, , \, \emptyset \, \right) \; \right).\end{tiny}$$ In contrast to the behavior of the Robinson-Schensted map in the setting of classical Weyl groups, the fact that ${\mathbf{P}}$ and ${\mathbf{Q}}$ coincide does not necessarily imply that the corresponding group element is an involution, as occurs here. We decompose $\tilde{w}$ as $u_0 \cdot u_1 \cdot u_2 \cdot u_3 \cdot u_4$ where $u_3=1$ and: $$\begin{aligned}
&u_0 = [1,3,2,4] \in S_{\{1,2,3,4\}}
&u_1 &= [6,5] \in S_{\{5,6\}}\\
&u_2 = [7,8] \in S_{\{7,8\}}
&u_4 &= [ 1, 2, 3, 4, \zeta^1 \, 5, \zeta^1 \, 6, \zeta^2 \, 7, \zeta^2 \, 8] \in W_8.\end{aligned}$$ Since $\operatorname{\mathit{sgn}}(u_0)=\operatorname{\mathit{sgn}}(u_1)=-1$, $\operatorname{\mathit{sgn}}(u_2)=1$, and $\operatorname{\mathit{sgn}}_i(u_4) = (\zeta^i)^2$, we compute $\operatorname{\mathit{sgn}}_i(\tilde{w}) = (\zeta^i)^2$. Furthermore, since $\operatorname{\mathit{inv}}({\mathbf{P}})=\operatorname{\mathit{inv}}({\mathbf{Q}})=1$, $e({\mathbf{P}})=2$, and $\operatorname{\mathit{spin}}({\mathbf{P}})=\operatorname{\mathit{spin}}({\mathbf{Q}})=3$, we compute $\pi_i(\tilde{w}) = (\zeta^i)^2$, which agrees with $\operatorname{\mathit{sgn}}_i(\tilde{w})$, as expected.
It is also possible to approach the present problem by examining the coplactic classes on $W_n$ defined by the Robinson-Schensted map instead of the equivalence classes used here. In lieu of an ascending class representative, one can use the unique element in each coplactic class whose left multitableau has no inversions. Using the generalized Knuth relations for $W_n$ described in [@iancu], one can then produce the same formula by following their action on $\operatorname{\mathit{sgn}}_i$, albeit with a little more work. The advantage of the equivalence classes described in this paper is that they reduce the problem to the symmetric group setting as completely as possible.
S. Ariki. Representation theory of a [H]{}ecke algebra of [$G(r,p,n)$]{} , 77(1):164–185, 1995.
S. Ariki. obinson-[S]{}chensted correspondence and left cells. *Combinatorial methods in representation theory* (Kyoto, 1998), , 28:1–20, Kinokuniya, Tokyo, 2000.
C. Bonnaf[é]{} and L. Iancu. Left cells in type $B\sb n$ with unequal parameters. , 7:587–609, 2003.
L. Iancu. Cellules de [K]{}azhdan-[L]{}usztig et correspondance de [R]{}obinson-[S]{}chensted. , 336(10):791–794, 2003.
A. Joseph. A characteristic variety for the primitive spectrum of a semisimple [L]{}ie algebra. Short version in: *Non-Commutative Harmonic Analysis* (Actes Colloq., Marseille-Luminy, 1976), , 587:102–118, Spinger, Berlin, 1977.
A. Reifegerste. Permutation sign under the [R]{}obinson-[S]{}chensted correspondence. , 8(1):103–112, 2004.
G. C. Shephard and T. A. Todd. Finite unitary reflection groups. , 6:274–304, 1954.
J. Sjöstrand. On the sign-imbalance of partition shapes. , 111(2):190–203, 2005.
R. Stanley. Some aspects of groups acting on finite posets. , 32(2):132–161, 1982.
D.-N. Verma. Möbius inversion for the [B]{}ruhat ordering on a [W]{}eyl group. , 4(4):393–398, 1971.
[^1]: The second author would like to thank Skidmore College for its hospitality during the completion of this work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A simplicial cell ball is a simplicial poset whose geometric realization is homeomorphic to a ball. Recently, Samuel Kolins gave a series of necessary conditions and sufficient conditions on $h$-vectors of simplicial cell balls, and characterized them up to dimension $6$. In this paper, we extend Kolins’ results. We characterize all possible $h$-vectors of simplicial cell balls in arbitrary dimension.'
address: ' Satoshi Murai, Department of Mathematical Science, Faculty of Science, Yamaguchi University, 1677-1 Yoshida, Yamaguchi 753-8512, Japan. '
author:
- Satoshi Murai
title: '$h$-vectors of simplicial cell balls'
---
Introduction
============
A *simplicial poset* is a finite poset $P$ with a minimal element $\hat 0$ such that every interval $[\hat 0, \sigma]$ for $\sigma \in P$ is a Boolean algebra. A typical example of a simplicial poset is the face poset of a simplicial complex, but not all simplicial posets come from simplicial complexes. Simplicial posets are CW-posets. Thus for any simplicial poset $P$ there is a regular CW-complex $\Gamma(P)$ whose face poset is isomorphic to $P$ (see [@Bj pp. 8–9]). A simplicial poset $P$ (or a CW-complex $\Gamma(P)$) is called a *simplicial cell $d$-ball (respectively $d$-sphere)* if $\Gamma(P)$ is homeomorphic to a $d$-ball (respectively $d$-sphere). Kolins [@Ko] studied $h$-vectors of simplicial cell balls and gave a characterization of $h$-vectors of simplicial cell balls up to dimension $6$. In this paper, we extend the results of Kolins and give a complete characterization of $h$-vectors of simplicial cell balls.
Let $P$ be a simplicial poset. We say that an element $\sigma \in P$ has *rank $i$*, denoted ${\ensuremath{\mathrm{rank}}\hspace{1pt}}\sigma =i$, if $[\hat 0,\sigma]$ is a Boolean algebra of rank $i$. The *dimension* of $P$ is $$\dim P= \max \{{\ensuremath{\mathrm{rank}}\hspace{1pt}}\sigma : \sigma \in P\}-1.$$ Let $d=\dim P +1$ and let $f_i=f_i(P)$ be the number of elements $\sigma \in P$ having rank $i+1$ for $i=-1,0,\dots,d-1$. Thus $f_i$ is the number of $i$-dimensional cells of $\Gamma (P)$. The vector $f(P)=(f_{-1},f_0,\dots,f_{d-1})$ is called the *$f$-vector (face vector) of $P$*. We define the *$h$-vector $h(P)=(h_0(P),h_1(P),\dots,h_d(P)) \in { \ensuremath{\mathbb{Z}}}^{d+1}$ of $P$* by the relation $$\sum_{i=0}^d f_{i-1} (1-t)^{d-i} = \sum_{i=0}^d h_i(P) t^{d-i}.$$ Then knowing $f(P)$ is equivalent to knowing $h(P)$.
On face vectors of simplicial cell spheres, Stanley [@St] and Masuda [@Ma] proved the following result, which characterizes all possible $h$-vectors of simplicial cell spheres.
\[1.1\] Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$. Then ${\mathbf{h}}$ is the $h$-vector of a simplicial cell $(d-1)$-sphere if and only if it satisfies the following conditions:
- $h_0=h_d=1$ and $h_i=h_{d-i}$ for $i=1,2,\dots,d-1$.
- $h_i \geq 0$ for $i=0,1,\dots,d$.
- If $h_n=0$ for some $1 \leq n \leq d-1$ then $\sum_{k=0}^d h_k$ is even.
For a vector ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$, we define $\partial {\mathbf{h}}=(\partial h_0,\partial h_1,\dots,\partial h_{d-1}) \in { \ensuremath{\mathbb{Z}}}^d$ by $$\partial h_i = (h_0+h_1+\cdots +h_i)-(h_d+h_{d-1}+\cdots +h_{d-i})$$ for $i=0,1,\dots,d-1$. It is known that if ${\mathbf{h}}$ is the $h$-vector of a simplicial cell $(d-1)$-ball $P$, then $\partial {\mathbf{h}}$ is the $h$-vector of the boundary sphere of $P$. In this paper, we prove the next result, which characterizes all possible $h$-vectors of simplicial cell balls.
\[1.2\] Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$. Then ${\mathbf{h}}$ is the $h$-vector of a simplicial cell $(d-1)$-ball if and only if it satisfies the following conditions:
- $h_0=1,$ $h_d=0$ and $h_k \geq 0$ for $k=1,2,\dots,d-1$.
- $\partial h_k \geq 0$ for $k=0,1,\dots,d-1$.
- If $d$ is odd and $\partial h_n=0$ for some $1 \leq n \leq d-2$ then $\sum_{k=0}^d h_k$ is even.
- If $\partial h_n=0$ for some $1 \leq n \leq d-2$ then $$h_k + h_{k-1} + \cdots +h_{k-n+1} \geq \partial h_k
\ \ \mbox{ for } k=n,n+1,\dots,d-1.$$
- If $\partial h_i=0$ and $h_j=0$ for some positive integers $i$ and $j$ with $i+j \leq d$ then $\sum_{k=0}^d h_k$ is even.
- Suppose $\partial h_n = 0$ for some $1 \leq n < \frac d 2$. If $(h_\ell + h_{\ell-1} + \cdots + h_{\ell-n+1}) -\partial h_\ell < n$ for some $ n \leq \ell \leq d-n$ then $\sum_{k=0}^d h_k$ is even.
- Suppose $\partial h_i=0$ and $h_j=0$ for some integers $i$ and $j$ with $0< i < \frac d 2$ and $d-i < j <d$. If $\partial h_\ell \leq \ell$ for some $\ell \leq d-j$ then $\sum_{k=0}^d h_k$ is even.
Note that conditions (5), (6) and (7) are unnecessary when $d$ is odd. We also note that some special cases of the above theorem are due to Kolins [@Ko]. For necessity, (1), (2), (3) and some special cases of (4), (5) and (7) (the condition (4) for $n=1$, the condition (5) for $i=1,2$ and $j=1$, and the condition (7) for $\ell=1$) were proved in [@Ko Sections 3, 4 and 5]. Also, when $\partial {\mathbf{h}}$ is positive or $\sum_{k=0}^d h_k$ is even, the sufficiency of the theorem was proved in [@Ko Theorem 16]. In particular, for even dimensional simplicial cell balls, the sufficiency of the theorem is due to Kolins.
A simplicial poset $P$ (or a CW-complex $\Gamma(P)$) is called a *simplicial cell decomposition of a topological space $X$* if $\Gamma(P)$ is homeomorphic to $X$. It was proved in [@Mu] that face vectors of simplicial cell decompositions of some manifolds without boundary, such as product of spheres and real projective spaces, are characterized by conditions similar to those in Theorem \[1.1\]. It would be of interest to find manifolds $M$ with boundary such that face vectors of simplicial cell decompositions of $M$ can be characterized by conditions similar to those in Theorem \[1.2\].
This paper is organized as follows. In Section 2, we prove the necessity of Theorem \[1.2\] by using the face ring of a simplicial poset. In Section 3, we prove the sufficiency of Theorem \[1.2\] by using constructibility and shellability.
Proof of necessity
==================
In this section, we prove the necessity of Theorem \[1.2\]. To prove this, we need the face ring of a simplicial poset introduced by Stanley [@St]. We assume familiarity with commutative algebra theory.
We recall some basic notations on commutative algebra. Let $K$ be an infinite field, $R=R_0 \bigoplus R_1 \bigoplus \cdots$ a finitely generated commutative graded $K$-algebra with $R_0=K$, where $R_k$ is the graded component of $R$ of degree $k$, and let $M$ be a finitely generated graded $R$-module. The *Hilbert series $H_M(t)$ of $M$* is the formal power series $H_M(t)=\sum _{k = 0}^\infty (\dim_K M_k )t^k$. The *Krull dimension* of $M$, denoted $\dim M$, is the minimal $m$ for which there exist homogeneous elements $\theta_1,\dots,\theta_m \in R$ of positive degrees such that $M/(\theta_1,\dots,\theta_m)M$ is a finite dimensional $K$-vector space. When $\dim M=d$, a sequence of homogeneous elements $\theta_1,\dots,\theta_d \in R$ of positive degrees such that $M/(\theta_1,\dots,\theta_d)M$ is a finite dimensional $K$-vector space is called a *homogeneous system of parameters* (*h.s.o.p. *for short) *of* $M$. Moreover, if $\theta_1,\dots,\theta_d$ have degree $1$, then we call $\theta_1,\dots,\theta_d$ a *linear system of parameters* (*l.s.o.p. *for short) *of* $M$. We say that $M$ is *Cohen-Macaulay* if, for every (equivalently some) h.s.o.p. $\theta_1,\dots,\theta_d$ of $M$, one has that $\theta_i$ is a non-zero divisor of $M/(\theta_1,\dots,\theta_{i-1})M$ for all $i=1,2,\dots,d$. Also, $R$ is said to be Cohen-Macaulay if it is a Cohen-Macaulay $R$-module.
Let $P$ be a $(d-1)$-dimensional simplicial poset and let $A=K[x_\sigma: \sigma \in P \setminus \{ \hat 0\}]$ be the polynomial ring over an infinite field $K$ in indeterminates indexed by the elements in $P \setminus \{\hat 0\}$. We define the grading of $A$ by $\deg x_\sigma = {\ensuremath{\mathrm{rank}}\hspace{1pt}}\sigma$. The *face ring of $P$* is the quotient ring $K[P]=A/I_P$, where $I_P$ is the ideal generated by the following elements:
- $x_\sigma x_\tau$, if $\sigma,\tau \in P$ have no common upper bounds in $P$.
- $x_\sigma x_\tau - x_{\sigma \wedge \tau} \sum_\rho x_\rho$, where the summation runs over the all minimal upper bounds of $\sigma$ and $\tau$ and where $\sigma \wedge \tau$ is the meet (largest lower bound) of $\sigma$ and $\tau$, otherwise. (We consider $x_{\sigma \wedge \tau}=1$ if $\sigma \wedge \tau=\hat 0$.)
Note that, for all $\sigma,\tau \in P$ such that $\sigma$ and $\tau$ have a common upper bound $\rho$, there is the unique largest lower bound of $\sigma$ and $\tau$ since $[\hat 0,\rho]$ is a Boolean algebra. Also, since generators of $I_P$ are homogeneous, the ring $K[P]$ is graded. In the special case when $P$ is the face poset of a simplicial complex, the ring $K[P]$ is isomorphic to the face ring (Stanley-Reisner ring) of a simplicial complex [@St2 p. 53]. See [@MMP p. 467].
A simplicial poset $P$ is said to be *Cohen-Macaulay* if the ring $K[P]$ is Cohen-Macaulay. Note that this is equivalent to saying that the order complex of $P \setminus \{\hat 0\}$ is a Cohen-Macaulay simplicial complex. On face rings of simplicial posets, the following properties are known.
- The Krull dimension of $K[P]$ is $\dim P +1$.
- $K[P]$ has an l.s.o.p.
- $H_{K[P]}(t)=(h_0+h_1t+ \cdots + h_dt^d) /(1-t)^d$, where $d=\dim P +1$ and where $(h_0,h_1,\dots,h_d)=h(P)$.
- If $\Gamma(P)$ is homeomorphic to a ball or a sphere then $P$ is Cohen-Macaulay.
See [@St pp. 325–326] for the first three properties. The last property follows from a topological criterion of Cohen-Macaulay simplicial complexes [@St2 II, Corollary 4.2 and Proposition 4.3].
A subset $I$ of a simplicial poset $P$ is said to be an *order ideal of* $P$ if $\sigma \in I$ and $\tau \leq \sigma$ imply $\tau \in I$. Thus an order ideal of a simplicial poset is again a simplicial poset. For a simplicial poset $P$ and elements $\sigma_1,\dots,\sigma_k \in P$, we write $\langle\sigma_1,\dots,\sigma_k \rangle$ for the order ideal of $P$ generated by $\sigma_1,\dots,\sigma_k$, in other words, $$\langle \sigma_1,\dots,\sigma_k\rangle=\{\tau \in P: \mbox{ there is $1 \leq j \leq k$ such that } \tau \leq \sigma_j\}.$$ For elements $f_1,\dots,f_m$ in a ring $A$, we write $(f_1,\dots,f_m)$ for the ideal of $A$ generated by $f_1,\dots,f_m$. We often use the following obvious fact.
\[subposet\] Let $P$ be a simplicial poset and $A=K[x_\sigma: \sigma \in P \setminus \{\hat 0\}]$. If $Q$ is an order ideal of $P$ then $A/(I_P+(x_\sigma: \sigma \not \in Q))$ is isomorphic to $K[Q]$ as a ring.
Let $P$ be a simplicial cell $(d-1)$-ball. Then each rank $d-1$ element of $P$ is covered by at most two elements. The *boundary $\partial P$ of $P$* is the order ideal $$\partial P=\langle \sigma \in P: {\ensuremath{\mathrm{rank}}\hspace{1pt}}\sigma=d-1,\ \sigma \mbox{ is covered by exactly one element in }P\rangle.$$ Since $\partial P$ is the face poset of the boundary cell complex of $\Gamma(P)$, $\partial P$ is a simplicial cell $(d-2)$-sphere. Moreover, it is known that if ${\mathbf{h}}=(h_0,h_1,\dots,h_d)$ is the $h$-vector of $P$ then $\partial {\mathbf{h}}=(\partial h_0,\partial h_1,\dots,\partial h_{d-1})$ is the $h$-vector of $\partial P$. See [@Ko Section 3]. On $h$-vectors of simplicial cell balls, the following result was proved in [@Ko Theorem 5].
\[kolins\] Let $P$ be a simplicial cell $(d-1)$-ball and ${\mathbf{h}}=h(P)=(h_0,h_1,\dots,h_d)$. Then
- $h_0=1,$ $h_d=0$ and $h_k \geq 0$ for $k=1,2,\dots,d-1$.
- $\partial h_k \geq 0$ for $k=0,1,\dots,d-1$.
- If $d$ is odd and $\partial h_n=0$ for some $1 \leq n \leq d-2$ then $\sum_{k=0}^d h_k$ is even.
We sketch a proof of the above theorem. The vectors ${\mathbf{h}}$ and $\partial {\mathbf{h}}$ are non-negative since $h$-vectors of Cohen-Macaulay simplicial posets are non-negative [@St Theorem 3.10]. Also, since $h_d(P)$ is equal to the reduced Euler characteristic of $P$ times $(-1)^{d-1}$, $h_d(P)=0$ if $P$ is a simplicial cell ball. Finally, (3) follows from Theorem \[1.1\] as follows: If $d$ is odd then $$\partial h_{\frac {d-1} 2}= (h_0+\cdots +h_{\frac {d-1} 2}) - (h_{\frac {d+1} 2}+ \cdots +h_d)$$ is equal to $\sum_{k=0}^d h_k$ mod $2$. Since $\sum_{k=0}^{d-1} \partial h_k$ is even by Theorem \[1.1\](3), by the symmetry $\partial h_i=\partial h_{d-1-i}$ of $\partial {\mathbf{h}}$, $\partial h_{\frac {d-1} 2}$ must be even.
In the rest of this section, we prove that conditions (4), (5), (6) and (7) in Theorem \[1.2\] are necessary conditions of $h$-vectors of simplicial cell balls.
Proof of (4)
------------
\[trivial\] Let $P$ be a $(d-1)$-dimensional simplicial poset and $h(P)=(h_0,\dots,h_d)$. Let $K[P]=A/I_P$ be the face ring of $P$ and $\theta_1,\dots,\theta_d \in A_1$ an l.s.o.p. of $K[P]$. If $K[P]$ is Cohen-Macaulay then, for all integers $n >0$ and $n \leq k \leq d$, one has $$\dim_K \big(A/\big(I_P+(\theta_1,\dots,\theta_{d-1},\theta_d^n)\big)\big)_k
=h_k+h_{k-1} + \cdots + h_{k-n+1}.$$
Let $R[i]=A/(I_P+(\theta_1,\dots,\theta_{i}))$ for $i=0,1,\dots,d-1$. Since $R[0]=K[P]$ is Cohen-Macaulay, we have the exact sequence $$0 \longrightarrow R[i-1] \stackrel{\times \theta_i}{\longrightarrow} R[i-1] \longrightarrow R[i] \longrightarrow 0$$ for $i=1,2,\dots,d-1$. The exact sequences show that $H_{R[i]}(t)=(1-t)H_{R[i-1]}(t)$ for $i=1,2,\dots,d-1$. Then, since $H_{K[P]}(t)=(h_0+h_1t+\cdots+h_dt^d)/(1-t)^d$, we have $H_{R[d-1]}(t)=(h_0+h_1t+\cdots+h_dt^d)/(1-t)$. This fact says $$\dim_K \big(R[d-1]\big)_k = h_k + h_{k-1} + \cdots + h_1 +h_0$$ for $k=0,1,\dots,d.$ Also, since $\theta_1,\dots,\theta_{d-1},\theta_d^n$ is an h.s.o.p. of $K[P]$, the sequence $$0 \longrightarrow R[d-1] \stackrel{\times \theta_d^n}{\longrightarrow} R[d-1] \longrightarrow A/\big(I_P+(\theta_1,\dots,\theta_{d-1},\theta_d^n)\big) \longrightarrow 0$$ is exact. Hence $$\begin{aligned}
\dim_K \big(A/\big(I_P+(\theta_1,\dots,\theta_{d-1},\theta_d^n)\big)\big)_k
&=& \dim_K \big(R[d-1] \big)_k - \dim_K \big( R[d-1] \big)_{k-n}\\
&=& h_k+h_{k-1} + \cdots + h_{k-n+1}\end{aligned}$$ for $k=n,n+1,\dots,d$, as desired.
\[necessary1\] Let $P$ be a simplicial cell $(d-1)$-ball and ${\mathbf{h}}=h(P)=(h_0,h_1,\dots,h_d)$. If $\partial h_n=0$ for some $1 \leq n \leq d-2$ then $$h_k + h_{k-1} + \cdots +h_{k-n+1} \geq \partial h_k
\ \ \mbox{ for } k=n,n+1,\dots,d-1.$$
Let $A=K[x_\sigma: \sigma \in P \setminus \{ \hat 0\}]$ and $J=I_P + (x_\sigma: \sigma \not \in \partial P) \subset A$. Then $A/J \cong K[\partial P]$ by Lemma \[subposet\]. For a general choice of linear forms $\theta_1,\dots,\theta_d \in A_1$, we have (e.g. use a criterion of an l.s.o.p. for face rings [@Ko Lemma 6])
- $\theta_1,\dots,\theta_{d-1}$ is an l.s.o.p. of $A/J$.
- $\theta_1,\dots,\theta_{d-1},\theta_d$ is an l.s.o.p. of $A/I_P$.
Let $\overline \Theta = \theta_1,\dots,\theta_{d-1}$. Since $A/J$ is a $(d-1)$-dimensional Cohen–Macaulay ring, $$\begin{aligned}
\label{B}
\dim_K \left(A/ \left( J+ (\overline \Theta) \right) \right)_k = h_k(\partial P)=\partial h_k\end{aligned}$$ for $k = 0,1,\dots,d-1$. In particular, since $\dim_K(A/(J+(\overline{\Theta})))_n=\partial h_n =0$ by the assumption, we have $\theta_d^n \in J+ (\overline {\Theta}).$ Then, since $J \supset I_P$, we have $J+(\overline\Theta) \supset I_P + (\overline\Theta)+(\theta_d^n)$. Thus $$\begin{aligned}
\dim_K \big( A/ \big( I_P + (\overline {\Theta})+ (\theta_d^n ) \big) \big)_ k \geq \dim_K \left( A/ \left( J+(\overline {\Theta}) \right) \right)_k=\partial h_k\end{aligned}$$ for all $k =0,1,\dots,d-1$. Then the statement follows from Lemma \[trivial\].
Proof of (5), (6) and (7)
-------------------------
Let $P$ be a $(d-1)$-dimensional simplicial poset. A rank $1$ element of $P$ is called a *vertex of $P$* and a maximal element (w.r.t. a partial order) of $P$ is called a *facet of $P$*. The *cone of $P$* is the product of the posets $$C(P)=P \times \{1,2\}=\{(\sigma,i): \sigma \in P,\ i=1,2\},$$ where the ordering is defined by $(\sigma,i) \geq (\tau,j)$ if $\sigma \geq \tau$ and $i \geq j$. Then $C(P)$ is simplicial and $\Gamma(C(P))$ is homeomorphic to the (topological) cone of $\Gamma(P)$. Also, a straightforward computation shows $h_i(P)=h_i(C(P))$ for all $i=0,1,\dots,d$ and $h_{d+1}(C(P))=0$.
Let $P$ be a simplicial cell ball. We define the simplicial poset $SP$ by $$SP= P \cup C(\partial P)/\! \sim$$ where $\sim$ is the equivalence relation defined by $\sigma \sim (\sigma,1)$ for all $\sigma \in \partial P$. In other words, $SP$ is a simplicial cell sphere such that $\Gamma(SP)$ is obtained from $\Gamma(P)$ by coning off its boundary.
We need the next result due to Masuda [@Ma Section 5] (see also [@MR] for a simplified proof).
\[2-2\] Let $P$ be a simplicial cell $(d-1)$-sphere, $K[P]=A/I_P$ the face ring of $P$ and $\Theta=\theta_1,\dots,\theta_d \in A_1$ an l.s.o.p. of $K[P]$. If $x_{v_1} x_{v_2} \cdots x_{v_d}$ is zero in $A/(I_P+(\Theta))$ for any sequence of $d$ distinct vertices $v_1,v_2,\dots,v_d$ of $P$ then the number of the facets of $P$ is even.
In the rest of this section, we fix the following notation: Let $P$ be a simplicial cell $(d-1)$-ball and $${\mathbf{h}}=h(P)=(h_0,h_1,\dots,h_d).$$ Let $Q=C(\partial P)$ and $SP=P \cup C(\partial P)/\! \sim$ be simplicial posets defined as above. Also, let $A=K[x_\sigma: \sigma \in SP \setminus\{\hat 0\}]$, $$J=I_{SP}+(x_\sigma: \sigma \in SP \setminus P)=I_{SP}+(x_\sigma: \sigma \in Q \setminus \partial P),$$ $$L= I_{SP}+(x_\sigma: \sigma \in SP \setminus Q)
=I_{SP}+(x_\sigma: \sigma \in P \setminus \partial P),$$ and let $\Theta=\theta_1,\theta_2,\dots,\theta_d \in A_1$ be a sequence of general linear forms. Then we have the following properties.
- $A/J \cong K[P]$, $A/L\cong K[Q]$ and $$A/(L+J)= A/\big(I_{SP} + (x_\sigma: \sigma \in SP \setminus \partial P)\big) \cong K[\partial P].$$
- $\overline \Theta =\theta_1,\dots,\theta_{d-1}$ is an l.s.o.p. of $A/(L+J)$ and $\Theta=\theta_1,\dots,\theta_{d}$ is a common l.s.o.p. of $A/I_{SP}$, $A/J$ and $A/L$.
Recall that $\sum_{k=0}^d h_k$ is equal to the number of the facets of $P$. The next lemma shows that, to prove (5)–(7), it is enough to study the number of the facets of $SP$.
\[facets\] Suppose $\partial h_i=0$ for some $1 \leq i \leq d-2$. If the number of the facets of $SP$ is even then $\sum_{k=0}^d h_k$ is even.
By the construction of $SP$, the number of the facets of $P$ is equal to that of $SP$ minus that of $\partial P$. Then the statement follows from Theorem \[1.1\](3).
\[int1\] Let $v_1,v_2,\dots,v_d$ be distinct vertices of $SP$. Then
- If $\partial h_i=0$ for some $1 \leq i \leq d-2$ then $x_{v_1}x_{v_2} \cdots x_{v_i} \in L+(\Theta)$.
- If $h_j=0$ for some $1 \leq j \leq d$ then $x_{v_d} x_{v_{d-1}} \cdots x_{v_{d-j+1}} \in
J+(\Theta)$.
Recall $A/J \cong K[P]$ and $A/L \cong K[Q]$. Since $Q$ is the cone of $\partial P$, $$\begin{aligned}
\label{XX1}
\dim_K \big( A/ \big( L+(\Theta) \big) \big)_k=h_k(Q)=h_k(\partial P)=\partial h_k\end{aligned}$$ for $k =0,1,\dots,d-1$. Thus, $\partial h_i=0$ implies $x_{v_{1}}x_{v_2} \cdots x_{v_i} \in L+(\Theta).$ Similarly, since $\dim_K (A/(J+(\Theta)))_j=h_j$, $h_j=0$ implies $x_{v_d} x_{v_{d-1}} \cdots x_{v_{d-j+1}} \in J+(\Theta)$.
Now we prove (5) and (6).
\[2-3\] If $\partial h_i=0$ and $h_j=0$ for some positive integers $i$ and $j$ with $i+j \leq d$ then $\sum_{k=0}^d h_k$ is even.
Let $v_1,v_2,\dots,v_d$ be distinct vertices of $SP$. By Lemmas \[2-2\] and \[facets\], it is enough to prove that the monomial $x_{v_1}x_{v_2}\cdots x_{v_d}$ is zero in $A/(I_{SP} +(\Theta))$.
Since $i+j \leq d$, Lemma \[int1\] shows $$x_{v_1} x_{v_2} \cdots x_{v_d} \in
LJ+(\Theta) \subset
I_{SP} +(\Theta)+(x_\sigma x_\tau :
\sigma \in P \setminus \partial P,\
\tau \in Q \setminus \partial P).$$ However, for all $\sigma \in P \setminus \partial P$ and $\tau \in Q \setminus \partial P$, we have $x_\sigma x_\tau \in I_{SP}$ since $\sigma$ and $\tau$ have no common upper bounds. Thus $x_{v_1} x_{v_2} \cdots x_{v_d} \in I_{SP}+(\Theta)$.
\[necessary7\] Suppose $\partial h_n = 0$ for some $1 \leq n < \frac d 2$. If $(h_\ell + h_{\ell-1} + \cdots + h_{\ell-n+1}) -\partial h_\ell < n$ for some $ n \leq \ell \leq d-n$ then $\sum_{k=0}^d h_k$ is even.
Let $v_1,v_2,\dots,v_d$ be distinct vertices of $SP$. Then $x_{v_1} x_{v_2} \cdots x_{v_n} \in L+(\Theta)$ by Lemma \[int1\](i). Since $n+ \ell \leq d$, by the same argument as in the proof of Theorem \[2-3\], it is enough to prove that $$x_{v_d}x_{v_{d-1}}\cdots x_{v_{d-\ell+1}} \in J+(\Theta).$$
Consider the exact sequence $$0 \longrightarrow
N\longrightarrow
A/\big(J+(\overline \Theta,\theta_d^n)\big)
\longrightarrow
A/\big(L+J+(\overline \Theta,\theta_d^n)\big)
\longrightarrow
0,$$ where $N=(L+J+(\overline \Theta,\theta_d^n))
/(J+(\overline \Theta,\theta_d^n))$. Since $A/(L+J)\cong K[\partial P]$ and since $\overline \Theta$ is an l.s.o.p. of $A/(L+J)$, we have $\dim_K (A/(L+J+(\overline \Theta)))_n=\partial h_n=0$ and $$A_n=\big(L+J+(\overline \Theta)\big)_n.$$ In particular, we have $L+J+(\overline \Theta)=L+J+(\overline \Theta,\theta_d^n)$. Then, by Lemma \[trivial\] we have $$\begin{aligned}
\label{qqq}
\hspace{16pt} \dim_K N_\ell&=&\dim_K \big(A/\big(J+(\overline \Theta,\theta_d^n)\big)
\big)_\ell
-
\dim_K \big(A/\big(L+J+(\overline \Theta,\theta_d^n)\big)
\big)_\ell\\
\nonumber & =&
(h_\ell+ h_{\ell-1}+\cdots + h_{\ell-n+1}) - \partial h_\ell \\
\nonumber &\leq& n-1,\end{aligned}$$ where the last inequality follows from the assumption.
Let $u_k=x_{v_d} x_{v_{d-1}} \cdots x_{v_{d-k+1}}$ for $k=1,2,\dots,\ell$. Since $A_n=(L+J+(\overline \Theta))_n$, any element in $A$ of degree $n$ is contained in $L+J+(\overline \Theta)$. Since $\ell \geq n$, this fact says that the elements $$u_\ell,u_{\ell-1} \theta_d,\dots,u_{\ell-n+1}\theta_d^{n-1}$$ are contained in $L+J+(\overline \Theta)$ since they are products of elements of degree $1$. Then says that the above $n$ elements are $K$-linearly dependent in $N$. Thus there are $\alpha_0,\alpha_1,\dots,\alpha_{n-1} \in K$ with $(\alpha_0,\alpha_1,\dots,\alpha_{n-1}) \ne (0,0,\dots,0)$ such that $$\alpha_0 u_\ell + \alpha_1 u_{\ell-1} \theta_d + \cdots + \alpha_{n-1} u_{\ell-n+1}\theta_d^{n-1}
\in J+(\overline \Theta,\theta_d^n).$$ Let $s = \min \{k : \alpha_k \ne 0\}$. Then there is an $h \in A$ such that $$\theta_d^s \alpha_s u_{\ell-s} + \theta_d^{s+1} h \in J+ (\overline \Theta).$$ Since $\Theta$ is an l.s.o.p. of a Cohen-Macaulay ring $A/J$, $\theta_d$ is a non-zero divisor of $A/(J+(\overline \Theta))$. Thus $$\alpha_s u_{\ell-s} + \theta_d h \in J +(\overline \Theta).$$ This fact implies $u_{\ell-s} \in J+(\Theta)$. Since $u_{\ell-s}$ divides $u_\ell$, we have $$x_{v_d}x_{v_{d-1}} \cdots x_{v_{d-\ell+1}}
= u_{\ell}
\in J+(\Theta),$$ as desired.
Next, we study the condition (7). To simplify the notation, we write $R=A/J \cong K[P]$ and let $$C=(L+J)/J=
\big((x_\sigma: \sigma \in P \setminus \partial P)+J \big)/J.$$ Thus $C$ is the ideal of $R$ generated by the interior faces of $P$.
\[2.3.1\] $C$ is a $d$-dimensional Cohen-Macaulay $A$-module such that $H_C(t)={(h_d+h_{d-1}t + \cdots + h_0t^d)} /{(1-t)^d}$.
Consider the exact sequence $$\begin{aligned}
\label{E2.4}
0 \longrightarrow
C \longrightarrow
R \longrightarrow
R/C \longrightarrow
0.\end{aligned}$$ Observe $R \cong K[P]$ and $R/C = A/(L+J) \cong K[\partial P]$. Then $$H_C(t)=H_R(t)-H_{R/C}(t)=H_{K[P]}(t)-H_{K[\partial P]}(t)
=\frac {(h_d+h_{d-1}t + \cdots + h_0t^d)} {(1-t)^d}.$$ Also, $C$ is a $d$-dimensional Cohen-Macaulay $A$-module since $R/C$ is a $(d-1)$-dimensional Cohen-Macaulay $A$-module and $R$ is a $d$-dimensional Cohen-Macaulay $A$-module (e.g., use [@E Corollary 18.6]).
The ideal $C$ is the canonical module of $R$ (see [@BH Chapter 3]). This fact can be proved in the same way as in the proof of [@BH Theorem 5.7.1], and gives an alternative proof of Lemma \[2.3.1\].
Since $C$ is a submodule of $R$ with $\dim C=\dim R$ and since $C$ is Cohen-Macaulay, $\Theta$ is an l.s.o.p. of $C$ and we have $$\dim_K \big(C/(\Theta C)\big)_i=h_{d-i}$$ for $i=0,1,\dots,d$ by Lemma \[2.3.1\]. This fact shows
\[E2.5\] If $h_j=0$ for some $1 \leq j \leq d-1$ then $C_{d-j}=\Theta C_{d-j-1}.$
\[2.3.2\] Suppose $h_j=0$ for some $1 \leq j \leq d-1$. If $\theta_d f \in (C+\overline \Theta R)_{d-j}$ then, for any linear form $l \in R_1$, one has $l f \in C+\overline \Theta R$.
By Lemma \[E2.5\], $\theta_d f \in \overline \Theta R + \theta_d C$. Thus there is $g \in C$ such that $\theta_d (f-g) \in \overline \Theta R$. Since $\theta_d$ is a non-zero divisor of $R/(\overline \Theta R)$, we have $f-g \in \overline \Theta R$ and $f \in C + \overline \Theta R$. Hence $l f \in C+\overline \Theta R$.
The next theorem completes the proof of the necessity of Theorem \[1.2\].
\[necessity6\] Suppose $\partial h_i=0$ and $h_j=0$ for some integers $i$ and $j$ with $0< i < \frac d 2$ and $d-i < j <d$. If $\partial h_\ell \leq \ell$ for some $\ell \leq d-j$ then $\sum_{k=0}^d h_k$ is even.
Let $v_1,v_2,\dots,v_d$ be distinct vertices of $SP$. Since $i<j$, in the same way as in the proof of Theorem \[2-3\], it is enough to prove that $$x_{v_d} x_{v_{d-1}}\cdots x_{v_{j+1}} \in J + (\Theta).$$
Let $u_k=x_{v_d} x_{v_{d-1}} \cdots x_{v_{d-k+1}}$ for $k=0,1,\dots,d-j$, where $u_0=1$. We regard each $u_k$ as an element in $R=A/J$. Recall that $\ell$ is an integer such that $\ell \leq d-j$ and $\partial h_\ell \leq \ell$. Since $\overline \Theta$ is an l.s.o.p. of $R/C=A/(L+J) \cong K[\partial P]$, $$\dim_K \big(R/(C+\overline \Theta R)\big)_\ell = \partial h_\ell \leq \ell.$$ Then the elements $$u_\ell, u_{\ell-1} \theta_d ,u_{\ell-2} \theta_d^2,\dots,u_0 \theta_d^\ell \in R_\ell$$ are $K$-linearly dependent in $R/(C+\overline \Theta R)$. Thus there are $\alpha_0,\alpha_1,\dots,\alpha_{\ell} \in K$ with $(\alpha_0,\alpha_1,\dots,\alpha_{\ell}) \ne 0$ such that $$\begin{aligned}
\alpha_0 u_\ell + \alpha_1 u_{\ell-1} \theta_d + \cdots + \alpha_\ell u_0 \theta_d^\ell
\in
C+\overline \Theta R.\end{aligned}$$ Let $s = \min\{ k : \alpha_k \ne 0\}$ and $h=(\sum_{k=s+1}^{\ell} \alpha_k u_{\ell-k} \theta_d^{k})/\theta_d^{s+1}$, where $h=0$ if $s=\ell$. Then $\theta_d^{s} (\alpha_{s} u_{\ell-s}+ \theta_d h ) \in (C+ \overline \Theta R)_\ell$ and $$\theta_d^{d-j-(\ell-s)} (\alpha_{s} u_{\ell-s}+ \theta_d h ) \in (C+ \overline \Theta R)_{d-j}.$$ By Lemma \[2.3.2\], we have $(u_{d-j}/u_{\ell-s}) (\alpha_s u_{\ell -s} + \theta_d h) \in C+\overline \Theta R$. Hence $$\alpha_s u_{d-j} \in (C+ \overline \Theta R+ \theta_d R)_{d-j} \subset \Theta R,$$ where the last inclusion follows from Lemma \[E2.5\]. Since $\alpha_s \ne 0$, $u_{d-j}$ is contained in $\Theta R$. This fact shows $u_{d-j}=x_{v_d} x_{v_{d-1}}\cdots x_{v_{j+1}} \in J + (\Theta)$ when we regard $u_{d-j}$ as an element of $A$.
Proof of sufficiency
====================
In this section, we prove the sufficiency of Theorem \[1.2\].
Since the proof is a bit long, we first give a short outline of the proof. The key tools of the proof are the two main lemmas outlined below.
- In Lemma \[3.A\], we introduce a simple way to make a new simplicial cell ball from a given simplicial cell ball by gluing a pair of simplexes along their boundaries.
- In Lemma \[3.B\], we construct a special simplicial cell ball whose boundary has a nice property.
Then we construct a simplicial cell ball with the desired $h$-vector as follows: First, starting from a simplex, we gradually make a larger simplicial cell ball by using (I) repeatedly; Then, at a certain point, we attach a simplicial cell ball in (II) to change the shape of the boundary; Finally, we make a simplicial cell ball with the desired $h$-vector by using (I) again. The first technique (I) appeared in [@Ko]. Indeed, if $\partial {\mathbf{h}}$ is positive or $\sum_{k=0}^d h_k$ is odd, then the first technique is enough to prove the desired statement. However, we need (II) when $\partial {\mathbf{h}}$ has a zero entry and $\sum_{k=0}^d h_k$ is even.
Before proving the main lemmas, we first study what happens to the $h$-vector when we glue two simplicial posets. Two finite posets $P$ and $Q$ are said to be *isomorphic* if there is a bijection $f:P \to Q$, called an *isomorphism*, such that, for all $ \sigma, \tau \in P$, $\sigma> \tau$ if and only if $f(\sigma) > f(\tau)$. We write $P \cong Q$ if $P$ and $Q$ are isomorphic.
Let $P$ and $Q$ be simplicial posets. Let $I \subset P$ and $J \subset Q$ be order ideals which are isomorphic as posets. For a given isomorphism $f : I \to J$, we define the equivalence relation $\sim_f$ on $P \cup Q$ by $\sigma \sim_f f(\sigma)$ for all $ \sigma \in I$, and define $$P \cup_f Q = (P \cup Q) /\! \sim_f.$$ Then $P \cup_f Q$ is again a simplicial poset. We write $P \cup _I Q$ (or $P {{}_I \cup_J} Q$) for this simplicial poset $P \cup_f Q$ when the isomorphism $f$ is clear.
\[3.1\] Let $P$ and $Q$ be $(d-1)$-dimensional simplicial posets and let $I \subset P$ and $J \subset Q$ be order ideals with $I \cong J$. If $\dim I=d-2$ then $$h_k(P\cup_I Q)=h_k(P) + h_k(Q) -h_k(I) + h_{k-1}(I)$$ for $k=0,1,\dots,d$, where $h_{-1}(I)=h_d(I)=0$.
The desired formula follows from $f_k(P \cup_I Q)=f_k(P)+f_k(Q)-f_k(I)$ by straightforward computations.
For a positive integer $i$, we write $[i]=\{1,2,\dots,i\}$. For integers $1 \leq k \leq d$, let $$\Delta_d(k)=
\big\{F \subset [d]: F \not \supset [k]\big\}.$$ By defining a partial order on $\Delta_d(k)$ by inclusion, we regard $\Delta_d(k)$ as a simplicial poset of dimension $d-2$. In other words, $\Delta_d(k)$ is the (abstract) simplicial complex generated by $\{ [d] \setminus \{i\}: i=1,2,\dots,k\}$. Let ${\mathbf{e}}_0,{\mathbf{e}}_1,\dots,{\mathbf{e}}_d \in { \ensuremath{\mathbb{Z}}}^{d+1}$ be the unit vectors of ${ \ensuremath{\mathbb{Z}}}^{d+1}$. Thus ${\mathbf{e}}_i$ is the vector such that $(i+1)$-th entry of ${\mathbf{e}}_i$ is $1$ and all other entries of ${\mathbf{e}}_i$ are $0$. It is easy to see that $h_i(\Delta_d(k))=1$ for $i < k$ and $h_i(\Delta_d(k))=0$ for $i \geq k$. Then the next lemma follows from Lemma \[3.1\].
\[3.2\] With the same notation as in Lemma \[3.1\], if $I \cong \Delta_d(k)$ then $$h(P \cup_I Q)=h(P) + h(Q) - {\mathbf{e}}_0 + {\mathbf{e}}_{k}.$$
To apply Lemmas \[3.1\] and \[3.2\] efficiently, it is convenient to have a nice way to construct simplicial cell balls by gluing two simplicial cell balls. Constructibility and shellability give such nice ways.
A $(d-1)$-dimensional simplicial poset $P$ is said to be *constructible* if
- $P$ is a Boolean algebra of rank $d$, or
- there are $(d-1)$-dimensional constructible simplicial posets $Q$ and $Q'$ such that $P=Q { {}_I \cup_{I'} } Q'$ for some order ideals $I \subset Q$ and $I' \subset Q'$ and that $I \cong I'$ is a $(d-2)$-dimensional constructible simplicial poset.
A simplicial poset is said to be *pure* if all its facets have the same rank. A $(d-1)$-dimensional pure simplicial poset $P$ is said to be *shellable* if there is an order $\sigma_1,\sigma_2,\dots,\sigma_r$ of the facets of $P$, called a *shelling of $P$*, such that, for $i=2,3,\dots,r$, $\langle \sigma_1,\sigma_2,\dots,\sigma_{i-1} \rangle \cap \langle \sigma_i \rangle$ is isomorphic to $\Delta_d(k_i)$ for some $1 \leq k_i \leq d$.
Note that shellable simplicial posets are constructible since $\Delta_d(k)$ is constructible. In the rest of this paper, a simplicial cell $d$-ball which is constructible (respectively shellable) will be called a *constructible $d$-ball* (respectively *shellable $d$-ball*).
We need the following well-known properties of constructibility and shellability. See e.g., [@Bj2 Theorem 11.4 and Section 12].
- Let $P$ be a simplicial cell sphere. If $Q \subsetneq P$ is a shellable simplicial poset with $\dim Q=\dim P$, then $Q$ is a simplicial cell ball.
- Boolean algebras and $\Delta_d(k)$ with $k <d$ are shellable balls.
- Let $P$ and $Q$ be constructible $d$-balls, and let $I \subsetneq \partial P$ and $J \subsetneq \partial Q$ be order ideals with $I \cong J$. If $I$ is a constructible $(d-1)$-ball then $P {_{I}\cup_J} Q$ is a constructible $d$-ball.
For a finite set $X$, we write $\# X$ for the cardinality of $X$. The next well-known lemma, which immediately follows from Lemma \[3.2\] and the fact that a Boolean algebra has the $h$-vector $(1,0,\dots,0)$, is useful to compute $h$-vectors of shellable simplicial posets.
\[shellablehvector\] Let $P$ be a $(d-1)$-dimensional shellable simplicial poset with a shelling $\sigma_1,\sigma_2,\dots,\sigma_r$. Then, for $i=1,2,\dots,d$, one has $$h_i(P)=\# \big\{ \ell \in \{2,3,\dots,r\}:
\langle \sigma_1,\sigma_2,\dots,\sigma_{\ell-1} \rangle \cap \langle \sigma_\ell \rangle
\cong \Delta_d(i)
\big\}.$$
Now we prove our first main lemma. We say that a simplicial poset $P$ contains a simplicial poset $Q$ if there is an order ideal $I \subset P$ which is isomorphic to $Q$.
\[3.A\] Fix integers $1 \leq i \leq m \leq d$ and $1 \leq j \leq d-i$. Let $P$ be a constructible $(d-1)$-ball such that $\partial P$ contains $\Delta_d(m)$ and let $\ell = \min\{m,i+j\}$. Then there is a constructible $(d-1)$-ball $Q$ such that
- $h(Q)=h(P) + {\mathbf{e}}_i + {\mathbf{e}}_{j}$, and
- $\partial Q$ contains $\Delta_d(\ell)$.
By the assumption, there is an order ideal $L \subset \partial P$ which is isomorphic to $\Delta_d(m)$. We write $$L=\big\{\mathcal L(F): F \subset [d],\ F \not \supset [m]\big\},$$ where the ordering is defined by $\mathcal L(F)>\mathcal L(G)$ if $F \supset G$. Since $i \leq m$, $$I=\big\{\mathcal L(F): F \subset [d],\ F \not \supset [i]\big\} \cong \Delta_d(i)$$ is an order ideal of $\partial P$. We prepare two Boolean algebras $$A=\big\{{\mathcal{A}}(F): F \subset [d]\big\} \mbox{ and } B=\big\{{\mathcal{B}}(F): F \subset [d] \big\},$$ where the ordering is defined by inclusion on $F$. Consider the order ideal $$J=\big\{ {\mathcal{A}}(F): F \subset [d],\ F \not \supset \{i+1,\dots,i+j\}\big\}
\cong \Delta_d(j).$$ Let $$Q=
(P \cup_I A) \cup_J B
=(P \cup A \cup B) /\! \sim,$$ where $\sim$ is the equivalence relation defined by $\mathcal L(F) \sim {\mathcal{A}}(F)$ if $F \not \supset [i]$ and ${\mathcal{A}}(F) \sim {\mathcal{B}}(F)$ if $F \not \supset \{i+1,\dots,i+j\}$. Then $Q$ is a constructible $(d-1)$-ball and has the desired $h$-vector by Lemma \[3.2\].
It remains to prove that $\partial Q$ contains $\Delta_d(\ell)$. Clearly, $\partial Q$ contains an order ideal $$\Delta
= \big\{{\mathcal{B}}(F): F \subset [d],\ F \not \supset [i] \big\}
\bigcup \big\{\mathcal L(F): F \subset [d],\ F \not \supset \{i+1,\dots,\ell\}\big\}/\!\sim.$$ We claim that $\Delta \cong \Delta_d(\ell)$. What we must prove is that if $F \not \supset [i]$ and $F \not \supset \{i+1,\dots,\ell\}$ then $\mathcal L(F) \sim {\mathcal{B}}(F)$. Indeed, $F \not \supset [i]$ implies $\mathcal L (F) \sim {\mathcal{A}}(F)$. Also, since $\ell \leq i+j$, $F \not \supset \{i+1,\dots,\ell\}$ implies ${\mathcal{A}}(F)\sim {\mathcal{B}}(F)$.
We note that Lemma \[3.A\] essentially appeared in the proof of [@Ko Theorem 16].
Next, we develop a way to construct simplicial cell balls by applying Lemma \[3.A\] repeatedly. Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}_{\geq 0}^{d+1}$ be a vector such that $\sum_{k=0}^d h_k$ is even. We write ${\mathbf{h}}={\mathbf{e}}_{i_1}+{\mathbf{e}}_{i_2}+ \cdots + {\mathbf{e}}_{i_a}$, where $a=\sum_{k=0}^d h_k$ and where $i_1 \leq i_2 \leq \cdots \leq i_a$. Consider the following decomposition of ${\mathbf{h}}$: $${\mathbf{h}}= ({\mathbf{e}}_{i_1} + {\mathbf{e}}_{i_a}) + ({\mathbf{e}}_{i_2} + {\mathbf{e}}_{i_{a-1}}) + \cdots + ({\mathbf{e}}_{i_{\frac a 2}}+{\mathbf{e}}_{i_{\frac a 2 +1}}).$$ We define the *initial number ${\ensuremath{\mathrm{in}}\hspace{1pt}}({\mathbf{h}})$ of ${\mathbf{h}}$* and the *width ${\mathrm{width}}({\mathbf{h}})$ of ${\mathbf{h}}$* by $${\ensuremath{\mathrm{in}}\hspace{1pt}}({\mathbf{h}})=i_{\frac a 2}$$ and $${\mathrm{width}}({\mathbf{h}})=\min \{ i_k+i_{a-k+1}: k=1,2,\dots, \mbox{${a \over 2}$}\}.$$ The following fact is straightforward.
\[dist\] With the same notation as above, one has $${\mathrm{width}}({\mathbf{h}})=
\max\{\ell : h_0+ h_1 +\cdots +h_k \leq h_d+h_{d-1} + \cdots +h_{\ell -k}
\mbox{ for all }k\}.$$
We need the following reformulation of Lemma \[3.A\].
\[kolinsrev\] Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}_{\geq 0}^{d+1}$ be such that $h_0=0$, $\partial h_i \geq 0$ for $i=0,1,\dots,d-1$ and $\sum_{k=0}^d h_k$ is even. Let $P$ be a constructible $(d-1)$-ball such that $\partial P$ contains $\Delta_d(\ell)$ with $\ell \geq {\ensuremath{\mathrm{in}}\hspace{1pt}}({\mathbf{h}})$. Then there is a constructible $(d-1)$-ball $Q$ such that $h(Q)=h(P)+{\mathbf{h}}$ and $\partial Q$ contains $\Delta_d(\min\{\ell,{\mathrm{width}}({\mathbf{h}})\})$.
We write $${\mathbf{h}}= ({\mathbf{e}}_{i_1} + {\mathbf{e}}_{i_a}) + ({\mathbf{e}}_{i_2} + {\mathbf{e}}_{i_{a-1}}) + \cdots + ({\mathbf{e}}_{i_{\frac a 2}}+{\mathbf{e}}_{i_{\frac a 2 +1}}),$$ where $a=\sum_{k=0}^d h_k$ and where $i_1 \leq i_2 \leq \cdots \leq i_a$. For $k=1,2,\dots,\frac a 2$, let $$w_k=\min \{\ell,i_1+i_a,i_2+i_{a-1},\dots,i_k + i_{a-k+1} \}$$ and let $w_0=\ell$. The assumption $\partial h_i \geq 0$ for $i=0,1,\dots,d-1$ implies that $$\begin{aligned}
\label{XYZ}
i_l+i_{a-l+1} \leq d\end{aligned}$$ for all $l=1,2,\dots, \frac a 2$.
We inductively prove that, for $k=0,1,\dots,\frac a 2$, there is a constructible $(d-1)$-ball $Q_k$ which satisfies the following two conditions:
- $h(Q_k)=h(P)+({\mathbf{e}}_{i_1} + {\mathbf{e}}_{i_a}) + ({\mathbf{e}}_{i_2} + {\mathbf{e}}_{i_{a-1}}) + \cdots + ({\mathbf{e}}_{i_k}+{\mathbf{e}}_{i_{a-k+1}})$;
- $\partial Q_k$ contains $\Delta_d(w_k)$.
Since $w_{\frac a 2}=\min\{\ell,{\mathrm{width}}({\mathbf{h}})\}$, if the above statement holds then $Q_{\frac a 2}$ is the simplicial poset with the desired properties.
For $k=0$, $Q_0=P$ satisfies the desired conditions. Suppose that such $Q_k$, where $0\leq k < \frac a 2$, exists. We prove the existence of $Q_{k+1}$. Observe $$i_{k+1} \leq \min\{\ell, i_{a-k}\} \leq w_k,$$ where the last inequality follows from $i_{a-k}\leq i_{a-k+1} \leq \cdots \leq i_a$. Recall that $i_{a-k} \leq d-i_{k+1}$ by and that $\partial Q_k$ contains $\Delta_d(w_k)$ by the induction hypothesis. By Lemma \[3.A\] (apply the case when $P=Q_k$, $i=i_{k+1}$, $m=w_k$ and $j=i_{a-k}$) there is a constructible $(d-1)$-ball $Q_{k+1}$ such that $h(Q_{k+1})=h(Q_k)+ {\mathbf{e}}_{i_{k+1}} + {\mathbf{e}}_{i_{a-k}}$ and $\partial Q_{k+1}$ contains $\Delta_d(\min\{w_k,i_{k+1}+i_{a-k}\})=\Delta_d(w_{k+1})$ as desired.
By considering the special case when $P$ is a Boolean algebra, we obtain the next corollary.
\[kolins2\] Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}_{\geq 0}^{d+1}$ be such that $h_0=0$, $\partial h_i \geq 0$ for $i=0,1,\dots,d-1$ and $\sum_{k=0}^d h_k$ is even. Then there is a constructible $(d-1)$-ball $P$ such that $h(P)={\mathbf{e}}_0+{\mathbf{h}}$ and $\partial P$ contains $\Delta_d({\mathrm{width}}({\mathbf{h}}))$.
The next lemma is our second main lemma.
\[3.B\] Let $n,m$ and $d$ be positive integers such that $n \leq \frac d 2$ and $d-n \leq m <d$. Let ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$ be a vector such that $\sum_{k=0}^d h_k=d$, $h_i=1$ for $0 \leq i <d-n$, $h_i>0$ for $d-n \leq i \leq m$ and $h_i=0$ for $i>m$. There is a constructible $(d-1)$-ball $P$ such that
- $h(P)={\mathbf{h}}$, and
- $\partial P$ contains two ideals $I_1$ and $I_2$ that have no common facets such that $I_1 \cong \Delta_d(n)$ and $I_2 \cong \Delta_d(d-n)$.
Consider the simplicial posets $$A = \big\{{\mathcal{A}}(F): F \subset [d+1],\ F \not \supset [n] \big\} \cong \Delta_{d+1}(n)$$ and $$B = \big\{{\mathcal{B}}(F): F \subset [d+1],\ F \not \supset \{n+1,\dots,d\} \big\} \cong \Delta_{d+1}(d-n),$$ where the ordering is defined by inclusion on $F$. Consider the simplicial complex $$\begin{aligned}
\Sigma\!\! &=&\!\! \big\langle [d+1]\setminus \{i,j\}: i\in [n],\ j\in \{n+1,\dots,d\} \big\rangle\\
&=&\!\! \big\{ F\! \subset\! [d+1]:\! \mbox{ $F \subset [d+1]\setminus \{i,j\}$ for some } i\! \in\! [n] \mbox{ and } j\! \in\! \{n\!+\!1,\dots,d\} \big\}.\end{aligned}$$ Let $$I_1=\big\langle {\mathcal{A}}\big([d+1]\setminus \{i,d+1\}\big): i=1,2,\dots,n\big\rangle \cong \Delta_d(n)$$ and $$I_2=\big\langle {\mathcal{B}}\big([d+1]\setminus \{i,d+1\}\big): i=n+1,n+2,\dots,d \big\rangle \cong \Delta_d(d-n).$$ Then $$\begin{aligned}
\partial A
&=&
\big\langle {\mathcal{A}}\big([d+1]\setminus\{i,j\}\big): i=1,2,\dots,n,\ j=n+1,\dots,d+1\big\rangle\\
&=&\big\{{\mathcal{A}}(F): F \in \Sigma\big\} \cup I_1,\end{aligned}$$ and similarly $$\partial B=\big\{{\mathcal{B}}(F): F \in \Sigma\big\} \cup I_2.$$
Let ${\boldsymbol g}=(g_0,g_1,\dots,g_d)= {\mathbf{h}}-\sum_{i=0}^{n-1} {\mathbf{e}}_i - \sum_{i=0}^{d-n-1} {\mathbf{e}}_i$. Observe that $A$ and $B$ are constructible $(d-1)$-balls with $h(A)=\sum_{i=0}^{n-1} {\mathbf{e}}_i$ and $h(B)=\sum_{i=0}^{d-n-1} {\mathbf{e}}_i$. By Lemma \[3.1\], if there is a constructible $(d-2)$-ball $\Omega \subset \Sigma$ such that $-h_i(\Omega)+h_{i-1}(\Omega)=g_i$ for all $i$, then the simplicial poset $$P=A \cup_\Omega B = (A \cup B) /\! \sim,$$ where $\sim$ is the equivalence relation defined by ${\mathcal{A}}(F) \sim {\mathcal{B}}(F)$ for $F \in \Omega$, satisfies the desired conditions (i) and (ii). Thus the next lemma completes the proof.
\[3.5\] With the same notation as in the proof of Lemma \[3.B\], there is a shellable $(d-2)$-ball $\Omega \subset \Sigma$ such that $-h_i(\Omega)+h_{i-1}(\Omega)=g_i$ for all $i$.
Observe $${\boldsymbol g}=(-1,\dots,-1,0,\dots,0,h_{d-n},h_{d-n+1},\dots,h_d),$$ where $-1$ appears in the first $n$ entries. Let $$\alpha_\ell=h_{(d-n)+\ell-1} + h_{(d-n)+\ell} +\cdots + h_d$$ for $\ell=1,2,\dots,n+1$. By the assumption on ${\mathbf{h}}=(h_0,h_1,\dots,h_d)$, $\alpha_1=n$ and $\alpha_\ell \leq n -(\ell-1) $ for all $\ell$.
Let $$\begin{aligned}
D&=&\big\{\{p,q\}: p\in [n],\ q \in \{n+1,\dots,d\},\ p+q \leq d \big\}\\
&&\bigcup \left[\bigcup_{\ell=1}^{m-(d-n)+1} \big\{ \{p,d+\ell-p\}: p=\ell,\ell+1,\dots,\ell+\alpha_{\ell}-1 \big\} \right].\end{aligned}$$ Since $\alpha_\ell \leq n-(\ell-1)$, $D \subset \{\{p,q\}: p \in [n],\ q\in\{n+1,\dots,d\} \}$. Let $$\Omega=\big\langle [d+1] \setminus \{p,q\}: \{p,q\} \in D \big\rangle \subset \Sigma.$$ We claim that $\Omega$ satisfies the desired conditions.
We first prove that $\Omega$ is shellable. We define the total order $\succ$ on $D$ by $\{p,q\} \succ \{p',q'\}$, where $p<q$ and $p'<q'$, if $p<p'$ or $p=p'$ and $q<q'$. We show that the total order on the facets of $\Omega$ induced by $\succ$ gives a shelling of $\Omega$. Since $h_{n-d},\dots,h_m>0$ and $h_{m+1}= \cdots =h_d=0$ by the assumption, we have $\alpha_1> \alpha_2> \cdots > \alpha _{m-(d-n)+2}= \cdots = \alpha _{n+1}=0$. Then, by the construction of $D$, for any $\{p,q\} \in D$ with $p<q$, we have $\{p-1,q\} \in D$ if $p \ne 1$ and $\{p,q-1\} \in D$ if $q \ne n+1$. This fact shows that, for any $\{p,q\} \in D$ with $\{p,q\} \ne \{1,n+1\}$, $$\begin{aligned}
&&\big\langle [d+1]\setminus \{s,t\}: \{s,t\} \in D,\ \{s,t\} \succ \{p,q\} \big\rangle
\bigcap \big\langle [d+1]\setminus \{p,q\} \big\rangle\\
&&= \big\langle [d+1] \setminus \{p,q,k\}:
k=1,2,\dots,p-1,n+1,n+2,\dots,q-1 \big \rangle\\
&& \cong \Delta_{d-1}(p+q-n-2).\end{aligned}$$ Hence $\Omega$ is shellable. (To see that the second line contains the first line, use the fact that if $F$ is contained in the first line then one has either $p'=\min([n] \setminus F) <p$ or $q'=\min(\{n+1,\dots,d\}\setminus F) <q$ since $[d+1]\setminus \{p',q'\}$ must be contained in $\{ [d+1]\setminus \{s,t\}: \{s,t\} \in D,\ \{s,t\} \succ \{p,q\} \}$.)
The above shelling and Lemma \[shellablehvector\] show $$h_i(\Omega)=\# \big\{\{p,q\} \in D: p+q-n-2=i \big\}$$ for $i=0,1,\dots,d-1$. Then a simple counting shows $$\begin{aligned}
h_i(\Omega)=
\left\{
\begin{array}{ll}
i+1, & \mbox{ if } i \leq n-1,\\
n, & \mbox{ if } n \leq i \leq d-n-1,\\
\alpha_{i-(d-n)+2}, & \mbox{ if } i \geq d-n.
\end{array}
\right.\end{aligned}$$ Note that $h_{d-n-1}(\Omega)=\alpha_1=n$. Then we have $-h_i(\Omega)+h_{i-1}(\Omega)=g_i$ for all $i$.
Finally, $\Omega$ is a simplicial cell $(d-2)$-ball since $\Omega$ can be identified with a $(d-2)$-dimensional subposet of a $(d-2)$-dimensional simplicial cell sphere $\partial A$.
The simplicial complex $\Sigma$ is the cone of the join of the boundaries of simplexes, and $h$-vectors of shellable subcomplexes of $\Sigma$ having the same dimension as $\Sigma$ are characterized in [@BFS]. This will give an alternative proof of Lemma \[3.5\].
\[sufficiency\] If ${\mathbf{h}}=(h_0,h_1,\dots,h_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$ satisfies conditions (1)–(7) in Theorem \[1.2\], then ${\mathbf{h}}$ is the $h$-vector of a simplicial cell $(d-1)$-ball.
If $\partial {\mathbf{h}}$ is positive or $\sum_{k=0}^d h_k$ is even, then the assertion was proved in [@Ko Theorem 16]. But we include a proof for completeness (our proof is essentially the same as that of Kolins).
[*Case 1.*]{} Suppose that all entries of $\partial {\mathbf{h}}$ are positive. If $\sum_{k=0}^d h_k$ is odd then, since $\partial({\mathbf{h}}-{\mathbf{e}}_0)$ is a non-negative vector, Corollary \[kolins2\] says that ${\mathbf{h}}={\mathbf{e}}_0+({\mathbf{h}}-{\mathbf{e}}_0)$ is the ${\mathbf{h}}$-vector of a simplicial cell $(d-1)$-ball. Suppose that $\sum_{k=0}^d h_k$ is even. We write $${\mathbf{h}}-{\mathbf{e}}_0=({\mathbf{e}}_{i_1}+{\mathbf{e}}_{i_a})+ ({\mathbf{e}}_{i_2}+{\mathbf{e}}_{i_{a-1}})+ \cdots + ({\mathbf{e}}_{i_{\frac {a-1} 2}}+{\mathbf{e}}_{i_{\frac {a+3} 2 }}) + {\mathbf{e}}_{i_{\frac {a+1} 2 }},$$ where $a=\sum_{k=0}^d h_k -1$ and where $i_1 \leq \cdots \leq i_a$. Let ${\mathbf{h}}'={\mathbf{h}}-{\mathbf{e}}_0 -{\mathbf{e}}_{i_{\frac {a+1} 2 }}$. Since $\partial({\mathbf{h}}-{\mathbf{e}}_0)$ is non-negative, we have $i_{a-l+1} \leq d-i_l$ for $l=1,2,\dots,\frac {a-1} 2$. Then $\partial {\mathbf{h}}'$ is non-negative and ${\mathrm{width}}({\mathbf{h}}') \geq \min\{i_{\frac {a+3} 2 },\dots,i_a\} \geq i_{\frac {a+1} 2}$. Thus, by Corollary \[kolins2\], there is a constructible $(d-1)$-ball $P$ such that $h(P)={\mathbf{e}}_0+{\mathbf{h}}'= {\mathbf{h}}-{\mathbf{e}}_{i_{\frac {a+1} 2}}$ and $\partial P$ contains $\Delta_d(i_{\frac {a+1} 2})$. Let $Q$ be a simplicial cell $(d-1)$-ball obtained from $P$ by gluing a Boolean algebra of rank $d$ along an order ideal $J \subset \partial P$ with $J \cong \Delta_d(i_{\frac {a+1} 2})$. Lemma \[3.2\] guarantees that the $h$-vector of $Q$ is $h(P)+{\mathbf{e}}_{i_{\frac {a+1} 2}}={\mathbf{h}}$.
[*Case 2.*]{} Suppose that $\partial {\mathbf{h}}$ has a zero entry and $\sum_{k=0}^d h_k$ is even. Let $$n=\min\{k :\partial h_k=0\}$$ and let ${\mathbf{h}}'=(h_0',h_1',\dots,h_d')={\mathbf{h}}-{\mathbf{e}}_0-{\mathbf{e}}_{d-n}$. Note that ${\mathbf{h}}'$ is a non-negative vector since $\partial h_{n-1}>0$ and $\partial h_n=0$ imply $h_{d-n} > 0$. By the same argument as in Case 1, it is enough to prove that there is a constructible $(d-1)$-ball $P$ such that $h(P)={\mathbf{h}}-{\mathbf{e}}_{d-n}$ and $\partial P$ contains $\Delta_d(d-n)$. By the choice of $n$, $\partial {\mathbf{h}}'$ is non-negative. Thus, by Corollary \[kolins2\], what we must prove is ${\mathrm{width}}({\mathbf{h}}') \geq d-n$.
By the condition (4), for $k \geq 0$, we have $$h_{k+n}+ \cdots + h_{k+1}
\geq
(h_0+ h_1+ \cdots + h_{k+n}) - (h_d+ h_{d-1}+ \cdots + h_{d-k-n})$$ (we substitute $k$ by $k+n$ to the inequality in (4)). Then we have $h_0+\cdots+h_{k} \leq h_d+ \cdots + h_{d-k-n}$ for all $k\geq 0$. This fact shows $$h_0'+h_1'+\cdots+h_{k}'
\leq
h_d'+ h_{d-1}'+ \cdots + h'_{d-k-n}$$ for all $k \geq 0$. Then Lemma \[dist\] guarantees ${\mathrm{width}}({\mathbf{h}}') \geq d-n$.
[*Case 3.*]{} Suppose that $\partial {\mathbf{h}}$ has a zero entry and $\sum_{k=0}^d h_k$ is odd. By the condition (3), $d$ must be even. Let $$n=\min\{k :\partial h_k=0\}.$$ By the symmetry of $\partial {\mathbf{h}}$, we have $n < \frac d 2$. Let $$m=\min\{k: h_k =0\}-1.$$ Thus $h_0,h_1,\dots,h_m>0$ and $h_{m+1}=0$. Then $m \leq d-1$ since $h_d=0$ and $m \geq d-n$ by the condition (5).
By the condition (7), $\partial h_{d-m-1} \geq d-m$. Recall that $\partial h_n=0$, $d-m \leq n$ and $\partial h_k > 0$ for $k<n$. Then, there is a sequence of integers $d-m \leq s_1 <s_2 < \cdots < s_p =n$ such that
- $d-m > \partial h_{s_1} > \cdots > \partial h_{s_p}=0$, and
- for any $s_{i-1} <j < s_i$ with $i \geq 2$ one has $\partial h_j \geq \partial h_{s_{i-1}}$, and for any $d-m -1 <j < s_1$ one has $\partial h_j \geq d-m$.
(The next figure explains what $s_1,s_2,\dots,s_p$ are.)
0.1in
( 50.1300, 17.5600)( 7.9800,-25.5600)
(14.2800,-10.5100)[(0,0)[$d\!-\!m$]{}]{}(16.2800,-6.9000)[(0,0)[$\partial h_i$]{}]{}
(14.2800,-13.0200)[(0,0)[$\partial h_{s_1}$]{}]{}(14.2800,-16.7800)[(0,0)[$\partial h_{s_2}$]{}]{}(14.2800,-20.5400)[(0,0)[$\partial h_{s_3}$]{}]{}(22.6400,-24.3000)[(0,0)[$d\!-\!m\!-\!1$]{}]{}(28.9200,-24.3000)[(0,0)[$s_1$]{}]{}(37.2800,-24.3000)[(0,0)[$s_2$]{}]{}(45.6400,-24.3000)[(0,0)[$s_3$]{}]{}(51.9100,-24.3000)[(0,0)[$s_p$]{}]{}(55.0100,-23.0000)[(0,0)[$i$]{}]{}(18.2000,-24.3200)[(0,0)[$\cdots$]{}]{}
Let $${\boldsymbol \gamma}=(d-m-\partial h_{s_1}) {\mathbf{e}}_{d-s_1} + \sum_{j=2}^p (\partial h_{s_{j-1}}- \partial h_{s_j} ){\mathbf{e}}_{d-s_j}$$ and $${\boldsymbol \delta}=(\delta_0,\delta_1,\dots,\delta_d) = \sum_{i=0}^m {\mathbf{e}}_i + {\boldsymbol \gamma}.$$ By the construction of ${\boldsymbol \delta}$, $\sum_{k=0}^d \delta_k=d+1$ and $\delta_{d-n} \geq 2$. Let $$\overline {\boldsymbol \delta} = {\boldsymbol \delta} -{\mathbf{e}}_{d-n}.$$ Then $\overline {\boldsymbol \delta}$ satisfies the assumption of Lemma \[3.B\]. Thus there is a constructible $(d-1)$-ball $P$ such that
- $h(P)=\overline {\boldsymbol \delta}$.
- $\partial P$ contains two ideals $I_1$ and $I_2$ that have no common facets such that $I_1 \cong \Delta_d(n)$ and $I_2 \cong \Delta_d(d-n)$.
We define ${\mathbf{h}}'=(h'_0,h_1',\dots,h'_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$ and ${\mathbf{h}}''=(h''_0,h_1'',\dots,h''_d) \in { \ensuremath{\mathbb{Z}}}^{d+1}$ by $${\mathbf{h}}'= (0,\dots,0,h_{n+1}-1,\dots,h_{d-n-1}-1,0,\dots,0)$$ and $${\mathbf{h}}''= (h_0-1,\dots,h_n-1,0,\dots,0,h_{d-n}-1,\dots,h_m-1,h_{m+1},\dots,h_d)
- {\boldsymbol \gamma}.$$ We claim that the following conditions hold.
- ${\mathbf{h}}={\mathbf{h}}'+{\mathbf{h}}''+ \overline {\boldsymbol \delta} + {\mathbf{e}}_{d-n}$;
- ${\mathbf{h}}'$ and ${\mathbf{h}}''$ are non-negative;
- $\sum_{k=0}^d h_k'$ and $\sum_{k=0}^d h''_k$ are even;
- $\partial {\mathbf{h}}'$ and $\partial {\mathbf{h}}''$ are non-negative and $\partial h''_n=0$;
- ${\mathrm{width}}({\mathbf{h}}') \geq d-n$.
We first prove that the above conditions (i)–(v) prove the desired statement.
By Corollary \[kolins2\] and conditions (ii)–(v), there is a constructible $(d-1)$-ball $Q$ with $h(Q)={\mathbf{e}}_0+{\mathbf{h}}'$ such that $\partial Q$ contains an order ideal $J$ which is isomorphic to $\Delta_d(d-n)$. Then the simplicial poset $$R=Q {{}_J \cup_{I_2}} P$$ is a constructible $(d-1)$-ball with $h(R)={\mathbf{h}}'+ \overline {\boldsymbol \delta} +{\mathbf{e}}_{d-n}$ by Lemma \[3.2\]. Also, $\partial R$ contains $\Delta_d(n)$ since $I_1$ and $I_2$ have no common facets.
Since ${\mathbf{h}}''$ is non-negative and $\sum_{k=0}^d h_k''$ is even, ${\mathbf{h}}''$ can be written in the form $${\mathbf{h}}''=({\mathbf{e}}_{i_1}+ {\mathbf{e}}_{i_a}) +({\mathbf{e}}_{i_2}+ {\mathbf{e}}_{i_{a-1}}) + \cdots +({\mathbf{e}}_{i_{\frac a 2}}+ {\mathbf{e}}_{i_{\frac a 2+1}})$$ where $a=\sum_{k=0}^d h_k''$ and $i_1 \leq \cdots \leq i_a$. Since $\partial h''_n=0$, we have $$i_1 \leq \cdots \leq i_{\frac a 2} \leq n \mbox{ and } d-n \leq i_{\frac a 2 +1} \leq \cdots \leq i_a.$$ Thus ${\ensuremath{\mathrm{in}}\hspace{1pt}}({\mathbf{h}}'') \leq n$. Since $\partial R$ contains $\Delta_d(n)$, Lemma \[kolinsrev\] guarantees the existence of a constructible $(d-1)$-ball whose $h$-vector is $h(R)+{\mathbf{h}}''={\mathbf{h}}$.
It remains to prove (i)–(v). Statement (i) is obvious.
[*Proof of (ii).*]{} Since $h_k > 0$ for $k \leq d-n$ (see the first paragraph of Case 3), ${\mathbf{h}}'$ is non-negative. To prove that ${\mathbf{h}}''$ is non-negative, it is enough to prove that $\delta_k \leq h_k$ for all $k$. For $k > m$, $h_k \geq 0=\delta_k$. Also, if $k \leq m$ and $k \not \in \{d-s_1,\dots,d-s_p\}$ then $h_k \geq 1 = \delta_k$ by the choice of $m$. Suppose $k=d-s_j$. Since $h_{s_j} - h_{d-s_j}= \partial h_{s_j} - \partial h_{s_j-1}$, $$h_{d-s_j} = h_{s_j} +(\partial h_{s_j-1}-\partial h_{s_j})
\geq 1+(\partial h_{s_{j-1}} -\partial h_{s_j}) = \delta_{d-s_j},$$ where the inequality follows from the choice of $s_1,\dots,s_p$. (If $j=1$ then we consider that $\partial h_{s_0}=d-m$.)
[*Proof of (iii).*]{} Since $\partial h_n=(h_0+\cdots+h_n) - (h_d + \cdots +h_{d-n})=0$ and since $h_0+h_1+ \cdots +h_d$ is odd, $$h_{n+1} + \cdots + h_{d-n-1}= (h_0+ \cdots +h_d) -(h_0+ \cdots+h_n)-(h_d+ \cdots + h_{d-n})$$ is odd. Then, since $d$ is even, $$(h_0'+ \cdots + h_d') = (h_{n+1} + \cdots + h_{d-n-1}) -(d-2n-1)$$ is even, as desired. Similarly, $\partial h_n=0$ implies that $$(h_0''+ \cdots + h_d'')=
(h_0+\cdots+h_n)+(h_d+ \cdots +h_{d-n})-2n-2$$ is even.
[*Proof of (iv).*]{} Observe $\partial h'_k= \partial h_k - \partial h_n$ for $k \geq n$. Since $\partial h_n=0$, we have $$\begin{aligned}
\partial h_k'=
\left\{
\begin{array}{lll}
0, & \mbox{ for } k=0,1,\dots,n,\\
\partial h_k, & \mbox{ for } k=n+1,\dots,\frac d 2 -1.\\
\end{array}
\right.\end{aligned}$$ Hence $\partial {\mathbf{h}}'$ is non-negative.
For $k=0,1,\dots,d-m-1$, $$\partial h''_k =\partial h_k -(k+1) \geq 0$$ by the condition (7). Also, for $k=n,\dots,\frac d 2 -1$, $$\partial h''_k=\partial h_n=0.$$ Let $k \in \{d-m,d-m+1,\dots,n-1\}$. Then, there is an $i$ such that $s_{i-1} \leq k < s_i$, where we consider that $s_0=d-m$ and $\partial h_{s_0}=d-m$. Then $$\begin{aligned}
\partial h''_k&=&
\partial h_k -(d\!-\!m) + \{ (d\!-\!m\!-\! \partial h_{s_1}) + (\partial h_{s_1}\! - \!\partial h_{s_2}) + \cdots + (\partial h_{s_{i-2}}\!-\!\partial h_{s_{i-1}})\} \\
& = & \partial h_k - \partial h_{s_{i-1}}\\
&\geq& 0,\end{aligned}$$ where the last inequality follows from the choice of $s_1,\dots,s_p$. Thus $\partial {\mathbf{h}}''$ is non-negative.
[*Proof of (v).*]{} The condition (6) says that, for $0 \leq k \leq d-2n$, $$\begin{aligned}
h_{k+n}+ \cdots + h_{k+1} -n
&\geq& \partial h_{k+n}\\
&=& \partial h_n + (h_{n+1}+ \cdots + h_{n+k}) -(h_{d-n-1}+ \cdots + h_{d-n-k})\end{aligned}$$ (we substitute $\ell$ by $k+n$ in the inequality in (6)). Since $\partial h_n=0$, for $n<k \leq d-2n$, we have $$h_{n+1} + \cdots + h_{k} +n\leq h_{d-n-1} + \cdots + h_{d-n-k}.$$ The above inequality says $$\begin{aligned}
h_0'+ \cdots + h_{k}' &=&h_{n+1} + \cdots + h_{k} -(k-n)\\
& \leq& h_{d-n-1} + \cdots + h_{d-n-k} -k
= h_d' + \cdots + h_{d-n-k}'\end{aligned}$$ for all $n<k<d-2n$. Also, if $k \leq n$ or $k \geq d-2n$ then it is clear that $$h_0'+ \cdots + h_{k}'\leq h_d' + \cdots + h_{d-n-k}'.$$ Then Lemma \[dist\] guarantees ${\mathrm{width}}({\mathbf{h}}')\geq d-n$.
**Acknowledgments**: I would like to thank Samuel Kolins for his valuable comments on my earlier work, which inspired me to find the complete necessary and sufficient conditions, and for pointing out a mistake in an earlier draft. This work was supported by KAKENHI 22740018.
[1]{}
A. Björner, Posets, regular CW complexes and Bruhat order, *European J. Combin.* **5** (1984), 7–16.
A. Björner, Topological methods, In: Handbook of combinatorics, R. Graham, M. Grotschel and L. Lovasz, Eds., Elsevier, Amsterdam, 1995, pp. 1819–1872.
A. Björner, P. Frankl and R.P. Stanley, The number of faces of balanced Cohen-Macaulay complexes and a generalized Macaulay theorem, *Combinatorica* **7** (1987), 23–34.
W. Bruns and J. Herzog, Cohen–Macaulay rings, Revised Edition, Cambridge University Press, Cambridge, 1998.
D. Eisenbud, Commutative algebra with a view toward algebraic geometry, Grad. Texts in Math., vol. 150, Springer-Verlag, New York, 1995.
S. Kolins, $f$-vectors of simplicial posets that are balls, *J. Algebraic Combin.*, to appear.
M. Masuda, $h$-vectors of Gorenstein$^\ast$ simplicial posets, *Adv. Math.* **194** (2005), 332–344.
H. Maeda, M. Masuda and T. Panov, Torus graphs and simplicial posets, *Adv. Math.* **212** (2007), 458–483.
E. Miller and V. Reiner, Stanley’s simplicial poset conjecture, after M. Masuda, *Comm. Algebra* **34** (2006), 1049–1053.
S. Murai, Face vectors of simplicial cell decompositions of manifolds, preprint, arXiv:1010.0319.
R.P. Stanley, $f$-vectors and $h$-vectors of simplicial posets, *J. Pure Appl. Algebra* **71** (1991), 319–331.
R.P. Stanley, Combinatorics and commutative algebra, Second edition, Progr. Math., vol. 41, Birkhäuser, Boston, 1996.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report the results of optical spectroscopy of the candidate evolved massive star MN44 revealed via detection of a circular shell with the [*Spitzer Space Telescope*]{}. First spectra taken in 2009 May–June showed the Balmer lines in emission as well as numerous emission lines of iron, which is typical of luminous blue variables (LBVs) near the visual maximum. New observations carried out in 2015 May–September detected significant changes in the spectrum, indicating that the star became hotter. We found that these changes are accompanied by significant brightness variability of MN44. In particular, the $I_{\rm c}$-band brightness decreased by $\approx$ 1.6 mag during the last six years and after reaching its minimum in 2015 June has started to increase. Using archival data, we also found that the $I_{\rm
c}$-band brightness increased by $\approx$3 mag in $\approx$30 yr preceding our observations. MN44 therefore represents the seventeenth known example of the Galactic bona fide LBVs. We detected a nitrogen-rich knot to the northwest of the star, which might represent an interstellar cloudlet interacting with the circumstellar shell. We discuss a possible association between MN44 and the [*INTEGRAL*]{} transient source of hard X-ray emission IGRJ16327$-$4940, implying that MN44 might be either a colliding-wind binary or a high-mass X-ray binary.
date: 'Accepted 2015 September 29. Received 2015 September 29; in original form 2015 August 11'
title: 'Discovery of a new bona fide luminous blue variable in Norma[^1]'
---
\[firstpage\]
line: identification – circumstellar matter – stars: emission-line, Be – stars: evolution – stars: individual: EM\*VRMF55 – stars: massive
Introduction {#sec:intro}
============
{width="14cm"}
Some massive stars evolve through the so-called luminous blue variable (LBV) phase in the course of their life (Conti 1984). During this phase, they experience episodes of enhanced (sometimes eruptive) mass loss, which is manifested in appearance of strong emission lines in their spectra. These episodes, lasting from years to decades and more, are accompanied by major brightness and spectral variability – the defining characteristics of the very rare class of objects called LBV stars (Humphreys & Davidson 1994). Recent searches for new members of this class though detection of their circumstellar nebulae (a common attribute of the LBVs; Clark, Larionov & Arkharov 2005; Kniazev, Gvaramadze & Berdnikov 2015) resulted in discovery of several dozens of compact mid-infrared nebulae of various morphology (Gvaramadze, Kniazev & Fabrika 2010a), whose central stars exhibit rich emission spectra, typical of bona fide LBVs (Gvaramadze et al. 2010a,b, 2012a; Wachter et al. 2010, 2011; Stringfellow et al. 2012a,b; Flagey et al. 2014). These discoveries nearly doubled the known population of Galactic candidate LBVs (cLBVs). To prove their LBV status a significant spectral and photometric variability has to be detected. Disclosure of new bona fide LBVs is of high importance for understanding the nature of their variability and the role of the LBV phase in the stellar evolution.
In this paper, we present recent results of our ongoing spectroscopic and photometric monitoring of the newly identified cLBVs, aimed to detect changes in their spectra and brightness. Our monitoring campaign, started in 2009, has already resulted in discovery of two new Galactic bona fide LBVs – Wray16-137 (Gvaramadze et al. 2014) and WS1 (Kniazev et al. 2015). Here, we report the discovery of the third Galactic bona fide LBV – MN44, whose LBV candidacy was suggested by the presence of a circular shell around it (Gvaramadze et al. 2010a). In Section\[sec:neb\], we present for the first time the images of the shell and its central star at several wavelengths, and review the existing data on the central star. In Section\[sec:obs\] we describe our spectroscopic and photometric observations. The results and implications are discussed in Section\[sec:dis\]. We summarize and conclude in Section\[sec:con\].
------------------------- ------------------ ------------- ------------------ ----------------------- --------------------- ---------------- -- -- -- --
Spectrograph Date Exposure Spectral scale Spatial scale Slit/Seeing Spectral range
(min) (Å pixel$^{-1}$) (arcsec pixel$^{-1}$) (arcsec) (Å)
Cassegrain (SAAO 1.9-m) 2009 May 31 3$\times$15 2.3 1.4 2.0$\times$180/1.0 4200$-$8100
Cassegrain (SAAO 1.9-m) 2009 June 2 3$\times$20 2.3 1.4 2.0$\times$180/1.3 4200$-$8100
RSS (SALT) 2015 May 6 1.5+15 0.97 0.51 1.25$\times$480/1.0 4200$-$7300
RSS (SALT) 2015 June 14 1+20 0.97 0.25 1.25$\times$480/2.5 4350$-$7450
RSS (SALT) 2015 August 2 1+25 0.97 0.25 1.25$\times$480/2.0 4200$-$7300
RSS (SALT) 2015 September 8 1+25 0.97 0.25 1.25$\times$480/1.4 4200$-$7300
------------------------- ------------------ ------------- ------------------ ----------------------- --------------------- ---------------- -- -- -- --
Circular shell MN44 and its central star {#sec:neb}
========================================
The circular shell MN44[^2] is located in the Norma constellation and it is one of many dozens of mid-infrared compact nebulae discovered by Gvaramadze et al. (2010a) in the $24 \,
\mu$m archival data of the [*Spitzer Space Telescope*]{} obtained with the Multiband Imaging Photometer for [*Spitzer*]{} (MIPS; Rieke et al. 2004) within the framework of the 24 and 70 Micron Survey of the Inner Galactic Disk with MIPS (Carey et al. 2009). In the MIPS 24$\mu$m image MN44 appears as a limb-brightened circular shell of radius and thickness of $\approx$15 and 6 arcsec, respectively, centred on a point-like source (see Fig.\[fig:neb\]). The central point source (star) is also visible in all (3.6, 4.5, 5.8 and $8.0\,\mu$m) images obtained with the [*Spitzer*]{} Infrared Array Camera (IRAC; Fazio et al. 2004) within the Galactic Legacy Infrared Mid-Plane Survey Extraordinaire (Benjamin et al. 2003), as well as in all ($J,H,K_{\rm s}$) Two-Micron All Sky Survey (2MASS) images (Skrutskie et al., 2006). The 2MASS coordinates of this star are: RA(J2000)=$16^{\rm h} 32^{\rm m} 39\fs95$, Dec.(J2000)=$-$$49\degr
42\arcmin 13\farcs8$ and $l$=$335\fdg0635$, $b$=$-$$1\fdg1557$. In the following, we will use the name MN44 only for the central star.
Fig.\[fig:neb\] shows that the nebula could also be discerned in the IRAC 8$\mu$m image and that there is the gleam of emission around the star in the H$\alpha$+\[N[ii]{}\] image obtained in the framework of the SuperCOSMOS H-alpha Survey (SHS; Parker et al. 2005). The SHS image also shows a knot of enhanced emission to the northwest of the star. This knot partly delineates the diffuse emission around the star and then curves in the northeast direction. Comparison of the 24 and 8$\mu$m images with the SHS one shows that the shell is brighter at the place of possible contact with the knot, which might be caused by interaction of the shell with a density inhomogeneity in the ambient medium. The nature of the optical emission around MN44 is further discussed in Section\[sec:ha\].
MN44 was identified as an emission-line star by Vega et al. (1980), who searched for H$\alpha$-emission objects using objective-prism spectra. In the SIMBAD data base this star is named EM\*VRMF55. According to Vega et al. (1980), in 1974 the $V$-band magnitude of MN44 was fainter than $15.58$. MN44 is located within the error circle of radius of 4.5 arcmin of the [*INTEGRAL*]{} transient source of hard X-ray emission IGRJ16327$-$4940 (Bird et al. 2010). Masetti et al. (2010) obtained an optical spectrum of MN44 on 2009 August 10 using the 1.9-m telescope of the South African Astronomical Observatory (SAAO) (see their fig.7) and broadly classified it as an OB star due to the strength of its H$\alpha$ emission line, “which is much larger than the typical values seen in blue supergiants". These authors tentatively associated MN44 with IGRJ16327$-$4940 and classified this X-ray source as a high-mass X-ray binary (HMXB) because of its “overall early-type star spectral appearance, which is typical of this class of objects". Possible implications of this association are discussed in Section\[sec:bin\].
Observations {#sec:obs}
============
First spectra of MN44 were obtained with the SAAO 1.9-m telescope on 2009 May 31 and June 2. The observations were performed with the Cassegrain spectrograph using a slit of 3 arcmin $\times$ 2 arcsec and grating with 300 lines mm$^{-1}$. This spectral setup covered a wavelength range of $\approx$4200–8100Å with a reciprocal dispersion of $\approx$2.3Åpixel$^{-1}$ and the spectral resolution full width at half-maximum (FWHM) of $\approx$7Å. The slit was oriented at the parallactic angle because there was no possibility to select a position angle (PA) at this telescope. Three exposures of 900s and 1200 s were taken during the first and second nights, respectively, with a seeing of $\approx$1.0–1.3 arcsec. The data were reduced in a standard way using the [midas]{} and [iraf]{} software. All one-dimensional (1D) spectra obtained during the two nights were then averaged to produce a final spectrum (see the lower curve in Fig.\[fig:spec\]). The spectrum shows the H$\alpha$ and H$\beta$ emission lines and numerous singly ionized iron emission lines, typical of LBVs near the brightness maximum (e.g. Stahl et al. 2001). In the red part of the spectrum we detected the prominent O[i]{} $\lambda$7771–4 Å triplet in absorption (not shown in Fig.\[fig:spec\]), which is a good indicator of luminosity (Merrill 1934). The equivalent width (EW) of this triplet of 1.93$\pm$0.18 Å implies that MN44 is a supergiant star of luminosity class Ia (e.g. Keenan & Hynek 1950; Osmer 1972). These findings along with the presence of the circular shell around the star allowed us to classify MN44 as a cLBV.
To search for possible spectral variability of MN44, we obtained four more spectra with the Southern African Large Telescope (SALT; Buckley, Swart & Meiring 2006; O’Donoghue et al. 2006) on 2015 May 6, June 14, August 2 and September 8 (see Table\[tab:log\] for the log of our spectroscopic observations of MN44). The spectra were taken with the Robert Stobie Spectrograph (RSS; Burgh et al. 2003; Kobulnicky et al. 2003) in the long-slit mode with a $1.25\arcsec$ slit width. The PG900 grating was used on May 6, August 2 and September 8 to cover the spectral range of 4200–7300 Å. We used the same grating on June 14 to cover the spectral range of 4350–7450 Å, which allowed us to fill gaps in the wavelength coverage of the three other spectra. The final reciprocal dispersion and FWHM spectral resolution of the spectra are $0.97$ Å pixel$^{-1}$ and 4.51$\pm$0.13 Å, respectively. The RSS uses a mosaic of three 2048$\times$4096 CCDs and the final spatial scales for observations were 0.51pix$^{-1}$ in the first night and 0.25pix$^{-1}$ in the next three nights. The corresponding seeing was of $\approx$1.0, 2.5, 2.0 and 1.4 arcsec. Two exposures were taken each night. The short exposures were used to avoid saturation of the extremely strong H$\alpha$ emission line. In all observations the slit was oriented at PA=43$\degr$ (measured from north to west) in order to cross the region of bright emission (knot) to the northwest of MN44. A Xe lamp arc spectrum was taken immediately after the science frames. A spectrophotometric standard star was observed during twilight time for relative flux calibration. The primary reduction of the data was done with the SALT science pipeline (Crawford et al. 2010). After that, the bias and gain corrected and mosaiced long-slit data were reduced in the way described in Kniazev et al. (2008). The resulting normalized spectra are presented in Fig.\[fig:spec\] (see the upper four curves). In the figure we show only parts of the spectra longward of $\approx$4800 Å because most of the SALT spectra were taken during bright time and their blue parts are very noisy and do not contain meaningful information.
Date $B$ $V$ $I_{\rm c}$ $V-I_{\rm c}$ $J$ $K_{\rm s}$
--------------------------- ---------------- ---------------- ---------------- --------------- --------------- ---------------
1974 May$^{(1)}$ – $>15.58$ – – – –
1980 May 23$^{(2)}$ – – 13.3$\pm$0.2 – – –
1982 August 8$^{(2)}$ 18.9$\pm$0.3 – – – – –
1998 July 1$^{(3)}$ – – 11.84$\pm$0.03 – 9.01$\pm$0.07 7.42$\pm$0.06
1998 November 3$^{(3)}$ – – 11.74$\pm$0.03 – 8.90$\pm$0.08 7.28$\pm$0.06
1999 June 18$^{(4)}$ – – – – 8.41$\pm$0.02 6.81$\pm$0.03
2009 May 31$^{(5)}$ – 14.41$\pm$0.04 10.20$\pm$0.10 4.21$\pm$0.11 – –
2012 May 5$^{(6)}$ – 14.75$\pm$0.01 10.46$\pm$0.01 4.29$\pm$0.01 – –
2013 January 13$^{(6)}$ 17.86$\pm$0.20 14.99$\pm$0.01 10.74$\pm$0.01 4.25$\pm$0.01 – –
2014 April 25$^{(6)}$ 18.58$\pm$0.08 15.67$\pm$0.03 11.47$\pm$0.01 4.20$\pm$0.03 – –
2015 May 19$^{(7)}$ – 15.70$\pm$0.03 11.77$\pm$0.05 3.93$\pm$0.06 – –
2015 June 14$^{(8)}$ – – 11.81$\pm$0.06 – –
2015 August 2$^{(8)}$ – – 11.70$\pm$0.05 – –
2015 September 8$^{(8)} $ – – 11.65$\pm$0.04 – –
\(1) Vega et al. (1980); (2) USNO B-1; (3) DENIS; (4) 2MASS; (5) 1.9-m telescope; (6) 76-cm telescope; (7) 1-m telescope; (8) SALT.
To detect photometric variability of MN44, we determined its $B,
V$ and $I_{\rm c}$ magnitudes on CCD frames obtained with the SAAO 76-cm telescope in 2012$-$2014. We used an SBIG ST-10XME CCD camera equipped with $BVI_{\rm c}$ filters of the Kron-Cousins system (see e.g. Berdnikov et al. 2012) to build the system of secondary standards in the field of MN44. With these standards, we derived $V$ and $I_{\rm c}$ magnitudes of MN44 using images obtained with the SAAO 1-m telescope on 2015 May 19, and an $I_{\rm c}$ magnitude using acquisition images obtained with the SALT on 2015 June 14, August 2 and September 8. We also calibrated the 1.9-m telescope spectrum and synthesized its $V$ magnitude in the way described in Kniazev et al. (2005). Using the same standards, we recalibrated the $I$ and $B$ magnitudes from the USNOB-1 catalogue (Monet et al. 2003). The results are presented in Table\[tab:phot\]. To this table we also added the lower limit on the $V$ magnitude given in Vega et al. (1980), $J$ and $K_{\rm s}$ magnitudes from 2MASS and two-epoch $I, J$ and $K_{\rm
s}$ photometry from the Deep Near Infrared Survey of the Southern Sky (DENIS; The DENIS Consortium, 2005).
Discussion {#sec:dis}
==========
MN44: a bona fide LBV {#sec:lbv}
---------------------
Fig.\[fig:spec\] shows a montage of five spectra obtained in 2009 and 2015 with the principal lines and most prominent diffuse interstellar bands (DIBs) indicated. All wavelengths are given in air. The lower curve represents the averaged spectrum based on two observations carried out on 2009 May 31 and June 2 with the SAAO 1.9-m telescope. Four other curves (from bottom to top) correspond to spectra taken with the SALT on 2015 May 6, June 14, August 2 and September 8. EWs, FWHMs and heliocentric radial velocities (RVs) of some lines in the spectra (measured applying the [midas]{} programs; see Kniazev et al. 2004 for details) are given in Table\[tab:inten\].
Comparison of the spectra shows that in 2015 the EWs of the H$\alpha$ and H$\beta$ lines increased by $\approx$3 and 6–9 times, respectively, compared to 2009 (see also Table\[tab:inten\]), while the broad Thomson scattering wings of these lines became even more broader. Some Fe[ii]{} lines already present in the 2009’s spectrum, e.g. at $\lambda\lambda$5317, 6384, 6456, became very prominent in 2015. The forbidden lines of \[N[ii]{}\] $\lambda$5755 and \[Fe[ii]{}\] $\lambda\lambda$5334, 5376 and 7155 increased their strength as well, and the \[N[ii]{}\] $\lambda$6584 line became very prominent on the red wing of the H$\alpha$ line. More importantly, the 2015’s spectra show emergence of the He[i]{} lines $\lambda\lambda$5876, 6676 and 7065. These lines were absent in the first epoch spectrum, but show up in 2015 May 6 as central absorptions accompanied by blue and red emission wings. The presence of these lines imply that the star became hotter during the last six years, which is expected in view of the significant decline of the stellar brightness during the same time period (see below). Fig.\[fig:HeI\] plots the evolution of the He[i]{} $\lambda$6676 line profile with time. One can see that the cental absorption became less deep in June 14, then returned to almost the same state as it was in May 6 and then again became less deep, while the emission wings remain almost intact.
The \[Fe[ii]{}\] $\lambda$5376 line in the spectra taken on 2015 May 6 and September 8 has a distinct flat-topped profile. This indicates that the line is formed in a region of constant expansion velocity and therefore its FWHM (respectively, 5.39$\pm$0.26 Å and 5.48$\pm$0.32 Å) could be used to estimate the terminal wind velocity, $v_\infty$ (Stahl et al. 1991). After correction for instrumental width, one finds $v_\infty$=165$\pm$29 and 174$\pm$33 ${{\rm\,km\,s^{-1}}}$. The wind velocity could also be estimated by using the FWHM of the \[N[ii]{}\] $\lambda$5755 line (e.g. Crowther, Hillier & Smith 1995). After correction for instrumental width, the FWHMs of 4.81$\pm$0.27 Å (2015 June 14), 5.04$\pm$0.26 Å (2015 August 2) and 5.05$\pm$0.20 (September 8) correspond to $v_\infty$ of 87$\pm$45, 117$\pm$33 and 118$\pm$21 ${{\rm\,km\,s^{-1}}}$, respectively. Note that although all five estimates agree with each other within the error margins, the lower values of $v_\infty$ based on the \[N[ii]{}\] line might be because this line originates closer to the star where the wind is still accelerated (Stahl et al. 1991). This supposition is supported by the observed \[N[ii]{}\] $\lambda\lambda$5755, 6584 line intensity ratio, which implies that the nitrogen forbidden lines originate in a dense ($\geq$$10^7 \, {\rm cm}^{-3}$) matter (e.g. Tarter 1969), probably close to the star (cf. Stahl et al. 1991). Note also that the low wind velocity of MN44 during its minimum light phase suggests that this star is close to the Eddington limit, probably because it already lost a significant fraction of its initial mass either in a single giant eruption or in a number of less catastrophic mass-loss episodes (Humphreys et al. 2014a,b).
The forbidden lines can also be used to derive the systemic velocity (Stahl et al. 2001). Inspection of Table\[tab:inten\], however, shows significant changes in RVs of the \[Fe[ii]{}\] and \[N[ii]{}\] lines from spectrum to spectrum. One can see that the RVs of most of these lines systematically reduced during the last four months and that the velocity decrease was more prominent during the first month. RVs of the Si[ii]{} $\lambda\lambda$6347, 6371 absorption doublet and the Balmer lines also demonstrate the same trend. The decrease in the RVs of the Balmer lines would be much more stronger if one takes into account the 2009’s spectrum. Although these changes suggest that MN44 might be a close eccentric binary system (cf. Section\[sec:bin\]), the limited available data did not allow us to prove this for sure.
.
The detected spectral variability of MN44 strongly suggests that this star is a bona fide LBV, currently evolving towards the hot state, i.e. towards the brightness minimum (e.g. Stahl et al. 2001). This implication is reinforced by photometric measurements given in Table\[tab:phot\], which show that MN44 has experienced strong brightness decline during the period of our spectroscopic observations. This is illustrated by Fig.\[fig:phot\], which plots $V$- and $I_{\rm c}$-band light curves of MN44 since its identification as a cLBV in 2009 with arrows indicating times when the spectra were obtained. For each data point (square) we give 1$\sigma$ error bars, which in most cases are less than the size of the squares themselves. Fig. \[fig:phot\] shows that MN44 became fainter in the $V$ and $I_{\rm c}$ bands, respectively, by $\approx$1.3 and 1.6 mag during the last 6 years. One can see also that the $I_{\rm
c}$-band brightness of MN44 has reached minimum in 2015 June and then starts to increase. On a longer time-scale, the $I_{\rm
c}$-band variability is even more striking. From Table\[tab:phot\] it follows that MN44 has brightened by $\approx$3 mag (!) during $\approx$30 yr preceding our monitoring campaign. The large magnitude of the $I_{\rm c}$-band excursion suggests that MN44 might belong to a rare group of LBVs exhibiting giant eruptions (as in $\eta$Car and PCyg) during which their bolometric luminosity changes (Humphreys & Davidson 1994). Table\[tab:phot\] also shows that MN44 became brighter in the $J$ and $K_{\rm s}$ bands by $\approx$0.6 mag during $\approx$7.5 months in 1998–1999.
By using the weighted least square approximation, we found that the $V$$-$$I_{\rm c}$ colour of MN44 has decreased by $\approx$0.3 mag during the last six years. On the other hand, as one can infer from Table\[tab:phot\], this decrease has occurred mostly during the last year, while in 2009–2014 the colour did not change much despite of significant ($>$1 mag) brightness decrease of the star. Such behaviour is not typical of LBV excursions at constant bolometric luminosity during which stars become bluer with the brightness decrease, but it was observed for LBVs showing changes in the bolometric luminosity (e.g. Clark et al. 2009; Kniazev et al. 2015). Whether this is the case for MN44 as well could be proved with a detailed spectral modelling, which is, however, beyond the scope of this paper.
Taken together, our observations of MN44 unambiguously show that this star is a new (seventeenth) member of the family of Galactic bona fide LBVs (see Kniazev et al. 2015 for a recent census of these objects).
H$\alpha$+\[N[ii]{}\] emission to the northwest of MN44 {#sec:ha}
-------------------------------------------------------
The SHS H$\alpha$+\[N[ii]{}\] image presented in Fig.\[fig:neb\] shows that there is a diffuse emission immediately around MN44 and further northwest from it. There is also a bright knot of emission apparently in contact with the circumstellar shell. To clarify the nature of the emission around MN44, we obtained long-slit SALT spectra with the slit oriented at PA=43$\degr$ in order to cross the knot. In two-dimensional (2D) spectra the knot appears as emission lines of H$\beta$, H$\alpha$, \[\] $\lambda\lambda$6548, 6584 and \[\] $\lambda\lambda$6717, 6731. In the following, we use the spectrum taken on 2015 May 6 because of its better quality. A part of this spectrum is presented in Fig.\[fig:2D\]. One can see two distinct velocity components in the emission lines. The lines produced in the inner region ($\leq$14 arcsec from MN44, i.e. within the confines of the 24$\mu$m shell) are shifted to the red end of the spectrum compared to the lines originated at larger angular distances from the star (outer region).
1D spectra of the inner and outer regions were extracted by summing up, without any weighting, all rows in angular distance intervals from $\approx$9 to 14 arcsec and from 14 to 40 arcsec, respectively (see Fig.\[fig:slit\]). The emission lines detected in these spectra were measured using the programs described in Kniazev et al. (2004). Table\[tab:int\] lists the observed intensities of the lines normalized to H$\beta$, F($\lambda$)/F(H$\beta$), the reddening-corrected line intensity ratios, I($\lambda$)/I(H$\beta$), the logarithmic extinction coefficients, $C$(H$\beta$), the colour excesses, $E(B-V)$, and the heliocentric radial velocities, $v_{\rm hel}$, of the two regions. The 1D spectrum of the inner region is shown in Fig.\[fig:Neb\].
The \[S[ii]{}\] $\lambda\lambda$6716, 6731 line intensity ratio can be used to derive the electron number density, $n_{\rm e}$, in the line-emitting gas (Krueger, Aller & Czyzak 1970; Saraph & Seaton 1970). In the inner region, the measured ratio of $1.89^{+0.82} _{-0.53}$ agrees at 1 sigma level with the theoretical upper limit on this ratio of $\approx 1.4$ (Krueger et al. 1970). Thus, one can only put an upper limit on $n_{\rm e}$ of $\leq 10 \, {\rm cm}^{-3}$. In the outer region, the line ratio of $1.24^{+1.19} _{-0.59}$ corresponds to $n_{\rm e}=190^{+2770}
_{-180}$.
The above estimates of $C$(H$\beta$) and $n_{\rm e}$(\[S[ii]{}\]) were derived for the CaseB recombination and under the assumption that the electron temperature, $T_{\rm e}$, is equal to $10^4$ K.
The \[N[ii]{}\] and \[S[ii]{}\] line intensities can be used to estimate the nitrogen to sulphur abundance ratio, $N({\rm N}^+
)/N({\rm S}^+ )$, which is almost independent of $n_{\rm e}$ and $T_{\rm e}$, provided that $n_{\rm e}\leq 1000 \, {\rm cm}^{-3}$ and $T_{\rm e}\leq 10^4$ K. In this case, it is given by (see Benvenuti, D’Odorico & Peimbert 1973 and references therein): $${N({\rm N}^+ )\over N({\rm S}^+ )} =3.61 {I(6584) \over
I(6716+6731)} \, . \label{eq:abud}$$ Using equation(\[eq:abud\]) and Table\[tab:int\], one finds $N({\rm N}^+ )/N({\rm S}^+ )$$\approx$15.55$^{+4.31}
_{-3.16}$ and 3.89$^{+3.20} _{-1.66}$ for the inner and outer regions, respectively. The former ratio is $\approx$3.0$^{+0.9}
_{-0.6}$ times larger than the solar value of 5.12 (Asplund et al. 2009), while the latter one is comparable to it. The high $N({\rm
N}^+ )/N({\rm S}^+ )$ ratio in the inner region could be owing to pollution by CNO-processed ejecta from MN44. Since the sulfur abundance is not expected to change during stellar evolution, one can argue that the inner region is enriched in N by a factor of $\approx$2–4. This provides further support to the physical association between the knot and the circumstellar shell around MN44. The solar abundance of N in the outer region suggests that it is an [H[ii]{} ]{}region.
Since the $E(B-V)$ values for the inner and outer regions are consistent with each other within the error margins, it is likely that both are located at the same distance and ionized by MN44 during its hot state. If so, then the difference in the heliocentric radial velocities of the two regions could be understood if the inner part of the knot was accelerated because of interaction with the circumstellar shell of MN44. This, in turn, would imply that the radial velocity of the shell (and MN44) relative to the [H[ii]{} ]{}region is 28$\pm$$13 \, {{\rm\,km\,s^{-1}}}$.
Reddening to and luminosity of MN44 {#sec:lum}
-----------------------------------
The colour excess towards the knot could be used to derive the $V$-band extinction, $A_V$, in the direction of MN44, and thereby to constrain the absolute visual magnitude, $M_V$, of this star. For the standard ratio of total to selective extinction of $R_V$=3.1, one finds $A_{\rm V}=4.43$ mag, while for $M_V$ one has the following equation: $$M_V=V-{\rm DM}-A_V \, , \label{eq:MV}$$ where DM is the distance modulus. Using equation(\[eq:MV\]) and the $V$-band magnitude of MN44 of 15.69 (measured on 2015 May 19), one has $M_V$=11.26$-$DM mag. To further constrain $M_V$, we note that the sightline towards MN44 first enters the Carina-Sagittarius arm (located at a distance of $\sim$2.1 kpc from the Sun), then (at $\sim$3.3 kpc) intersects the Crux-Scutum arm, then twice crosses the Norma arm (at $\sim$5.5 and 10.8 kpc), and then again crosses the Crux-Scutum and Carina-Sagittarius arms (at $\sim$15 and 19 kpc, respectively) (e.g. Cordes & Lazio 2002). The distances at which the sightline intersects the spiral arms correspond to DM values of $\approx$11.61, 12.59, 13.70, 15.17, 15.88 and 16.39 mag, respectively. One can see that $M_V$ of MN44 would be at the upper end of values typical of B supergiants even if this star is located in the more distant part of the Carina-Sagittarius arm (DM=16.39 mag), i.e. $M_V$$\approx$$-$5.1 mag. To derive the bolometric luminosity of MN44, we note that the 2015’s spectra of this star are very similar to those of the LBV WS1 in the cool state, for which we derived the effective temperature, $T_{\rm eff}$, of $\approx$12000K (Kniazev et al. 2015). At this temperature, the bolometric correction of MN44 is equal to $\approx$$-$0.8 mag. Correspondingly, the bolometric luminosity of MN44 (for DM=16.39 mag) would be only of $\log(L_{\rm bol}/{\rm\,L_\odot})$$\approx$4.3, which is much smaller than the minimum value of 5.1 ever derived for known bona fide or cLBVs (Humphreys et al. 2014a).
The reddening towards MN44 could also be estimated by matching the dereddened spectral slope of this star with those of stars of similar $T_{\rm eff}$. Using the Stellar Spectral Flux Library by Pickles (1998), we found $E(B-V)$=3.30$\pm$0.10 mag (this estimate only slightly depends on the assumed $T_{\rm eff}$; see Gvaramadze et al. 2012a) and $A_V$=10.23 mag. Correspondingly, one has $M_V$=5.46$-$DM mag. One can see that MN44 would has a reasonable luminosity, i.e. $\log(L_{\rm bol}/{\rm\,L_\odot})$$\geq$5.1, if the star is located at $d$$\geq$3.3 kpc, i.e. in the near segment of the Crux-Scutum arm or further out (DM$\geq$12.59 mag; see Table\[tab:DM\]). At these distances, MN44 would be located on the cool side of the SDor instability strip (Wolf 1989; Groh et al. 2009) on the Hertzsprung-Russell diagram.
------- ------- --------- ------------------------------------ -------------------- ------- ------ -------------------------
$d$ DM $M_V$ $\log(L_{\rm bol}/{\rm\,L_\odot})$ $M_{\rm init}$ $t$ $h$ $h/t$
(kpc) (mag) (mag) (${\rm\,M_\odot}$) (Myr) (pc) (${{\rm\,km\,s^{-1}}}$)
2.1 11.61 $-$6.2 4.7 15 12 40 3
3.3 12.59 $-$7.1 5.1 20 9 70 8
5.5 13.70 $-$8.2 5.5 35 6 110 18
10.8 15.17 $-$9.7 6.1 60 4 220 55
15 15.88 $-$10.4 6.4 120 3 300 100
19 16.39 $-$10.9 6.6 150 2.5 380 150
------- ------- --------- ------------------------------------ -------------------- ------- ------ -------------------------
The $A_V$ value derived from the spectral slope of MN44 exceeds by $\approx$5.8 mag the extinction derived towards the knot. A trivial explanation of this discrepancy is that MN44 and the knot are located at different distances and are simply projected by chance along the same line-of-sight. This explanation, however, is unlikely because there are strong indications that the knot is in a contact with the circumstellar shell of MN44. Instead, one can envisage two other possible explanations.
Our preferred explanation of the large discrepancy in the $A_V$ values is that MN44 is shrouded by dense circumstellar material and that this material significantly contributes to the extinction of the star. For example, the dusty material of the Homunculus nebula is responsible for $\approx$4 mag of the $V$-band extinction towards $\eta$Car (Humphreys, Davidson & Smith 1999; cf. van Genderen, de Groot & The 1994). It is, therefore, possible that the dusty ejecta from recent outbursts of MN44 causes the excess extinction towards this star as well. If this explanation is correct, then one can expect that the changes in the brightness of MN44 might be caused not only by the intrinsic variability of the star itself, but also by the changeable circumstellar extinction (cf. van Genderen et al. 1994). Or else the extra extinction might be due to a dense circumbinary disk if MN44 is a binary system (see Section\[sec:bin\]).
An alternative explanation of the discrepancy is that the intrinsic Balmer decrement in the spectrum of the knot is much shallower than what follows from the standard CaseB recombination model, i.e. the intrinsic value of the $I$(H$\alpha$)/$I$(H$\beta$) ratio is significantly smaller than $\approx$3. For this explanation to work, the H emission should arise in a very dense medium with $n_{\rm e}\ga10^{13} \, {\rm
cm}^{-3}$ (Drake & Ulrich 1980). This requirement does not contradict to the low electron density indicated by the \[S[ii]{}\] $\lambda\lambda$6716, 6731 line intensity ratio because these forbidden lines could originate in a low-density halo around the dense core of the knot. If the H emission indeed originates in the very dense medium, then the $I$(H$\alpha$)/$I$(H$\beta$) ratio could be as small as $\approx$1 (see fig.7 in Drake & Ulrich 1980), which for the line intensities observed in the spectrum of the knot would result in a factor of $\approx$2 larger values of $E(B-V)$ and $A_{\rm V}$, thereby bringing them closer to the values derived from the dereddening the spectral slope of MN44. To check this explanation, one has to measure more Balmer line ratios in the spectrum of the knot, which requires much more deep spectroscopic observations.
An additional constraint on the distance to MN44 comes from the empirical amplitude-luminosity relation of LBVs, which implies that the amplitude of photometric variability increases with $L_{\rm bol}$ (Wolf 1989). The strong changes in the brightness of MN44 suggest that the luminosity of this star should be well above the lower end of the range of luminosities derived for (c)LBVs. This means that the star should be located at least in the Norma arm or in the next arms out, i.e. at $d$$\geq$5.5 kpc. At these distances, MN44 would lie at $\geq$110 pc below the Galactic plane. This separation from the plane is comparable to or larger than the exponential scale height of runaway O stars of $\approx$90 pc (Stone 1979) and is at least a factor of two larger than that of “normal" OB stars (Stone 1979; Reed 2000). Thus, MN44 might be a runaway star; we further discuss this possibility in Section\[sec:run\].
On the other hand, the larger the distance to MN44, the larger its luminosity and initial mass, $M_{\rm init}$, the shorter the lifetime of this star, $t$, and the larger its separation from the Galactic plane, $h=d\sin b$ (see Table\[tab:DM\]). This means that the more massive the star the higher the peculiar velocity perpendicular to the Galactic plane, $\sim$$h/t$, it should have (see Table\[tab:DM\]). For example, if MN44 is located at $d$=15 kpc ($M_{\rm init}\approx$$120 \, {\rm\,M_\odot}$) then its short lifetime of $t\sim$3 Myr would imply that to reach the height of $h\approx$300 pc, the star should be ejected from the parent cluster with a velocity of $\approx$$100 \, {{\rm\,km\,s^{-1}}}$ (provided that the ejection event has occurred soon after the star was born). Although ejection of such massive and high-velocity stars is not impossible, the probability of these events is very low (Gvaramadze & Gualandris 2011). It would be even much lower if MN44 is a massive binary system (see next section).
Finally, the distance to MN44 could also be constrained through the kinematic distance to the [H[ii]{} ]{}region northwest of the star (in Sect.\[sec:ha\], we argue that they are likely located at the same distance). Using the heliocentric radial velocity of the [H[ii]{} ]{}region of $-$39$\pm$$12 \, {{\rm\,km\,s^{-1}}}$, adopting the radial velocity dispersion of [H[ii]{} ]{}regions within an arm of $\sim$$10 \,
{{\rm\,km\,s^{-1}}}$ (Georgelin & Georgelin 1976), and assuming the distance to the Galactic Centre of $R_0$=8.0 kpc and the circular rotation speed of the Galaxy of $\Theta _0 =240 \, {{\rm\,km\,s^{-1}}}$ (Reid et al. 2009), and the solar peculiar motion $(U_{\odot},V_{\odot},W_{\odot})=(11.1,12.2,7.3) \, {{\rm\,km\,s^{-1}}}$ (Schönrich, Binney & Dehnen 2010), one finds a kinematic distance of $\approx$$2.3^{+1.1} _{-1.4}$ or $12.1^{+1.4} _{-1.1}$ kpc. For the distance range implied by the error margins of the former distance estimate, the luminosity of MN44 would be unrealistically low (see above). The error margins of the latter distance estimate allow the possibility that MN44 is located in the outer segment of the Norma arm. In this case, the bolometric luminosity of MN44 of $\log(L_{\rm bol}/{\rm\,L_\odot})$=6.1 would exceed the Humphreys-Davidson luminosity limit (Humphreys & Davidson 1979), i.e. the star would be located in a region of the Hertzsprung-Russell diagram where the massive stars are expected to experience unsteady (eruptive) high mass-loss episodes. We caution however, that the possible association between the optically visible knot and the circumstellar shell is rather indicative of moderate ($\approx$4 mag) interstellar extinction towards MN44, which in turn suggests that this star is located at a shorter distance, i.e. in the inner segment of the Norma arm. In fact, the above kinematic distance estimates would be meaningless if the [H[ii]{} ]{}region has a peculiar radial velocity of several tens of ${{\rm\,km\,s^{-1}}}$, e.g. because of interaction with the stellar wind/ejecta during the previous mass-loss episodes.
MN44: a massive X-ray binary? {#sec:bin}
-----------------------------
In Section\[sec:neb\], we noted that Masetti et al. (2010) identified MN44 as a counterpart of the [*INTEGRAL*]{} source of hard X-ray emission IGRJ16327$-$4940 and tentatively classified this source as a HMXB. Let us discuss whether the LBV classification of MN44 is consistent with the possibility that this object is a source of hard X-ray emission.
First, we list some relevant data on IGRJ16327$-$4940 given in the Fourth soft gamma-ray source catalogue obtained with the IBIS gamma-ray imager on board the [*INTEGRAL*]{} satellite (Bird et al. 2010). In this catalogue, IGRJ16327$-$4940 is indicated as a transient (strongly variable) source with the 20–40 and 40–100 keV band fluxes time-averaged over the total exposure time of 3319.3 ks (or $\approx$38.4 d) of $F$(20–40)$<$0.2 mCrab and $F$(40–100)=0.4$\pm$0.1 mCrab, respectively. The catalogue also gives the peak 20–40 keV band flux of $F_{\rm
peak}$(20–40)=2.1$\pm$0.7 mCrab, i.e. the mean flux measured during the largest observed outburst, in which the source was detected at a maximum level of significance of 5.6-$\sigma$.
The above fluxes translate to the X-ray-to-bolometric luminosity ratios of $\log[L_{\rm X}(20-40)/L_{\rm bol}]<-5.30$, $\log[L_{\rm
X}(40-100)/L_{\rm bol}]=-4.91$ and $\log[L_{\rm X}(20-40)/L_{\rm
bol}]=-4.28$, which are independent of the actual distance to MN44. These ratios should be compared with $\log[L_{\rm X}/L_{\rm
bol}]$ values observed for other (c)LBVs.
Nazé, Rauw & Hutsemékers (2012) conducted a search for X-ray emission from the Galactic bona fide and cLBVs using available at that time data from the [*XMM-Newton*]{} and [*Chandra*]{} X-ray observatories, which covered 31 of 67 known objects of this type. Besides two already know X-ray sources $\eta$Car and NAMEVICYG12, they reported detection of X-ray emission from two additional (c)LBVs, Cl\*Westerlund1W243 and GAL026.47+00.02, as well as possible detections of two more (c)LBVs near the Galactic Centre, GCIRS34W and GCIRS33SE. For four (c)LBVs with confidently detected X-ray emission the following $\log[L_{\rm X}/L_{\rm bol}]$ values were derived: $\sim$$-$5 ($\eta$Car), $-$6.1 (NAMEVICYG12), $-$5.95 (GAL026.47+00.02) and $-$7.0...7.3 (Cl\*Westerlund1W243). These values imply that the $L_{\rm X}/L_{\rm bol}$ ratios of the first three stars are one or two orders of magnitude larger than those typical of single O stars (i.e. $\approx$$10^{-7}$; e.g. Pallavicini et al. 1981; Sana et al. 2006), whose X-ray emission is believed to originate because of shocks inside stellar winds (e.g. Owocki, Castor & Rybicki 1988; Feldmeier, Puls & Pauldrach 1997). Since the wind velocities of the majority of (c)LBVs ($\sim$$100 \, {{\rm\,km\,s^{-1}}}$) are about an order of magnitude lower than those of O stars, the bright X-ray emission of the three stars cannot be caused by intrinsic wind shocks, but rather should originate in colliding winds (Usov 1992; Stevens, Blondin & Pollock 1992), i.e. these stars should be members of close massive binaries (Nazé et al. 2012). This assertion conforms with the observed hardness and variability of the X-ray emission from two of these stars – $\eta$Car and NAMEVICYG12. Cl\*Westerlund1W243 is also likely to be a massive binary system because its slow wind cannot produce wind shocks strong enough to account for the observed X-ray emission. If so, then the X-ray emission from this object should originate in a companion O star (Nazé et al. 2012).
For the remaining 23 (c)LBVs observed with [*Chandra*]{} and [*XMM-Newton*]{}, only upper limits on their X-ray fluxes were obtained. In most cases, these limits imply that $\log[L_{\rm
X}/L_{\rm bol}]<-5...6$, which leaves the possibility that the corresponding stars might be X-ray bright close massive binaries. For five (c)LBVs very strong constraints on the X-ray emission were derived, $\log[L_{\rm X}/L_{\rm bol}]<-8.2...9.4$, indicating that some (c)LBVs are intrinsically weak X-ray sources. Possible reasons for this are that these (c)LBVs are single stars and/or their X-ray emission is significantly attenuated by dense circumstellar material (Nazé et al. 2012).
The high values of $\log[L_{\rm X}/L_{\rm bol}]$ of MN44 along with the hardness and variability of its X-ray emission point to the possibility that this star is a colliding-wind binary. In this case, the observed X-ray emission should originate at the shock interface between colliding winds, while its variability could be due to changes in the strength of the shocks caused by the eccentricity of the binary orbit and/or the SDor-like variability of wind parameters (velocity, mass-loss rate) of the LBV star. Interestingly, at the distance of 5.5 kpc, the 20–100 keV band X-ray luminosity of MN44 of 1.4$\times 10^{34} \, {\rm
erg} \, {\rm s}^{-1}$ would be comparable to that of $\eta$Car of 0.7$\times$$10^{34} \, {\rm erg} \, {\rm s}^{-1}$ (Leyder, Walter & Rauw 2008). In this energy range, the X-ray emission originates mostly from non-thermal processes. In the case of $\eta$Car, it is believed that its 20–100 keV band X-ray emission is dominated by inverse Compton scattering of stellar UV photons by electrons accelerated in the wind collision zone (Leyder et al. 2008). The same mechanism could be at work in MN44 as well.
If subsequent (optical or infrared) spectroscopic monitoring will prove that MN44 forms a binary system with another massive star, then dedicated X-ray observations of this system would be desirable to search for orbital modulation of its X-ray emission (e.g. Nazé et al. 2007; Hamaguchi et al. 2014). The X-ray light curve, however, could not necessarily be periodic because the orbital modulation of the X-ray emission might be affected by changes in the wind of the LBV component of the binary (unless the SDor-like activity of MN44 is triggered by the effect of the companion star). Also, an analysis of the [*INTEGRAL*]{} data would be of interest to check if the changes in the hard X-ray flux of MN44 are related to any particular orbital phase.
Alternatively, as suggested by Masetti et al. (2010), MN44 could be a HMXB, i.e. a binary system composed of a massive star and a compact object, either a neutron star or black hole (van den Heuvel & Heise 1972; Tutukov & Yungelson 1973). In these systems, the hard X-ray emission is generated because of accretion of the stellar wind material onto the compact object, while its variability could be due to changeable accretion rate. In the case of MN44 the accretion rate could vary because of changes in the wind velocity and mass-loss rate of the LBV companion star and/or because of the eccentricity of the binary orbit. If the HMXB nature of MN44 would be confirmed (e.g. by detection of pulsed X-ray emission), then this system would represent a first known example of a HMXB with a bona fide LBV donor star (cf. Mason et al. 2012; Clark et al. 2013).
MN44: a runaway star? {#sec:run}
---------------------
MN44, like the majority of other bona fide and cLBVs, is located in the field and, therefore, is most probably a runaway star (Gvaramadze et al. 2012a,b). The runaway status of MN44 is also suggested by separation of this star from the Galactic plane (see Sect.\[sec:lum\]). To check this possibility, we searched for proper motion measurements for MN44 using the VizieR catalogue access tool[^3]. We found four catalogues that provide proper motions for MN44, namely, UCAC2 (Zacharias et al. 2004), PPMXL (Röser, Demleitner & Schilbach 2010), SPM4.0 (Girard et al. 2011) and UCAC4 (Zacharias et al. 2013). Since the measurement uncertainties in the first two catalogues are larger than the measurements themselves, we will use only the later two catalogues, which give $\mu _\alpha \cos \delta =-7.86\pm2.25, \, \mu _\delta
=-2.59\pm2.41$ and $\mu _\alpha \cos \delta =-7.5\pm3.0, \, \mu
_\delta =-5.6\pm3.0$, respectively.
Using the same Galactic constants and the solar peculiar motion as in Section\[sec:lum\], we calculated the peculiar transverse velocity of MN44, $v_{\rm tr}=(v_{\rm l} ^2+v_{\rm b} ^2)^{1/2}$, where $v_{\rm l}$ and $v_{\rm b}$ are the velocity components in the Galactic coordinates. For the sake of illustration, we adopted two plausible distances: 5.5 and 10.8 kpc. For the error calculation, only the errors of the proper motion measurements were considered. The results are summarized in Table\[tab:prop\].
Taken at face value, the obtained peculiar velocities imply that MN44 is a runaway star moving towards the Galactic plane, i.e. in the “incorrect" direction (recall that MN44 is located below the Galactic plane). However, the large error margins of the velocity components allow the possibility that MN44 is moving away from the Galactic plane at 2$\sigma$ and 1$\sigma$ significance level for the SPM4.0 and UCAC4 proper motion measurements, respectively. Note also that the derived $v_{\rm tr}$ differs from zero only at $\approx$2$\sigma$ or lower significance. Thus, to unambiguously prove the runaway status of MN44, one needs more precise proper motion measurements, which could be achieved with the space astrometry mission Gaia.
---------------- ------ ------------------------- ------------------------- -------------------------
Sources for $d$ $v_l$ $v_b$ $v_{\rm tr}$
proper motions kpc (${{\rm\,km\,s^{-1}}}$) (${{\rm\,km\,s^{-1}}}$) (${{\rm\,km\,s^{-1}}}$)
SPM4.0 5.5 $-67$$\pm$61 99$\pm$61 120$\pm$61
UCAC4 5.5 $-118$$\pm$78 38$\pm$78 124$\pm$78
SPM4.0 10.8 36$\pm$120 198$\pm$120 201$\pm$120
UCAC4 10.8 $-64$$\pm$154 79$\pm$154 102$\pm$154
---------------- ------ ------------------------- ------------------------- -------------------------
: Peculiar transverse velocity and its components (in Galactic coordinates) for two adopted proper motion measurements and distances.[]{data-label="tab:prop"}
Summary and conclusion {#sec:con}
======================
In this paper, we reported the discovery of a new (seventeenth) Galactic bona fide LBV. The discovery was made through the detection of a circular mid-infrared shell (of diameter of $\approx$30 arcsec) with the [*Spitzer Space Telescope*]{} and follow-up spectroscopic and photometric observations of its central star – MN44. The first epoch (2009) optical spectroscopy of MN44 revealed a rich emission spectrum, typical of LBVs near the visual maximum. The spectra taken six yr later showed the emergence of He[i]{} emission lines, which indicates that the star became hotter ($T_{\rm eff}$$\approx$12000 K). Besides, the EWs of the Balmer lines increased by $\approx$3–9 times, while the forbidden lines of \[N[ii]{}\] and \[Fe[ii]{}\], and some Fe[ii]{} lines became much more prominent. Using the flat-topped \[Fe[ii]{}\] $\lambda$5376 line, we derived the terminal wind velocity of MN44 of $\approx$$170 \, {{\rm\,km\,s^{-1}}}$ during the current hot state of this star, which could be considered as an indication that MN44 is close to its Eddington limit.
Our photometric observations showed that the spectral variability of MN44 is accompanied by the brightness decrease ($\sim$2 mag during the last six yr), which is typical of LBVs recovering from SDor-type outbursts. From archival photometric data, we also derived that MN44 has brightened in the $I_{\rm c}$ band by $\approx$3 mag in $\approx$30 yr preceding our observations. The large amplitude of photometric variability and the separation of MN44 from the Galactic plane were used to suggest that this star is located in the Norma arm, i.e. at $\sim$5.5 or 10.8 kpc. At these distances, MN44 lies, respectively, $\approx$110 and 220 pc below the Galactic plane, which implies that MN44 is a runaway star.
Comparison of the [*Spitzer*]{} and SHS images revealed a knot of H$\alpha$+\[N[ii]{}\] emission to the northwest of MN44. The knot partially delineates the mid-infrared circumstellar shell, which is brighter in the place of possible contact with the knot. We suggested that the shell is interacting with a density inhomogeneity (knot) in the ambient medium. This suggestion was reinforced by the detection of enhanced nitrogen abundance in the knot, which might be caused by pollution from CNO-processed ejecta from MN44. Using the Balmer decrement in the spectrum of the knot, we derived the $V$-band extinction towards MN44 of $4.43$ mag. This value turns out to be smaller by $\approx$6 mag than the extinction derived from dereddening the spectral slope of MN44. Two possible explanations of this discrepancy were discussed. The first one (our preference) is that MN44 is shrouded by dusty circumstellar material, which is responsible for most of the reddening of the star. Alternatively, the intrinsic Balmer decrement in the spectrum of the knot might be much shallower than what follows from the CaseB recombination model. This could happen if the number density of the knot is $\ga$$10^{13} \, {\rm
cm}^{-3}$.
We discussed possible association of MN44 with the [*INTEGRAL*]{} transient source of hard X-ray emission IGRJ16327$-$4940. If real, this association would imply that MN44 is either a colliding-wind binary (similar to $\eta$Car and several other Galactic (c)LBVs) or a first known example of a HMXB with a bona fide LBV donor star.
To conclude, further spectroscopic and photometric monitoring of MN44 with better resolution and cadence would be of great importance for understanding the driving mechanisms of variability of this star, and, potentially, for revealing its duplicity. If the binary status of MN44 will be proved, then dedicated X-ray observations of this system would be highly desirable to search for orbital modulation of its X-ray flux (if the system is a wind-colliding binary) or X-ray pulsations from the compact object (if MN44 is a HMXB).
Acknowledgements
================
We are grateful to R.M.Humphreys (the referee) for useful comments and suggestions on the manuscript. Some observations reported in this paper were obtained with the Southern African Large Telescope (SALT). VVG and LNB acknowledge the Russian Science Foundation grants 14-12-01096 and 14-22-00041, respectively. AYK acknowledges support from the National Research Foundation (NRF) of South Africa. We thank Dave Kilkenny for getting for us $V$- and $I_{\rm c}$-band images of MN44 with the SAAO 1-m telescope. This work is based in part on archival data obtained with the [*Spitzer*]{} Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA, and has made use of the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration, the SIMBAD database and the VizieR catalogue access tool, both operated at CDS, Strasbourg, France.
Asplund M., Grevesse N., Sauval A. J., Scott P., 2009, ARA&A, 47, 481 Benjamin R. A. et al., 2003, PASP, 115, 953 Benvenuti P., D’Odorico S., Peimbert M., 1973, A&A, 28, 447 Berdnikov L. et al., 2012, Astronomy Reports, 56, 290 Bird A. J. et al., 2010, ApJS, 186, 1 Buckley D. A. H., Swart G. P., Meiring J. G., 2006, in Stepp L. M., ed., Proc. SPIEConf. Ser.Vol. 6267, Ground-based and AirborneTelescopes. SPIE, Bellingham, p. 62670Z Burgh E. B., Nordsieck K. H., Kobulnicky H. A., Williams T. B., O’Donoghue D., Smith M. P., Percival J. W., 2003, in Iye M., Moorwood A. F. M., eds, Proc. SPIE Conf. Ser. Vol. 4841, Instrument Design and Performance for Optical/Infrared Ground-based Telescopes. SPIE, Bellingham, p. 1463 Carey S. J. et al., 2009, PASP, 121, 76 Clark J. S., Larionov V. M., Arkharov A., 2005, A&A, 435, 239 Clark J. S., Crowther P. A., Larionov V. M., Steele I. A., Ritchie B. W., Arkharov A. A., 2009, A&A, 507, 1555 Clark J. S., Bartlett E. S., Coe M. J., Dorda R., Haberl F., Lamb J. B., Negueruela I., Udalski A., 2013, A&A, 560, A10 Conti P. S., 1984, in Maeder A., Renzini A., eds, Observational Tests of the Stellar Evolution Theory. Reidel, Dordrecht, p. 233 Cordes J. M., Lazio T. J. W., 2002, astro-ph/0207156 Crawford S. M. et al., 2010, in Silva D. R., Peck A. B., Soifer B. T., Proc. SPIE Conf. Ser. Vol. 7737, Observatory Operations: Strategies, Processes, and Systems III. SPIE, Bellingham, p. 773725 Crowther P. A., Hillier D. J., Smith L. J., 1995, A&A, 293, 172 Davidson K. et al., 1999, AJ, 118, 1777 Drake S. A., Ulrich R. K., 1980, ApJS, 42, 351 Fazio G. G. et al., 2004, ApJS, 154, 10 Feldmeier A., Puls J., Pauldrach A. W. A., 1997, A&A, 322, 878 Flagey N., Noriega-Crespo A., Petric A. O., Geballe T. R., 2014, AJ, 148, 34 Georgelin Y. M., Georgelin Y. P., 1976, A&A, 49, 57 Girard T. M. et al., 2011, AJ, 142, 15 Groh J. H. et al., 2009, ApJ, 705, L25 Gvaramadze V. V., Gualandris A., 2011, MNRAS, 410, 304 Gvaramadze V. V., Kniazev A. Y., Fabrika S., 2010a, MNRAS, 405, 1047 Gvaramadze V. V., Kniazev A. Y., Fabrika S., Sholukhova O., Berdnikov L. N., Cherepashchuk A. M., Zharova A. V., 2010b, MNRAS, 405, 520 Gvaramadze V. V. et al., 2012a, MNRAS, 421, 3325 Gvaramadze V. V., Weidner C., Kroupa P., Pflamm-Altenburg J., 2012b, MNRAS, 424, 3037 Gvaramadze V. V., Kniazev A. Y., Berdnikov L. N., Langer N., Grebel E. K., Bestenlehner J. M., 2014, MNRAS, 445, L84 Hamaguchi K. et al., 2014, ApJ, 795, 119 Humphreys R. M., Davidson K., 1979, ApJ, 232, 409 Humphreys R. M., Davidson K., 1994, PASP, 106, 1025 Humphreys R. M., Davidson K., Smith N., 1999, PASP, 111, 1124 Humphreys R. M., Weis K., Davidson K., Bomans D. J., Burggraf B., 2014a, ApJ, 790, 48 Humphreys R. M., Davidson K., Gordon M. S., Weis K., Burggraf B., Bomans D. J., Martin J. C., 2014b, ApJ, 782, L21 Keenan P. C., Hynek J. A., 1950, ApJ, 111, 1 Kniazev A. Y., Gvaramadze V. V., Berdnikov L. N., 2015, MNRAS, 449, L60 Kniazev A. Y., Pustilnik S. A., Grebel E. K., Lee H., Pramskij A. G., 2004, ApJS, 153, 429 Kniazev A. Y. et al., 2005, AJ, 130, 1558 Kniazev A. Y. et al., 2008, MNRAS, 388, 1667 Kobulnicky H. A., Nordsieck K. H., Burgh E. B., Smith M. P., Percival J. W., Williams T. B., O’Donoghue D., 2003, in Iye M., Moorwood A. F. M., eds, Proc. SPIE Conf. Ser. Vol. 4841, Instrument Design and Performance for Optical/Infrared Ground-based Telescopes. SPIE, Bellingham, p. 1634 Krueger T.K., Aller L.H., Czyzak S.J., 1970, ApJ, 160, 921 Leyder J.-C., Walter R., Rauw G., 2008, A&A, 477, L29 Masetti N. et al., 2010, A&A, 519, A96 Mason A. B., Clark J. S., Norton A. J., Crowther P. A., Tauris T. M., Langer N., Negueruela I., Roche P., 2012, MNRAS, 422, 199 Merrill P. W., 1934, ApJ, 79, 183 Monet D. G. et al., 2003, AJ, 125, 984 Nazé Y., Rauw G., Hutsemékers D., 2012, A&A, 538, A47 Nazé Y., Corcoran M. F., Koenigsberger G., Moffat A. F. J., 2007, ApJ, 658, L25 O’Donoghue D. et al., 2006, MNRAS, 372, 151 Osmer P. S., 1972, ApJS, 24, 247 Owocki S. P., Castor J. I., Rybicki G. B., 1988, ApJ, 335, 914 Pallavicini R., Golub L., Rosner R., Vaiana G. S., Ayres T., Linsky J.L., 1981, ApJ, 248, 279 Parker Q. A. et al., 2005, MNRAS, 362, 689 Pickles A. J., 1998, PASP, 110, 863 Reed B. C., 2000, AJ, 120, 314 Reid M. J., Menten K. M., Zheng X. W., Brunthaler A., Xu Y., 2009, ApJ, 705, 1548 Rieke G. H. et al., 2004, ApJS, 154, 25 Röser S., Demleitner M., Schilbach E., 2010, AJ, 139, 2440 Sana H., Rauw G., Nazé Y., Gosset E., Vreux, J.-M., 2006, MNRAS, 372, 661 Saraph H.E., Seaton M.J., 1970, MNRAS, 148, 367 Schönrich R., Binney J., Dehnen W., 2010, MNRAS, 403, 1829 Skrutskie M. F. et al., 2006, AJ, 131, 1163 Stahl O., Mandel H., Szeifert Th., Wolf B., Zhao F., 1991, A&A, 244, 467 Stahl O. et al., 2001, A&A, 375, 54 Stevens I. R., Blondin J. M., Pollock A. M. T., 1992, ApJ, 386, 265 Stone R. C., 1979, ApJ, 232, 520 Stringfellow G. S., Gvaramadze V. V., Beletsky Y., Kniazev A. Y., 2012a, in Richards M. T., Hubeny I., eds, Proc. IAU Symp. 282, From Interacting Binaries to Exoplanets: Essential Modeling Tools. Cambridge Univ. Press, Cambridge, p. 267 Stringfellow G. S., Gvaramadze V. V., Beletsky Y., Kniazev A. Y., 2012b, in Drissen L., St-Louis N., Robert C., Moffat A. F. J., eds, ASP Conf. Ser. Vol. 465, Four Decades of Massive Star Research – A Scientific Meeting in Honor of Anthony J. Moffat. Astron. Soc. Pac., San Francisco, p. 514 Tarter C. B., 1969, ApJS 18, 1 The DENIS Consortium, 2005, VizieR Online Data Catalog, 2263, 0 Tutukov A., Yungelson L. R., 1973, Nauchn. Inform., 27, 86 Usov V. V., 1992, ApJ, 389, 635 van den Heuvel E. P. J., Heise J., 1972, Nature Phys. Sci., 239, 67 van Genderen A. M., de Groot M. J. H., The P. S., 1994, A&A, 283, 89 Vega E. I., Rabolli M., Feinstein A., Muzzio J. C., 1980, AJ, 85, 1207 Wachter S., Mauerhan J. C., van Dyk S. D., Hoard D. W., Kafka S., Morris P. W., 2010, AJ, 139, 2330 Wachter S., Mauerhan J., van Dyk S., Hoard D. W., Morris P., 2011, Bull. Soc. R. Sci. Liège, 80, 291 Wolf B., 1989, A&A, 217, 87 Zacharias N., Urban S. E., Zacharias M. I., Wycoff G. L., Hall D. M., Monet D. G., Rafferty T. J., 2004, AJ, 127, 3043 Zacharias N., Finch C.T., Girard T.M., Henden A., Bartlett J.L., Monet D.G., Zacharias M.I., 2013, AJ, 145, 44
[^1]: Based on observations made with the Southern African Large Telescope (SALT) under the program (PI: A.Y. Kniazev).
[^2]: In the Set of Identifications, Measurements and Bibliography for Astronomical Data (SIMBAD) data base this shell is named \[GKF2010\]MN44.
[^3]: http://webviz.u-strasbg.fr/viz-bin/VizieR
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Matthias R. Gaberdiel$^1$, Stefan Hohenegger$^2$ and Roberto Volpato$^1$\
\
${}^{1}$ Institut für Theoretische Physik\
ETH Zurich\
CH-8093 Zürich\
Switzerland\
\
${}^{2}$ Max-Planck-Institute for Physics\
Föhringer Ring 6\
D-80805 München\
Germany
title: |
[MPP-2011-69]{}
Symmetries of K3 sigma models
---
Introduction
============
Recently, a hidden ${\mathbb{M}}_{24}$ symmetry of the elliptic genus of K3 has attracted some attention. This development started with the observation of Eguchi, Ooguri and Tachikawa [@EOT] who noted that the first few multiplicities with which the ${\cal N}=4$ characters appear in the elliptic genus of K3 are sums (with integer coefficients) of dimensions of representations of the Mathieu group ${\mathbb{M}}_{24}$. The appearance of these dimensions suggests that the underlying vector space (consisting of the states that contribute to the elliptic genus) carries an action of ${\mathbb{M}}_{24}$. Assuming this group action one can then also define the ‘twining genera’[^1], [*i.e.*]{} the elliptic genus with the insertion of a group element $g\in {\mathbb{M}}_{24}$, \_[RR]{} ( g y\^[J\_0]{} q\^[L\_0-]{} (-1)\^F |[q]{}\^[|[L]{}\_0-]{} (-1)\^[|[F]{}]{} ) , and the usual string arguments suggest that these twining genera must have good modular properties under some congruence subgroup of $SL(2,\mathbb{Z})$. Using these modular properties as well as the explicit knowledge of the first few coefficients that follow from the observation of [@EOT], all twining genera could be determined [@Cheng:2010pq; @Gaberdiel:2010ch; @Gaberdiel:2010ca; @Eguchi:2010fg]. In turn, this leads to a stringent test of the proposal: knowing all twining genera one can deduce the decomposition of all multiplicity spaces into ${\mathbb{M}}_{24}$ representations, and it was found that at least for the first 1000 coefficients, non-negative integer multiplicities appear [@Gaberdiel:2010ca; @Eguchi:2010fg].[^2] This analysis therefore gives very convincing evidence for a hidden ${\mathbb{M}}_{24}$ symmetry underlying the elliptic genus of K3. Further support for this conjecture was given in [@Govindarajan:2010cf] (see also [@Govindarajan:2010fu; @Govindarajan:2009qt]).
Part of this symmetry can be understood geometrically. First of all, the elliptic genus is independent of the specific point in the moduli space of K3 that is considered, and thus the symmetries of the elliptic genus are in some sense the union of all symmetries that are present at different points in moduli space. The geometrical symmetries of K3, [*i.e.*]{} the symplectic automorphisms, have been studied some time ago by Mukai and Kondo [@Mukai; @Kondo], and they found that at any point in moduli space these symmetries form a subgroup of ${\mathbb{M}}_{23}$. The Mathieu group ${\mathbb{M}}_{23}$ is a maximal subgroup of ${\mathbb{M}}_{24}$, and thus this argument ‘explains’ part of the observation of [@EOT].
As is familiar for example from T-duality, string theory typically has more than just the geometric symmetries, and one may therefore expect that the remaining symmetries of ${\mathbb{M}}_{24}$ may be accounted for by ‘stringy symmetries’ (see also [@Taormina:2010pf]). In order to analyse this question, we study in this paper the stringy version of the Mukai-Kondo theorem. More specifically, we classify the spacetime supersymmetry preserving automorphisms of the non-linear $\sigma$-model at an arbitrary point in the moduli space of K3. From the point of view of the worldsheet, these symmetries are characterised by the property that they preserve the ${\cal N}=(4,4)$ superconformal algebra, as well as the spectral flow operators. Given the observation of [@EOT], one may have expected that all these symmetries should form a subgroup of ${\mathbb{M}}_{24}$, but actually the answer is more complicated:
[**Theorem:**]{}
*Let $G$ be the group of symmetries of a non-linear $\sigma$-model on $K3$ preserving the ${\mathcal{N}}=(4,4)$ superconformal algebra as well as the spectral flow operators. Then one of the following possibilities holds:*
[[()]{}]{}
$G=G'. G''$, where $G'$ is a subgroup of ${\mathbb{Z}}_2^{11}$, and $G''$ is a subgroup of ${\mathbb{M}}_{24}$ with at least four orbits when acting as a permutation on $\{1,\ldots,24\}$
$G=5^{1+2}.{\mathbb{Z}}_4$
$G={\mathbb{Z}}_3^4.A_6$
$G=3^{1+4}.{\mathbb{Z}}_2.G''$, where $G''$ is either trivial, ${\mathbb{Z}}_2$, ${\mathbb{Z}}_2^2$ or ${\mathbb{Z}}_4$.
Here $p^{1+2n}$ denotes an extra special group of order $p^{1+2n}$, and $N.Q$ denotes a group $G$ for which $N$ is a normal subgroup such that $G/N\cong Q$ (for an exposition of our mathematical notation and conventions see Appendix \[s:notation\]). Note that except for case (i) with $G'$ trivial, these groups are not subgroups of ${\mathbb{M}}_{24}$; in particular, for cases (ii)-(iv) this follows from the fact that their order does not divide |\_[24]{}|=2\^[10]{}3\^3571123 . On the other hand, all groups in (i)–(iv) [*are*]{} subgroups of the Conway group $Co_1$, and thus the analogue of the Mukai-Theorem is that the stringy symmetries all lie in $Co_1$. One may take this as evidence that the elliptic genus of K3 should in fact have a hidden $Co_1$ symmetry, but from its decomposition in terms of ${\mathcal{N}}=4$ elliptic genera, we have not seen any hint for this. In any case, the result of the Theorem means that the explanation of the ${\mathbb{M}}_{24}$ symmetry appearing in the elliptic genus of K3 must be more subtle.
Sketch of Proof
---------------
Let us briefly sketch the proof of the Theorem, before returning to more general considerations below; the details of this argument as well as the underlying assumptions will be spelled out in Section \[symmetries\] and Appendix \[s:latticeproofs\].
The basic strategy of the proof follows closely the proof of the Mukai-theorem, given by Kondo. The moduli space of sigma-models on K3 has the form \_[K3]{} = O(\^[4,20]{})\\O(4,20)/(O(4)O(20)) . Here the Grassmannian $O(4,20)/(O(4)\times O(20))$ parametrises the choice of a positive definite four-dimensional subspace $\Pi\subset {\mathbb{R}}^{4,20}$, and $O(\Gamma^{4,20})$ is the group of automorphisms of the even unimodular lattice $\Gamma^{4,20}\subset {\mathbb{R}}^{4,20}$ of signature $(4,20)$. We may think of $\Gamma^{4,20}$ as the integral homology of K3, [*i.e.*]{} as the D-brane charge lattice, while the position of $\Pi$ is specified by the choice of a Ricci-flat metric and a $B$-field on K3. In particular, it therefore determines the four left- and right-moving supercharges.
The supersymmetry preserving automorphisms of the non-linear $\sigma$-model characterised by $\Pi$ generate the group $G\equiv G_\Pi$ that consists of those elements of $O(\Gamma^{4,20})$ that leave $\Pi$ pointwise fixed. We denote by $L^G$ the sublattice of $G$-invariant vectors of $L\equiv \Gamma^{4,20}$, and define $L_G$ to be its orthogonal complement. By construction, $\Pi$ is a subspace of the real vector space $L^G\otimes {\mathbb{R}}\subset {\mathbb{R}}^{4,20}$, and since $\Pi$ has signature $(4,0)$, the orthogonal complement $L_G$ must be a negative definite lattice of rank at most $20$. The basic idea is now to embed $L_G(-1)$ — the $(-1)$ means that we change the sign of its intersection matrix — into the Leech lattice $\Lambda$. Such an embedding exists, provided we assume that $L_G(-1)$ does not contain any vectors of length squared two (which would signal some gauge enhancement and thus would lead to a singular CFT). Since the action of $G$ fixes all vectors of $\Lambda$ orthogonal to $L_G(-1)$, it follows that $G$ must be a subgroup of $Co_1 \subset Co_0={\rm Aut}(\Lambda)$ that fixes pointwise a sublattice of the Leech lattice of rank at least $4$. A more careful analysis then leads to the separate cases (i)-(iv) above.
Comments and Outline
--------------------
Since the result of the Theorem is somewhat contrary to expectations, we have also studied a few explicit conformal field theories describing K3 at different points in moduli space in detail. In particular, we have done this for (A) the orbifold point ${\mathbb T}^4/{\mathbb Z}_2$; (B) the orbifold point ${\mathbb T}^4/{\mathbb Z}_4$ which is equivalent to the Gepner model $(2)^4$; and (C) the Gepner model $(1)^6$. Given that these descriptions are very explicit, it is possible to identify (at least some of) the supersymmetry-preserving automorphisms. In each case we have computed the resulting symmetry group, and compared it with the possibilities allowed for by the Theorem. We find that (C) realises case (iii), while both (A) and (B) correspond to case (i) with $G'$ non-trivial. In particular, all of these cases therefore describe K3s for which the stringy symmetries do not lie inside ${\mathbb{M}}_{24}$.
From the point of view of the argument leading to the Theorem, the symmetries of the worldsheet theory should have an interpretation as lattice symmetries. Actually, this point of view can also be directly understood in conformal field theory: as mentioned before, the lattice $\Gamma^{4,20}$ can be identified with the D-brane charge lattice, and $\Pi$ describes the four left- and right-moving supercharges. Thus the supersymmetry preserving automorphisms of the CFT should be in one-to-one correspondence with the symmetries of the D-brane charge lattice that leave $\Pi$ pointwise invariant. Using conformal field theory methods, it is fairly straightforward to determine the D-brane charge lattice, as well as the $\Pi$-preserving symmetries. For each of the three cases (A)–(C) we have verified that the resulting symmetry groups reproduce precisely those obtained from the explicit construction of the symmetry generators above. Incidentally, this method to determine the supersymmetry preserving automorphisms by analysing the D-brane charge lattice constitutes a nice general approach that can be applied to any non-linear $\sigma$-model on K3.
The paper is organised as follows. In the following section (Section \[symmetries\]) we give a more detailed description of the main Theorem and the assumptions that go into its proof. Section \[s:orbifold\] is devoted to the study of the orbifold point ${\mathbb T}^4/{\mathbb Z}_2$. Among other things, we calculate the twining genera for the various symmetries and find the twining genus of the 2B conjugacy class of ${\mathbb{M}}_{24}$ that does not lie inside ${\mathbb{M}}_{23}$. (The corresponding symmetry is the stringy 4-fold T-duality symmetry ${\mathcal{T}}$, see (\[Z2T\]).) However, we also find symmetries that do not lie inside ${\mathbb{M}}_{24}$, and whose twining genus does not agree with the twining genus of any conjugacy class in ${\mathbb{M}}_{24}$. (The simplest example is the ‘quantum symmetry’ ${\mathcal{Q}}$, see (\[Z2Q\]).) In Section \[s:16\], the same analysis is done for the $(1)^6$ Gepner point, while Section \[s:24\] deals with the $(2)^4$ Gepner model. Our notation and some basic mathematical background is described in Appendix \[s:notation\], while some of the details of the proof of the Theorem have been delegated to Appendix \[s:latticeproofs\]. Finally, Appendix \[s:Gepner\] contains some of the details of the D-brane charge analysis for the Gepner models.
Symmetries of Non-linear $\sigma$-models on K3 {#symmetries}
==============================================
In the following we shall consider two-dimensional theories with ${\mathcal{N}}=(4,4)$ superconformal symmetry and central charge $c=6$. Theories of this type can be classified according to their elliptic genus $\phi(\tau,z)$. The symmetries of the theory constrain the elliptic genus to be a weak Jacobi form of weight $0$ and index $1$, and the only possibilities are $\phi(\tau,z)=0$, which corresponds to the case of the target space being ${\mathbb T}^4$, or \[K3eg\] (,z)=8\_[i=2]{}\^4=2y+20+2y\^[-1]{}+O(q) , where $\vartheta_i$ are the Jacobi theta functions. This second case arises if the target space is K3 and it is the main focus of the present paper.
Moduli Space of Non-linear $\sigma$-models on K3
------------------------------------------------
As mentioned in the introduction, the moduli space of ${\mathcal{N}}=(4,4)$ theories with elliptic genus is believed to be the quotient (see for example [@Aspinwall:1996mn; @NahmWend]) \[K3moduli\] \_[K3]{} = O(\^[4,20]{})\\O(4,20)/(O(4)O(20)) . Here the Grassmannian $O(4,20)/(O(4)\times O(20))$ parametrises the choice of a positive definite four-dimensional subspace in ${\mathbb{R}}^{4,20}$, and $O(\Gamma^{4,20})$ is the group of automorphisms of the even unimodular lattice $\Gamma^{4,20}$ with signature $(4,20)$.
Geometrically, we can think of $\Gamma^{4,20}$ as the integral homology lattice $H_{\rm even}(X,{\mathbb{Z}})$ of the K3 manifold $X$, with the bilinear form given by the intersection number. The space ${\mathbb{R}}^{4,20}$ is interpreted as the real even cohomology $H^{\rm even}(X,{\mathbb{R}})$ endowed with the cup product, and the embedding $\Gamma^{4,20}\subset {\mathbb{R}}^{4,20}$ is realised through Poincaré duality $H_{\rm even}(X,{\mathbb{Z}})\cong H^{\rm even}(X,{\mathbb{Z}})\subset H^{\rm even}(X,{\mathbb{R}})$. The non-linear $\sigma$-model is determined by choosing a Ricci-flat metric and a B-field on the manifold $X$; this corresponds to the choice of the $4$-dimensional subspace $\Pi\subset H^{\rm even}(X,{\mathbb{R}})$, though the relationship is rather involved (see *e.g.* [@Aspinwall:1996mn]).
In string theory, the homology lattice can be identified with the lattice of D-brane charges, and the intersection number for $\alpha,\beta\in H_{\rm even}(X,{\mathbb{Z}})$ is reproduced by the overlap \[inter\] = q\^[(L\_0+L\_0)-]{}(-1)\^[F\_L]{} \_[RR]{} . Here only the RR part of the boundary states $\|\alpha{\rangle\!\!\!\;\rangle},\|\beta{\rangle\!\!\!\;\rangle}$ corresponding to the D-branes wrapping the cycles $\alpha$ and $\beta$ contribute [@Brunner:1999jq]. (Alternatively, (\[inter\]) is the Witten index in the R-sector of the relative open string.) In this picture, the dual space of real cohomology is naturally identified with the space of $24$ (anti-)chiral RR ground states with $h=\bar h=\tfrac{1}{4}$. Under the action of ${\rm SU}(2)_L\times {\rm SU}(2)_R$, which is part of the ${\mathcal{N}}=(4,4)$ superconformal symmetry, the $24$ RR ground states split into a four-dimensional $({\bf 2},{\bf 2})$-representation and $20$ singlets. The four-dimensional subspace $\Pi\subset {\mathbb{R}}^{4,20}$ is then to be identified with the subspace of RR states transforming in the $({\bf 2},{\bf 2})$-representation.
Characterisation of ${\mathcal{N}}=(4,4)$ Preserving Symmetries
---------------------------------------------------------------
Our goal is to classify the discrete symmetries $\tilde G_\Pi$ of a given ${\mathcal{N}}=(4,4)$ theory parametrised by $\Pi$ in the moduli space ${\mathcal{M}}_{\rm K3}$. Any symmetry $g\in \tilde{G}_\Pi$ must obviously leave $\Pi$ invariant. It must therefore either be an element of $O(\Gamma^{4,20})$, or it must act trivially on the Grassmannian in (\[K3moduli\]). However, the latter case would imply that the symmetry exists everywhere in moduli space, and we know (for example from studying deformations of the orbifold line) that this is not possible. Thus we conclude that the symmetries $\tilde{G}_\Pi$ of the theory at $\Pi$ is precisely the subgroup of $O(\Gamma^{4,20})\subset O(4,20,{\mathbb{R}})$ that leaves $\Pi$ (setwise) fixed.
A general symmetry $g\in \tilde G_\Pi$ will preserve the ${\mathcal{N}}=(4,4)$ superconformal algebra only up to an automorphism. From now on we want to focus on the subgroup $G_\Pi\subset \tilde G_\Pi$ that actually leaves the ${\mathcal{N}}=(4,4)$ superconformal algebra invariant, [*i.e.*]{} for which this automorphism is trivial, and that preserve the spectral flow operators. These symmetries are characterised by the condition that they preserve space-time supersymmetry [@Banks:1988yz]. They are also relevant for the analysis of ‘Mathieu Moonshine’ [@EOT] that was reviewed at the beginning of the Introduction. Indeed, ${\mathbb{M}}_{24}$ appears to act on the multiplicity spaces with which the ${\cal N}=4$ representations appear in the elliptic genus, and hence must commute with the left-moving ${\mathcal{N}}=4$ superconformal algebra. Furthermore, according to the proposal of [@Cheng:2010pq; @Gaberdiel:2010ch], the four RR ground states that transform in the $({\bf 2},{\bf 2})$ of the ${\rm SU}(2)_L\times {\rm SU}(2)_R$ — these are the spacetime supercharges — sit in a singlet representation of ${\mathbb{M}}_{24}$. Thus the symmetries that are described by ${\mathbb{M}}_{24}$ should leave the full ${\mathcal{N}}=(4,4)$ superconformal algebra invariant, and preserve the spectral flow operators. Obviously, ${\mathbb{M}}_{24}$ is not a subgroup of $O(\Gamma^{4,20})$, and thus we cannot explain the full ${\mathbb{M}}_{24}$ symmetry by looking at just one point in moduli space. However, the elliptic genus is constant over moduli space, and one may therefore expect that we can account for the entire ${\mathbb{M}}_{24}$ by putting information from different points in moduli space together. This is one of the main motivations for classifying the symmetry groups at different points in moduli space.
In any case, since $\Pi\subset {\mathbb{R}}^{4,20}$ can be identified with the subspace of RR states transforming in the $({\bf 2},{\bf 2})$-representation, we conclude that
> \[th:generic\] [*The subgroup $G_\Pi$ of symmetries of the K3 $\sigma$-model characterised by $\Pi\in{\mathbb{R}}^{4,20}$ that leave the ${\mathcal{N}}=(4,4)$ superconformal algebra invariant and that preserve the spectral flow operators, is the subgroup of $O(\Gamma^{4,20})\subset O(4,20,{\mathbb{R}})$ that leaves $\Pi$ (pointwise) fixed.*]{}
In the following, by a symmetry of an ${\mathcal{N}}=(4,4)$ theory in ${\mathcal{M}}_{\rm K3}$, we will always mean a transformation with these properties. Note that this restriction excludes some very interesting symmetries, for example mirror symmetry in a self-mirror theory. On the other hand, the extension of our arguments to more general cases is fairly straightforward.
Classification of the Groups of Symmetries {#s:classif}
------------------------------------------
In this section we will classify the possible groups $G_\Pi$. The related problem in classical geometry has been previously considered by Mukai [@Mukai], who classified the groups of symplectic automorphisms of K3 surfaces. Mukai proved that the symplectic automorphisms of any K3 surface form a subgroup of ${\mathbb{M}}_{23}$, which in turn is a maximal subgroup of ${\mathbb{M}}_{24}$. The Mathieu group ${\mathbb{M}}_{23}$ is finite and its subgroups are well studied, so that the Mukai theorem provides a very explicit description of all symplectic automorphisms. In the following, we will extend the Mukai theorem to the classification of the symmetry groups of the $\sigma$-models. In particular, we will show that $G_\Pi\subset Co_1$, the Conway group $Co_1$. Note that $Co_1$ contains ${\mathbb{M}}_{23}$ (as well as ${\mathbb{M}}_{24}$) as a subgroup.
As argued in the previous section, the symmetries of interest form the subgroup of $O(\Gamma^{4,20})\subset O(4,20,{\mathbb{R}})$ that fix (pointwise) the positive-definite four-dimensional subspace $\Pi\subset {\mathbb{R}}^{4,20}$, characterising the relevant point in moduli space. However, not all choices of $\Pi$ correspond to well-defined conformal field theories. In particular, when $\Pi$ is orthogonal to a vector $v\in\Gamma^{4,20}$ of norm $v^2=-2$ (usually called a root of $\Gamma^{4,20}$), the corresponding non-linear $\sigma$-model is not well defined [@Aspinwall:1995xy]. This subtlety can be understood by considering the model as an internal CFT in type IIA superstring theory. This theory is dual to heterotic string theory compactified on ${\mathbb T}^4$. Generically, the corresponding low energy effective field theory contains an abelian gauge group $U(1)^{24}$. However, when $\Pi$ is orthogonal to a root $v$, the gauge group is enhanced to a non-abelian gauge group, and $v$ is interpreted as a root of the corresponding Lie algebra. The additional states in type IIA superstring theory are interpreted as D-branes becoming massless at this point of the moduli space. This means that the corresponding perturbative superconformal field theory cannot describe correctly all massless degrees of freedom of the theory, and hence the non-linear $\sigma$-model is expected to be inconsistent [@Strominger:1995cz]. Therefore, in the following, we shall exclude the points in the moduli space where $\Pi$ is orthogonal to a root. It is believed that these are the only singular points in the moduli space.
Our strategy to characterise the groups $G_\Pi$ is inspired by the proof of the Mukai theorem given by Kondo [@Kondo]. Let $\Pi\in{\mathbb{R}}^{4,20}$ be a $4$-dimensional positive definite space, not orthogonal to any root (vector of norm $-2$) in $L\equiv \Gamma^{4,20} \subset {\mathbb{R}}^{4,20}$, and let $G\equiv G_\Pi\subset O(\Gamma^{4,20})$ be the subgroup of lattice automorphisms fixing $\Pi$ pointwise. We denote by $L^G$ the sublattice of vectors fixed by $G$ \[LuG\] \^[4,20]{}L\^G:={v\^[4,20]{}|g(v)=vgG} , and by $L_G$ its orthogonal complement \[LG\] \^[4,20]{}L\_G:={w\^[4,20]{}|wv=0vL\^G} . By definition, the real vector space $L^G\otimes {\mathbb{R}}\subset {\mathbb{R}}^{4,20}$ will contain $\Pi$, $\Pi\subset L^G\otimes {\mathbb{R}}$, and since $\Pi$ has signature $(4,0)$, the orthogonal complement $L_G$ must be negative definite and have rank at most $20$. Furthermore, every vector in $L_G$ is orthogonal to $\Pi$, so that, by our assumption on $\Pi$, $L_G$ contains no roots. The proof of the Theorem then proceeds as follows — the relevant details are given in Appendix \[s:latticeproofs\]:
- First (see Appendix \[s:GinOGamma\]), we prove that $L_G(-1)$, can be embedded in the even unimodular lattice $\Gamma^{25,1}$. The action of $G$ on $L_G$ extends to an action on $\Gamma^{25,1}$, which fixes all vectors orthogonal to $L_G(-1)$ in $\Gamma^{25,1}$. Thus, $G$ is a subgroup of ${\rm Aut}(\Gamma^{25,1})$.
- Next (see Appendix \[s:proofConway\]), using the fact that $L_G(-1)$ contains no vectors of norm $2$ and the properties of ${\rm Aut}(\Gamma^{25,1})$, we show that $L_G(-1)$ must be contained in a positive definite sublattice of $\Gamma^{25,1}$, namely the Leech lattice $\Lambda$. This is the unique $24$-dimensional even unimodular lattice with no roots and its group of automorphisms is the Conway group $Co_0$ [@Conway]. This group can be obtained by extending the sporadic finite simple group $Co_1$ [@Atlas] of order |Co\_1|=2\^[21]{}3\^95\^47\^2111323\~410\^[18]{} , by the ${\mathbb{Z}}_2$ symmetry that changes the sign of all vectors in $\Lambda$. Once again, the action of $G$ on $L_G(-1)$ can be extended to an action on $\Lambda$ which fixes all vectors orthogonal to $L_G(-1)$. This means, in particular, that the ${\mathbb{Z}}_2$ symmetry that changes the sign of all vectors in $\Lambda$ is not an element of $G$, since it has no non-trivial fixed vectors. It follows that $G$ is a subgroup of the finite simple group $Co_1$.
Combining these results thus leads to the natural analogue of the Mukai-Theorem:
[**Proposition.**]{} [*The group $G_\Pi$ of symmetries of any non-linear $\sigma$-model on K3 is a subgroup of the Conway group $Co_1 \subset Co_0={\rm Aut}(\Lambda)$ that fixes pointwise a sublattice of the Leech lattice $\Lambda$ of rank at least $4$.*]{}
In order to give a more precise description of the groups of symmetries $G_\Pi$, one needs a detailed classification of the subgroups of $Co_0$ that fix a sublattice of rank $4$ in $\Lambda$. The result of this somewhat technical analysis — the details are explained in Appendix \[s:proofs\] — is the Theorem stated in the Introduction.
We should mention that our analysis does not actually prove that $Co_1$ is the smallest possible group containing all these symmetry groups. However, some simple considerations on the order of the groups $|G_\Pi|$ are sufficient to exclude all maximal subgroups of $Co_1$, except for $Co_2$. Furthermore, all the cases (i)–(iv) in the Theorem are actually realised by some ${\mathcal{N}}=(4,4)$ model, provided we assume that every four-dimensional subspace $\Pi\subset {\mathbb{R}}^{4,20}$, not orthogonal to any vector of norm $-2$ in $\Gamma^{4,20}$, leads to a consistent conformal field theory (see Appendix \[B.4\]). The characterisation of the symmetry groups as given in the Theorem is therefore optimal.
In the following we shall describe in detail specific examples that realise some of the possibilities of the Theorem. In particular, the ${\mathbb T}^4/{\mathbb{Z}}_2$ orbifold model to be discussed in Section \[s:orbifold\] and the $(2)^4$ Gepner model of Section \[s:24\] are examples of case (i) and have symmetry groups that are not subgroups of ${\mathbb{M}}_{24}$, while the Gepner model $(1)^6$ that will be studied in Section \[s:16\] realises precisely case (iii).
The ${\mathbb Z}_2$ Orbifold Model {#s:orbifold}
==================================
In order to illustrate the general predictions of the Theorem let us consider a few specific examples. We begin with the ${\mathbb T}^4/{\mathbb{Z}}_2$ orbifold model where we take ${\mathbb T}^4$ to be the orthogonal torus at the self-dual radius, [*i.e.*]{} the four radii take on the self-dual value, and set the $B$-field on the torus to zero. We write ${\mathbb{T}}^4= {\mathbb{T}}^2\times {\mathbb{T}}^2$, and label the two ${\mathbb{T}}^2$s by $i=1,2$. For each ${\mathbb{T}}^2$ we use complex coordinates, and thus the left-moving bosonic and fermionic modes are $$\alpha^{(i)}_{n} \ , \qquad \bar{\alpha}^{(i)}_{n} \ , \qquad \qquad
\psi^{(i)}_{n} \ , \qquad \bar{\psi}^{(i)}_{n} \ , \qquad i=1,2 \ ,$$ and similarly for the right-movers. (The right-movers are denoted by a tilde.) The ${\mathbb Z}_2$ orbifold ${\cal I}$ acts as $-1$ an all of these modes. In addition it maps the momentum ground states ${\cal I} (p_L,p_R) = (-p_L,-p_R)$.
The Spectrum in the RR sector {#sec:2.1}
-----------------------------
We denote the states in the untwisted sector of ${\mathbb{T}}^4={\mathbb{T}}^2\times {\mathbb{T}}^2$ by |p\_L,p\_R;N,;s;s , where $p_L=({n}+{w})/\sqrt{2}$, $p_R=({n}-{w})/\sqrt{2}$ are the left and right momenta at the self dual radius, with ${n}\in{\mathbb Z}^4$ the momentum and ${w}\in{\mathbb Z}^4$ the winding numbers. Furthermore, $N$ and $\tilde{N}$ denote the left and right oscillator contributions, while $s,\tilde s$ label the Ramond ground states. More specifically, $$(s;\tilde s)=(s_1,s_2;\tilde s_1,\tilde s_2) \ ,$$ where each $s_i$, $i=1,2$, can take the two values $\pm \tfrac{1}{2}$, and the zero modes $\bar\psi_0^{(i)}$ and $\psi_0^{(i)}$ map the states with $s_i=\pm \tfrac{1}{2}$ into one another; the analogous statement holds for the $\tilde s_i$ in the right-moving sector. The RR ground states have charge $(S,\tilde{S})$ with respect to the left- and right-moving $U(1)$-current, where $$S = \sum_i s_i \ , \qquad \tilde{S} = \sum_i \tilde{s}_i \ .$$ To obtain the spectrum of the ${\mathbb{Z}}_2$-orbifold theory we have to project onto states that are even under the operator ${\mathcal{I}}$ acting as |p\_L,p\_R;N,;s;s= (-1)\^[|N|+||]{}(-1)\^[S+]{} |-p\_L,-p\_R;N,;s;s , where $|N|$ is the total number of oscillators appearing in $N$, and likewise for $|\tilde{N}|$.
In the twisted sector the states are labelled by |i;N, , where $i=1,\ldots, 16$ distinguishes the $16$ different fixed points of the orbifold, while $N,\tilde{N}$ denote again the oscillator numbers. The ground states do not carry any charge with respect to the left- and right-moving $U(1)$-currents. The orbifold projection in the twisted sector acts as |i;N,=(-1)\^[|N|+||]{} |i;N, . It is straightforward to calculate the elliptic genus from this description. In the untwisted sector one finds \[Z2U\] \^[(U)]{}(,z)=(2y+4+2y\^[-1]{})\_[n=1]{}\^= 8 , while the contribution of the twisted sector equals \[Z2Tw\] \^[(T)]{}(,z)= 8( +) . It is easy to see that their sum, $\phi^{(U)}(\tau,z) + \phi^{(T)}(\tau,z)$, reproduces precisely (\[K3eg\]).
Symmetries and Twining Genera {#s:3.2}
-----------------------------
The orbifold theory possesses various symmetries that leave the ${\mathcal{N}}=(4,4)$ superconformal algebra invariant and that preserve the spectral flow operators; in particular, we have
[[()]{}]{}
${\cal R}$: rotation of the two ${\mathbb T}^2$’s by $90$ and $-90$ degrees, respectively.
${\cal E}$: exchanging the two ${\mathbb T}^2$’s, together with an inversion acting on the second ${\mathbb T}^2$, say.
$H_a$: half-period translations, that act as $(-1)^{p\cdot a}$, $a\in ({\mathbb{Z}}/2{\mathbb{Z}})^4$, in the untwisted sector, and by a permutation on the $16$ twisted sectors.
${\cal T}$: 4-fold T-duality. (Note that the 2-fold T-duality induces a non-trivial automorphism of the ${\mathcal{N}}=(4,4)$ superconformal algebra.)
${\cal Q}$: the quantum symmetry that acts as $+1$ on the untwisted, and as $-1$ on the twisted sector.
These symmetries generate the group $G=2^{1+8}\rtimes {\mathbb{Z}}_2^3$. The normal subgroup acting trivially on the untwisted sector RR ground states with $p_L=p_R=0$ is the extra special group $2^{1+8}$ of order $2^9$, containing $\mathcal{Q}$, the center of $G$. It is generated by the half-shifts $H_a$ that form an abelian ${\mathbb{Z}}_2^4$, and by their dual symmetries $G_a=\mathcal{T}H_a\mathcal{T}$ that change the sign of half of the twisted sectors. The only non-trivial commutators are $H_aG_bH_aG_b={\cal Q}^{a\cdot b}$, where $a,b\in({\mathbb{Z}}/2{\mathbb{Z}})^4$. The quotient group $G/2^{1+8}\cong {\mathbb{Z}}_2^3$ is generated by ${\cal R}$, ${\cal E}$ and ${\cal T}$. Conjugation by ${\cal R}$ and ${\cal E}$ yields a permutation of the half shifts (and the analogous permutations for $G_a$), while conjugation by $\mathcal{T}$ exchanges $G_a$ and $H_a$. The group $G=2^{1+8}\rtimes {\mathbb{Z}}_2^3$ realises case (i) of the Theorem with $G'={\mathbb{Z}}_2^6$ and $G''={\mathbb{Z}}_2^4 \rtimes {\mathbb{Z}}_2^2$. Here $G''\subset {\mathbb{M}}_{24}$ is generated by $H_a$, ${\cal E}$ and ${\cal R}$, while $G'\subset {\mathbb{Z}}_2^{11}$ is generated by ${\mathcal{Q}}$ (one ${\mathbb{Z}}_2$ factor), by the composition ${\mathcal{T}}{\cal R} {\cal E}$ (another ${\mathbb{Z}}_2$), and by the elements of the form $H_a {\cal T} {\cal R} {\cal E} H_a$ (giving ${\mathbb{Z}}_2^4$).
Given the explicit description of the RR sector from above, it is straightforward to calculate the corresponding twining genera, \_g(,z)=\_[RR]{}(g q\^[L\_0-]{}|q\^[|L\_0-]{}y\^[J\_0]{}(-1)\^[F+|F]{}) , the elliptic genus with the insertion of the symmetry $g$.[^3] In particular, we can check how these twining genera compare with the Mathieu twining genera that were worked out in [@Cheng:2010pq; @Gaberdiel:2010ch; @Gaberdiel:2010ca; @Eguchi:2010fg]. Our explicit results are for example $$\begin{aligned}
\phi_{\mathcal{R}}(\tau,z) & = & 4\frac{\vartheta_2(2\tau,2z)}{\vartheta_2(2\tau,0)}
+4\frac{\vartheta_3(2\tau,2z)}{\vartheta_3(2\tau,0)} = \phi_{\rm 2A}(\tau,z) \label{orbR}\\
\phi_{\cal E}(\tau,z)& = & 4\frac{\vartheta_2(2\tau,2z)}{\vartheta_2(2\tau,0)}
+4\frac{\vartheta_3(2\tau,2z)}{\vartheta_3(2\tau,0)} = \phi_{\rm 2A}(\tau,z)\label{orbE} \\
\phi_{H_a}(\tau,z)& = & 8\frac{\vartheta_2(\tau,z)^2}{\vartheta_2(\tau,0)^2}=\phi_{\rm 2A}(\tau,z)\label{orbH} \\
\phi_{\mathcal{T}}(\tau,z) & = &
-2\vartheta_{4}(2\tau)^4\frac{\vartheta_1(\tau,z)^2}{\eta(\tau)^6} = \phi_{\rm 2B}(\tau,z)\label{Z2T} \\
\phi_{{\mathcal{T}}H_{a}}(\tau,z) & = & \phi_{\rm 4A}(\tau,z) \qquad\qquad\qquad \quad
[\hbox{$a^2$ even, say $a= (1100)$} ]\\
\phi_{{\mathcal{R}}H_{a}}(\tau,z) & = & \phi_{\rm 4B}(\tau,z) \qquad \qquad \qquad \quad
[\hbox{{\it e.g.}\ for $a=(1111)$}]\\
\phi_{{\mathcal{T}}H_{a}}(\tau,z) & = & \phi_{\rm 4C}(\tau,z) \qquad \qquad \qquad \quad
[\hbox{{\it e.g.}\ for $a=(1000)$}] \\
\phi_{{\mathcal{Q}}}(\tau,z) &= &
8\Bigl(\frac{\vartheta_{2}(\tau,z)^2}{\vartheta_{2}(\tau,0)^2}
- \frac{\vartheta_{3}(\tau,z)^2}{\vartheta_{3}(\tau,0)^2}
-\frac{\vartheta_{4}(\tau,z)^2}{\vartheta_{4}(\tau,0)^2}\Bigr) \nonumber \\
& = & 2 \phi_{\rm 2A}(\tau,z) - \phi_{\rm 1A} (\tau,z) \qquad \label{Z2Q} \\[4pt]
\phi_{{\mathcal{Q}}{\mathcal{T}}H_a G_b}(\tau,z) & = &
-2 \, \frac{\vartheta_3(2\tau)^4\, \vartheta_1(\tau,z)^2}{\eta(\tau)^6}
- 2 \frac{\vartheta_3(\tau,z)^2}{\vartheta_3(\tau)^2}
- 2 \frac{\vartheta_4(\tau,z)^2}{\vartheta_4(\tau)^2} \label{Z24p} \\
& = & \tfrac{1}{2}(-\phi_{\rm 1A}+\phi_{\rm 2A}+2\phi_{\rm 4B}) \quad \;
[\hbox{{\it e.g.}\ for $a=(1100)$, $b=(0011)$}] \ . \nonumber\end{aligned}$$ It is worth pointing out that the generators $H_a$, ${\cal E}$ and ${\cal R}$ that generate $G''\subset {\mathbb{M}}_{24}$ lead to twining genera that directly agree with ${\mathbb{M}}_{24}$ twining genera. On the other hand, the twining genera of the quantum symmetry ${\mathcal{Q}}$, see (\[Z2Q\]), and of the group elements $g ={\mathcal{Q}}{\mathcal{T}}H_{a}G_{b}$ for suitable choices of $a$ and $b$, see (\[Z24p\]), do not equal [*any*]{} Mathieu twining genus (but can only be expressed in terms of linear combinations of such twining genera). Finally, certain twining genera involving the 4-fold T-duality, namely ${\mathcal{T}}$ and ${\mathcal{T}}H_a$ for suitable $a$, give rise to twining genera, whose conjugacy classes (2B, 4A and 4C) do not lie inside ${\mathbb{M}}_{23}$. This ties in with the fact that T-duality is a non-geometric symmetry.
The D-brane Charge Lattice
--------------------------
As suggested by the proof of the Theorem (and explained in Section \[symmetries\]), we should also be able to characterise the symmetry group of the model as the symmetries of the D-brane charge lattice that leave the 4-dimensional subspace $\Pi$ (corresponding to the 4 supercharges) invariant. For the case at hand, the D-brane charge lattice can be computed straightforwardly since the primitive branes may be taken to be fractional D0, D2 and D4 branes. They can be constructed as explained, for example, in [@Gaberdiel:2000jr] (see also [@Bergman:1999kq]). If we denote the Ishibashi states in the untwisted and the $i^{\rm th}$ twisted sector by $| {\rm Dp,U} \rangle\!\rangle$ and $| {\rm Dp},i \rangle\!\rangle$, respectively, the fractional branes have the structure $$\begin{aligned}
|\!| {\rm D4},\epsilon, \delta \rangle\!\rangle_{\rm f} & = & \tfrac{1}{2} | {\rm D4,U} \rangle\!\rangle
+ \tfrac{\epsilon}{4} \sum_{i=1}^{16} \delta_i \, | {\rm D4},i \rangle\!\rangle \\
|\!| {\rm D2(jk)},\epsilon,\delta \rangle\!\rangle_{\rm f} & = & \tfrac{1}{2} | {\rm D2(jk),U} \rangle\!\rangle
+ \tfrac{\epsilon}{2} \sum_{i \in P_{jk}} \delta_i\, | {\rm D2}(jk),i \rangle\!\rangle \\
|\!| {\rm D0},i,\epsilon \rangle\!\rangle_{\rm f} & = & \tfrac{1}{2} | {\rm D0,U} \rangle\!\rangle
+ \epsilon| {\rm D0},i \rangle\!\rangle \ ,\end{aligned}$$ where $|\!| {\rm D0,i,f} \rangle\!\rangle$ is the fractional D0-brane at the $i^{\rm th}$ fixed point, while $|\!| {\rm D2(jk),f} \rangle\!\rangle$ denotes the fractional D2-branes oriented along the (jk) direction, with $P_{jk}$ the set of four fixed points between which the D2-brane is ‘spanned’. Furthermore, $\epsilon$ and $\delta_i$ are signs, and the configurations of signs $\delta_i$ arise from Wilson lines, [*i.e.*]{} not all configurations of signs are allowed. The complete even self-dual charge lattice is spanned by[^4]
[[()]{}]{}
9 fractional D4-branes: one has no Wilson line and all twisted charges $+\tfrac{1}{4}$ ($\epsilon=+1$); one has no Wilson line and all twisted charges $-\tfrac{1}{4}$ ($\epsilon=-1$); the remaining 7 fractional D-branes have twisted charge $+\tfrac{1}{4}$ at the origin, and different choices of Wilson lines.
9 fractional D0-branes: one sits at the origin and has twisted charge $+1$; one sits at the origin and has twisted charge $-1$; the remaining 7 fractional D0-branes have charge $+1$ and sit at different fixed points.
6 fractional D2-branes, oriented along all $6= {4 \choose 2}$ 2-planes, without any Wilson lines ([*i.e.*]{} $\delta_i=+1$) and positive twisted charge ($\epsilon=+1$).
The non-zero entries of the intersection form are $$\begin{aligned}
& & \langle\!\langle {\rm D4}, \! {\rm U}| {\rm D0}, \! {\rm U} \rangle \!\rangle = 2 \ , \quad \nonumber \\
& & \langle\!\langle {\rm D2(12)}, \!{\rm U}\, | {\rm D2(34)}, \!{\rm U} \rangle \!\rangle =
\langle\!\langle {\rm D2(13)},\! {\rm U}\, | {\rm D2(24)}, \!{\rm U} \rangle \!\rangle =
\langle\!\langle {\rm D2(14)}, \!{\rm U}\, | {\rm D2(23)}, \!{\rm U} \rangle \!\rangle = 2 \nonumber \\
&& \langle\!\langle {\rm D}*,i | {\rm D}*,j\rangle\!\rangle = - 2 \delta_{ij} \ .\end{aligned}$$ It is then straightforward to check that the intersection matrix of the above D-branes has determinant $1$, [*i.e.*]{} that these $24$ branes generate indeed the full charge lattice. The RR charges that transform in the $({\bf 2},{\bf 2})$-representation of ${\rm SU}(2)_L\times {\rm SU}(2)_R$ and hence span $\Pi$ are precisely those carried by the bulk brane combinations ( [D0]{} + [D4]{} ) , ([D2(12)]{} + [D2(34)]{} ) , ([D2(13)]{} + [D2(24)]{} ) , ([D2(14)]{} + [D2(23)]{} ) , where D2(ij) denotes the bulk D2-brane oriented along the (ij) direction; this can be deduced from the analysis of [@Brunner:2006tc], see in particular Appendix A.2. The orthogonal complement (with respect to the intersection form) then turns out to be a lattice of rank 20. Upon changing the sign of its quadratic form, this lattice can be embedded into the Leech lattice $\Lambda$. Its orthogonal complement in $\Lambda$ is generated by four vectors $y_1,\ldots,y_4\in\Lambda$ with $y_i\cdot y_j=4\delta_{ij}$. According to our general argument, the group of symmetries $G$ of the model is isomorphic to the group of automorphisms of $\Lambda$ that act trivially on $y_1,\ldots,y_4$. Since $G$ fixes some vectors of norm $8$ in $\Lambda$ (for example $y_1+y_2$), it must be a subgroup of ${\mathbb{Z}}_2^{12}\rtimes{\mathbb{M}}_{24}$, see Appendix \[s:proofs\]. A more detailed analysis shows that $G=2^{1+8}\rtimes {\mathbb{Z}}_2^3$, thus matching the results from above.
The Gepner Model $(1)^6$ {#s:16}
========================
Next we consider the $(1)^6$ Gepner model which will turn out to realise case (iii) of the Theorem. It is constructed by taking six tensor powers of the ${\mathcal{N}}=2$ minimal model at $k=1$, subject to the ${\mathbb{Z}}_3$ orbifold projection generated by \_[i=1]{}\^6\^l\_[m\_i,s\_i;|m\_i,|s\_i]{}e\^[\_i m\_i]{}\_[i=1]{}\^6\^l\_[m\_i,s\_i;|m\_i,|s\_i]{} . For a short review of Gepner models, as well as an outline of our notations and conventions see Appendix \[s:Gepner\].
The Spectrum
------------
The orbifold theory has an untwisted (${\mathcal{H}}^{(0)}$) and two twisted (${\mathcal{H}}^{(1)}$ and ${\mathcal{H}}^{(-1)}$) sectors, with spectrum \^[(n)]{}=\_[i=1]{}\^6\_[l\_i,m\_i+n,s\_i]{}|\_[l\_i,m\_i-n,|s\_i]{} . Invariance under the orbifold symmetry requires \_[i=1]{}\^6 m\_i03 . For $k=1$, we may take $l=0$, with $m\in{\mathbb{Z}}/6{\mathbb{Z}}$ and $s\in{\mathbb{Z}}/4{\mathbb{Z}}$. The only states that contribute to the elliptic genus are the RR states with $\bar h=\frac{1}{4}$. It is easy to see that the condition $\bar h=\tfrac{1}{4}$ together with the $U(1)$-charge integrality condition (that follows from orbifold invariance) is only satisfied if the right-moving ground state is of the form (0,1,1)\^[6]{},(0,-1,-1)\^[6]{} ,(0,1,1)\^[3]{}(0,-1,-1)\^[3]{} . In the last case, all the $20$ different permutations of the factors should be considered. Thus, the RR states for which the right-movers are R ground states are explicitly $$\begin{array}{llr}
n=0 : \qquad & \otimes_i^6 \Phi^0_{1,s_i;1,1} & (1\text{ state}) \\[2pt]
&\otimes_i^6 \Phi^0_{-1,s_i;-1,-1} & (1\text{ state}) \\[2pt]
&\bigl(\otimes_i^3 \Phi^0_{1,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^3 \Phi^0_{-1,s_i;-1,-1}\bigr)
\text{ and permutations} \quad & (20\text{ states}) \\[4pt]
n=1: \qquad &\otimes_i^6 \Phi^0_{3,s_i;1,1},& (1\text{ state}) \\[2pt]
&\otimes_i^6 \Phi^0_{1,s_i;-1,-1} & (1\text{ state}) \\[2pt]
&\bigl(\otimes_i^3 \Phi^0_{3,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^3 \Phi^0_{1,s_i;-1,-1}\bigr)
\text{ and permutations} & (20\text{ states})\\[4pt]
n=-1: \qquad &\otimes_i^6 \Phi^0_{-1,s_i;1,1} & (1\text{ state}) \\[2pt]
&\otimes_i^6 \Phi^0_{3,s_i;-1,-1} & (1\text{ state}) \\[2pt]
&\bigl(\otimes_i^3 \Phi^0_{-1,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^3 \Phi^0_{3,s_i;-1,-1}\bigr)
\text{ and permutations} & (20\text{ states})
\end{array}$$ where $s_i=1,3$. In each case, states with $s_i=1$ and $s_i=3$, $i=1,\ldots,6$, are mapped into one another under the action of the ${\mathcal{N}}=2$ algebra of the $i^{\rm th}$ model. The contribution to the elliptic genus of the ${\mathcal{N}}=2$ representation containing $\otimes_i^6\Phi^0_{m_i,s_i;\bar m_i,\bar s_i}$ is \_i\^6I\^0\_[m\_i]{}(,z)I\^0\_[|m\_i]{}(|,0) , where $I^l_m(\tau,z)$ are the ${\mathcal{N}}=2$ ‘characters’ that are defined in (\[CharacterN2rep\]). Since \[signs\] I\^0\_[-1]{}(|,0)=-1 ,I\^0\_[1]{}(|,0)=1 ,I\^0\_[3]{}(|,0)=0 , we obtain (,z)=2 \_I\^0\_m(,z)\^6 -20\_ I\^0\_m(,z)\^3I\^0\_[m+2]{}(,z)\^3 , which reproduces indeed (\[K3eg\]).
Symmetries and Twining Genera {#symmetries-and-twining-genera}
-----------------------------
Let us first describe the symmetries that preserve the ${\mathcal{N}}=2$ superconformal symmetry. For $i\in\{1,\ldots, 6\}$, we denote by $e_i$ the ${\mathbb{Z}}_3$ symmetry that acts as e\_i(\_[j=1]{}\^6 \^0\_[m\_j,s\_j;|m\_j,|s\_j]{}) =e\^(\_[j=1]{}\^6 \^0\_[m\_j,s\_j;|m\_j,|s\_j]{}) . These symmetries generate a group ${\mathbb{Z}}_3^5$ (because the product of all of them is the orbifold symmetry, under which all states are invariant by construction). There are also the right-moving analogs $\bar e_i$, where the phase depends on $\bar m_i$ instead of $m_i$. In addition we have the quantum symmetry ${\mathcal{Q}}$ which acts by multiplication by $e^{\frac{2\pi i n}{3}}$ on the $n^{\rm th}$ twisted sector ${\mathcal{H}}^{(n)}$. This gives an additional ${\mathbb{Z}}_3$. Note that in the $n^{\rm th}$ twisted sector $\bar m_i=m_i-2n$, and thus |e\_i=e\_i , implying that the transformations ${\mathcal{Q}},e_i,\bar e_i$ form a group ${\mathbb{Z}}_3^6$. Finally, there are permutations of the six factors of $(1)^6$, where we define the action of $\pi\in S_6$ by \[Jpm\] (\_[i=1]{}\^6 \^l\_[m\_i+n,0;m\_i-n,0]{}) =[sgn]{}()\^[n]{} (\_[i=1]{}\^6 \^l\_[m\_[(i)]{}+n,0;m\_[(i)]{}-n,0]{}) , with additional signs for $s_i,\bar s_i\neq 0$, given by the usual bosonic/fermionic statistics.
Thus, the group of symmetries preserving the ${\mathcal{N}}=2$ superconformal symmetry is ${\mathbb{Z}}_3\times {\mathbb{Z}}_3^5.S_6$. The subgroup preserving the ${\mathcal{N}}=4$ superconformal algebra is then generated by the transformations that leave the currents $J^+$ and $J^-$ of the ${\mathcal{N}}=4$ superconformal algebra (\^0\_[-2,2;0,0]{})\^[6]{} \^[(- 1)]{} ,(\^0\_[2,2;0,0]{})\^[6]{} \^[(+1)]{} invariant. This group has the structure ${\mathbb{Z}}_3^4\rtimes A_6$ and is generated by[^5]
[[()]{}]{}
The phase transformations \_[i=1]{}\^6 e\_i\^[n\_i]{} \_[i=1]{}\^6 n\_i03 , where the constraint assures invariance of the currents. Because of the orbifold invariance relation $\prod_{i=1}^6 e_i=1$, they generate the normal subgroup ${\mathbb{Z}}_3^4$.
The even permutations, since the odd permutations act on the states (\[Jpm\]) with a minus sign. The even permutations form the alternating group $A_6$.
The resulting group is the semidirect product ${\mathbb{Z}}_3^4\rtimes A_6$, where we have the obvious action $\pi(\prod_{i=1}^6 e_i^{n_i})=\prod_{i=1}^6 e_{\pi(i)}^{n_i}$ of $A_6$ on the generators of ${\mathbb{Z}}_3^4$. The $(1)^6$ Gepner model is therefore an example of case (iii) of the Theorem.
With this description it is now straightforward to calculate the corresponding twining genera, and for the convenience of the reader we have collected them in Table \[Tab:TwiningGenera16\]. For example, we have for the phase transformation $e_1e_2e_3$ $$\begin{aligned}
\phi_{e_1e_2e_3}(\tau,z)&=2\sum_{m=-1,1,3}I_{m}^6-(2+9e^{2\pi i/3}+9e^{-2\pi i/3})
\sum_{m=-1,1,3}I_m^3I_{m+2}^3\\
&= 2\sum_{m=-1,1,3}I_{m}^6+7\sum_{m=-1,1,3}I_m^3I_{m+2}^3
=-\tfrac{1}{2}\phi_{{\rm 1A}}+\tfrac{3}{2}\phi_{{\rm 3A}} \equiv \hat\phi_{\rm 3a}\ ,\end{aligned}$$ which corresponds to the fourth line of the table. The other twining genera that do
-------------- ---------------------- --------- --------- --------------------------- ---------------------
perm. conditions on phases **$N$** **\#** $\text{Tr}_{\mathbf{24}}$ $\phi_g$
$ijklmn$ $1$ $1$ $24$ $\phi_{\rm 1A}$
$\phi_{\rm 2A}$
$2220$ $\phi_{\rm 3A}$
$ijklmn$ $20$ $-3$ $\hat\phi_{\rm 3a}$
$(ijkl)(mn)$ $7290$ $\phi_{\rm 4B}$
$(ijklm)n$ $11664$ $\phi_{\rm 5A}$
$1620$ $\phi_{\rm 6A}$
$\hat\phi_{\rm 6a}$
$\hat\phi_{\rm 9a}$
$1080$ $\hat\phi_{\rm 9b}$
-------------- ---------------------- --------- --------- --------------------------- ---------------------
: Twining genera of the $(1)^6$ model. The symmetry generators have been labelled by the structure of the permutations $\{i,j,k,l,m,n\}$ of the minimal models and the individual phase shifts $e_i^{n_i}e_j^{n_j}e_k^{n_k}e_l^{n_l}e_m^{n_m}e_n^{n_n}$. The multiplicity (labelled by $\#$) is the number of ‘independent’ generators within each class of symmetries which are not identified through the action of the orbifold. The order of each generator is denoted by $N$ and $\text{Tr}_{24}$ gives the trace over the $24$-dimensional representation. Finally, the twining genera $\hat\phi_{\rm 3a}$, $\hat\phi_{\rm 6a}$ and $\hat\phi_{\rm 9ab}$ are not Mathieu twining genera and are defined in the main body of the text.[]{data-label="Tab:TwiningGenera16"}
not directly agree with ${\mathbb{M}}_{24}$ twining genera — since we are in case (iii) of the Theorem, there is no reason to expect any ${\mathbb{M}}_{24}$ twining genera — are $$\begin{aligned}
\hat\phi_{\rm 6a}(\tau,z)&=\tfrac{1}{2}\left(\phi_{\rm 2A}(\tau,z)+\phi_{\rm 6A}(\tau,z)\right) \\
\hat\phi_{\rm 9a}(\tau,z)& =\tfrac{1}{2}\left(\phi_{\rm 3A}(\tau,z)+\phi_{\rm 3B}(\tau,z)\right) \\
\hat\phi_{\rm 9b}(\tau,z)&=\tfrac{1}{4}\left[\phi_{0,1}(\tau,z)+6\bigl(2\psi^{(3)}(\tau)+3\psi^{(9)}(\tau)
+3E_2^{(9)}(\tau)\bigr)\phi_{-2,1}(\tau,z)\right] \ ,\end{aligned}$$ where $\phi_{0,1}$ and $\phi_{-2,1}$ are the standard Jacobi forms of ${\rm SL}(2,\mathbb{Z})$ of index one and weight 0 and $-2$, respectively, and $\{\psi^{(3)},\psi^{(9)},E_2^{(9)}\}$ form a basis of weight $2$ modular forms of $\Gamma_0(9)$ (for more information and explicit definitions see [@Gaberdiel:2010ca]).
The D-brane Charge Lattice
--------------------------
As before, we can also determine the symmetry group of this model from an analysis of the D-brane charge lattice. The construction of Gepner model D-branes is standard, and is briefly sketched in Appendix \[s:Gepner\]. The tensor product A-type branes with $L_i=S_i=0$ and $M_1,\ldots,M_6\in {\mathbb{Z}}_6$ generate a charge lattice of rank $22$ with signature $(2,20)$. This is as expected, since the A-type tensor product branes only couple to the $22$ RR ground states in the untwisted sector.
The other charges are carried by B-type permutation branes of the type described in (\[16Bper\]). As we vary $M_1,\ldots,M_6,\hat M\in{\mathbb{Z}}_6$, the intersection form of the B-type branes gives a matrix of rank $10$ with signature $(2,8)$. Again, this is what we expect since these permutation branes couple to the $2$ RR ground states in the twisted sectors $n=\pm 1$, and to the $8$ RR ground states in the untwisted sector with $m_1=-m_2$, $m_3=-m_4$, and $m_5=-m_6$.
In order to obtain the full charge lattice we have to combine these two constructions; for example, a set of $22$ A-type D-branes with $L_i=0=S_i$, $i=1,\ldots,6$ and suitable values for $M_1,\ldots,M_6$, and two B-type D-branes with $L_i=M_i=0$, $S_i=0$ and $\hat M=\pm2$ generate the full unimodular lattice $\Gamma^{4,20}$ (see the LaTeX source code for details). Next, we denote that four RR ground states in the $({\bf 2},{\bf 2})$ representation of ${\rm SU}(2)_L\times {\rm SU}(2)_R$ according to their $J_0^3,\tilde{J}_0^3$ charges as $\Phi_{1,\bar 1}$, $\Phi_{1,-\bar 1}$, $\Phi_{-1,\bar 1}$, and $\Phi_{-1,-\bar 1}$, Let us consider the sublattice $(\Gamma^{4,20})^\perp$ of D-branes that are neutral under these four states. The $22$ A-type branes generate the sublattice of D-branes that are neutral with respect to $\Phi_{1,-\bar 1}$ and $\Phi_{-1,\bar 1}$, while their charge with respect to $\Phi_{1,\bar 1}$ and $\Phi_{-1,-\bar 1}$ is given by $$\begin{aligned}
Q_{\Phi_{1,\bar 1}}(\|0,M_i,0{\rangle\!\!\!\;\rangle}_{\rm A})&= \frac{1}{3}e^{\frac{\pi i}{3}\sum_i M_i}\ ,\\
Q_{\Phi_{-1,-\bar 1}}(\|0,M_i,0{\rangle\!\!\!\;\rangle}_{\rm A})&= \frac{1}{3}e^{-\frac{\pi i}{3}\sum_i M_i}\ .\end{aligned}$$ With the redefinition \^a=3 (\_[1,|1]{}+\_[-1,-|1]{}) ,\^b=2 i(\_[1,|1]{}-\_[-1,-|1]{}) , the states $\Phi^a,\Phi^b$ correspond to elements of the dual lattice $(\Gamma^{4,20})^*\cong \Gamma^{4,20}$ so that the sublattice generated by A-type D-branes and orthogonal to these elements has maximal rank, (\^[4,20]{})\^=20 . The discriminant group of this $20$-dimensional lattice is ${\mathbb{Z}}_9\times{\mathbb{Z}}_3^2$, and its discriminant form is the same as the one of the S-lattice $2^93^6$ (see Appendix \[s:proofs\]), whose quadratic form is \[Q2923\] Q\_[2\^[9]{}3\^[6]{}]{}=
4 & 1 & 1 & -2\
1 & 4 & 1 & -2\
1 & 1 & 4 & 1\
-2 & -2 & 1 & 4
. By the general lattice gluing procedure [@Conway], it follows that the S-lattice $2^93^6$ and $(\Gamma^{4,20})^\perp(-1)$ are orthogonal sublattices of a positive definite even unimodular lattice of rank $24$. This procedure also provides explicitly the quadratic form for this unimodular lattice. With the help of some computer algorithm we have shown that the resulting lattice has no vectors of norm $2$, thus proving that it is indeed the Leech lattice $\Lambda$. All sublattices of $\Lambda$ with quadratic form are related to the S-lattice $2^93^6$ by some Leech lattice automorphism, and the pointwise stabiliser of each of them is the group $G={\mathbb{Z}}_3^4\rtimes A_6$ [@Curtis], thus matching the results from the previous subsection.
The $(2)^4$ Model {#s:24}
=================
Our last example is the ‘quartic’ model $(2)^4$ which is constructed by taking a $\mathbb{Z}_4$ orbifold of four tensor powers of the ${\mathcal{N}}=2$ minimal model of level $k=2$. Our notation and conventions are the same as in the previous section and are again summarised in Appendix \[s:Gepner\].
The Spectrum
------------
For the $\mathbb{Z}_4$ orbifold we have in addition to the untwisted sector $\mathcal{H}^{(0)}$ three twisted sectors ($n=1,2,3$), with spectrum \^[(n)]{}=\_[i=1]{}\^4\_[l\_i,m\_i+n,s\_i]{}|\_[l\_i,m\_i-n,|s\_i]{} , where invariance under the $\mathbb{Z}_4$-orbifold enforces $\sum_{i=1}^4 m_i\equiv 0\mod 4$. At $k=2$ there are six ${\mathcal{N}}=2$ R-sector representations, which we may label by $(l=0, m=\pm 1, \pm 3)$ and $(l=1,m=0,2)$. For the elliptic genus we are again only interested in those RR states for which the right-moving states are ground states with $\bar h=\frac{1}{4}$. The relevant coset representations are (compare also [@Brunner:2006tc; @Brunner:2009mn]) $$\begin{aligned}
&\hspace{1.2cm}(0,1,1)^{\otimes 4}\,,\hspace{1.5cm} (1,2,1)^{\otimes 4}\,,\hspace{1.5cm} (0,-1,-1)^{\otimes 4}\,,\\
&(0,1,1)^{\otimes 2}\otimes (0,-1,-1)^{\otimes 2}\,,\hspace{1cm} (0,1,1)\otimes (0,-1,-1)\otimes (1,2,1)^{\otimes 2}\,,\end{aligned}$$ where in the second line all $6$ and $12$ permutations are included, respectively. Explicitly, the RR states that can contribute to the elliptic genus are thus of the form $$\begin{array}{llll}
n=0: & n=1: & n=2: & n=3: \\
\otimes_i^4 \Phi^0_{1,s_i;1,1} \qquad & \otimes_i^4 \Phi^0_{3,s_i;1,1} \qquad &
\otimes_i^4 \Phi^0_{-3,s_i;1,1} \qquad & \otimes_i^4 \Phi^0_{-1,s_i;1,1} \\
\otimes_i^4 \Phi^1_{2,s_i;2,1} \qquad & \otimes_i^4 \Phi^1_{0,s_i;2,1} \qquad &
\otimes_i^4 \Phi^1_{-2,s_i;2,1} \qquad & \otimes_i^4 \Phi^1_{0,s_i;2,1} \\
\otimes_i^4 \Phi^0_{-1,s_i;-1,-1} \qquad & \otimes_i^4 \Phi^0_{1,s_i;-1,-1} \qquad &
\otimes_i^4 \Phi^0_{3,s_i;-1,-1} \qquad & \otimes_i^4 \Phi^0_{-3,s_i;-1,-1} \ .
\end{array}$$ In addition, there are the states (again written in the order $n=0$, $n=1$, $n=2$, and $n=3$) $$\begin{array}{ll}
\bigl(\otimes_i^2 \Phi^0_{1,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^2 \Phi^0_{-1,s_i;-1,-1}\bigr) \qquad &
\bigl(\otimes_i^2 \Phi^0_{3,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^2 \Phi^0_{1,s_i;-1,-1}\bigr) \\
\bigl(\otimes_i^2 \Phi^0_{-3,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^2 \Phi^0_{3,s_i;-1,-1}\bigr) \qquad &
\bigl(\otimes_i^2 \Phi^0_{-1,s_i;1,1}\bigr)\otimes\bigl(\otimes_i^2 \Phi^0_{-3,s_i;-1,-1}\bigr)\ ,
\end{array}$$ where in each case there are 6 different permutations, as well as the states $$\begin{array}{ll}
\bigl(\otimes_i^2 \Phi^1_{2,s_i;2,1}\bigr)\otimes \Phi^0_{-1,s_i;-1,-1}\otimes \Phi^0_{1,s_i;1,1} \quad &
\bigl(\otimes_i^2 \Phi^1_{0,s_i;2,1}\bigr)\otimes \Phi^0_{1,s_i;-1,-1}\otimes \Phi^0_{3,s_i;1,1} \\
\bigl(\otimes_i^2 \Phi^1_{-2,s_i;2,1}\bigr)\otimes \Phi^0_{3,s_i;-1,-1}\otimes \Phi^0_{-3,s_i;1,1} \quad &
\bigl(\otimes_i^2 \Phi^1_{0,s_i;2,1}\bigr)\otimes \Phi^0_{-3,s_i;-1,-1}\otimes \Phi^0_{-1,s_i;1,1} \ ,
\end{array}$$ where now there are 12 different permutations each. Again, the states with $s_i=1,3$ are mapped into one another under the action of the ${\mathcal{N}}=2$ algebra of the $i^{\rm th}$ model. Since I\^0\_[1]{}(|,0)=1 , I\^0\_[3]{}(|,0)=0 , I\^1\_[0]{}(|,0)=0 , I\^1\_[2]{}(|,0) =1 the total contribution to the elliptic genus is then = \_ , which agrees indeed with (\[K3eg\]).
Symmetries and Twining Genera {#symmetries-and-twining-genera-1}
-----------------------------
In the quartic $(2)^4$ model the currents $J^\pm$ of the left- and right moving $\mathcal{N}=4$ superconformal algebra are given by $$J^\pm= (0,\pm2,2)^{\otimes4}\otimes \overline{(0,0,0)}^{\otimes 4} \ , \qquad \hbox{and} \qquad
\bar{J}^\pm=(0,0,0)^{\otimes4}\otimes\overline{(0,\pm2,2)}^{\otimes 4}\ .\label{N4current24}$$ The symmetries that leave these currents invariant are
[[()]{}]{}
Phase shifts, which are generated by \_[i=1]{}\^[4]{} e\_i\^[a\_i]{} ( \^2 (-1)\^[F\_s]{} )\^ , A = \_[i=1]{}\^[4]{} a\_i A 0 2 . Here each $e_{i}^{a_i}$ acts as e\_[i]{}\^[a\_i]{}:\^\_[m\_i,s\_i;|[m]{}\_i,|[s]{}\_i]{} e\^\^\_[m\_i,s\_i;|[m]{}\_i,|[s]{}\_i]{} , ${\mathcal{Q}}$ is the quantum symmetry of the Gepner orbifold (that acts as a phase $e^{\frac{\pi i n}{2}}$ on the states of the $n^{\text{th}}$ twisted sector), and $(-1)^{F_s}$ is the left-moving spacetime fermion number operator that acts as $+1$ ($-1$) on the left-moving NS (R) sector. (The inclusion of $(-1)^{F_s}$ is required in order to preserve the spectral flow operators.) Taking into account the overall $\mathbb{Z}_4$ invariance coming from the Gepner orbifold, these phase shifts generate the group $\mathbb{Z}_4^2\times {\mathbb{Z}}_2$.
The permutations \[Jpm4\] (\_[i=1]{}\^4 \^l\_[m\_i+n,0;m\_i-n,0]{})=[sgn]{}()\^[n]{} \^[ [sgn]{}()-1]{} (\_[i=1]{}\^4 \^l\_[m\_[(i)]{}+n,0;m\_[(i)]{}-n,0]{}) . They generate the symmetric group $S_4$.
These symmetries generate the group $({\mathbb{Z}}_2\times {\mathbb{Z}}_4^2)\rtimes S_4$, thus realising case (i) of the Theorem with $G'={\mathbb{Z}}_2^3$ and $G''={\mathbb{Z}}_2^2.S_4$. Here $G'\subset {\mathbb{Z}}_2^{11}$ is generated by the phases $e_1^2e_2^2$, $e_2^2e_3^2$ and $e_3^2{\mathcal{Q}}^2(-1)^{F_s}$, while $G''\subset {\mathbb{M}}_{24}$ is generated by the permutations, giving the $S_4$ factor, as well as by the phases $e_1e_2{\mathcal{Q}}^2(-1)^{F_s}$ and $e_2^3e_3^3{\mathcal{Q}}^2(-1)^{F_s}$, giving the ${\mathbb{Z}}_2^2$ factor.[^6]
We can also calculate the associated twining genera, and our results are collected in Table \[Tab:TwiningGenera\]. Again, we see that some of the generators in $G'$ lead to the twining genus \_[4a]{} =-(\_[1A]{}-\_[2A]{}-2\_[4B]{}) , that does not coincide with any twining genus of ${\mathbb{M}}_{24}$.
------------ ---------------------- ------- ------- --------- -------- --------------------------- -------------------
perm. conditions on phases $n_Q$ $n_F$ **$N$** **\#** $\text{Tr}_{\mathbf{24}}$ $\phi_g$
$ijkl$ $0$ $0$ $1$ $1$ $24$ $\phi_{\rm 1A}$
$79$ $\phi_{\rm 2A}$
$(ij)(kl)$ $2$ $1$ $2$ $24$ $0$ $\phi_{\rm 2B}$
$(ijk)l$ $0$ $0$ $3$ $128$ $6$ $\phi_{\rm 3A}$
$ijkl$ $2$ $1$ $4$ $6$ $-4$ $\hat{\phi}_{4a}$
$306$ $\phi_{\rm 4B}$
$(ijk)l$ $2$ $1$ $6$ $128$ $2$ $\phi_{\rm 6A}$
$(ij)kl$ $2$ $0$ $8$ $96$ $2$ $\phi_{\rm 8A}$
------------ ---------------------- ------- ------- --------- -------- --------------------------- -------------------
: The twining genera of the $(2)^4$ model. Here the symmetries have been labelled by the structure of the permutations of $\{i,j,k,l\}$, the phase shifts of the individual symmetries $e_i^{n_i}e_j^{n_j}e_k^{n_k}e_l^{n_l}$, the power of the operator $Q^{n_Q}$ and the spacetime fermion number $(-1)^{n_F F_s}$. The remaining part of the notation is the same as for Table 1.[]{data-label="Tab:TwiningGenera"}
The D-brane Charge Lattice
--------------------------
The derivation of a set of D-branes generating the lattice of RR charges is analogous to the construction for the $(1)^6$ model. The A-type tensor product branes (see Appendix \[s:Dbranes\]) are now only charged under the $21$ RR ground states in the untwisted sector. The remaining charges can be accounted for in terms of B-type permutation branes. Taking $21$ A-type branes with $L_i=S_i=0$ and suitable combinations for $M_i$, as well as $3$ B-type permutation branes with $L_i=M_i=S_i=0$ and suitable values of $\hat{M}$ leads indeed to the full charge lattice $\Gamma^{4,20}$, [*i.e.*]{} the resulting intersection form has determinant one.[^7]
The sublattice $(\Gamma^{4,20})^\perp$ of D-branes that are neutral with respect to the RR ground states in the $({\bf 2},{\bf 2})$ representation of ${\rm SU}(2)_L\times {\rm SU}(2)_R$ has maximal rank $20$. Upon changing the sign of its quadratic form, it can be embedded into the Leech lattice $\Lambda$, and its orthogonal complement $\Lambda^G$ has quadratic form
6 & 2 & 0 & 0\
2 & 4 & -2 & 4\
0 & -2 & 6 & -4\
0 & 4 & -4 & 8
. Since $\Lambda^G$ contains a vector of norm $8$, its point-wise stabiliser must be a subgroup of ${\mathbb{Z}}_2^{12}\rtimes{\mathbb{M}}_{24}$ (see Appendix \[s:proofs\]). More precisely, we have shown that the stabiliser turns out to be isomorphic to $G=({\mathbb{Z}}_2\times{\mathbb{Z}}_4^2)\rtimes S_4$, which is the group of symmetries we have found in the previous subsection.
Conclusions
===========
In this paper, we have shown that the symmetries of a non-linear $\sigma$-model on K3 that preserve the ${\mathcal{N}}=(4,4)$-superconformal algebra as well as the spectral flow operators, form a subgroup of the Conway group $Co_1$. This provides a stringy analogue of the Mukai theorem in algebraic geometry that shows that the symplectic automorphisms of any K3 form a subgroup of the Mathieu group ${\mathbb{M}}_{23}$. The specific subgroups that can actually arise in our case are spelled out in the Theorem stated in the Introduction.
Our result is somewhat unexpected in view of the recent observation of [@EOT], relating the elliptic genus of K3 to the Mathieu group ${\mathbb{M}}_{24}$. In particular, it follows from our Theorem (as well as the explicit examples) that the symmetries of a given K3 model are not, in general, subgroups of ${\mathbb{M}}_{24}$.[^8] As a consequence, their twining genera do not, in general, agree with those appearing in the context of Mathieu moonshine [@Cheng:2010pq; @Gaberdiel:2010ch; @Gaberdiel:2010ca; @Eguchi:2010fg], and we have seen explicit examples of this. In particular, this therefore means that the naive idea that ${\mathbb{M}}_{24}$ arises as the ‘union’ of all symmetries from different points in moduli space needs to be refined.
At least on the face of it, our Theorem seems to suggest that the elliptic genus of K3 could exhibit some sort of moonshine based on $Co_1$ or $Co_2$, but we have seen no evidence of this since the dimensions of their representations do not match the coefficients of the elliptic genus. It is intriguing that a connection between $Co_1$ and the BKM algebras arising in ${\mathbb{T}}^6$-compactifications of the heterotic string has recently been observed in [@Govindarajan:2011mp]; given that the heterotic string on ${\mathbb{T}}^6$ is dual to type IIA on $K3\times {\mathbb{T}}^2$ this could be related to our findings.
Our analysis also provides useful tools for the general understanding of non-linear $\sigma$-models on K3. For example, our Theorem suggests the existence of models with some large symmetry groups, and gives precise predictions for their lattice of D-brane charges. These predictions were nicely verified in the three examples we considered. In particular, we showed that the $(1)^6$ Gepner model realises case (iii) of the Theorem. Some preliminary investigations suggest that the groups described in case (iv) might be realised in terms of ${\mathbb{T}}^4/{\mathbb{Z}}_3$ torus orbifolds for different choices of metric and B-field, while it is more difficult to guess which model realises case (ii). For case (i), the Theorem predicts the existence of a model with symmetry group ${\mathbb{Z}}_2^8\rtimes{\mathbb{M}}_{20}$, that might correspond to a certain ${\mathbb{T}}^4/{\mathbb{Z}}_2$ orbifold.
On more general grounds, the sublattice of D-branes that are neutral under the RR ground states in the $({\bf 2},{\bf 2})$ representation of SU$(2)_L\times {\rm SU}(2)_R$ is, in a certain sense, the stringy analogue of the Picard lattice in algebraic geometry. Since the groups of symmetries have a genuine action on this sublattice, it would be interesting to understand for which models this lattice has maximal rank $20$. For example, one can show that for a Gepner model of type $(k_1)\cdots (k_r)$, a necessary condition for this to happen is that the greatest common divisor $\gcd(k_1+2,k_2+2,\ldots)$ of their shifted levels is $3$, $4$ or $6$.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Michael Douglas, Martin Fluder, Daniel Persson and Yuji Tachikawa for useful discussions and correspondences. The research of MRG is supported by the Swiss National Science Foundation.
Notation and Mathematical Background {#s:notation}
====================================
Group Theory
------------
Let us give a brief summary of our conventions regarding finite groups.
--------------------------- --
$A\times B$
\[10pt\] $N\rtimes H$
\[10pt\]
\[20pt\] ${\mathbb{Z}}_n$
\[10pt\]
\[40pt\]
--------------------------- --
---------------- --
$S_n$
\[10pt\] $A_n$
\[10pt\]
\[35pt\]
\[20pt\]
\[20pt\]
\[15pt\]
---------------- --
Lattices {#s:lattices}
--------
For any lattice $L$, we denote by $L^*$ its dual lattice and by $L(n)$, $n\in{\mathbb{R}}$, the lattice obtained from $L$ by multiplying the quadratic form by $n$. If $L$ is integral, then $L\subseteq L^*$, and the finite abelian group $A_L=L^*/L$ is called the *discriminant group*. We denote by $l(L)$ the minimal number of generators of $A_L$ (notice that $l(L)\le \operatorname{rk}L$, with $\operatorname{rk}L$ the rank of the lattice ).
Let $A_L$ be the discriminant group of an integral lattice $L$, and $q_L$ the associated *discriminant quadratic form*, [*i.e.*]{} the form q\_L:A\_L/2induced by the quadratic form on $L$. More generally, we denote by $A_q$ a finite abelian group with a quadratic form $q:A_q\to {\mathbb{Q}}/2{\mathbb{Z}}$. The quadratic form $q$ determines a bilinear form on $A_q$ which takes values in ${\mathbb{Q}}/{\mathbb{Z}}$; we denote it by $a\cdot b$, where $a,b\in A_q$.
If $L$ is a sublattice of a unimodular lattice $\Gamma$ and $L^\perp$ is its orthogonal complement in $\Gamma$, then there is an isomorphism between the discriminant groups \[A.2\] :A\_L A\_[L\^]{} , that flips the sign of the quadratic form \[A.3\] q\_L=-q\_[L\^]{} . More precisely, $\bar x\cong \bar y$, with $\bar x\in A_L$ and $\bar y\in A_{L^\perp}$, if and only if $x+y\in\Gamma$ for any choice $x\in L^*$, $y\in(L^\perp)^*$ of representatives of $\bar x,\bar y$. Conversely, given two even lattices $L_1$, $L_2$ with isomorphic discriminant groups $\gamma:A_{L_1}\stackrel{\cong}{\to} A_{L_2}$ and opposite discriminant quadratic forms $q_{L_1}=-q_{L_2}\circ\gamma$, one can construct an even unimodular lattice $\Gamma$ by ‘gluing’ $L_1$ and $L_2$, [*i.e.*]{} ={xyL\_1\^\*L\_2\^\*|x|y} , where $\bar x,\bar y$ are the images in $A_{L_1}$, $A_{L_2}$ of $x\in L_1^*$ and $y\in L_2^*$, respectively.
A sublattice $L'$ of a lattice $L$ is called *primitive* if $L/L'$ is a free group; in other words, $L'=(L'\otimes {\mathbb{Q}})\cap L$. Correspondingly, a primitive embedding of a lattice $L'$ in $L$ is an embedding such that the image is primitive.
The Leech Lattice and the Golay Code
------------------------------------
The Leech lattice can be defined in terms of the binary Golay code ${\mathcal{C}}_{24}$, a $12$-dimensional subspace of the vector space ${\mathbb{F}}_2^{24}$, where ${\mathbb{F}}_2=\{0,1\}$ is the field with two elements. To each vector $f=(f_1,\ldots,f_{24})$ of ${\mathcal{C}}_{24}$ ([*i.e.*]{} to each codeword) we associate a subset $X_f$ of $\Omega=\{1,\ldots,24\}$, corresponding to the non-zero coordinates of $f$, [*i.e.*]{} $X_f=\{i\in\Omega\mid f_i\neq 0\}$. This collection of $2^{12}=4096$ subsets of $\Omega$ (called ${\mathcal{C}}$-sets) includes the empty set, $\Omega$ itself, 759 ${\mathcal{C}}$-sets with $8$ elements (special octads), $2576$ with $12$ elements and 759 with $16$ elements (the complements in $\Omega$ of the special octads). Furthermore, for any choice of $5$ distinct elements in $\Omega$ there is a unique special octad containing them. These properties are sufficient to determine the collection of ${\mathcal{C}}$-sets, and thus the Golay code, up to permutations of the objects in $\Omega$. The Mathieu group ${\mathbb{M}}_{24}$ can be defined as the group of automorphisms of the Golay code, or, equivalently, as the subgroup of $S_{24}$ stabilising the collection of ${\mathcal{C}}$-sets.
An explicit description of the Leech lattice $\Lambda\subset{\mathbb{R}}^{24}$ can be given as follows: the vector $v=\frac{1}{\sqrt{8}}(v_1,\ldots,v_{24})$ is an element of the Leech lattice provided that
- the $v_i$, $i=1,\ldots,24$, are all integers of the same parity;
- $\sum_{i=1}^{24}v_i\equiv 0$ or $4\mod 8$ according to $v_i\equiv 0$ or $1\mod 2$, respectively; and
- for each $\nu\in\{0,1,2,3\}$, the set $\{i\in \Omega\mid v_i\equiv \nu\mod 4\}$ is a ${\mathcal{C}}$-set.
Proof of the Theorem {#s:latticeproofs}
====================
In this appendix, we give the remaining details of the arguments of Section \[symmetries\], leading to the proof of the Theorem.
$G$ as a Subgroup of $O(\Gamma^{25,1})$ {#s:GinOGamma}
---------------------------------------
For each lattice $L$ and group $G\subset {\rm Aut}(L)$, we define as in (\[LuG\]) the $G$-invariant lattice $L^G$ as $L^G=\{x\in L\mid g(x)=x,\forall g\in G\}$. Furthermore, $L_G$ is its orthogonal complement in $L$, $L_G=\{x\in L\mid x\cdot y=0,\forall y\in L^G\}$. We take $L=\Gamma^{4,20}\subset{\mathbb{R}}^{4,20}$, and consider the case where $G$ is the subgroup of ${\rm Aut}(\Gamma^{4,20})\subset O(4,20,{\mathbb{R}})$, fixing a positive-definite four plane $\Pi\subset{\mathbb{R}}^{4,20}$.
\[p:lemma\] For any choice of a positive 4-plane $\Pi$, $L_G$ is a negative definite lattice of rank $\operatorname{rk}L_G\le 20$. $G$ acts trivially on $A_{L_G}$, and $l(L_G)\le 24-\operatorname{rk}(L_G)$.
The first part is obvious. By definition $G$ acts trivially on $L^G$, and hence its induced action on $A_{L^G}$ is also trivial. Since $L^G$ and $L_G$ are orthogonal primitive sublattices of the self-dual lattice $L$, it follows that for each $y\in (L_G)^*$ there exists a vector $v=x+y\in L$ with $x\in (L^G)^*$. For all $w\in L^G$ and any lattice automorphism $g\in G$, $w\cdot g(v)= g(w)\cdot g(v) = w\cdot v$, so that $g(v)-v\in(L^G)^\perp=L_G$. Since $g$ is linear and fixes $x\in(L^G)^*$, we have $g(v)-v=g(x+y)-(x+y)=g(y)-y$. It follows that $g(y)\equiv y\mod L_G$, so that $G$ acts trivially on $A_{L_G}=(L_G)^*/L_G$. Finally, $l(L_G)=l(L^G)\le \operatorname{rk}(L^G)$, and the last statement follows.
Up to isomorphism, there is a unique even unimodular lattice $\Gamma^{25,1}$ of signature $(25,1)$. The lattice $\Gamma^{25,1}$ can be defined as the (additive) subgroup of ${\mathbb{R}}^{25,1}$ with elements $(x_0,\ldots,x_{24};x_{25})$ such that x\_0+…+x\_[24]{}-x\_[25]{}2 , where either $x_i\in{\mathbb{Z}}$ for all $i$, or $x_i\in{\mathbb{Z}}+\frac{1}{2}$ for all $i$. In the rest of this subsection, we will prove that $L_G(-1)$ can be embedded into $\Gamma^{25,1}$.
Recall that, for any prime $p$, a $p$-group is a group whose order is a power of $p$. A Sylow $p$-subgroup of a group $G$ is a maximal $p$-subgroup, [*i.e.*]{} a $p$-subgroup of $G$ which is not a proper subgroup of any other $p$-subgroup. For abelian groups, there is a unique Sylow $p$-subgroup for each prime $p$, the subgroup of elements whose order is a power of $p$. For more general finite groups, for each given $p$ the Sylow $p$-subgroups are all isomorphic and related by conjugation.
Let $A_q$ be a finite abelian group with quadratic form $q:A_q\to {\mathbb{Q}}/2{\mathbb{Z}}$. For any prime $p$, let $A_{q_p}$ be the Sylow $p$-subgroup of $A_q$, and $q_p$ the restriction of $q$ to $A_{q_p}$. Note that if $a\in A_{q_p}$, then for any $b\in A_q$ p\^n (ab)0 , where $p^n$ is the order of $a$, and $a\cdot b$ is the bilinear form induced by $q$. In particular, if $b\in A_{q_{p'}}$, with $p'\neq p$, this implies $a\cdot b\equiv 0\mod {\mathbb{Z}}$. Thus, we have an orthogonal decomposition of the quadratic form $q=\oplus_pq_p$, where each $q_p$ is a quadratic form on an abelian $p$-group.
It follows from Theorem 1.12.2 of [@Nikulin] that an even lattice $L$ of signature $(t^+,t^-)$ and discriminant group $A_q$ can be primitively embedded into some even unimodular lattice of signature $(d^+,d^-)$, provided that
\[embedcond\]$$\begin{aligned}
&d^+-d^-\equiv 0\mod 8\label{embedcond1}\\
&d^--t^-\ge 0,\quad d^+-t^+\ge 0,\quad d^++d^--t^--t^+\ge l(A_q),\label{embedcond2}\\
&d^-+d^+-t^--t^+> l(A_{q_p})\quad \text{ for all odd primes $p$},\label{embedcond3}\\
&d^-+d^+-t^--t^+> l(A_{q_2})\quad\text{or}\quad q_2=q_\theta^{(2)}(2)\oplus q_2'\
\text{ for some }q_2' \ ,\label{embedcond4} \end{aligned}$$
where $q_\theta^{(2)}(2)$ is the discriminant quadratic form of the $\mathfrak{su}(2)$ root lattice $A_1\cong{\mathbb{Z}}(2)$.
\[th:embed\] The lattice $L_G(-1)$ can be primitively embedded into $\Gamma^{25,1}$. The action of $G$ can be extended to an action on $\Gamma^{25,1}$ such that the $G$-invariant sublattice $(\Gamma^{25,1})^G$ is the orthogonal complement of $L_G(-1)$ in $\Gamma^{25,1}$, and such that $(\Gamma^{25,1})^G$ contains an element of norm $2$.
Let us prove that we can embed $L_G(-1)\oplus A_1$ into $\Gamma^{25,1}$, where $A_1$ denotes the root lattice of the $\mathfrak{su}(2)$ Lie algebra. Condition obviously holds. By proposition \[p:lemma\], we have $\operatorname{rk}(L_G(-1)\oplus A_1)+l(L_G(-1)\oplus A_1)\le 26$, so that also is satisfied. Let $A_q$ be the discriminant group of $L_G(-1)\oplus A_1$, with discriminant quadratic form $q$. Let us consider the decomposition $q=\oplus_{p}q_p$ where, for each prime $p$, $q_p$ is the restriction of $q$ to the $p$-Sylow subgroup $A_{q_p}$. Since the discriminant group of $A_1$ is ${\mathbb{Z}}/2{\mathbb{Z}}$, we have $l(A_{q_p})\le l(L_G)<26-\operatorname{rk}(L_G\oplus A_1)$ for all odd $p$, and holds. Finally, it is clear that $q_2=q_\theta^{(2)}(2)\oplus q_2'$, where $q_\theta^{(2)}(2)$ is the discriminant form of $A_1$ and $q_2'$ is the restriction of $q_{L_G}$ to the $2$-Sylow subgroup. Thus, it follows from that $L_G(-1)\oplus A_1$ can be primitively embedded into $\Gamma^{25,1}$.
Since $G$ acts trivially on $A_{L_G}$, the action of $G$ on $L_G(-1)$ can be extended to an action on $\Gamma^{25,1}$ which acts trivially on the orthogonal complement of $L_G(-1)$ in $\Gamma^{25,1}$. Thus, $(\Gamma^{25,1})^G\cong (L_G(-1))^\perp$, and $A_1$ is a sublattice of $(\Gamma^{25,1})^G$, so that $(\Gamma^{25,1})^G$ contains a vector of norm $2$.
$G$ as a Subgroup of $Co_0$ {#s:proofConway}
---------------------------
The automorphism group ${\rm Aut}(\Gamma^{25,1})$ of $\Gamma^{25,1}$ is generated by the sign flip $x_{i}\mapsto -x_{i}$, together with the subgroup of autochronous transformations ${\rm Aut}^+(\Gamma^{25,1})$ which stabilise the cone of positive time vectors in ${\mathbb{R}}^{25,1}$ (see chapter 27 of [@Conway]). The group ${\rm Aut}^+(\Gamma^{25,1})$ contains a normal subgroup $W$ (the Weyl group), generated by the reflections $R_r$ with respect to the hyperplanes $r^\perp$ R\_r(x)=x-(xr)r ,x\^[25,1]{} , where $r$ is any root in $r\in \Gamma^{25,1}$, [*i.e.*]{} satisfies $r\cdot r=2$. The complement in ${\mathbb{R}}^{25,1}$ of the union $\bigcup_{r\cdot r=2}r^\perp$ of the corresponding hyperplanes has infinitely many connected components, and the closure of each component is called a Weyl chamber.
The group of autochronous transformations is the semidirect product $W\rtimes Co_{\infty}$, where $Co_\infty$ is the automorphism group of the Dynkin diagram of $W$. The groups $W$ and $Co_{\infty}$ can be described more explicitly upon choosing a set of generators for $W$, [*i.e.*]{} a set of fundamental roots. One convenient choice is given by the set of Leech roots, [*i.e.*]{} by the vectors $r\in \Gamma^{25,1}$ with rr=2rw=-1 , where $w$ is the null (Weyl) vector w=(0,1,2,3,…,23,24;70)\^[25,1]{} . The sublattice $\Gamma^{25,1}\cap w^\perp$ is degenerate, while its quotient $(\Gamma^{25,1}\cap w^\perp)/w$ is the Leech lattice $\Lambda$ (see chapter 26 of [@Conway]), the unique positive even unimodular lattice of rank $24$ containing no roots. The automorphism group $Co_\infty$ of the Dynkin diagram of $W$ contains ${\rm Aut}(\Lambda)=Co_0$, the automorphism group of the Leech lattice, as the subgroup which fixes a given reference Leech root $\bar r$. The group $Co_{\infty}$ is generated by $Co_0$, together with the translations of the Leech roots by vectors in $\Lambda$.
Let us consider the embedding of $L_G(-1)\oplus A_1$ into $\Gamma^{25,1}$. Clearly, the sign flip does not fix any linear sublattice of $\Gamma^{25,1}$, so that $G\subseteq {\rm Aut}^+(\Gamma^{25,1})$. Note that the lattice $(\Gamma^{25,1})^G$ always contains a vector in the interior of some Weyl chamber. For if this was not true, then $(\Gamma^{25,1})^G$ would be contained in one of the hyperplanes orthogonal to some root $r$, and thus $r\in (\Gamma^{25,1})_G=L_G(-1)$. But this would contradict our assumption that $L_G$ contains no vector of norm $-2$. Since $W$ acts transitively on the Weyl chambers, we can choose our embedding of $L_G(-1)$ into $\Gamma^{25,1}$ such that $(\Gamma^{25,1})^G$ contains a vector in the interior $K$ of the fundamental Weyl chamber containing $w$. Since $K\cap t(K)=\emptyset$, for all non-trivial $t\in W$, it follows that $G$ must be contained in $Co_\infty$. Since $w$ is fixed by $Co_\infty$, we have $(\Gamma^{25,1})_G\subset (\Gamma^{25,1}\cap w^\perp)$, and the projection $(\Gamma^{25,1}\cap w^\perp)\to (\Gamma^{25,1}\cap w^\perp)/w\cong\Lambda$ induces an embedding of $(\Gamma^{25,1})_G$ into the Leech lattice $\Lambda$ L\_G(-1)(\^[25,1]{})\_G(\^[25,1]{}w\^) (\^[25,1]{}w\^)/w . As in proposition \[th:embed\], the action of $G$ on $(\Gamma^{25,1})_G$ can be extended to an action on $\Lambda$, such that $L_G(-1)\cong(\Gamma^{25,1})_G\cong\Lambda_G$ is the orthogonal complement of the sublattice $\Lambda^G\subset \Lambda$ fixed by $G$. We conclude that $G$ is a subgroup of ${\rm Aut}(\Lambda)\cong Co_0$ fixing a sublattice $\Lambda^G$ of rank at least $4$, thus proving the Proposition in Section \[symmetries\].
The Proof of the Theorem {#s:proofs}
------------------------
The stabilisers of sublattices of the Leech lattice have been classified [@Curtis; @Atlas]. We will use this classification to prove now the Theorem stated in the Introduction. The action of ${\rm Aut}(\Lambda)\cong Co_0$ is well defined on the classes of the quotient $\Lambda/2\Lambda$, because $2\Lambda$ is stable under lattice automorphisms. In particular, if $v\in \Lambda$ is fixed by the action of $G$, then $G$ must be in the stabiliser of the class of $\Lambda/2\Lambda$ containing $v$. Note that opposite vectors $x, -x\in\Lambda$ are contained in the same class in this quotient. More generally, if we define the *short vectors* $x\in\Lambda$ to be the vectors of norm $x^2\le 8$, then for each non-trivial class in $\Lambda/2\Lambda$, one of the following mutually exclusive alternatives holds [@Conway]:
1. the class contains exactly one pair of short vectors $\pm x\in\Lambda$ of type 2 ($x^2=4$);
2. the class contains exactly one pair of short vectors $\pm x\in\Lambda$ of type 3 ($x^2=6$);
3. the class contains exactly $24$ pairs of short vectors $\pm x_1,\ldots,\pm x_{24}\in\Lambda$ of type $4$, that are mutually orthogonal ($x_i\cdot x_j=8\delta_{ij}$).
Thus, each primitive vector $v\in \Lambda^G$ is congruent modulo $2\Lambda$ to some short vector $v_s\in\Lambda$, $v_s^2\le 8$. We can now distinguish the following cases:
[**Case 1:**]{} Suppose there is a vector $v\in\Lambda^G$ which is congruent modulo $2\Lambda$ to a short vector of norm $8$. Then the class of $v$ in $\Lambda/2\Lambda$ contains $24$ mutually orthogonal pairs $\pm x_1,\ldots,\pm x_{24}\in\Lambda$ of norm $8$. The subgroup of $Co_0$ stabilising such a class is the semidirect product $N= {\mathbb{Z}}_2^{12}\rtimes {\mathbb{M}}_{24}$. Here, ${\mathbb{M}}_{24}$ acts by permutations of the $24$ pairs $\pm x_1,\ldots,\pm x_{24}$, while each element $\epsilon_f\in {\mathbb{Z}}_2^{12}$ is associated to a codeword $f\equiv (f_1,\ldots,f_{24})\in ({\mathbb{Z}}/2{\mathbb{Z}})^{24}$ in the binary Golay code ([@Conway], chapter 11) and acts by \[Golayaction\] \_f(x\_i)= (-1)\^[f\_i]{}x\_i ,i=1,…,24 , on the vectors of the class. Thus, $G$ is a subgroup of ${\mathbb{Z}}_2^{12}\rtimes {\mathbb{M}}_{24}$ that fixes a subspace of dimension at least $4$. This realises case (i) of the Theorem. In particular, $G''\subset{\mathbb{M}}_{24}$ can be any subgroup with at least four orbits $\Omega_1,\ldots, \Omega_4$ when acting on $\{1,\ldots,24\}$, and $G'$ is generated by the $\epsilon_f\in{\mathbb{Z}}_2^{12}$ such that $f_i=0$ for all $i\in\Omega_1\sqcup\ldots \sqcup\Omega_4$.
[**Case 2:**]{} If Case 1 does not apply, then each primitive vector $v\in\Lambda^G$ is congruent modulo $2\Lambda$ to a pair of short vectors $\pm v_s\in\Lambda$ with $v_s^2\le 6$. Let us assume that, for each $v\in \Lambda^G$, the corresponding short vector $v_s$ is also contained in $\Lambda^G$. Since $\Lambda^G$ is primitive, $(v-v_s)/2\in\Lambda$ is also contained in $\Lambda^G$. The sublattices $S\subset \Lambda$ containing only short vectors of norm $4$ and $6$ and such that each primitive vector is congruent to a short one mod $2S$ are called ${\mathcal{S}}$-lattices, and they have been completely classified [@Curtis]. In particular, up to automorphisms, there are only three ${\mathcal{S}}$-lattices of rank at least $4$ [@Atlas]: $$\begin{array}{cccc}
S \qquad & \operatorname{rk}S \qquad & {\rm Stab}(S) \qquad & {\rm Aut}(S)\\
2^93^6 \qquad & 4 \qquad & {\mathbb{Z}}_3^4\rtimes A_6\qquad & {\mathbb{Z}}_2\times(S_3\times S_3).{\mathbb{Z}}_2\\
2^{5}3^{10} \qquad & 4 \qquad & 5^{1+2}.{\mathbb{Z}}_4 \qquad & {\mathbb{Z}}_2\times S_5\\
2^{27}3^{36} \qquad & 6 \qquad & 3^{1+4}.{\mathbb{Z}}_2 \qquad & {\mathbb{Z}}_2\times U_4(2).{\mathbb{Z}}_2 \ .
\end{array}$$ Since $\Lambda^G$ is a ${\mathcal{S}}$-lattice of rank at least $4$, $G$ must be one of the groups ${\rm Stab}(S)$ from above, corresponding to the cases (ii), (iii) and (iv) (with $G''$ trivial) of the Theorem.
[**Case 3:**]{} The last case arises if each primitive vector $v\in\Lambda^G$ is congruent modulo $2\Lambda$ to a pair of short vectors $\pm v_s\in\Lambda$ with $v_s^2\le 6$, but some of these short vectors are not contained in $\Lambda^G$. In this case, we define a finite chain of sublattices of $\Lambda$ \[latticechain\] \^G=S\_0S\_1…S\_N=S , where each $S_{i+1}$ is defined in terms of $S_i$ as follows [@Allcock]:
1. If $S_i$ is contained in an ${\mathcal{S}}$-lattice of the same rank or if $S_i$ contains a vector congruent modulo $2\Lambda$ to a short vector of norm $8$, then we set $S=S_i$ and the procedure stops.
2. Otherwise, if $S_i$ contains a vector $v\in 2\Lambda$ with $v/2\notin S_i$, then $S_{i+1}$ is obtained by adjoining $v/2$ to $S_i$. There are only a finite number of vectors to check, namely one representative for each class in $S_i/2S_i$. Note that $S_i\otimes{\mathbb{Q}}=S_{i+1}\otimes{\mathbb{Q}}$, so that $\operatorname{rk}S_i=\operatorname{rk}S_{i+1}$ and ${\rm Stab}(S_i)={\rm Stab}(S_{i+1})$.
If neither of these cases applies, then there is a nonempty set of short vectors $v_s\notin S_i\otimes{\mathbb{Q}}$, with $0< v_s^2\le 6$, congruent modulo $2\Lambda$ to some $v\in S_i$. Then $S_{i+1}$ is defined as follows:
3. If one of the short vectors is such that $v_s\cdot w\neq 0$ for some $w\in S_{i}$, then $S_{i+1}$ is obtained by adjoining $v_s$ to $S_i$. Any element $g\in {\rm Stab}(S_i)$ must preserve the class of $v_s$ in $\Lambda/2\Lambda$, so that $g(v_s)\in\{\pm v_s\}$, and also the product $v_s\cdot w\neq 0$. Thus, the only possibility is $g(v_s)=v_s$, so that ${\rm Stab}(S_{i+1})={\rm Stab}(S_i)$. Furthermore, we have a strong inequality $\operatorname{rk}S_{i+1}>\operatorname{rk}S_i$.
4. If all the short vectors $v_s$ are orthogonal to $S_i$, we choose one of them and define $S_{i+1}$ as the ${\mathbb{Z}}$-linear span of $S_i$ and $v_s$. The only non-trivial action of an element $g\in {\rm Stab}(S_i)$ on $S_{i+1}$ is $g(v_s)=-v_s$, so that $|{\rm Stab}(S_i):{\rm Stab}(S_{i+1})|\le 2$. Furthermore, $\operatorname{rk}S_{i+1}>\operatorname{rk}S_i$.
The stabiliser ${\rm Stab}(S)$ of the lattice $S$ at the end of the chain must be a subgroup of ${\mathbb{Z}}_2^{12}\rtimes{\mathbb{M}}_{24}$ or one of the stabilisers of the ${\mathcal{S}}$-lattices above. However, if the case (4) of the above procedure occurs for some intermediate $S_i$, ${\rm Stab}(S)$ might be just a normal subgroup of $G$. Our analysis is greatly simplified by the following result, which is an immediate consequence of Lemma 4.8 of [@Allcock].
If $\operatorname{rk}S_i>3$ and none of the cases (1), (2) and (3) applies, then $S_i$ is a sublattice of the ${\mathcal{S}}$-lattice $2^{27}3^{36}$.
It is easy to see that if $S_i$ is a sublattice of an ${\mathcal{S}}$-lattice also $S_{i+1}$ is. Since the starting point $S_0=\Lambda^G$ of the chain has already rank greater than $3$, this proposition implies that if $\Lambda^G$ is not a sublattice of $2^{27}3^{36}$, then case (4) above never occurs. Thus, in this case, $G\cong {\rm Stab}(S)$ and we reobtain the groups (i), (ii) or (iii) of the Theorem. If $\Lambda^G$ is a sublattice of $2^{27}3^{36}$, then we have an inverse chain of inclusions for the stabilisers (S)…(S\_1)G=[Stab]{}(S\_0) , where each stabiliser group is a subgroup of index at most $2$ in the previous one |[Stab]{}(S\_i):[Stab]{}(S\_[i+1]{})|2 . Furthermore, whenever $|{\rm Stab}(S_i):{\rm Stab}(S_{i+1})|=2$, we have $\operatorname{rk}S_{i+1}>\operatorname{rk}S_i$, and since $\operatorname{rk}S_{2^{27}3^{36}}-\operatorname{rk}\Lambda^G\le 2$, we have \[indexbound\] 2\^2 . Thus, $G$ is the extension of ${\rm Stab}(S_{2^{27}3^{36}})$ by a group $G''$ of order at most $4$, and all the possibilities are considered in the case (iv) of the Theorem. This completes the proof of the Theorem.
Realising all Symmetry Groups of K3 {#B.4}
-----------------------------------
Let $G\subset {\rm Aut}(\Lambda)$ be the pointwise stabiliser of a sublattice $S\cong \Lambda^G$ of rank $\operatorname{rk}S\ge 4$. In this section, we will prove that, for each such $G$, there exists a non-linear $\sigma$-model on K3 having $G$ as its group of symmetries. This result is based on the assumption that for every choice of a positive definite $4$-dimensional subspace $\Pi\subset {\mathbb{R}}^{4,20}$ such that no root $v\in\Gamma^{4,20}$ is orthogonal to $\Pi$, the corresponding non-linear $\sigma$-model is well defined.
Let $d+4$, with $d\ge 0$, be the rank of $S$ and let $S^\perp\equiv \Lambda_G$ be its orthogonal complement in $\Lambda$. Suppose that $S^\perp(-1)$, obtained from $S^\perp$ by changing the sign of its quadratic form, can be primitively embedded into $\Gamma^{4,20}$. Then, by a reasoning similar to proposition \[th:embed\], we conclude that the action of $G$ on $S^\perp(-1)$ can be extended to $\Gamma^{4,20}$, in such a way that $(\Gamma^{4,20})_G\cong S^{\perp}(-1)$, [*i.e.*]{} that the sublattice $(\Gamma^{4,20})^G$ invariant under $G$ is the orthogonal complement of $S^\perp(-1)$. Since $(\Gamma^{4,20})^G$ has signature $(4,d)$, one can always find a positive definite four dimensional subspace $\Pi\subset{\mathbb{R}}^{4,20}$ such that $(\Gamma^{4,20})_G=\Gamma^{4,20}\cap\Pi^\perp$. Furthermore, $(\Gamma^{4,20})_G$ contains no vectors of norm $-2$, and $G$ is the subgroup of $O(\Gamma^{4,20})$ fixing $\Pi$ pointwise. Following the arguments of Section \[symmetries\], $\Pi$ corresponds to a well defined non-linear $\sigma$-model with symmetry group $G$.
Thus it remains to prove that a primitive embedding of $S^\perp(-1)$ into $\Gamma^{4,20}$ always exists. The ‘gluing’ construction described in Appendix \[s:GinOGamma\] shows that, for an even lattice with signature $(t_+,t_-)$ and discriminant form $q$ — we will denote them as a triple $(t_+,t_-;q)$ — the existence of an embedding into some even unimodular lattice with signature $(l_+,l_-)$ is equivalent to the existence of an even lattice with signature and discriminant form $(l_+-t_+,l_--t_-,-q)$. In [@Nikulin Theorem 1.12.4], the following sufficient conditions are proved for the existence of such an embedding:
1. $l_+-l_-\equiv 0\mod 8$
2. $t_+\le l_+$, $t_-\le l_-$ and $t_++t_-\le \frac{1}{2}(l_++l_-)$.
The lattice $S$ has signature $(4+d,0)$, so that it can be embedded into an even unimodular lattice $\Gamma^{8+d,d}\cong E_8\oplus U^d$ of signature $(8+d,d)$, where $U$ is the unimodular lattice of signature $(1,1)$. Let $S'$ be the orthogonal complement of $S$ in $\Gamma^{8+d,d}$. If $q$ denotes the discriminant form of $S$, then $S'$, $S^\perp$ and $S^\perp(-1)$ have signature and discriminant form $(4,d;-q)$, $(20-d,0;-q)$ and $(0,20-d;q)$, respectively. By comparing the signatures and discriminant forms of $S^\perp(-1)$ and $S'$, we conclude that these two lattices can be ‘glued’ together to form the even unimodular lattice $\Gamma^{4,20}$.
Gepner Models {#s:Gepner}
=============
Here we collect, following [@Brunner:2005fv], our conventions for the description of ${\mathcal{N}}=2$ minimal models and Gepner models.
${\mathcal{N}}=2$ Minimal Models and Gepner Models at $c=6$ {#App:GepnerModel}
-----------------------------------------------------------
The ${\mathcal{N}}=2$ minimal model at level $k$ has central charge $c=\tfrac{3k}{k+2}$, and can be described in terms of the coset that captures the bosonic subalgebra of the ${\mathcal{N}}=2$ superconformal algebra. The coset representations are labeled by (l,m,s) ,l=0,…,k,m\_[2k+4]{} ,s\_4 , subject to the condition l+m+s=02 , and with the field identification (l,m,s)\~(k-l,m+k+2,s+2) . We denote by $[l,m,s]$ the class corresponding to $(l,m,s)$. In terms of the ${\mathcal{N}}=2$ algebra, the irreducible representations are of the form ${\mathcal{H}}_{(l,m,s)}\oplus {\mathcal{H}}_{(l,m,s+2)}$ with $s$ even (odd) for the NS (R) sector, since the fermionic generators of the ${\mathcal{N}}=2$ algebra map ${\mathcal{H}}_{(l,m,s)}$ to ${\mathcal{H}}_{(l,m,s+2)}$. The conformal weight and $U(1)$-charge of the ground state in the $(l,m,s)$ sector are given by h\_[l,m,s]{}=+ ,q\_[l,m,s]{} =-2 . The character of the $(l,m,s)$ coset representation equals (with $q=e^{2\pi i\tau}$, $y=e^{2\pi i z}$) \_[\[l,m,s\]]{}(,z)=\_[[H]{}\_[\[l,m,s\]]{}]{} (q\^[L\_0-]{} y\^[J\_0]{}) = \_[j]{} c\^l\_[m+4j-s]{}() q\^[(-+2j)\^2]{}y\^[-+2j]{} . Here, $c^l_m(\tau)$ can be obtained from the identity \_[m/2k]{} c\^l\_m() \_[m,k]{}(,z)= , where $\theta_{m,k}(\tau,z)$ is the $\mathfrak{su}(2)$ theta function \_[m,k]{}(,z)=\_[n]{}q\^[k(n+)\^2]{}y\^[k(n+)]{} with $m\in {\mathbb{Z}}/2k{\mathbb{Z}}$. For the calculation of the elliptic genus we are interested in the trace with the insertion of $(-1)^F$; for the ${\mathcal{N}}=2$ representation corresponding to $(l,m,s)$, this leads to I\^l\_m(,z)=\_[\[l,m,s\]]{}(,z)-\_[\[l,m,s+2\]]{}(,z) ,\[CharacterN2rep\] where $s=0$ in the NS sector and $s=1$ in the Ramond sector.
A Gepner model at $c=6$ is defined as a ${\mathbb{Z}}_H$ orbifold of a tensor product $(k_1)\cdots(k_r)$ of ${\mathcal{N}}=2$ minimal models, where \_[i=1]{}\^r =6 , H={k\_i+2} . The spectrum of the orbifold theory is given by \_ \_[i=1]{}\^r \_[\[l\_i,m\_i+n,s\_i\]]{}|\_[\[l\_i,m\_i-n,|s\_i\]]{} , and the sum is over the sectors $[l_i,m_i,s_i]$ satisfying the orbifold conditions $$\begin{aligned}
&\sum_{i=1}^{r}\frac{m_i}{k_i+2}\in{\mathbb{Z}}&\text{$r$ even}\ ,\\
&\sum_{i=1}^{r}\frac{m_i}{k_i+2}-\frac{s_1}{2}\in{\mathbb{Z}}&\text{$r$ odd}\ ,\end{aligned}$$ together with the spin alignment condition s\_i-s\_j2 ,i,j=1,…,r . We denote a state transforming in the ${\mathcal{H}}_{l,m,s}\otimes \bar{\mathcal{H}}_{l,\bar m,\bar s}$ coset representations by \^l\_[m,s;|m,|s]{}=\_[l,m,s]{}|\_[l,|m,|s]{} .\[GepnerStates\] Imposing the various constraints, it is then clear that the elliptic genus of these Gepner models is given by [@Kawai:1993jk] $$\begin{aligned}
\phi(\tau,z)&=\frac{1}{2^{r-1}\, H}\sum_{a,b=0}^{H}\prod_{i=1}^r
\sum_{l_i=0}^{k_i} \sum_{m_i= -k_i-1}^{k_i+2}\,
e^{\frac{2\pi i(m_i+a)b}{k_i+2}}
I^{l_i}_{m_i}(\tau,z)I^{l_i}_{m_i+2a}(\bar{\tau},0)\ .\label{EllGenGepner}\end{aligned}$$ Using $$\begin{aligned}
I^l_m(\tau,0)=\delta_{m,l+1}-\delta_{m,-l-1}\end{aligned}$$ we can directly evaluate (\[EllGenGepner\]). For the two cases considered in this paper, namely $(1)^6$ and $(2)^4$, we then find indeed (\[K3eg\]).
D-branes in Gepner Models {#s:Dbranes}
-------------------------
A-type (B-type) D-branes satisfy the gluing conditions $$\begin{array}{llll}
(L_n-\bar L_{-n})|\!|{\rm A}{\rangle\!\!\!\;\rangle}=0 \quad
&(J_n-\bar J_{-n})|\!|{\rm A}{\rangle\!\!\!\;\rangle}=0 \quad
&(G^{\pm}_r+i\eta\bar G^{\pm}_{-r})|\!|{\rm A}{\rangle\!\!\!\;\rangle}=0 \quad
& \text{(A-type)} \label{Atype}\\
(L_n-\bar L_{-n})|\!|{\rm B}{\rangle\!\!\!\;\rangle}=0
&(J_n+\bar J_{-n})|\!|{\rm B}{\rangle\!\!\!\;\rangle}=0
& (G^{\pm}_r+i\eta\bar G^{\pm}_{-r})|\!|{\rm B}{\rangle\!\!\!\;\rangle}=0
&\text{(B-type)}\ ,
\end{array}$$ where $\eta=\pm 1$. In our construction we shall always consider tensor product A-type D-branes, as well as permutation B-type D-branes. The former are described by the boundary states [@Recknagel:1997sb] $$\begin{aligned}
|\!|L_i,M_i,S_i{\rangle\!\!\!\;\rangle}_{{\rm A},s}=&{\mathcal{N}}e^{-\frac{-\pi i}{2}s\sum_iS_i}
\sum_{l_i=0}^{k_i}\sum_{m_i=0}^{k_i+1}\sum_{\nu_i=0}^1
\prod_i\Bigl(\frac{1+(-1)^{l_i+m_i+s}}{2}\Bigr)
\Bigl(\frac{1}{H}\sum_{t\in{\mathbb{Z}}_H}e^{2\pi it\sum_i\frac{m_i}{k_i+2}}\Bigr)\nonumber\\
& \cdot (-1)^{\sum_iS_i\nu_i}e^{\pi i\sum_i M_i\frac{m_i}{k_i+2}}
\prod_i\frac{S_{L_il_i}}{\sqrt{S_{0l_i}}}|l_i,m_i,s+2\nu_i{\rangle\!\!\!\;\rangle}_{\rm A}\ ,\end{aligned}$$ where $H=\operatorname{lcm}\{k_i+2\}$, ${\mathcal{N}}=\sqrt{H}\prod_i\bigl(\frac{2}{k_i+2}\bigr)^{1/4}$ and L\_i=0,…, k\_i , M\_i\_[2k\_i+4]{} , S\_i\_4 , L\_i+M\_i+S\_i , S\_i+S\_j Here $s=0,1$ labels the NSNS or the RR closed string sector, respectively, and S\_[L\_il\_i]{}=() is the $S$-matrix of the $\mathfrak{su}(2)$ affine algebra at level $k_i$. Furthermore |l\_i,m\_i,s+2\_i\_[A]{}\_i\_[\[l\_i,m\_i,s+2\_i\]]{} |\_[\[l\_i,m\_i,s+2\_i\]]{} is the Ishibashi state satisfying the A-type gluing conditions (\[Atype\]) for each minimal model factor separately. The left-moving world-sheet fermion number operator acts on this boundary state as (-1)\^[F\_L]{}L\_i,M\_i,S\_i\_[[A]{},s]{}=L\_i,M\_i+1,S\_i+1\_[[A]{},s]{} . The overlap between two such boundary states is given by $$\begin{aligned}
{}_{{\rm A},s}{\langle\!\!\!\;\langle}L_i',M_i',S_i'|\!|
&q^{\frac{1}{2}(L_0+\bar L_0)-\frac{c}{24}}(-1)^{F_L} |\!|L_i,M_i,S_i{\rangle\!\!\!\;\rangle}_{{\rm A},s}\\
&=\sum_{t\in{\mathbb{Z}}_H}\sum_{(l_i,m_i,s_i)}e^{-\frac{\pi i s}{2}
\sum_i(S_i+1-S_i'+ s_i)}\prod_i\delta^{(2)}(S_i+1-S_i'+s_i)\notag\\
&\qquad \cdot \prod_i
\Bigl[\delta^{(k_i+2)}\Bigl(\frac{M_i+1-M_i'+ m_i}{2}+t\Bigr)\,
N_{L_i l_i}^{L_i'}\, \chi_{[l_i,m_i,s_i]}(\tilde q)\Bigr] \ ,\notag\end{aligned}$$ where $\tilde{q}$ is the open string parameter.
Given a permutation $\pi\in S_r$, a permutation brane $|\!|{\rm B}{\rangle\!\!\!\;\rangle}^\pi$ satisfies the $\pi$-twisted B-type boundary conditions $$\begin{aligned}
\label{Bpbound}
(L_n^{(i)}-\bar L_{-n}^{(\pi(i))})\|B{\rangle\!\!\!\;\rangle}^\pi=0\ , \qquad
(J_n^{(i)}+\bar J_{-n}^{\pi(i)})\|B{\rangle\!\!\!\;\rangle}^\pi=0\ ,\end{aligned}$$ and similarly for the fermionic gluing conditions. For example, for the $(1)^6$ model we consider $\pi\in S_6$ with cycle decomposition $(12)(34)(56)$. The corresponding permutation brane is defined as [@Recknagel:2002qq] $$\begin{aligned}
|\!|L_1,&L_2,L_3,M_1,M_2,M_3,\hat M,S_i{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s} \notag \\
&=\frac{1}{2^3}\frac{1}{\sqrt{3}}\sum_{n\in{\mathbb{Z}}_3}e^{-\frac{\pi in\hat M}{3}}\sum_{l_1,l_2,l_3=0}^1\,
\sum_{m_1,m_2,m_3=0}^5\, \sum_{\nu_1,\ldots,\nu_6\in{\mathbb{Z}}_2}
\prod_{i}\Bigl(\frac{1+(-1)^{l_i+m_i+n+s}}{2}\Bigr) \notag \\
&\quad \cdot\prod_i
\frac{S_{L_il_i}}{S_{0l_i}}e^{i \frac{\pi}{3} \sum_i^3 m_iM_i}(-1)^{\sum_i^6 S_i\nu_i}
e^{- i \frac{s \pi}{2} \sum_i^6S_i} |[l_i,m_i+n,s+2\nu_i]{\rangle\!\!\!\;\rangle}^{\pi,g^n}_{{\rm B}}\ , \label{16Bper}\end{aligned}$$ where $\hat M\in{\mathbb{Z}}_6$, while $L_i+M_i$, $\hat M+\sum_iM_i$ and $S_i+S_j$ are even, so that the boundary state is invariant under $n\mapsto n+3$, $m_i\mapsto m_i+3, i=1,\ldots, 3$. Here, $|[l_i,m_i+n,s_i]{\rangle\!\!\!\;\rangle}^{\pi,g^n}_{{\rm B}}$ is the Ishibashi state in the $n^{\rm th}$ twisted sector \_[i=1]{}\^3 ( (\_[\[l\_i,m\_i+n,s\_[2i-1]{}\]]{}|\_[\[l\_i,m\_i-n,-s\_[2i]{}\]]{} ) ( \_[\[l\_i,-m\_i+n,s\_[2i]{}\]]{}|\_[\[l\_i,-m\_i-n,-s\_[2i-1]{}\]]{} ) ) , that is uniquely characterised (up to normalisation) by the B-type boundary conditions (together with the corresponding fermionic formulae). Note that (-1)\^[F\_L]{} ||L\_i,M\_i,M,S\_i\^\_[[B]{},s]{}= ||L\_i,M\_i,M,S\_i+1\^\_[[B]{},s]{} . For the case at hand, the overlap between two such branes with $L_i=0$ and $s=1$ (RR sector) is $$\begin{aligned}
&{}^{\pi}_{{\rm B},s=1}{\langle\!\!\!\;\langle}0,0,0,M_1',M_2',M_3',\hat M',0|\!|
q^{\frac{1}{2}(L_0+\bar L_0)-\frac{c}{24}}(-1)^{F_L}
|\!|0,0,0,M_1,M_2,M_3,\hat M,0{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s=1}\notag\\
&=\sum_{ m_i=\pm 1}
e^{-\frac{\pi i }{2}\sum_i^6(1+ m_i)}
\delta^{(3)}\Bigl(\frac{\hat M'-\hat M+\sum_i m_i}{2}\Bigr)
\prod_{i=1}^3\delta^{(3)}\Bigl(\frac{M_i-M_i'+ m_{2i-1}- m_{2i}}{2}\Bigr) \ .\end{aligned}$$
In order to determine the relative intersection number between the A-type and the B-type D-branes, we also need to understand the overlap between the different types of branes. For the $(1)^6$ model, the only non-vanishing overlaps are between Ishibashi states in the $n=0$ sectors of the form \_[i=1]{}\^3 (\_[\[l\_i,m\_i,s\_i\]]{}|\_[\[l\_i,m\_i,s\_i\]]{}\_[\[l\_i,-m\_i,-s\_[i]{}\]]{}|\_[\[l\_i,-m\_i,-s\_i\]]{}) . Taking into account the implicit relative phase between the Ishibashi states as in [@Brunner:2005fv], one finds that the overlap between the relevant Ishibashi states is \_[A]{}l\_i,m\_i,s\_i| q\^[(L\_0+|L\_0)-]{} |\[l\_i,m\_i+n,s\_i\]\^\_[B]{} =\_[i=1]{}\^3e\^[m\_i-s\_i]{}\_[\[l\_i,m\_i,s\_i\]]{}(q\^2) , and hence the overlap between the relevant boundary states is given by $$\begin{aligned}
{}_{{\rm A},s=1}&{\langle\!\!\!\;\langle}0,M_i',0|\!| q^{\frac{1}{2}(L_0+\bar L_0)-\frac{c}{24}}(-1)^{F_L}
|\!|0,0,0,M_1,M_2,M_3,\hat M,0{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s=1}\notag\\
=&\sum_{m_1, m_2, m_3=\pm 1}-e^{-\frac{\pi i}{2}(1+\sum_i m_i)}\prod_i^3
\delta^{(3)}\Bigl(\frac{M_i-M_{2i-1}'+M_{2i}'+ m_i+1}{2}\Bigr)\ .\end{aligned}$$
A similar computation can be done for the $(2)^4$ model. The B-type permutation brane for $\pi\in S_4$ with cycle decomposition $(12)(34)$ is $$\begin{aligned}
|\!|L_1,&L_2,M_1,M_2,\hat M,S_i{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s} \notag \\
&=\frac{1}{2^2}\frac{1}{\sqrt{4}}\sum_{n\in{\mathbb{Z}}_4}e^{-i \frac{\pi n}{4} \hat M}\sum_{l_1,l_2=0}^2\,
\sum_{m_1,m_2=0}^7\, \sum_{\nu_1,\ldots,\nu_4=0}^1
\prod_{i}\Bigl(\frac{1+(-1)^{l_i+m_i+n+s}}{2}\Bigr) \label{24Bper} \\
& \quad\cdot \prod_i
\frac{S_{L_il_i}}{S_{0l_i}}e^{i \frac{\pi}{4} \sum_im_iM_i}(-1)^{\sum_i^4 S_i\nu_i}
e^{- i \frac{\pi s}{2} \sum_i^4S_i}|[l_i,m_i+n,s+2\nu_i]{\rangle\!\!\!\;\rangle}^{\pi,g^n}_{{\rm B}}\ , \notag \end{aligned}$$ where $\hat M\in{\mathbb{Z}}_8$, and $L_i+M_i$, $\hat M+\sum_iM_i$ and $S_i+S_j$ are all even. The left-moving fermionic number operator acts now by (-1)\^[F\_L]{} ||L\_1,L\_2,M\_1,M\_2,M,S\_i\^\_[[B]{},s]{}= ||L\_1,L\_2,M\_1,M\_2,M+4,S\_i+1\^\_[[B]{},s]{} , and the overlap between two branes with $L_1,L_2=0$ is $$\begin{aligned}
&{}^{\pi}_{{\rm B},s=1}{\langle\!\!\!\;\langle}0,0,M_1',M_2',\hat M',0|\!|
q^{\frac{1}{2}(L_0+\bar L_0)-\frac{c}{24}}(-1)^{F_L}
|\!|0,0,M_1,M_2,\hat M,0{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s=1} \nonumber \\
&=\frac{1}{4} \sum_{(l_i, m_i,s_i)}\delta_{ l_1 l_2}\delta_{ l_3 l_4}\,
\prod_{i=1}^{4}\delta^{(2)}(s_i+1)\,
e^{-\frac{\pi i }{2}\sum_i^4(1+s_i)}
\delta^{(4)}\Bigl(\frac{\hat M'-\hat M+4+\sum_i^4 m_i}{2}\Bigr) \nonumber\\
&\hspace{2.5cm}\cdot
\prod_{i=1}^2\delta^{(4)}\Bigl(\frac{M_i-M_i'+m_{2i-1}- m_{2i}}{2}\Bigr)\,
\chi_{[l_i m_i s_i]}(\tilde{q}) \notag\\
&=\sum_{\substack{ m_1, m_2\in\{1,2,3\}\\ s_1, s_2=\pm 1}}e^{-\frac{\pi i }{2}(2+s_1+s_2)}\,
\delta^{(4)}\Bigl(\frac{\hat M'-\hat M+4+\sum_i^2 m_i( s_i+1)}{2}\Bigr) \nonumber\\
&\hspace{2.5cm}\cdot \prod_{i=1}^2\delta^{(4)}\Bigl(\frac{M_i-M_i'+ m_{i}(s_i-1)}{2}\Bigr)\ .\end{aligned}$$ Finally, the RR-overlap between A-type and permutation B-type branes in this model is given by $$\begin{aligned}
{}_{{\rm A},s=1}&{\langle\!\!\!\;\langle}0,M_i',0|\!| q^{\frac{1}{2}(L_0+\bar L_0)-\frac{c}{24}}(-1)^{F_L}
|\!|0,0,0,M_1,M_2,\hat M,0{\rangle\!\!\!\;\rangle}^{\pi}_{{\rm B},s=1}\notag\\
=&\sum_{m_1, m_2=\pm 1}e^{-\frac{\pi i}{2}(2+\sum_i m_i)}
\delta^{(4)}\Bigl(\frac{M_1-M_{1}'+M_{2}'+ m_1+1}{2}\Bigr)\nonumber\\
&\hspace{2.5cm}
\cdot\delta^{(4)}\Bigl(\frac{M_2-M_{3}'+M_{4}'+ m_2+1}{2}\Bigr)\ .\end{aligned}$$
[99]{}
T. Eguchi, H. Ooguri and Y. Tachikawa, [*Notes on the K3 surface and the Mathieu group $M_{24}$*]{}, Exper. Math. [**20**]{} (2011) 91 . T. Gannon, [*Moonshine beyond the Monster: The bridge connecting algebra, modular forms and physics*]{}, Cambridge University Press (2006).
M.C.N. Cheng, [*K3 Surfaces, N=4 dyons, and the Mathieu group $M_{24}$*]{}, Commun. Number Theory Phys. [**4**]{} (2010) 623 . M.R. Gaberdiel, S. Hohenegger and R. Volpato, [*Mathieu twining characters for K3*]{}, JHEP [**1009**]{} (2010) 058 . M.R. Gaberdiel, S. Hohenegger and R. Volpato, [*Mathieu Moonshine in the elliptic genus of K3*]{}, JHEP [**1010**]{} (2010) 062 . T. Eguchi and K. Hikami, [*Note on twisted elliptic genus of K3 surface*]{}, Phys. Lett. B [**694**]{} (2011) 446 . S. Govindarajan, [*Brewing moonshine for Mathieu*]{}, [arXiv:1012.5732 \[math.NT\]]{}. S. Govindarajan, [*BKM Lie superalgebras from counting twisted CHL dyons*]{}, [arXiv:1006.3472 \[hep-th\]]{}.
S. Govindarajan and K. Gopala Krishna, [*BKM Lie superalgebras from dyon spectra in Z(N) CHL orbifolds for composite N*]{}, JHEP [**1005**]{} (2010) 014 .
S. Mukai, [*Finite groups of automorphisms of $K3$ surfaces and the Mathieu group*]{}, Invent. Math. [**94**]{} (1988) 183.
S. Kondo, [*Niemeier lattices, Mathieu groups and finite groups of symplectic automorphisms of K3 surfaces*]{}, Duke Math. Journal [**92**]{} (1998) 593, appendix by S. Mukai.
A. Taormina and K. Wendland, [*The symmetries of the tetrahedral Kummer surface in the Mathieu group $M_{24}$*]{}, [arXiv:1008.0954 \[hep-th\]]{}. P.S. Aspinwall, [*K3 surfaces and string duality*]{}, [arXiv:hep-th/9611137]{}. W. Nahm and K. Wendland, [*A hiker’s guide to K3: Aspects of N = (4,4) superconformal field theory with central charge c = 6*]{}, Commun. Math. Phys. [**216**]{} (2001) 85 . I. Brunner, M.R. Douglas, A.E. Lawrence and C. Romelsberger, [*D-branes on the quintic*]{}, JHEP [**0008**]{} (2000) 015 . T. Banks and L.J. Dixon, [*Constraints on string vacua with space-time supersymmetry*]{}, Nucl. Phys. B [**307**]{} (1988) 93. P.S. Aspinwall, [*Enhanced gauge symmetries and Calabi-Yau threefolds*]{}, Phys. Lett. B [**371**]{} (1996) 231 . A. Strominger, [*Massless black holes and conifolds in string theory*]{}, Nucl. Phys. B [**451**]{} (1995) 96 . M.R. Gaberdiel, [*Lectures on non-BPS Dirichlet branes*]{}, Class. Quant. Grav. [**17**]{} (2000) 3483 . O. Bergman and M.R. Gaberdiel, [*Non-BPS states in heterotic type IIA duality*]{}, JHEP [**9903**]{} (1999) 013 . I. Brunner, M.R. Gaberdiel and C.A. Keller, [*Matrix factorisations and D-branes on K3*]{}, JHEP [**0606**]{} (2006) 015 . J.H. Conway and N.J.A. Sloane, [*Sphere packings, lattices and groups*]{}, Grundlehren der Mathematischen Wissenschaften [**290**]{} 3rd edition, Springer-Verlag, New York (1999).
R.T. Curtis, [*On subgroups of [$\cdot 0$]{}. [I]{}. [L]{}attice stabilizers*]{}, J. Algebra [**27**]{} (1973) 549.
I. Brunner, M.R. Gaberdiel, S. Hohenegger and C.A. Keller, [*Obstructions and lines of marginal stability from the world-sheet*]{}, JHEP [**0905**]{} (2009) 007 . S. Govindarajan, D.P. Jatkar and K.G. Krishna, [*BKM superalgebras from counting dyons in N=4 supersymmetric type II compactifications*]{}, [arXiv:1106.1318 \[hep-th\]]{}.
V.V. Nikulin, [*Integer symmetric bilinear forms and some of their geometric applications*]{}, Izv. Akad. Nauk SSSR Ser. Mat. [**43**]{} (1979) 111.
J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker and R.A. Wilson, [*Atlas of finite groups*]{}, Oxford University Press (1985).
D. Allcock, [*Orbits in the Leech lattice,*]{} Exper. Math. [**14**]{}(4) (2005) 491.
I. Brunner and M.R. Gaberdiel, [*Matrix factorisations and permutation branes*]{}, JHEP [**0507**]{} (2005) 012 . T. Kawai, Y. Yamada and S.K. Yang, [*Elliptic genera and N=2 superconformal field theory*]{}, Nucl. Phys. B [**414**]{} (1994) 191 . A. Recknagel and V. Schomerus, [*D-branes in Gepner models*]{}, Nucl. Phys. B [**531**]{} (1998) 185 . A. Recknagel, [*Permutation branes*]{}, JHEP [**0304**]{} (2003) 041 .
[^1]: These are the analogues of the so-called MacKay-Thompson series for Monstrous Moonshine, see [@Gannon] for a modern review.
[^2]: We thank Yuji Tachikawa for informing us that he has now checked the decomposition for the first $1000$ coefficients.
[^3]: The condition that the symmetry preserves the full ${\mathcal{N}}=(4,4)$ superconformal algebra guarantees that the resulting twining genus still defines a weak Jacobi form. Symmetries that only preserve the ${\mathcal{N}}=(2,2)$ superconformal algebra typically act non-trivially on $J^\pm$, and then the resulting twining genera do not have the shift symmetry under $z\mapsto z+\tau$.
[^4]: The explicit description of the relevant branes and their intersection matrix is given in the LaTeX source.
[^5]: We have also checked that these symmetries preserve the spectral flow operators.
[^6]: Note that these last two phases are order $2$ only after taking the quotient by $G'$; in fact, their squares are non-trivial elements in $G'$, so that as elements of the whole group $G$ they are order 4.
[^7]: See the LaTeX source code for details; there we also give further details about the sublattice $(\Gamma^{4,20})^\perp$ and its embedding into the Leech lattice.
[^8]: Apparently this was also independently noted by the authors of [@EOT]; we thank Yuji Tachikawa for discussions about this point.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The AMS experiment onboard the International Space Station has recently provided cosmic ray electron and positron data with unprecedented precision in the range from 0.5 to 350GeV. The observed rise in the positron fraction at energies above 10GeV remains unexplained, with proposed solutions ranging from local pulsars to TeV-scale dark matter. Here, we make use of this high quality data to place stringent limits on dark matter with masses below $\sim$300GeV, annihilating or decaying to leptonic final states, essentially independent of the origin of this rise. We significantly improve on existing constraints, in some cases by up to two orders of magnitude.'
author:
- Lars Bergström
- Torsten Bringmann
- Ilias Cholis
- Dan Hooper
- Christoph Weniger
date: 'October 25, 2013'
title: New limits on dark matter annihilation from AMS cosmic ray positron data
---
#### Introduction. {#introduction. .unnumbered}
The AMS (Alpha Magnetic Spectrometer) collaboration has very recently announced the results of its first data collected from the International Space Station [@ams], consisting of a high precision measurement of the cosmic ray (CR) positron fraction [@Aguilar:2013qda]. This new data provides a confirmation of the rise of this quantity above 10GeV, as previously observed by PAMELA [@Adriani:2008zr] and Fermi [@FermiLAT:2011ab] (and with earlier hints provided by HEAT [@Barwick:1997ig] and AMS-01 [@Aguilar:2007yf]). Such a rise is not predicted in the standard scenario, in which CR positrons are mostly produced as [*secondary*]{} particles, as a result of collisions of CR protons with the interstellar medium (ISM). Instead, the large positron fraction seems to require the existence of at least one additional nearby [*primary*]{} source of high energy positrons. Local pulsars have emerged as the leading astrophysical candidates [@Hooper:2008kg; @Profumo:2008ms], although it has also been argued that strong local sources might not actually be needed when taking into account the spiral structure of the Milky Way in full 3-D propagation models [@Gaggero:2013rya] and that even a secondary production mechanism in the shock waves of supernovae remnants [@Blasi:2009hv; @Mertsch:2009ph] could provide a viable mechanism to explain the data [@Serpico:2011wg].
A more exotic possibility is that the observed positrons may be produced in the annihilations or decays of TeV-scale dark matter (DM) particles. Such scenarios, however, require unexpectedly large annihilation rates into predominantly leptonic final states [@Bergstrom:2009fa; @Finkbeiner:2010sm; @Yuan:2013eja; @Cholis:2013psa; @Jin:2013nta] and are subject to significant constraints from CR antiproton, gamma-ray and synchrotron data [@Donato:2008jk; @Bertone:2008xr; @Bergstrom:2008ag; @Cirelli:2009dv; @Evoli:2011id; @Kappl:2011jw; @Wechakama:2012yb; @DeSimone:2013fia]. Upcoming AMS data may help to settle this open issue not only by increasing statistics and extending their study to higher energies, but also by providing high precision measurements of other CR particle spectra (likely breaking degeneracies in the propagation parameters [@breakdegen]). Fermi and AMS will also further constrain any anisotropy in the positron/electron flux (where current limits are already close to discriminating between some of the scenarios described above [@Ackermann:2010ip; @DiBernardo:2010is]).
In this Letter, we do not make any attempt to explain the origin of the rise in the positron fraction. Instead, we focus on using the AMS data to derive limits on subdominant exotic contributions to the observed CR positron spectrum, in particular from DM with masses below $\sim$300GeV. While positrons have been used in the past to probe DM annihilation or decay [@Ellis:1988qp; @Rudaz:1987ry; @Kamionkowski:1990ty; @Baltz:1998xv; @Baltz:2001ir; @Kopp:2013eka], we exploit here for the first time the extremely high quality of the AMS data to search for pronounced [*spectral features*]{} in the positron flux predicted in some DM models [@Tylka:1989xj; @Turner:1989kg; @Baltz:2002we; @Hooper:2004xn; @Baltz:2004ie; @Bergstrom:2008gr; @Hooper:2012gq]. Much as exploiting spectral features can significantly improve the sensitivity of indirect DM searches using gamma rays [@Bringmann:2011ye], we demonstrate that the same is true for positrons, despite energy losses and other complicating factors. We derive limits that exceed the currently most stringent results on DM annihilation into leptons [@Ackermann:2011wa; @cmb] by up to two orders of magnitude.
This Letter is organized as follows. We first briefly review various astrophysical sources of leptons and how they manifest themselves in the observed CR flux, and then discuss possible contributions from DM. We continue with a description of the statistical treatment implemented here, before moving on to present our main results and conclusions. In an Appendix [@supp], we collect further technical details of our procedure for deriving limits on a possible DM signal, discussing in particular the impact of systematic uncertainties in the background modeling.
#### Astrophysical origins of cosmic ray leptons. {#astrophysical-origins-of-cosmic-ray-leptons. .unnumbered}
The origin of high energy electrons can be traced back to i) supernova explosions that accelerate the ISM to produce what are typically referred to as primary CRs, ii) inelastic collisions of primary CR protons and nuclei with the ISM (resulting in charged mesons, which decay, producing secondary electrons [*and*]{} positrons), and iii) individual sources such as pulsars that produce $e^\pm$ pairs. The averaged spectrum of propagated primary CR electrons (originating from many supernovae) is expected to be harder than that of the secondary $e^{\pm}$ component because the primary CR progenitors of the secondaries have also experienced propagation effects; both spectra are well described by power-laws, with spectral indices of about $3.3$ to $3.5$ ($3.7$) for primary electrons (secondary $e^{\pm}$) at energies above $\sim$10GeV [@secondaries]. The contribution from all galactic pulsars can be approximated by a power-law with an exponential cut-off at high energies, with a propagated spectral index of 2.0 $\pm$ 0.5 [@Hooper:2008kg; @Profumo:2008ms].
The Galactic Magnetic Field at scales $\gtrsim100$pc has a random and a regular component [@Jansson:2009ip]. As CR leptons propagate away from their sources, they follow the field lines and scatter off $B$-field irregularities. The net effect can be approximated as a random walk diffusion within a zone surrounding the Galactic Disk [@LongairDiff; @Strong:2007nh]. Further away from the disk the magnetic fields become weak, essentially leading to freely propagating CRs. During their propagation throughout the Galaxy, electrons and positrons also experience significant energy losses due to synchrotron and inverse Compton scattering on the galactic radiation field and the cosmic microwave background. The impact of other effects such as convective winds, ionization losses, or positron annihilation in collisions with matter are not significant for leptons in the energy range considered here [@Strong:2007nh; @LongairEloss] and are therefore ignored. We do, however, include bremsstrahlung emission, diffusive re-acceleration, and solar modulation inside the heliosphere (using the force-field approximation [@Gleeson1968]), which have an impact on CR $e^{\pm}$ spectra below 5–10GeV [@Strauss2012; @Maccione:2012cu].
For the propagation of CR leptons, we use the standard numerical tool GALPROP v54 [@galprop], which includes up-to-date implementations of the local interstellar radiation field and galactic gas distribution. These are relevant for both the production of secondary leptons and energy losses. GALPROP assumes a diffusion zone with cylindrical symmetry within which CRs diffuse and beyond which they escape. Its scale height, $L$, and other diffusion parameters, notably the diffusion time-scale and local diffusion properties, are constrained by observed CR ratios, including $\bar{p}/p$, B/C and $^{10}$Be/$^9$Be. As reference values we assume $L=4$ kpc, corresponding to the value best-fit by CR data [@Trotta:2010mx] and favored by radio observations [@Bringmann:2011py], and the standard default GALPROP assumptions for the local radiation and magnetic field energy densities, corresponding to $U_{\rm rad} + U_{B} = 1.7$eVcm$^{-3}$ [@galprop]. For the diffusion zone scale height, values of $L<2$kpc are in tension with a combined analysis of CR and gamma-ray data [@Cholis:2011un], while increasing $L$ beyond $8$kpc does not significantly alter our results.[^1] The propagation of high-energy leptons is actually dominated by energy losses rather than diffusion, implying that more conservative limits would arise for larger values of the local radiation and magnetic field energy densities. In our subsequent discussion, we will allow for an increase of $U_{\rm rad} + U_{B}$ by up to $50\%$ with respect to the reference value, which is still compatible with gamma-ray and synchrotron data [@Jaffe:2011qw; @Bringmann:2011py].
#### Positrons from dark matter. {#positrons-from-dark-matter. .unnumbered}
DM particles annihilating or decaying in the Galactic Halo may also contribute to the CR lepton spectrum, producing equal numbers of positrons and electrons. For annihilating DM, the injected spectrum of CR leptons per volume and time is given by $Q=\frac12\langle\sigma v\rangle\left(\rho_\chi/m_\chi\right)^2 dN/dE$ (divided by 2 if the DM particle is not self-conjugate), while for decaying DM, this is instead $Q=\Gamma\rho_\chi/m_\chi dN/dE$, where $\Gamma$ is the decay rate. Here, $\langle\sigma v\rangle$ is the velocity-averaged annihilation rate, $\rho_{\chi}$ is the DM density, $m_{\chi}$ is the DM mass, and $dN/dE$ is the spectrum of leptons produced per annihilation or decay. As our default choice, we adopt a DM distribution which follows an Einasto profile [@Merritt:2005xc], normalized to a local density of $\rho^\odot_\chi=0.4$GeV [@Catena:2009mf; @Salucci:2010qr].
Positrons from DM annihilation or decay typically result from the decay of $\pi^+$ (for hadronic final states), or the leptonic decay of $\tau^+$ or $\mu^+$. Owing to the high multiplicity of such processes, the resulting $e^+$ energy distribution at injection (which we take from Ref. [@Cirelli:2010xx]) is typically very soft. If DM annihilates directly into $e^\pm$, however, these are produced nearly monochromatically. Even after accounting for energy losses from propagation, a very characteristic spectrum arises in this case, with a sharp edge-like feature at $E\!=\!m_\chi$ (or at $E\!=\!m_\chi/2$ for decaying DM). A comparably distinct spectral feature arises from the annihilation of Majorana DM into $e^+e^-\gamma$ final states. Popular examples for DM models with large annihilation rates into $e^\pm$ final states include Kaluza-Klein DM [@Hooper:2004xn], while the supersymmetric neutralino is a possible candidate for producing a spectrum dominated by $e^+e^-\gamma$ final states [@Bergstrom:2008gr].[^2]
![The $e^{\pm}$ spectrum from annihilating DM, after propagation, for different annihilation final states, assuming $\langle \sigma v\rangle$$=3\times 10^{-26}$ cm$^{3}$s$^{-1}$. Solid lines refer to reference diffusion zone ($L$=4kpc) and energy loss assumptions ($U_{\rm
rad}+U_{B} = 1.7$eVcm$^{-3}$). Dashed (dotted) lines show the effect of a different scale height $L$=8(2) kpc. The dash-dotted line shows the impact of increasing the local radiation plus magnetic field density to $U_{\rm rad}+U_{B} = 2.6$eVcm$^{-3}$.[]{data-label="fig:spectra"}](spectra.eps){width="\linewidth"}
We illustrate this in Fig. \[fig:spectra\] by showing the propagated $e^\pm$ spectra for various final states and an annihilation rate that corresponds to the “thermal" cross section of $\langle \sigma v \rangle_{\rm
therm} \equiv 3\times 10^{-26}\rm \ cm^3 s^{-1}$ (which leads to the correct relic density in the simplest models of thermally produced DM). As anticipated, the $e^+e^-$ and $e^+e^-\gamma$ final states result in the most pronounced spectral features – a fact which helps considerably, as we will see, to distinguish them from astrophysical backgrounds. For the case of $e^+e^-$ final states, we also show how the spectrum depends on our local diffusion and energy loss assumptions within the range discussed above. Increasing $L$ enables CR leptons to reach us from greater distances due to the larger diffusion volume and therefore results in softer propagated spectra. While the peak normalization of the spectrum depends only marginally on $L$, it may be reduced by up to a factor of $\sim$2 when increasing the assumed local energy losses via synchrotron radiation and inverse Compton scattering by 50%. In Fig. \[fig:fraction\], we show a direct comparison of the DM signal with the AMS data, for the case of $e^+e^-$ final states contributing at the maximum level allowed by our constraints (see below) for two fiducial values of $m_\chi$. Again, it should be obvious that the shape of the DM contribution differs at all energies significantly from that of the background.
![The AMS positron fraction measurement [@Aguilar:2013qda] and background+signal fit for DM annihilating directly to $e^+ e^-$, for $m_\chi=10$GeV and 100GeV. The normalization of the DM signal in each case was chosen such that it is barely excluded at the $95\%$ CL. For better visibility, the contribution from DM (lower lines) has been rescaled as indicated.[]{data-label="fig:fraction"}](plot_fraction_combined.eps){width="\linewidth"}
#### Statistical treatment. {#statistical-treatment. .unnumbered}
We use the likelihood ratio test [@Rolke:2004mj] to determine the significance of, and limits on, a possible DM contribution to the positron fraction measured by AMS. As likelihood function, we adopt a product of normal distributions $\mathcal{L}=\prod_i N(f_i|\mu_i, \sigma_i)$; $f_i$ is the measured value, $\mu_i$ the positron fraction predicted by the model, and $\sigma_i$ its variance. The DM contribution enters with a single degree of freedom, given by the non-negative signal normalization. Upper limits at the $95\%$CL on the DM annihilation or decay rate are therefore derived by increasing the signal normalization from its best-fit value until $-2\ln\mathcal{L}$ is changed by 2.71, while profiling over the parameters of the background model.
We use data in the energy range 1–350GeV; the variance $\sigma_i$ is approximated by adding the statistical and systematic errors of the measurement in quadrature, $\sigma_i=(\sigma_{i,\rm stat}^2 + \sigma_{i,\rm
sys}^2)^{1/2}$. Since the total relative error is always small (below 17%), and at energies above 4GeV dominated by statistics, we expect this approximation to be very reliable. The binning of the published positron fraction follows the AMS energy resolution, which varies between 10.4% at 1GeV and 1.5% at 350GeV. Although we do not account for the finite energy resolution of AMS in our analysis, we have explicitly checked that this impacts our results by no more than 10%.
As our nominal model for the part of the $e^\pm$ spectrum that does not originate from DM, henceforth simply referred to as the astrophysical background, we use the same phenomenological parameterization as the AMS collaboration in their analysis [@Aguilar:2013qda]. This parameterization describes each of the $e^\pm$ fluxes as the sum of a common source spectrum – modeled as a power-law with exponential cutoff – and an individual power-law contribution (only the latter being different for the $e^+$ and $e^-$ fluxes). After adjusting normalization and slope of the secondary positrons such that the overall flux reproduces the Fermi $e^+\!+\!e^-$ measurements [@Ackermann:2010ij], the five remaining model parameters are left unconstrained. This phenomenological parameterization provides an extremely good fit (with a $\chi^2/{\rm d.o.f.} =
28.5/57$), indicating that no fine structures are observed in the AMS data. For the best-fit spectral slopes of the individual power-laws we find $\gamma_{e^-} \simeq 3.1$ and $\gamma_{e^+} \simeq 3.8$, respectively, and for the common source $\gamma_{e^\pm} \simeq 2.5$ with a cutoff at $E_c\simeq$800GeV, consistent with Ref. [@Aguilar:2013qda]. Subsequently, we will keep $E_c$ fixed to its best-fit value.
#### Results and Discussion. {#results-and-discussion. .unnumbered}
![Upper limits ($95\%$ CL) on the DM annihilation cross section, as derived from the AMS positron fraction, for various final states (this work), WMAP7 (for $\ell^+\ell^-$) [@cmb] and Fermi LAT dwarf spheroidals (for $\mu^+ \mu^-$ and $\tau^+ \tau^-$) [@Ackermann:2011wa]. The dotted portions of the curves are potentially affected by solar modulation. We also indicate $\langle \sigma v
\rangle_{\rm therm} \equiv 3\times 10^{-26}\rm \ cm^3 s^{-1}$. The AMS limits are shown for reasonable reference values of the local DM density and energy loss rate (see text), and can vary by a factor of a few, as indicated by the hatched band (for clarity, this band is only shown around the $e^+ e^-$ constraint).[]{data-label="fig:results"}](limits.eps){width="\linewidth"}
Our main results are the bounds on the DM annihilation cross section, as shown in Fig. \[fig:results\]. No significant excess above background was observed. For annihilations proceeding entirely to $e^+e^-$ final states, we find that the “thermal” cross section is firmly excluded for $m_\chi\lesssim 90$GeV. For $m_\chi\sim10$GeV, which is an interesting range in light of recent results from direct [@Bernabei:2010mq; @Aalseth:2011wp; @Angloher:2011uu; @Agnese:2013rvf; @Frandsen:2013cna] and indirect [@Hooper:2011ti; @Hooper:2013rwa; @Abazajian:2012pn] DM searches, our upper bound on the annihilation cross section to $e^+ e^-$ is approximately two orders of magnitude below $\langle \sigma v\rangle_{\rm therm}$. If only a fraction $f$ of DM annihilates like assumed, limits would scale like $f^{-2}$ (and, very roughly, $\langle\sigma v\rangle_{\rm therm}\propto f^{-1}$). We also show in Fig. \[fig:results\] the upper bounds obtained for other leptonic final states. As expected, these limits are weaker than those found in the case of direct annihilation to electrons – both because part of the energy is taken away by other particles (neutrinos, in particular) and because they feature broader and less distinctive spectral shapes. These new limits on DM annihilating to $\mu^+ \mu^-$ and $\tau^+ \tau^-$ final states are still, however, highly competitive with or much stronger than those derived from other observations, such as from the cosmic microwave background [@cmb] and from gamma-ray observations of dwarf galaxies [@Ackermann:2011wa]. Note that for the case of $e^+e^-\gamma$ final states even stronger limits can be derived for $m_\chi\gtrsim50$GeV by a spectral analysis of gamma rays [@Bringmann:2012vr]. We do not show results for the $\bar b b$ channel, for which we nominally find even weaker limits due to the broader spectrum (for $m_\chi\simeq100$GeV, about $\langle\sigma v\rangle\lesssim 1.1\cdot 10^{-24}\,\rm cm^3 s^{-1}$). In fact, due to degeneracies with the background modeling, limits for annihilation channels which produce such a broad spectrum of positrons can suffer from significant systematic uncertainties. For this reason, we consider our limits on the $e^+e^-$ channel to be the most robust.
Uncertainties in the $e^{\pm}$ energy loss rate and local DM density weaken, to some extent, our ability to robustly constrain the annihilation cross sections under consideration in Fig. \[fig:results\]. We reflect this uncertainty by showing a band around the $e^+e^-$ constraint, corresponding to the range $U_{\rm rad} + U_{B} = (1.2-2.6)$eVcm$^{-3}$, and $\rho_\chi^\odot=(0.25-0.7)$GeVcm$^{-3}$ [@Salucci:2010qr; @Iocco:2011jz] (note that the *form* of the DM profile has a much smaller impact). Uncertainty bands of the same width apply to each of the other final states shown in the figure, but are not explicitly shown for clarity. Other diffusion parameter choices impact our limits only by up to $\sim$10%, except for the case of low DM masses, for which the effect of solar modulation may be increasingly important [@DellaTorre:2012zz; @Maccione:2012cu]. We reflect this in Fig. \[fig:results\] by depicting the limits derived in this less certain mass range, where the peak of the signal $e^+$ flux (as shown in Fig. \[fig:spectra\]) falls below a fiducial value of 5GeV, with dotted rather than solid lines.
For comparison, we have also considered a collection of physical background models in which we calculated the expected primary and secondary lepton fluxes using GALPROP, and then added the contribution from all galactic pulsars. While this leads to an almost identical description of the background at high energies as in the phenomenological model, small differences are manifest at lower energies due to solar modulation and a spectral break [@StrongPapers; @Ptuskin:2005ax; @Trotta:2010mx] in the CR injection spectrum at a few GeV (both neglected in the AMS parameterization). We cross-check our fit to the AMS positron fraction with lepton measurements by Fermi [@Ackermann:2010ij]. Using these physical background models in our fits, instead of the phenomenological AMS parameterization, the limits do not change significantly. The arguably most extreme case would be the appearance of dips in the background due to the superposition of several pulsar contributions, which might conspire with a hidden DM signal at almost exactly the same energy. We find that in such situations, the real limits on the annihilation rate could be weaker (or stronger) by up to roughly a factor of 3 for any individual value of $m_\chi$. See the Appendix [@supp] for more details and further discussion of possible systematics that might affect our analysis.
Lastly, we note that the upper limits on $\langle\sigma v\rangle(m_\chi)$ reported in Fig. \[fig:results\] can easily be translated into upper limits on the decay width of a DM particle of mass $2m_\chi$ via $\Gamma \simeq
\langle \sigma v \rangle \rho^\odot_\chi / m_\chi$. We checked explicitly that this simple transformation is correct to better than 10% for the $L=$4 kpc propagation scenario and $e^+e^-$ and $\mu^+\mu^-$ final states over the full considered energy range.
#### Conclusions. {#conclusions. .unnumbered}
In this Letter, we have considered a possible dark matter contribution to the recent AMS cosmic ray positron fraction data. The high quality of this data has allowed us for the first time to successfully perform a spectral analysis, similar to that used previously in the context of gamma ray searches for DM. While we have found no indication of a DM signal, we have derived upper bounds on annihilation and decay rates into leptonic final states that improve upon the most stringent current limits by up to two orders of magnitude. For light DM in particular, our limits for $e^+e^-$ and $\mu^+\mu^-$ final states are significantly below the cross section naively predicted for a simple thermal relic. When taken together with constraints on DM annihilations to hadronic final states from gamma rays [@Ackermann:2011wa] and antiprotons [@Evoli:2011id], this new information significantly limits the range of models which may contain a viable candidate for dark matter with $m_\chi\sim\mathcal{O}(10)$GeV.
The AMS mission is planned to continue for 20 years. Compared to the 18 months of data [@Aguilar:2013qda] our analysis is based on, we expect to be able to strengthen the presented limits by at least a factor of three in the energy range of 6–200GeV with the total data set, and by more in the likely case that systematics and the effective acceptance of the instrument improve.
This work makes use of SciPy [@SciPy], Minuit [@James:1975dr] and Matplotlib [@Hunter:2007]. The research of L.B. was carried out under Swedish Research Council (VR) contract no. 621-2009-3915. T.B. acknowledges support from the German Research Foundation (DFG) through the Emmy Noether grant BR 3954/1-1. I.C., C.W. and D.H. thank the *Kavli Institute for Theoretical Physics* in Santa Barbara, California, for their kind hospitality. This work has been supported by the US Department of Energy.
[00]{}
M. Aguilar [*et al.*]{} \[AMS Collaboration\], Phys. Rev. Lett. [**110**]{}, no. 14, 141102 (2013). O. Adriani [*et al.*]{} \[PAMELA Collaboration\], Nature [**458**]{}, 607 (2009) \[arXiv:0810.4995 \[astro-ph\]\]. M. Ackermann [*et al.*]{} \[Fermi LAT Collaboration\], Phys. Rev. Lett. [**108**]{}, 011103 (2012) \[arXiv:1109.0521 \[astro-ph.HE\]\]. S. W. Barwick [*et al.*]{} \[HEAT Collaboration\], Astrophys. J. [**482**]{}, L191 (1997) \[astro-ph/9703192\]. M. Aguilar [*et al.*]{} \[AMS-01 Collaboration\], Phys. Lett. B [**646**]{}, 145 (2007) \[astro-ph/0703154 \[ASTRO-PH\]\]. D. Hooper, P. Blasi and P. D. Serpico, JCAP [**0901**]{}, 025 (2009) \[arXiv:0810.1527 \[astro-ph\]\]. S. Profumo, Central Eur. J. Phys. [**10**]{}, 1 (2011) \[arXiv:0812.4457 \[astro-ph\]\]. D. Gaggero, L. Maccione, G. Di Bernardo, C. Evoli and D. Grasso, Phys. Rev. Lett. [**111**]{}, 021102 (2013) \[arXiv:1304.6718 \[astro-ph.HE\]\]. P. Blasi, Phys. Rev. Lett. [**103**]{}, 051104 (2009) \[arXiv:0903.2794 \[astro-ph.HE\]\]. P. Mertsch and S. Sarkar, Phys. Rev. Lett. [**103**]{}, 081104 (2009) \[arXiv:0905.3152 \[astro-ph.HE\]\]. For a review, see P. D. Serpico, Astropart. Phys. [**39-40**]{}, 2 (2012) \[arXiv:1108.4827 \[astro-ph.HE\]\]. L. Bergström, J. Edsjö and G. Zaharijas, Phys. Rev. Lett. [**103**]{}, 031103 (2009) \[arXiv:0905.0333 \[astro-ph.HE\]\]. D. P. Finkbeiner, L. Goodenough, T. R. Slatyer, M. Vogelsberger and N. Weiner, JCAP [**1105**]{}, 002 (2011) \[arXiv:1011.3082 \[hep-ph\]\]. Q. Yuan, X. -J. Bi, G. -M. Chen, Y. -Q. Guo, S. -J. Lin and X. Zhang, arXiv:1304.1482 \[astro-ph.HE\]. I. Cholis and D. Hooper, Phys. Rev. D [**88**]{}, 023013 (2013) \[arXiv:1304.1840 \[astro-ph.HE\]\]. H. -B. Jin, Y. -L. Wu and Y. -F. Zhou, arXiv:1304.1997 \[hep-ph\]. F. Donato, D. Maurin, P. Brun, T. Delahaye and P. Salati, Phys. Rev. Lett. [**102**]{}, 071301 (2009) \[arXiv:0810.5292 \[astro-ph\]\]. G. Bertone, M. Cirelli, A. Strumia and M. Taoso, JCAP [**0903**]{}, 009 (2009) \[arXiv:0811.3744 \[astro-ph\]\]. L. Bergström, G. Bertone, T. Bringmann, J. Edsjö and M. Taoso, Phys. Rev. D [**79**]{}, 081303 (2009) \[arXiv:0812.3895 \[astro-ph\]\]. M. Cirelli, P. Panci and P. D. Serpico, Nucl. Phys. B [**840**]{}, 284 (2010) \[arXiv:0912.0663 \[astro-ph.CO\]\]. C. Evoli, I. Cholis, D. Grasso, L. Maccione and P. Ullio, Phys. Rev. D [**85**]{}, 123511 (2012) \[arXiv:1108.0664 \[astro-ph.HE\]\]. R. Kappl and M. W. Winkler, Phys. Rev. D [**85**]{}, 123522 (2012) \[arXiv:1110.4376 \[hep-ph\]\]. M. Wechakama and Y. Ascasibar, arXiv:1212.2583 \[astro-ph.CO\]. A. De Simone, A. Riotto and W. Xue, JCAP [**1305**]{}, 003 (2013) \[arXiv:1304.1336 \[hep-ph\]\]. M. Pato, D. Hooper and M. Simet, JCAP [**1006**]{}, 022 (2010) \[arXiv:1002.3341 \[astro-ph.HE\]\]. M. Ackermann [*et al.*]{} \[Fermi-LAT Collaboration\], Phys. Rev. D [**82**]{}, 092003 (2010) \[arXiv:1008.5119 \[astro-ph.HE\]\]. G. Di Bernardo, C. Evoli, D. Gaggero, D. Grasso, L. Maccione and M. N. Mazziotta, Astropart. Phys. [**34**]{}, 528 (2011) \[arXiv:1010.0174 \[astro-ph.HE\]\]. J. R. Ellis, R. A. Flores, K. Freese, S. Ritz, D. Seckel and J. Silk, Phys. Lett. B [**214**]{}, 403 (1988). S. Rudaz and F. W. Stecker, Astrophys. J. [**325**]{}, 16 (1988). M. Kamionkowski and M. S. Turner, Phys. Rev. D [**43**]{}, 1774 (1991). E. A. Baltz and J. Edsjö, Phys. Rev. D [**59**]{}, 023511 (1998) \[astro-ph/9808243\]. E. A. Baltz, J. Edsjö, K. Freese and P. Gondolo, Phys. Rev. D [**65**]{}, 063511 (2002) \[astro-ph/0109318\]. J. Kopp, arXiv:1304.1184 \[hep-ph\]. A. J. Tylka, Phys. Rev. Lett. [**63**]{}, 840 (1989) \[Erratum-ibid. [**63**]{}, 1658 (1989)\]. M. S. Turner and F. Wilczek, Phys. Rev. D [**42**]{}, 1001 (1990). E. A. Baltz and L. Bergström, Phys. Rev. D [**67**]{}, 043516 (2003) \[hep-ph/0211325\]. D. Hooper and G. D. Kribs, Phys. Rev. D [**70**]{}, 115004 (2004) \[hep-ph/0406026\]. E. A. Baltz and D. Hooper, JCAP [**0507**]{}, 001 (2005) \[hep-ph/0411053\]. L. Bergström, T. Bringmann and J. Edsjö, Phys. Rev. D [**78**]{}, 103520 (2008) \[arXiv:0808.3725 \[astro-ph\]\]. D. Hooper and W. Xue, Phys. Rev. Lett. [**110**]{}, 041302 (2013) \[arXiv:1210.1220 \[astro-ph.HE\]\]. T. Bringmann, F. Calore, G. Vertongen and C. Weniger, Phys. Rev. D [**84**]{}, 103525 (2011) \[arXiv:1106.1874 \[hep-ph\]\]. M. Ackermann [*et al.*]{} \[Fermi-LAT Collaboration\], Phys. Rev. Lett. [**107**]{}, 241302 (2011) \[arXiv:1108.3546 \[astro-ph.HE\]\]. S. Galli, F. Iocco, G. Bertone and A. Melchiorri, Phys. Rev. D [**80**]{}, 023505 (2009) \[arXiv:0905.0003 \[astro-ph.CO\]\]; T. R. Slatyer, N. Padmanabhan and D. P. Finkbeiner, Phys. Rev. D [**80**]{}, 043526 (2009) \[arXiv:0906.1197 \[astro-ph.CO\]\]; S. Galli, F. Iocco, G. Bertone and A. Melchiorri, Phys. Rev. D [**84**]{}, 027302 (2011) \[arXiv:1106.1528 \[astro-ph.CO\]\]; T. R. Slatyer, arXiv:1211.0283 \[astro-ph.CO\]; J. M. Cline and P. Scott, JCAP [**1303**]{}, 044 (2013) \[Erratum-ibid. [**1305**]{}, E01 (2013)\] \[arXiv:1301.5908 \[astro-ph.CO\]\]; L. Lopez-Honorez, O. Mena, S. Palomares-Ruiz and A. C. Vincent, JCAP [**1307**]{}, 046 (2013) \[arXiv:1303.5094 \[astro-ph.CO\]\]. See Appendix.
See, e.g., I. V. Moskalenko and A. W. Strong, Astrophys. J. [**493**]{}, 694 (1998) \[astro-ph/9710124\]; R. Schlickeiser, “Cosmic Ray Astrophysics” Springer (2002), and references therein. For an alternative discussion, see also B. Katz, K. Blum and E. Waxman, Mon. Not. Roy. Astron. Soc. [**405**]{}, 1458 (2010) \[arXiv:0907.1686 \[astro-ph\]\]. See R. Jansson, G. R. Farrar, A. H. Waelkens and T. A. Ensslin, JCAP [**0907**]{}, 021 (2009) \[arXiv:0905.2228\[astro-ph\]\]. and references therein.
M. S. Longair, “High Energy Astrophysics”, Vol. [**1**]{} pp. 346-350, Cambridge University Press (2004).
A. W. Strong, I. V. Moskalenko and V. S. Ptuskin, Ann. Rev. Nucl. Part. Sci. [**57**]{}, 285 (2007) \[astro-ph/0701517\]. M. S. Longair, “High Energy Astrophysics”, Vol. [**2**]{} pp. 274-286, Cambridge University Press (2004).
L. J. Gleeson and W. I. Axford, Astrophys. J. [**154**]{}, 1011 (1968). R. D. Strauss, M. S. Potgieter, I. Büsching and A. Kopp, Astrophys. J. [**735**]{}, 83 (2011).
L. Maccione, Phys. Rev. Lett. [**110**]{}, 081101 (2013) \[arXiv:1211.6905\[astro-ph.HE\]\]. A. W. Strong, I. V. Moskalenko, T. A. Porter, E. Orlando, S. W. Digel and A. E. Vladimirov “GALPROP v.54: Explanatory Supplement (2011)”; http://galprop.stanford.edu/ .
R. Trotta, G. Johannesson, I. V. Moskalenko, T. A. Porter, R. R. de Austri and A. W. Strong, Astrophys. J. [**729**]{}, 106 (2011) \[arXiv:1011.0037\[astro-ph.HE\]\]. T. Bringmann, F. Donato and R. A. Lineros, JCAP [**1201**]{}, 049 (2012) \[arXiv:1106.4821 \[astro-ph.GA\]\]. I. Cholis, M. Tavakoli, C. Evoli and P. Ullio, JCAP [**1205**]{}, 004 (2012) \[arXiv:1106.5073\[astro-ph.HE\]\]. T. R. Jaffe, A. J. Banday, J. P. Leahy, S. Leach and A. W. Strong, \[arXiv:1105.5885\[astro-ph.GA\]\]. D. Merritt, J. F. Navarro, A. Ludlow and A. Jenkins, Astrophys. J. [**624**]{}, L85 (2005) \[astro-ph/0502515\]. R. Catena and P. Ullio, JCAP [**1008**]{}, 004 (2010) \[arXiv:0907.0018 \[astro-ph.CO\]\]. P. Salucci, F. Nesti, G. Gentile and C. F. Martins, Astron. Astrophys. [**523**]{}, A83 (2010) \[arXiv:1003.3101 \[astro-ph.GA\]\]. M. Cirelli, G. Corcella, A. Hektor, G. Hutsi, M. Kadastik, P. Panci, M. Raidal and F. Sala [*et al.*]{}, JCAP [**1103**]{}, 051 (2011) \[Erratum-ibid. [**1210**]{}, E01 (2012)\] \[arXiv:1012.4515 \[hep-ph\]\]. W. A. Rolke, A. M. Lopez and J. Conrad, Nucl. Instrum. Meth. A [**551**]{}, 493 (2005) \[physics/0403059\]. M. Ackermann [*et al.*]{} \[Fermi LAT Collaboration\], Phys. Rev. D [**82**]{}, 092004 (2010) \[arXiv:1008.3999 \[astro-ph.HE\]\]. R. Bernabei [*et al.*]{} \[DAMA and LIBRA Collaborations\], Eur. Phys. J. C [**67**]{}, 39 (2010) \[arXiv:1002.1028 \[astro-ph.GA\]\]. C. E. Aalseth, P. S. Barbeau, J. Colaresi, J. I. Collar, J. Diaz Leon, J. E. Fast, N. Fields and T. W. Hossbach [*et al.*]{}, Phys. Rev. Lett. [**107**]{}, 141301 (2011) \[arXiv:1106.0650 \[astro-ph.CO\]\]. G. Angloher, M. Bauer, I. Bavykina, A. Bento, C. Bucci, C. Ciemniak, G. Deuter and F. von Feilitzsch [*et al.*]{}, Eur. Phys. J. C [**72**]{}, 1971 (2012) \[arXiv:1109.0702 \[astro-ph.CO\]\]. R. Agnese [*et al.*]{} \[CDMS Collaboration\], \[arXiv:1304.4279 \[hep-ex\]\]. M. T. Frandsen, F. Kahlhoefer, C. McCabe, S. Sarkar and K. Schmidt-Hoberg, JCAP [**1307**]{}, 023 (2013) \[arXiv:1304.6066 \[hep-ph\]\]. D. Hooper and T. Linden, Phys. Rev. D [**84**]{}, 123005 (2011) \[arXiv:1110.0006 \[astro-ph.HE\]\]. D. Hooper and T. R. Slatyer, arXiv:1302.6589 \[astro-ph.HE\]. K. N. Abazajian and M. Kaplinghat, Phys. Rev. D [**86**]{}, 083511 (2012) \[arXiv:1207.6047 \[astro-ph.HE\]\]. T. Bringmann, X. Huang, A. Ibarra, S. Vogl and C. Weniger, JCAP [**1207**]{}, 054 (2012) \[arXiv:1203.1312 \[hep-ph\]\]. F. Iocco, M. Pato, G. Bertone and P. Jetzer, JCAP [**1111**]{}, 029 (2011) \[arXiv:1107.5810 \[astro-ph.GA\]\]. S. Della Torre, P. Bobik, M. J. Boschini, C. Consolandi, M. Gervasi, D. Grandi, K. Kudela and S. Pensotti [*et al.*]{}, Adv. Space Res. [**49**]{}, 1587 (2012). A. W. Strong, I. V. Moskalenko and O. Reimer, Astrophys. J. [**537**]{}, 763 (2000) \[Erratum-ibid. [**541**]{}, 1109 (2000)\] \[astro-ph/9811296\]; A. W. Strong, I. V. Moskalenkoand O. Reimer, Astrophys. J. [**613**]{}, 962 (2004) \[astro-ph/0406254\]. V. S. Ptuskin, I. V. Moskalenko, F. C. Jones, A. W. Strong and V. N. Zirakashvilli, Astrophys. J. [**642**]{}, 902 (2006) \[astro-ph/0510335\]. E. Jones, T. Oliphant, P. Peterson, [*et. al.*]{}, [*[SciPy]{}: Open source scientific tools for [Python]{}*]{}, 2001.
F. James and M. Roos, Comput. Phys. Commun. [**10**]{}, 343 (1975). J. D. Hunter, Matplotlib: A 2D graphics environment, Comput. in Science & Engineering [**9**]{}, 90 (2007).
O. Adriani, [*et al.*]{} \[PAMELA Collaboration\], Phys. Rev. Lett. [**106**]{}, 201101 (2011). R. D. Strauss, M. S. Potgieter, I. Büsching and A. Kopp, Astrophys. & Sp. Sc. [**339**]{}, 223 (2012).
J. Arons, Space Sci. Rev. [**75**]{}, 235 (1996) \[arXiv:0811.3894 \[astro-ph\]\]. D. A. Green, \[arXiv:0905.3699 \[astro-ph\]\]. D. Malyshev, I. Cholis and J. Gelfand, Phys. Rev. D [**80**]{}, 063005 (2009) \[arXiv:0903.1310 \[astro-ph.HE\]\]. T. Kobayashi, Y. Komori, K. Yoshida and J. Nishimura, Astrophys. J. [**601**]{}, 340 (2004) \[astro-ph/0308470 \[astro-ph\]\]. D. Grasso [*et al.*]{} \[Fermi-LAT Collaboration\], Astropart. Phys. [**32**]{}, 140 (2009) \[arXiv:0905.0636 \[astro-ph.HE\]\]. R. N. Manchester, G. B. Hobbs, A. Teoh and M. Hobbs, Astrophys. J. [**129**]{}, 1993 (2005); http://www.atnf.csiro.au/people/pulsar/psrcat/ .
P. Gondolo, J. Edsjö, P. Ullio, L. Bergström, M. Schelke and E. A. Baltz, JCAP [**0407**]{}, 008 (2004) \[astro-ph/0406204\]; P. Gondolo, J. Edsjö, L. Bergström, P. Ullio, M. Schelke, E. A. Baltz, T. Bringmann and G. Duda, [http://www.darksusy.org]{}.
[@supp] Appendix {#appendix .unnumbered}
================
Here, we describe additional tests carried out in order to estimate the degree to which our DM limits might vary under alternative assumptions pertaining to the astrophysical background and cosmic ray propagation. In addition, we quantify the significance of spectral features in the observed positron fraction.
{width="0.49\linewidth"} {width="0.49\linewidth"}
In deriving our main results, as shown in Fig. 3, we used the phenomenological parameterization of the AMS collaboration [@Aguilar:2013qda] for the astrophysical contribution to the positron fraction, and adopted our reference assumptions of $L=4$ kpc and $U_{rad} + U_{B}
=1.7$ eV cm$^{-3}$. In Fig. \[fig:BackPropVariation\], for the case of direct DM annihilation to $e^{+}e^{-}$, we show in the *left panel* the impact of different propagation parameters when treating the astrophysical background in the same way as in Fig. 3. Changing the diffusion conditions ($L=2 - 8$ kpc) in the Galaxy in that case only affects our limits by $O(10\%)$, while allowing for higher energy losses ($U_{rad} + U_{B} =2.6$ eV cm$^{-3}$) can alter our limits by a factor of $\sim$2, with higher losses resulting in weaker limits (see also Fig. 1). In the *right panel*, we repeat this exercise, but replace the AMS background parametrization with physically motivated models for the primary $e^{-}$, secondary $e^{\pm}$, and pulsar originated $e^{\pm}$ fluxes (see discussion in the main text), calculated with the same galactic propagation model as used in determining the spectrum of CR leptons from DM. In this case, our results can be further altered by a factor of up to $\sim$3. The reason for this change is that our physically motivated models describe the individual components by power-laws with breaks at a few GeV. These spectral features in the background can be the result of different energy loss mechanisms kicking in,[^3] or from individual local and recent supernovae affecting the high energy $e^{-}$ spectrum. Also, observations at microwave and radio frequencies suggest a different spectral power-law for the CR $e^{\pm}$ at $\sim$1GeV [@StrongPapers; @Bringmann:2011py; @Jaffe:2011qw] compared to CR $e^{\pm}$ flux measurements at higher energies [@FermiLAT:2011ab; @Adriani:2011xv]. While changes in the spectral power-law describing these components are motivated by the reasons just described, sharp breaks used to implement them are theoretically less accurate and fit slightly worse the AMS positron fraction spectrum.
In addition, our physically motivated models include the impact of solar modulation by using the force field approximation. Solar modulation modifies the position and normalization of the dark matter signal flux, but is negligible at energies $>$5GeV.[^4] We do not expect solar modulation to significantly smoothen a sharp spectral peak at higher energies.
Given that we consider a population of pulsars as one possible source of the rising positron fraction above 10GeV (with TeV-scale DM or a single dominant pulsar being alternative possibilities), we will briefly discuss the impact of their modeling on our limits. For pulsars that eventually inject equal amounts of $e^{\pm}$ into the ISM, their injection spectra can be estimated from gamma-ray and synchrotron observations towards known pulsars, such as the Crab.[^5] Typical injection power-law values for the differential spectrum are expected to be in the range of 1-2 leading to propagated spectra with power-laws in the range of 2.0 $\pm$ 0.5. Our fits for the averaged pulsar contribution agree with these expectations.
![ The *black line* shows our nominal limit on $e^+e^-$ final states, obtained by adopting the power-law background model. The *gray lines*, in contrast, show limits obtained when the contribution from many pulsars is taken into account (for 15 different realizations).[]{data-label="fig:LimitsRand"}](limits_rand.eps){width="\linewidth"}
In addition, as suggested by Refs. [@Kobayashi:2003kp; @Malyshev:2009tw; @Grasso:2009ma], the total contribution from many pulsars – each with a different age, distance, initial rotational energy, injected energy into $e^{\pm}$, and unique environmental surroundings affecting energy losses and diffusion – is expected to give a spectrum with many peaks and dips, especially at higher energies where fewer pulsars significantly contribute. With fine enough energy resolution and high statistics, one should be able to observe such spectral features. By using the data from the ATNF pulsar catalogue [@ATNF] and implementing the parametrization of Ref. [@Malyshev:2009tw], we ran multiple realizations of such combined spectra to study the impact of possible dips and peaks in the background spectrum on the derived DM limits. In particular, we include in these realizations all pulsars within 4kpc from us, except for millisecond pulsars and pulsars in binary systems. While we keep their individual locations and ages in all realizations fixed, we vary i) the local CR diffusion properties and energy-losses, ii) the cuts on the current spin-down power of pulsars as recorded in the ATNF catalogue and, most importantly, iii) the fraction $\eta$ of initial rotational energy of the individual neutron stars that is injected into the ISM in the form of $e^{\pm}$.
We then fit to the AMS data the injection spectral properties (taken to be the same for all pulsars), the *averaged* value of $\eta$, the primary SNe $e^{-}$, secondary $e^{\pm}$ CR flux normalizations and the solar modulation potential. Even though the injection $e^{\pm}$ spectrum is taken to be the same for simplicity, however, the propagated pulsar spectra differ because of the different ages, distances and energy outputs; this is clearly seen in the inset of Fig. \[fig:LimitsRand\] where we plot the resulting positron fraction. For DM channels that give broad continuous spectra, such as muons and taus (see Fig. 1), the presence of multiple peaks and dips is unimportant. For hard spectra such as from monochromatic $e^{\pm}$, however, our limits can be modified by a factor of up to $\sim$3, as also shown in Fig. \[fig:LimitsRand\].
![Significance for a contribution from a $e^+e^-$ DM signal to the AMS-02 positron fraction, for different DM energies, in units of Gaussian sigma. Negative values correspond to negative (but unphysical) signal normalizations.[]{data-label="fig:TS"}](TS.eps){width="\linewidth"}
In Fig. \[fig:TS\] we show, for the case of $e^+e^-$ final states, the local significance for a DM signal as function of the DM mass. The significance is plotted in units of Gaussian sigma, and given by the square-root $\sqrt{TS}$ of the Test Statistics $TS=-2\log \mathcal{L}_{\rm null}/\mathcal{L}_{\rm
alt}$. Here, $\mathcal{L}_{\rm alt/null}$ denote respectively the likelihood of the alternative (DM signal) and null (no DM signal) hypothesis. For illustration, we also allow negative (obviously unphysical) signal normalizations in the fit, which are mapped onto the negative y-axis. As background model in the fit we use the reference power-law model from Ref. [@Aguilar:2013qda]. We do not find any indications for local, edge-like, features in the AMS data.
Lastly, as a simple cross-check, we have also run DarkSUSY [@Gondolo:2004sc] with standard parameters for propagation (based on the prescription given in Ref. [@Baltz:1998xv]) and an NFW profile normalized to 0.4 GeV cm$^{-3}$. For the electron spectrum, we used a simple broken power law which agrees with the PAMELA electron data [@Adriani:2011xv] for $E>5$GeV. Knowing that the AMS positron fraction measurement is well described by a simple background model, we then just demand that the DM signal does not exceed the reported $2\sigma$ error bars at the energy of the feature. The resulting limit curve agrees well with the more sophisticated treatment described in the main text.
[^1]: For $L=$ 2, 4, 8 kpc and rigidity $R$, we adopt a diffusion coefficient $D( R)= D_{0}
(\frac{R}{1 GV})^{0.5}$, with $D_{0} =$ 0.81, 1.90, 2.65 ($\times 10^{28}$ cm$^{2}$s$^{-1}$), and Alfvén velocities 9, 10, 10 km s$^{-1}$.
[^2]: By $e^+e^-\gamma$ we will always refer to that specific situation, dominated by photon emission off virtual selectrons $\tilde e$. We assume at least one of the $\tilde e$ to be degenerate in mass with the neutralino.
[^3]: At a few GeV the $e^{\pm}$ energy losses due to bremsstrahlung emission, dominant at lower energies, equal locally those due to synchrotron radiation and ICS (dominant at higher energies). Since the energy loss rate $dE/dt$ due to bremsstrahlung radiation scales as $E$ while the $dE/dt$ due to synchrotron and ICS as $E^2$ (at the Thompson cross-section regime), a spectral change in the propagated $e^{\pm}$ around that energy is expected (see, e.g., Ref… [@Bringmann:2011py]).
[^4]: In certain models, solar modulation can also affect the observed height of the peak in the positron fraction by changing the ratio of electrons-to-positrons of same energy before entering the Heliosphere [@Strauss2012; @Strauss2012B].
[^5]: Yet the uncertainties are still large due to a lack of exact understanding of local environments or the type of relevant supernova remnants (within which the pulsars exist); typically, the $e^\pm$ also get further accelerated at the termination shock between the magnetosphere and the pulsar wind nebulae (PWN), and the $e^{\pm}$ injected to the ISM are dominantly coming from middle aged pulsars after their respective PWN have been disrupted (see Refs. [@Arons; @Green:2009qf; @Malyshev:2009tw])
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Extensive Langevin dynamics simulations are used to characterize the adsorption transition of a flexible magnetic filament grafted onto an attractive planar surface. Our results identify different structural transitions at different ratios of the thermal energy to the surface attraction strength: filament straightening, adsorption and the magnetic flux closure. The adsorption temperature of a magnetic filament is found to be higher in comparison to an equivalent nonmagnetic chain. The adsorption has been also investigated under the application of a static homogeneous external magnetic field. We found that the strength and the orientation of the field can be used to control the adsorption process, providing a precise switching mechanism. Interestingly, we have observed that the characteristic field strength and tilt angle at the adsorption point are related by a simple power law.'
author:
- 'Pedro A. Sánchez'
- 'Ekaterina V. Novak'
- 'Elena S. Pyanzina'
- 'Sofia S. Kantorovich'
- 'Joan J. Cerdà'
- Tomás Sintes
title: Adsorption transition of a grafted ferromagnetic filament controlled by external magnetic fields
---
\[sec:intro\]Introduction
=========================
The combination of polymers and micro- or nanoparticles is one of the most successful available approaches for the design of novel materials with highly tunable properties [@2006-balazs; @2019-alexandrov-ptrsa]. One of the main and simpler examples is the filling of polymer matrices with magnetic particles to create magnetoresponsive gels and elastomers whose mechanical properties can be changed on the fly by means of external fields[@2007-filipcsei; @2016-odenbach-aam; @2017-lopez-lopez-pre; @2017-lopez-lopez-pa; @2018-shamonin]. At the nanoscale, polymer coatings are broadly used to stabilize magnetic nanoparticles in suspension as an essential ingredient for the synthesis of ferrofluids and the fine tuning of their static and dynamic properties[@2006-odenbach; @2009-vekas; @2018-ivanov-pre; @2019-lebedev-pre].
A more sophisticated approach for the synthesis of hybrid polymer magnetic materials is the polymer crosslinking of assembled magnetic particles in order to stabilize specific structures. The simplest case is the stabilization of the linear chains that magnetic micro- and nanoparticles tend to form under uniform static external fields. Such linear micro- and nanostructures, often addressed as magnetic filaments (MF), can be used in numerous applications that take advantage of their high magnetic response and shape anisotropy [@2005-cebers; @2011-wang; @2016-cebers-afm]. For instance, MFs have been used for the design of magnetically actuated artificial propellers and swimmers [@2005-dreyfus; @2009-belovs-pre; @2010-benkoski], micromechanical sensors [@2003-goubault], microchannel actuators and mixers [@2009-fahrni; @2020-zaben-sm] or magnetic resonance contrast agents [@2008-corr].
A large part of the aforementioned applications of MFs involve two important aspects: a certain degree of flexibility of the chain backbone and its sensible interaction with rigid surfaces, being both particularly important for micro- and nanofluidic applications. Regarding the first aspect, current synthesis techniques allow polymer crosslinking of paramagnetic microparticles with a high control of the degree of flexibility of the resulting filament [@2014-byrom]. Due to their smaller size, control on the crosslinking of monodomain ferromagnetic nanoparticles is much more difficult and most attempts to date have achieved rather rigid structures only [@2007-benkoski-jacs; @2009-zhou; @2014-hill]. However, it has been shown already that it is possible to create flexible noncrosslinked chains of polymer coated ferromagnetic nanoparticles [@2014-townsend] and cutting-edge synthesis techniques, such as polymer templating [@2012-sarkar] or DNA directed self-assembly [@2015-tian-nn; @2016-tian-nm], are paving the way to the creation of highly flexible nanofilaments of monodomain superparamagetic and ferromagnetic nanoparticles. Interactions of MFs with surfaces are also widespread among their technological applications. In many cases, the filaments are end-grafted to the surface of a larger particle (for instance, in artificial swimmers) or wall (pumpers, mixers). In addition, the structural similarity of MFs with molecular polymers has inspired their use in dense polymer brush-like arrangements in order to create magnetoresponsive coatings [@2008-choi; @2010-ye; @2015-sanchez-mm1; @2016-sanchez-fd; @2017-pyanzina-sm; @2019-cerda-sm]. To this regard, permanently stabilized and grafted nanoscale flexible filaments can broaden the already promising potential for applications of brush-like systems of simple magnetically assembled chains, hybrid polymer microfibers or rigid magnetic micropillars[@2010-vilfan; @2014-tokarev; @2015-sun-fms; @2016-orlandi-pre; @2018-hanasoge-mn].
Following many of the applications mentioned above, most theoretical studies on surface grafted MFs to date have been focused on their magnetoelastic response and hydrodynamic interactions with the background fluid, whereas the grafting surface only played the role of an inert geometric constraint [@2003-goubault; @2009-fahrni; @2009-gauger; @2013-chen; @2011-babataheri; @2017-vazquez-montejo-prm; @2017-dempster-pre]. However, one can think in very interesting applications involving noninert surfaces. For instance, attractive surfaces in the walls of a microfluidic channel can adsorb substances carried by the flowing liquid, acting as a filter. The presence of grafted MFs also experiencing the attraction to the surface would provide a switching mechanism on the adsorbing properties of the walls: in absence of external fields the filaments would remain adsorbed, becoming a steric barrier for the adsorption of free flowing components; application of adequate external fields that could force the desorption of the MFs under given flow conditions would activate the adsorption of the former. In order to design a system as such, first it is essential to understand the adsorption process of a grafted MF on an attractive surface and how this can be controlled by means of external fields. This is the main goal of this work.
Several years ago we presented the first theoretical study on the adsorption transition of a free semiflexible MF on a flat surface in absence of external fields [@2011-sanchez-sm]. Even closely related to the topic studied here, to our best knowledge the field-induced adsorption/desorption of MFs has never been addressed before.
Here we employ computer simulations with a mesoscale model to study the equilibrium behavior of a flexible filament made of ferromagnetic particles grafted to an attractive flat surface. By means of molecular dynamics in the canonical ensemble, we first study the adsorption transition that takes place as the ratio of the thermal fluctuations to the strength of the surface attraction decreases, discussing how it compares to the adsorption of an equivalent nonmagnetic filament. Second, we study how such transition is affected by the application of homogeneous magnetic fields of different strengths and orientations with respect to the surface plane. We show that both, the strength and the orientation of the field can be used to drive the adsorbed or desorbed state of the filament, providing a precise switching mechanism.
\[sec:model\]Simulation model and methods
=========================================
![Scheme of the grafted magnetic filament model, showing a configuration equilibrated under a strong applied field, $\vec H$, tilted an angle $\theta_H$ with respect to the normal of the grafting surface. Ferromagnetic particles of the filament are represented as two-color spheres, with colors indicating the orientation of their central dipole moment. See the main text for further details.[]{data-label="fig:model"}](model.pdf){width="5.5cm"}
Due to the characteristic lengths and time scales involved in hybrid materials that combine polymers and micro- or nanoparticles, simulation models for such systems have to rely on coarse-grained approximations. Numerous computer simulations of magnetic gels, elastomers and filaments are based on bead-spring representations with different levels of detail [@2011-wood-pre; @2013-annunziata-jcp; @2014-tarama-pre; @2015-weeber-jmmm; @2016-pessot-jcp; @2018-weeber-jpcm]. Regarding particle-based simulation of MFs, since the main role of the polymer components is to provide the permanent linking of the magnetic particles in the chain, the most convenient approach is to represent them implicit by means of simple elastic bonding potentials, whereas the particles are usually simulated as beads with point magnetic dipoles [@2011-sanchez-sm; @2015-pshenichkov-pre; @2016-wei-lm; @2018-kuznetsov-jmmm; @2020-mostarac-jml]. Our mesoscale model of the grafted ferromagnetic filament is based on such approach and is very similar to those used in our previous studies on these systems [@2011-sanchez-sm; @2015-sanchez-mm1; @2016-sanchez-fd; @2017-pyanzina-sm; @2019-cerda-sm].
Figure \[fig:model\] shows a sketch of our model system and its interactions. We consider only the case of a filament formed by $N_p=30$ identical ferromagnetic particles. The latter are modeled as spherical beads of characteristic diameter $\sigma$, carrying a point magnetic dipole $\vec \mu$ at their centers. Note that this is an accurate description for spherical monodomain ferromagnetic particles. In addition, we adopt the limit of infinite magnetic anisotropy for these particles, so that their dipole moments have not only a constant modulus but also a fixed orientation with respect to the particle’s body frame.
The steric interactions between these particles, as well as their attractive interaction with the grafting surface, are represented by means of truncated Lennard-Jones (LJ) type potentials, shifted to make them smoothly vanish at the truncation point, $r_{\mathrm{cut}}$: $$\begin{gathered}
U_{\mathrm{{t-s}}}(r; \epsilon, \sigma, r_{\mathrm{cut}})=\\
=\left\{ \begin{array}{ll}
U_{\mathrm{{LJ}}}(r; \epsilon, \sigma)-U_{\mathrm{{LJ}}}(r_{\mathrm{cut}}; \epsilon, \sigma), & r<r_{\mathrm{{cut}}}\\
0, & r\geq r_{\mathrm{{cut}}}
\end{array}\right. .
\label{eq:LJts}\end{gathered}$$ where $\epsilon$ is the energy scale of the interaction and $r$ is the characteristic separation distance (center-to-center for particle pairs or center-to-surface for particle-surface interactions). In order to model a purely repulsive steric interaction between the particles as the one produced by a soft polymer coating, we take the conventional Lennard-Jones potential, $$\begin{gathered}
U_{\mathrm{LJ}}(r; \epsilon_{\mathrm{c}}, \sigma)= \\
= U^{12-6}_{\mathrm{LJ}}(r; \epsilon_{\mathrm{c}}, \sigma) = 4 \epsilon_{\mathrm{c}} \left [ \left ( \frac{\sigma}{ r}\right )^{12} - \left ( \frac{\sigma} {r}\right )^{6} \right ],
\label{eq:LJ}\end{gathered}$$ truncated at the position of its minimum, $r_{\mathrm{cut}} = 2^{1/6} \sigma$. The combination (\[eq:LJts\]) and (\[eq:LJ\]) corresponds to the soft-core interaction known as Weeks-Chandler-Andersen (WCA) potential [@1971-weeks]. For the interaction with the surface we apply expression (\[eq:LJts\]) to a 9–3 LJ potential, $$\begin{gathered}
U_{\mathrm{LJ}}(r; \epsilon_{\mathrm{s}}, \sigma)= \\
= U^{9-3}_{\mathrm{LJ}}(r; \epsilon_{\mathrm{s}}, \sigma) = \frac{3\sqrt{3}}{2} \epsilon_{\mathrm{s}} \left [ \left ( \frac{\sigma}{ r}\right )^{9} - \left ( \frac{\sigma} {r}\right )^{3} \right ],
\label{eq:LJ93}\end{gathered}$$ which is the result of integrating the conventional potential (\[eq:LJ\]) over an infinite plane. In this case we make the interaction attractive by taking $r_{\mathrm{cut}} = 3.5 \sigma$. Despite the truncation and shifting of attractive Lennard-Jones type potentials introduces a discontinuity in their derivatives at the truncation point, the use of such large cutoff, correponding to the maximum value in the range most frequently used in simulations, provides a discontinuity small enough to ensure that its effects will be negligible in front of the thermal fluctuations, at least for the range of temperatures of interest sampled here. Note that the resulting potential, which we label as $U_{\mathrm{s}}$, has a well whose minimum is located at $r_{\mathrm{min}}=3^{1/6} \sigma$ and, after applying the shift corresponding to the selected truncation, its depth is $\bar \epsilon_{\mathrm{s}} = -\epsilon_{\mathrm{s}} \left [ 1 - 3\sqrt{3}(3.5^{-9} - 3.5^{-3} )/2\right ] \approx -1.06 \epsilon_{\mathrm{s}}$. In the following we will discuss the strength of the attraction to the surface in terms of $\bar \epsilon_{\mathrm{s}}$.
We assume the chain structure of our filament to be stabilized by long polymer crosslinks attached to a very narrow region of the surface of the linked particles. Under these conditions the filament backbone is flexible and the crosslinks can be modeled as a simple finitely extensible nonlinear elastic (FENE) potential. This bonding potential is defined as [@1986-grest] $$U_{\mathrm{FENE}}(r; K, r_{\mathrm{max}}) = -\frac{1}{2} K r_{\mathrm{max}}^2 \ln \left [ 1 - \left ( \frac{r}{r_{\mathrm{max}}} \right ) ^2\right ] ,$$ where $K$ defines the elastic strength of the bond and $r_{\mathrm{max}}$ its maximum extension. As shown in the sketch of Figure \[fig:model\], the FENE springs are attached to points of the particle’s surfaces located at the projections of the head and the tail of their central dipoles. This corresponds to the crosslinking of a chain of particles assembled into a head-to-tail configuration by the presence of a homogeneous external field. Therefore, the orientation of the dipole moments of the particles is coupled to the filament backbone due to the crosslinks.
Each pair of magnetic dipoles $\vec{\mu}_{i}$ and $\vec{\mu}_{j}$ experiences the conventional long-range dipole-dipole pair interaction $$U_{\mathrm{dip}}(\vec r_{ij};\ \vec \mu_i,\ \vec \mu_j)=\frac{\vec{\mu}_{i}\cdot\vec{\mu}_{j}}{r^{3}}-\frac{3\left[\vec{\mu}_{i}\cdot\vec{r}_{ij}\right]\left[\vec{\mu}_{j}\cdot\vec{r}_{ij}\right]}{r^{5}},
\label{eq:dipdip}$$ where $r = \| \vec r_{ij} \|$ is the displacement vector between the the centers of the corresponding particles. Finally, each dipole moment also experiences a Zeeman interaction with applied external fields. In general, the interaction of a point magnetic dipole with a net magnetic flux density $\vec B$ at the position of the former is $U_{\mathrm{Z}}(\vec \mu, \vec B)=-\vec \mu \cdot \vec B$. Here, however, we only need to consider the contribution to $\vec B$ of an applied external field of strength $\vec H$. The approximations assumed in the modeling of the magnetic properties of our particles (point dipoles of fixed modulus, independent of the applied field, and infinite magnetic anisotropy) allows us to simply write $$U_{\mathrm{Z}}(\vec \mu, \vec H)=-\vec \mu \cdot \vec H,
\label{eq:zeeman}$$ provided a convenient unit rescaling is used for these parameters. In this way, in Eqs. (\[eq:dipdip\]) and (\[eq:zeeman\]) we expressed the magnetic interactions in our system in terms of an external control parameter, the applied field intensity $\vec H$, and an extensive effective parameter, the dipole moment $\vec \mu$, which incorporates the specific properties of the material forming the particles.
As is usual in simulations with mesoscale models, we use a set of reduced, *i.e.*, dimensionless units. By choosing scales that keep the numerical values not too far from unity the stability of the calculations is enhanced and, importantly, in this way the same model may represent very different systems, as long as the ratios between the distinct interaction strengths remain the same. Here we define lengths and masses in units of the diameter and mass of the beads, so we take $\sigma=1$ and $m=1$. Energy scale, $\epsilon^*$, is given by the strength of the thermal fluctuations at room temperature, $T^*$, so that $\epsilon^*=kT^*$, being $k$ the Boltzmann constant. In order to simplify the notation, henceforth we will use the reduced temperature, $T$, to represent the strength of the thermal fluctuations. Therefore, $T=1$ under room conditions. Time scale is related to the parameters above as $\tau^* = \sigma^* (m^* / \epsilon^*)^{1/2}$, however, since here we are only interested in equilibrium properties, this scale is not relevant for the discussion. The strength of the dipole-dipole interaction is defined naturally by the squared dipole moment of the particles, which we set either to $\mu^2=0$ (nonmagnetic particles) or $\mu^2=5$. Note that scales of dipole moment and applied field strengths are defined by $(4\pi\epsilon^* (\sigma^*)^3/\mu_0)^{1/2}$ and $m^* (\tau^*)^{-2}(\mu_0 \sigma^* /4\pi\epsilon^*)^{1/2}$, respectively. The parameters of the bonding interaction are set to $K=30$, $r_{\mathrm{max}}=0.5$, which provide an average center-to-center distance between linked particles of approximately $\sim$1 at $T=1$. We sample strengths of the attraction to the surface within the interval $\bar \epsilon_{\mathrm{s}} \in [0.14,\, 2.89]$. Finally, we analyze the influence of the temperature and external field on the adsorption of the filament on the grafting surface by sampling respective ranges of temperatures and external field strengths $T\in [0.25,\, 5]$ and $H \in [0, 2] $. This choice of parameters could correspond, for instance, to filaments made of magnetite nanoparticles of $\sim$35 nm in diameter coated with a repulsive soft layer of $\sim$6.5 nm and exposed to external fields of up to $\sim$3 kA/m.
The parameters described above have been sampled by performing molecular dynamics (MD) simulations with a Langevin thermostat. The latter treats implicitly the effects of the thermal fluctuations of the background fluid by introducing stochastic forces and friction terms, satisfying the conventional fluctuation-dissipation rules, in the translational and rotational Newtonian equations of motion [@1986-grest; @1987-allen]. The latter have been integrated by means of a velocity Verlet scheme. Since here we are interested in equilibrium properties only, hydrodynamic interactions have not been taken into account.
Despite the simplicity of the system studied here, statistical sampling of transitions of polymer-like structures is in general rather demanding. In order to efficiently improve the statistics, we used the replica exchange molecular dynamics (REMD) method [@1999-sugita; @2001-mitsutake]. In this approach, $N$ independent simulations of the same system (replicas) are run in parallel, each with one value of a given parameter, $A$, from an ordered set: $A=\{ A_1,\, A_2,\, \dots,\, A_N\}$, where $A_1 < A_2 < \dots < A_N$. After equilibration of each replica, an attempt to exchange the configurations with adjacent values $A_i$, $A_{i+1}$ is performed according to the Boltzmann probability $P(A_i,\, A_{i+1})$, defined as $$\begin{gathered}
P(A_i,\, A_j) =
\min\left (1,\, \exp \left [\frac{U_{i}(A_{i})}{A_{i}} + \right. \right.\\ + \frac{U_{i+1}(A_{i+1})}{A_{i+1}}
\left. \left. - \frac{U_i(A_{i+1})}{A_{i+1}} - \frac{U_{i+1}(A_{i})}{A_{i}} \right ] \right )
\label{eq:exchangeprob}\end{gathered}$$ where $U_a(A_b)$ is the potential energy of the configuration equilibrated under $A=A_a$ but calculated by taking $A=A_b$. This exchange procedure, which is intended to prevent the system to get trapped into local minima, requires the energy histograms of adjacent replicas to have a significant overlap in order to be effective. However, in general one wants to run the least possible number of replicas that span the range of interest of $A$ in order to minimize the computational load. As a reasonable compromise, we chose sets of parameter values that provided overlaps of about 30% of the histograms area. Even though REMD simulations are mainly used to simulate systems under different temperatures, this technique can be used to study any parameter affecting the internal energy. In reference [@2013-sanchez-a] we used REMD to study the influence of the dipole moment on the equilibrium configurations of a filament in bulk. Here, we applied this approach for the sampling of the sets of temperatures and external fields mentioned above. Specifically, we performed REMD simulations separately for different temperatures at zero field and for different fields strengths and orientations at $T=1$.
REMD simulations are naturally well suited for an additional statistical refinement: the weighted histogram analysis method (WHAM) [@1989-ferrenberg-prl; @1992-kumar]. This technique combines statistics from simulations at different values of the parameters by weigthing them according to their thermodynamic probability. Its most widespread application corresponds to the combination of statistics from a set of canonical simulations performed at different inverse temperatures. In our notation this corresponds to $\beta_i=T_i^{-1}=\lbrace \beta_1, \,, \dots,\, \beta_N \rbrace$. After equilibration, each simulation provides a set of $M_i$ measures of the internal energy of the system, $E$, with a correlation length $\tau_i$. The histograms of such measures are estimates of the probability distributions of energy values: $$p_i(E) = \frac{h_i(E)}{ \Delta E M_i},$$ where $h_i(E)$ is the number of measurements of energies within the interval $[E-\Delta E /2,\, E+\Delta E/2)$. The true canonical distribution is actually $$p_i(E) = g(E) e^{-\beta_i E + f_i},$$ where $g(E)$ is the density of states, a probability density function that describes the number of configurations with energy $E$ that the system may adopt, and $f_i= - \log Z_{\beta_i}$ is the dimensionless free energy. Therefore, each simulation provides an estimate of the density of states. Individually, such estimates are only accurate within the limited range of energies on which the corresponding histogram has significant values. However, it is possible to make a weighted combination of all the estimates in order to obtain a better approximation: $$g(E) = \sum_{i=1}^N w_i(E) p_i(E) e^{-\beta_i E + f_i},$$ where the weights $w_i(E)$ should fulfill $\sum_{i=1}^N w_i(E) =1$ for all $E$. Such weights have to be chosen to minimize the uncertainty associated to the histograms. In the simplest scheme, the minimization assumes a Poisson distribution for the histograms and takes into account the correlations of the measures, leading to the pair of expressions $$g(E) = \frac{\sum_{i=1}^N l_i h_i(E)}{\sum_{j=1}^N \Delta E M_j l_j e^{-\beta_jE + f_j}}$$ and $$e^{-f_i} = \sum_E g(E) e^{-\beta_i E},$$ where $l_i = (1+2\tau_i)^{-1}$. This set of equations can be solved numerically to determine $g(E)$, for instance by self-consistent iteration. Once the density of states is known, expectation values of any observable of the system, $\langle O \rangle$, at any inverse temperature $\beta^*$ can be calculated as $$\langle O \rangle_{\beta^*} = \frac{\sum_E g(E) O(E) e^{-\beta^* E}}{\sum_E g(E) e^{-\beta^* E}}.$$ Note that this expression will provide good estimations for any arbitrary $\beta^*$ within the range of sampled values, *i.e.*, its estimations are not limited to the discrete set of temperatures used in the simulations. This method will be applied in next Section to obtain finely resolved curves of the adsorption energy and the structural parameters of the filament as a function of the temperature.
The simulation protocol consisted of different MD steps. First, random initial configurations of the grafted chain were prepared for each replica by performing $5\cdot10^6$ integrations of damped dynamics in absence of magnetic interactions and at temperature $T=2$. The damping, achieved by setting the translational and rotational friction constants, $\Gamma_T$ and $\Gamma_R$, to $\Gamma_T=\Gamma_R=50$, helps to fastly relax artificial initial configurations without the need of a very small time step. The latter was fixed for the whole protocol at $\delta t=10^{-3}$. At this point is important to underline that, since we are only interested in equilibrium properties, the choice of $\Gamma_T$ and $\Gamma_R$ is physically irrelevant and, therefore, it can obey to considerations of simplicity and computational efficiency. Except for the initial damped MD step, we chose to fix $\Gamma_T=\Gamma_R=1$ for the subsequent cycles, as these values are known for providing a fast equilibration in this type of systems. After such initial setup, the final parameters for each replica (temperature and external field) were set, the calculations of the magnetic interactions were enabled and a large set of main equilibration-measures-exchange cycles was started. In each of these cycles, the equilibration consisted in $10^6$ integrations. Measures of the system configurations from each replica were stored during the next $2\cdot10^6$ integrations at intervals of $5\cdot10^5$ integrations. This large amount of integrations between measures ensures small correlations even at low temperatures. Finally, the attempt of configuration exchange for adjacent replicas was carried out. For each simulation set, at least $10^3$ cycles of equilibration-measures-exchange were performed. All the simulations were made with the simulation package [ESPResSo]{} 3.3.1 [@2013-arnold].
\[sec:results\]Results and discussion
=====================================
In the next two Sections we present the simulation results of the equilibrium properties of our model MF grafted to an attractive surface. First, we characterize its adsorption transition on cooling in absence of external field, comparing its behavior to the case of an equivalent nonmagnetic chain. Next, we analyze how the transition of the MF can be favored or hindered by a homogeneous static external field depending on its strength and orientation.
### Adsorption transition at zero field
Computer simulations allow to easily access every component of the system internal energy. Therefore, we can characterize the adsorption transition by computing the normalized total energy of interaction with the surface, $$\bar U_{\mathrm{s}} = \frac{1}{N_p \bar \epsilon_\mathrm{s}} \sum_{i=1}^{N_p} U_{\mathrm{s}}(r_i; \bar \epsilon_{\mathrm{s}}),$$
where the sum applies over each particle forming the filament and $U_{\mathrm{s}}(r_i; \bar \epsilon_{\mathrm{s}})$ is given by Eqs. (\[eq:LJts\]) and (\[eq:LJ93\]). Note that the normalization makes this parameter to have a strict lower boundary $\bar U_{\mathrm{s}} \ge -1$. Another useful quantity we can compute is the ratio of the fluctuations of the adsorption energy to the squared thermal energy, normalized by the number of filament beads, $$c_V = \frac{\langle \bar U^2_{\mathrm{s}}\rangle - \langle \bar U_{\mathrm{s}}\rangle^2 }{N_p T^2}.$$ We label such quantity as $c_V$ due to its analogy with a specific heat. We expect $c_V$ to have a maximum at the characteristic temperature of each transition.
{width="18.1cm"}
Figure \[fig:ads-M5\] shows adsorption transition curves of the magnetic chain ($\mu^2=5$) for different strengths of surface attraction, $\bar \epsilon_{\mathrm{s}}$. They correspond to the average surface energy, $\langle \bar U_{\mathrm{s}} \rangle$, measured as a function of the temperature by means of WHAM calculations on the simulation data. We can see that at high temperatures $\langle \bar U_{\mathrm{s}} \rangle$ tends to display a plateau whose value increases with decreasing $\bar \epsilon_{\mathrm{s}}$. Under such conditions the thermal fluctuations are too strong to let the adsorption take place and the most entropically favorable configurations for the filament are those that minimize its contact with the surface. However, the grafted end particle is constrained to remain within its interaction range in any case. This produces the $\bar \epsilon_{\mathrm{s}}$-dependent bias in the surface energy observed for the high temperature limit. With decreasing $T$, the onset of the adsorption transition can be identified by the corresponding drop of $\langle \bar U_{\mathrm{s}} \rangle$, being more abrupt and taking place at lower temperatures as the surface interaction strength weakens. The scaled fluctuations corresponding to each curve, shown in the inset of Figure \[fig:ads-M5\], have a well defined peak in all cases but for the weakest surface attraction. The same exception can be observed in the trend of the main curves within the region of low temperatures: $\langle U_{\mathrm{s}} \rangle$ tends to approach its lower boundary in all cases except for $\bar \epsilon_{\mathrm{s}} \approx 0.14$, which is significantly lower than the lowest sampled strength of thermal fluctuations, $T=0.25$. This simply reflects the fact that the adsorption transition can only take place at $T$ values comparable to $\bar \epsilon_{\mathrm{s}}$, which can be evidenced by obtaining the characteristic adsorption temperature, $T_{\mathrm{ads}}$, corresponding to each $\bar \epsilon_{\mathrm{s}}$ from the positions of the peaks of $c_V$.
The dependence of $T_{\mathrm{ads}}$ on $\bar \epsilon_{\mathrm{s}}$ for both, magnetic and nonmagnetic chains, is presented in Figure \[fig:ads-diagram\]. We can see that, at least for the range of values sampled here, there is a rather linear relationship between these parameters independently of the magnetic or nonmagnetic nature of the filament. However, the adsorption of the MF is observed at slightly higher temperatures than its nonmagnetic counterpart. This can be attributed to the increased backbone stiffness led by the dipole-dipole interactions [@2013-sanchez-a]. Without regard the origin of the backbone rigidity, semiflexible polymer-like chains are known to adsorb on attractive surfaces at higher temperatures than their flexible counterparts [@1979-birshtein-bp; @2001-sintes; @2013-hsu-mm] due to their lower configurational entropy and, thus, to their lower average entropic repulsion with walls. Here, the anisotropic nature of the dipole-dipole interaction and the coupling between the dipole orientations and the chain backbone can lead only to a decrease of the chain configurational entropy with respect to the nonmagnetic case.
The adsorption transition of a filament involves the change from three- to two-dimensional structures. However, this is not the only structural change that this type of chain-like systems experience on cooling. First, the drecrease of thermal fluctuations tends to reduce the stretching and bending of the bonds, making the backbone locally smoother [@2011-sanchez-sm]. As pointed above, in MFs the latter is favored by the dipole-dipole interactions. Besides, MFs also may experience a magnetic flux closure transition on cooling, changing from open to ring-like structures [@2011-sanchez-sm; @2013-sanchez-a]. In order to determine how these three effects interact in our system, we computed two standard structural parameters: the radius of gyration $$R_g = \left [ \frac{1}{N_p} \sum_{i=1}^{N_p} (\vec r_i - \langle \vec r \rangle)^2\right ]^{1/2} ,
\label{eq:Rg}$$ where $\langle \vec r \rangle = \sum_{k=1}^{N_p} \vec r_k / N_p$, and the end-to-end distance $$R_{ee} = \left \| \vec r_1 - \vec r_{N_p}\right \| ,$$ calculated from the positions of the filament particles, $\vec r_i$, being $\vec r_1$ and $\vec r_{N_p}$ those corresponding to the chain ends. Note that for a MF with dipole moments coupled to the chain backbone, $R_{ee}$ is basically proportional to its net magnetization [@2011-sanchez-sm; @2013-sanchez-a]. Figure \[fig:rads-nofield\] shows the dependence of the averages of these parameters on $T$ for both cases, $\mu^2=0$ and $\mu^2=5$, and for different surface attraction strengths. Such averages have been obtained also from WHAM calculations. At high temperatures both, MFs and nonmagnetic chains in a desorbed state, adopt a random coil structure. As the temperature is decreased the chains experience an important straightening, which is evidenced by the significant growth of $\langle R_g \rangle$ and $\langle R_{ee} \rangle$ and illustrated respectively by the snapshots (iv), (iii) and (ii) in Figure \[fig:rads-nofield\]. Nonmagnetic chains keep experiencing such straightnening as the temperature is further reduced to its minimum sampled value, even after the characteristic adsorption temperature, $T_{\mathrm{ads}}$, is reached (see snapshot (i) in the same Figure). Straightening of MFs, however, happens only for temperatures above a certain value. At temperatures below such limit, however, they show an abrupt drop of both, $\langle R_g \rangle$ and $\langle R_{ee} \rangle$, associated to a prominent peak in the fluctuations of these parameters (shown in the insets). Such drop signals the adoption of a closed loop structure driven by the dipole-dipole interactions, as illustrated by snapshots (v) and (vi). Taking the position of the fluctuation peaks as the characteristic temperature of such closure transition, $T_{\mathrm{dip}}$, we can see that its value is practically constant for all adsorbed configurations, $T_{\mathrm{dip}} \approx 0.6$. In the case of the weakest sampled adsorption strength, $\bar \epsilon_{\mathrm{s}} \approx 0.14$, for which the full adsorption was not reached within the sampled interval of temperatures, the desorbed MF also experiences a closure transition but at a lower temperature, $T_{\mathrm{dip}} \approx 0.5$. This can be explained by the higher configurational entropy of the desorbed filament compared to its adsorbed state.
![Average parallel (top panel) and perpendicular (bottom panel) components of the radius of gyration of the grafted MF ($\mu^2=5$) as a function of the temperature for different surface attraction strengths. Inset in top panel shows the fluctuations of the parallel component.[]{data-label="fig:splitrads"}](splitrads-nofield.eps){width="8.5cm"}
The separation of the two transitions of the MF, adsorption and closure, found under not too weak surface attraction conditions, can be better visualized by splitting the components of the radius of gyration parallel and perpendicular to the surface, $R_{g,\parallel}$ and $R_{g,\perp}$. These parameters are obtained from Eq. (\[eq:Rg\]) by taking the corresponding components of the position vectors. Figure \[fig:splitrads\] shows their WHAM averages as a function of the temperature for each surface attraction strength. We can see that for $\bar \epsilon_{\mathrm{s}} \gtrsim 0.72$ the parallel component captures the initial chain straightening and its subsequent closure as $T$ is decreased, whereas the perpendicular component reflects the adsorption transition. The comparison of these curves evidences that both, straigthening and adsorption happen simultaneously for this range of parameters. Interestingly, the fluctuations of each $R_{g,\parallel}$ curve show two peaks, which signal the characteristic adsorption and closure temperatures. As expected, the trace of the adsorption transition is absent from the curves of system $\bar \epsilon_{\mathrm{s}} \approx 0.14$. On cooling, it also displays a much weaker straightening before the onset of its closure.
In summary, our results indicate that the closure transition of the MF is independent of its adsorption transition for values of $\bar \epsilon_{\mathrm{s}}$ large enough to impose $T_{\mathrm{ads}} > T_{\mathrm{dip}}$. Under such conditions the closure takes place with a two-dimensional constraining of the filament and we observe $T_{\mathrm{dip}} = \left ( T_{\mathrm{dip}} \right )_{\mathrm{adsorbed}} \approx$ const. For low values of $\bar \epsilon_{\mathrm{s}}$ one can expect the adsorption transition to take place for a MF already in its closed state, so that $T_{\mathrm{ads}} < T_{\mathrm{dip}} = \left ( T_{\mathrm{dip}} \right )_{\mathrm{desorbed}}$, with $\left ( T_{\mathrm{dip}} \right )_{\mathrm{desorbed}} < \left ( T_{\mathrm{dip}} \right )_{\mathrm{adsorbed}}$. Besides these qualitative considerations, the accurate characterization of structural transitions at very low temperatures may require more refined simulation approaches, being out of the scope of this work.
### Adsorption transition under tilted fields
Once the structural behavior displayed by this system on cooling has been characterized in detail, we address the main point of this work: the control of the adsorption/desorption of the MF at constant temperature by means of static homogeneous external fields. In general, such fields will tend to align each individual dipole in their direction, leading to an overall straightening and orientation of the chain backbone, thus, reducing its configurational entropy and increasing its effective stiffness. Qualitatively, in one hand one can expect the external field to hinder the closure transition of the MF while, on the other hand, its presence may favor or even force the adsorption or desorption of the MF depending on its strength and orientation: a field with strong enough component pointing out of the plane of the attractive surface can force the desorption, whereas a strong field component pointing into or parallel to such plane may favor the adsorption. However, the decrease in the configural entropy of the filament induced by the field makes difficult to anticipate its quantitative effects on the adsorption/desorption transition. Thus, it is necessary to characterize such effects in order to understand how the field can be used to control such transition.
In the following discussion, we take $T=1$ as fixed reference temperature and we consider only surface attraction strengths that led to an adsorption within the sampled range of temperatures, *i.e.*, $\bar \epsilon_{\mathrm{s}} \gtrsim 0.72$. All parameters presented below are calculated from direct sampling averages.
![Average adsorption energy curves as a function of the field strength, $H$, for different field tilting angles, $\theta_H$, and two selected values of $\bar \epsilon_{\mathrm{s}}$. Dotted lines connecting the symbols and horizontal dashed lines are a guide to the eye. The latter correspond to the values of $\langle \bar U_{\mathrm{s}} \rangle$ at zero field.[]{data-label="fig:ads-field"}](adsorption-field-eps.eps){width="8.5cm"}
We start our analysis by examining the average surface energy as a function of field strength and orientation. Figure \[fig:ads-field\] shows such results for two selected values of $\bar \epsilon_{\mathrm{s}}$. From the position of the inflection points, we can see that, as expected, fields with small tilt angles easily force the complete desorption of the MF, requiring a weaker field to achieve it. The latter is signaled by $\langle \bar U_{\mathrm{s}} \rangle$ reaching its maximum saturation value, which in all cases is very close but not exactly equal to zero. The latter is a consequence of having the position of the end grafted particle permanently constrained within the range of interaction of the surface. Comparing the two absorption strengths, the saturation value of $\langle \bar U_{\mathrm{s}} \rangle$ is closer to zero for the stronger attraction, whereas the bigger deviation from zero at $\bar \epsilon_{\mathrm{s}} \approx 0.14$ simply reflects the stronger fluctuations of the position of the grafted end particle due to the smaller depth of the surface potential well. From the same comparison, one can also observe that the weaker is the adsorbed state at zero field, as signaled by a relatively high surface energy, the weaker is the field strength required to force the desorption. As $\theta_H$ increases, such forced desorption requires stronger fields up to a point in which the field inverts its effect and starts to favor the adsorption. Importantly, such characteristic angle decreases with increasing values of $\bar \epsilon_{\mathrm{s}}$.
The results shown in Figure \[fig:ads-field\] confirm the expected monotonous dependence of the field effects on its tilting angle. However, what happens for a fixed angle when the field strength is changed is a more subtle question, as the nonmonotonic profile of the curve corresponding to $\bar \epsilon_{\mathrm{s}} \approx 0.72$, $\theta_H = 70^{\circ}$ evidences: for weak strengths the adsorption is favored as $H$ grows, but only up to a certain point after which any further increase favors the desorption. Such dependence on the field strength can be explained by the interplay between the filament configurational entropy and the two main terms of the energy, *i.e.*, the magnetic and the surface interaction terms. Since at $\theta_H = 70^{\circ}$ the main component of the field is parallel to the surface, magnetic interactions tend to not only decrease the overall configurational entropy but also to penalize the exploration of configurations occupying regions far from the surface. Note, however, that lateral configurational fluctuations with respect to the axis defined by the field direction can not be fully hindered at any field strength, which means that under large tilt angles the surface remains entropically reachable by the filament. However, any large tilt angle $0 \ll \theta_H < 90^{\circ}$ still puts a magnetic energy penalty on fully adsorbed states due to the misalignment of the field and the surface. At weak fields such penalty is relatively small and the interaction with the surface can dominate. The latter benefits from the initial decrease in the fluctuations led by the growth of the field strength, being able to overcompensate the increase of magnetic energy for the adsorbed state. This corresponds to the observed initial enhacement of the adsorption led by weak growing fields. At some field strength such energy balance saturates and finally inverts, becoming the interaction with the field the dominant one at high $H$. Importantly, even though for the set of parameters sampled here we observe this nonmonotonic behavior only in one case, the reasoning of its explanation can also be applied to other large values of $\theta_H$ and even to systems with different $\bar \epsilon_{\mathrm{s}}$. For instance, one can expect the curve for $\theta_h = 70^{\circ}$, $\bar \epsilon_{\mathrm{s}} \approx 1.45$ to also invert its trend at very large values of $H$.
In order to analyze the switching of the field effects on the adsorption behavior, here we focus on the regions of monotonous response to weak fields and determine the boundaries between the adsorption and desorption regimes as a function of the tilt. Analogously to the analysis of the transitions in temperature discussed above, we take the position of the maximum of the fluctuations of $\langle \bar U_{\mathrm{s}} \rangle$ as the point that represents the characteristic boundary between the adsorbed and desorbed states. Symbols in Figure \[fig:ads-fields\] show the field strength at such characteristic boundary, $H_{\mathrm{ads}}$, obtained for each $\bar \epsilon_{\mathrm{s}}$ as a function of the tilting angle. We can observe that $H_{\mathrm{ads}}(\theta_H)$ exhibits a regular trend in all cases, with a slight growth for small tilting angles that becomes very steep for larger ones. This suggests that $H_{\mathrm{ads}}$ may follow a simple unique function of $(\theta_H)$. According to the discussion at the beginning of this Section, it is reasonable to assume that the relevant magnitude controlling the desorbing effect of the field is its component perpendicular to the adsorbing surface. Therefore, we assumed the following power law for $H_{\mathrm{ads}}$: $$H_{\mathrm{ads}} = H_{\mathrm{ads}}^0 (\cos \theta_H)^{-\nu},
\label{eq:powerlaw}$$ where $H_{\mathrm{ads}}^0$ is the characteristic field for $\theta_H=0$. We performed a least-squares fit of Eq. (\[eq:powerlaw\]) to all simulation datasets in Figure \[fig:ads-fields\], obtaining a single fitted exponent $\nu=3.67 \pm 0.34$. The results of this fitting for each $\bar \epsilon_{\mathrm{s}}$ are shown in the same figure as solid lines. Note that curves corresponding to $\bar \epsilon_{\mathrm{s}} \gtrsim 1.45$ separate states of strong adsorption from desorbed configurations, whereas the curve for $\bar \epsilon_{\mathrm{s}} \approx 0.72$ bounds weakly adsorbed states only. In general, one can not expect a strong adsorption taking place for $\bar \epsilon_{\mathrm{s}} < T$. Such limiting condition is represented by the dashed curve in Figure \[fig:ads-fields\], which shows Eq. (\[eq:powerlaw\]) with the value of $H_{\mathrm{ads}}^0$ corresponding to $\bar \epsilon_{\mathrm{s}} = T = 1$, as obtained by interpolation. Finally, it is worth mentioning that the extrapolation of these fitted curves can be considered valid as far as they do not reach relatively large tilts. This is the case of $\bar \epsilon_{\mathrm{s}} \gtrsim 1.45$ but not of $\bar \epsilon_{\mathrm{s}} \approx 0.72$. As discussed above, the latter already exhibits a nonmonotonic response at $\theta_H = 70^{\circ}$, thus for bigger tilts Eq. (\[eq:powerlaw\]) is not expected to hold for such a weak surface attraction strength.
The results presented above suggest that, under control of the external field, adsorption/desorption behavior can be described by a single master curve. This can be evidenced by performing a rescaling of all the transition curves obtained in our simulations according to fitted Eq. (\[eq:powerlaw\]). Figure \[fig:ads-fields-master\] shows the result of such rescaling for all available datasets, *i.e.*, for all measured fields strengths, desorption-inducing tilting angles and the four considered surface attraction strengths. Apart from the expected $\bar \epsilon_{\mathrm{s}}$-dependent differences in the saturation values, there is a collapse of all data into a characteristic transition master curve that captures the effects of the external field on this system. Finally, as shown in the inset of Figure \[fig:ads-fields-master\], $H^0_{\mathrm{ads}}$ values obtained from the fitting of Eq. (\[eq:powerlaw\]) show a linear dependence on $\bar \epsilon_{\mathrm{s}}$.
Conclusions
===========
In this work we have studied, by means of computer simulations with a mesoscale model, the parameters that determine the equilibrium structural behavior of a flexible magnetic filament made of ferromagnetic nanoparticles and grafted to an attractive flat surface.
First, we characterized the different structural transitions that the filament experiences on cooling: backbone straightening, magnetic flux closure transition and adsorption on the attractive surface. We have shown that, for surface attraction strengths that compare to the thermal fluctuations at room temperature or above, the straightening and the adsorption transition take place simultaneously, whereas the closure transition requires lower temperatures. We also evidenced that the magnetic filament adsorbs at slightly higher temperatures than its nonmagnetic equivalent chain due to its effective increased rigidity induced by dipole-dipole interactions.
Finally, we studied the conditions to control the adsorption/desorption of the magnetic filament by means of static homogeneous magnetic fields. We have shown that the state of the filament can be effectively controlled by both, the strength and the orientation of the applied field. Filament desorption can be easily forced by fields perpendicular to the adsorbing surface. As the field tilting angle with respect to the normal increases, stronger fields are needed to force the desorption, until a maximum angle is reached and the effect of the field is inverted, forcing the adsorption. Importantly, the characteristic field strengths and tilting angles that separate the adsorbed and desorbed states are related by a simple power law whose prefactor depends linearly on the surface attraction strength. Therefore, the field-induced adsorption/desorption of the filament is fully represented by a transition master curve. This fundamental characterization may be essential for the future design of field-switchable micro- or nanofluidic devices based on magnetic filaments.
We consider the results presented here as a preliminary step on the way to the design of magnetically controlled filtering microfluidic devices based on MFs. Future works with this perspective will require to include hydrodynamic interactions in order to study the dynamic response and nonequilibrium properties of these systems.
Acknowledgements {#acknowledgements .unnumbered}
================
This research was supported by the Russian Science Foundation, Grant No.19-12-00209. T.S. acknowledges support by the Spanish AEI/MCI/FEDER(UE), Grant No. RTI2018-095441-B-C22 and The Maria de Maeztu R$\&$D Program (MDM-2017-0711). Simulations were carried out at the Vienna Scientific Cluster (VSC).
[68]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1126/science.1130557) [****, ()](\doibase 10.1098/rsta.2018.0353) “,” (, , ) Chap. , pp. [****, ()](\doibase 10.1007/s00419-015-1092-6) [****, ()](\doibase 10.1103/PhysRevE.96.022605) [****, ()](\doibase 10.1016/j.physa.2017.06.007) in [**](\doibase 10.1016/B978-0-12-813594-5.00007-2), , (, ) pp. in [**](\doibase 10.1016/S1567-2719(05)16003-X), Vol. , (, ) pp. in [**](\doibase 10.1007/978-3-540-49661-8_25), (, ) pp. [****, ()](\doibase 10.1103/PhysRevE.98.050602) [****, ()](\doibase
10.1103/PhysRevE.100.032605) [****, ()](\doibase 10.1016/j.cocis.2005.07.002) [****, ()](\doibase
10.1142/S1793292011002305) [****, ()](\doibase 10.1002/adfm.201502696) [****, ()](\doibase
10.1038/nature04090) [****, ()](\doibase 10.1103/PhysRevE.79.051503) [****, ()](\doibase 10.1039/b918215b) [****, ()](\doibase
10.1103/PhysRevLett.91.260802) [****, ()](\doibase 10.1039/B908578E) [ ()](\doibase 10.1039/D0SM00403K) [****, ()](\doibase 10.1021/ja710172z) [****, ()](\doibase
10.1021/la5009939) [****, ()](\doibase
10.1021/ja070779d) [****, ()](\doibase
10.1021/nn8005366) [****, ()](\doibase 10.1021/am405786u) [****, ()](\doibase 10.1021/nn501787v) [****, ()](\doibase 10.1021/jp208020z) [****, ()](\doibase
10.1038/nnano.2015.105) [****, ()](\doibase 10.1038/nmat4571) [****, ()](\doibase 10.1002/adma.200801423) [****, ()](\doibase 10.1002/polb.22062) [****, ()](\doibase
10.1021/acs.macromol.5b01086) [****, ()](\doibase
10.1039/c5fd00133a) [****, ()](\doibase 10.1039/C6SM02606K) [****, ()](\doibase
10.1039/C9SM01285K) [****, ()](\doibase
10.1073/pnas.0906819106) [****, ()](\doibase 10.1002/adfm.201303358) [****, ()](\doibase 10.1007/s11706-015-0291-y) [****, ()](\doibase 10.1103/PhysRevE.93.062604) [****, ()](\doibase 10.1038/s41378-018-0010-9) [****, ()](\doibase 10.1140/epje/i2008-10388-1) [****, ()](\doibase 10.1039/C3LC50407G) [****, ()](\doibase 10.1017/S002211201100005X) [****, ()](\doibase 10.1103/PhysRevMaterials.1.064402) [****, ()](\doibase
10.1103/PhysRevE.95.052606) [****, ()](\doibase
10.1039/C0SM00772B) [****, ()](\doibase 10.1103/PhysRevE.83.011402) [****, ()](\doibase 10.1063/1.4807003) [****, ()](\doibase
10.1103/PhysRevE.90.042311) [****, ()](\doibase 10.1016/j.jmmm.2015.01.018) [****, ()](\doibase 10.1063/1.4962365) [****, ()](\doibase
10.1088/1361-648X/aaa344) [****, ()](\doibase 10.1103/PhysRevE.92.042303) [****, ()](\doibase 10.1021/acs.langmuir.6b02268) [****, ()](\doibase 10.1016/j.jmmm.2017.10.091) [****, ()](\doibase 10.1016/j.molliq.2020.112761) [****, ()](\doibase 10.1063/1.1674820) [****, ()](\doibase 10.1103/PhysRevA.33.3628) @noop [**]{}, ed., Oxford Science Publications (, , ) [****, ()](\doibase 10.1016/S0009-2614(99)01123-9) [****, ()](\doibase 10.1002/1097-0282(2001)60:2<96::AID-BIP1007>3.0.CO;2-F) [****, ()](\doibase 10.1063/1.4815915) [****, ()](\doibase 10.1103/PhysRevLett.63.1195) [****, ()](\doibase 10.1002/jcc.540130812) in [**](\doibase
10.1007/978-3-642-32979-1_1), , Vol. , (, ) pp. [****, ()](\doibase 10.1002/bip.1979.360180511) [****, ()](\doibase 10.1021/ma000493s) [****, ()](\doibase 10.1021/ma400112q)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the concept of $H^2$-gradient flow for the Willmore energy and other functionals that depend at most quadratically on the second fundamental form is well-defined in the space of immersions of Sobolev class $W^{2,p}$ from a compact, $n$-dimensional manifold into Euclidean space, provided that $p \geq 2$ and $p>n$. We also discuss why this is *not* true for Sobolev class $H^2=W^{2,2}$. In the case of equality constraints, we provide sufficient conditions for the existence of the projected $H^2$-gradient flow and demonstrate its usability for optimization with several numerical examples.'
author:
- 'Henrik Schumacher[^1]'
bibliography:
- 'Literature.bib'
title: 'On $H^2$-gradient Flows for the Willmore Energy'
---
[^1]: <[email protected]>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Understanding the hydrodynamical processes and conditions at the interface between the Magellanic Stream (MS) and the Galactic halo is critical to understanding the MS and by extension, gaseous tails in other interacting galaxies. These processes operate on relatively small scales and not only help shape this clumpy stream, but also affect the neutral gas dynamics and transfer of mass from the stream to the halo, thus affecting metal enrichment and gas replenishment of the Galaxy. We describe an observational program to place constraints on these processes through high-resolution measurements of HI emission, HI absorption and H$\alpha$ emission with unprecedented sensitivity. Methods will include structural analysis, searching for cold gas cores in clumps and analyzing gas kinematics as it transitions to the halo. The latter method includes sophisticated spatial integration techniques to deeply probe the neutral gas, which we apply to a new HI map obtained from the Green Bank Telescope with the highest sensitivity HI observations of the MS to date. We demonstrate that the integration techniques enhance sensitivity even further, thus allowing detection of apparent MS gas components with density approaching that of the Galactic halo.'
author:
- |
Lou Nigra,$^{,}$ Snežana Stanimirović, J. S. Gallagher, III,\
Felix J. Lockman, David L. Nidever, Steven R. Majewski
title: 'The Magellanic Stream to Halo Interface: Processes that shape our nearest gaseous Halo Stream'
---
Introduction
============
The Magellanic Stream (MS) is our nearest major gaseous interaction remnant. It trails across much of the southern galactic sky for $>100^{\circ}$ behind the Magellanic Clouds (MCs), passing near the Southern Galactic Pole at its midpoint. The MS head is presumed to be nominally between $52$ and $61$ kpc away, the distances to the MCs established by @koerwer09 and @hilditch05. The distance to the tip is not well-constrained but estimates are of the same order as distances to the MCs. These distances are about 15 times closer than the nearest extragalactic tidal features of the Andromeda system, therefore the MS presents us with the opportunity to closely study star formation potential (or lack thereof), mass transfer, and kinematics of a major interaction feature on fine scales inaccessible elsewhere. No signs of stars or even star formation have been found in the MS to date. The mean column density of the MS ranges from $4\times10^{18}$ to $4\times10^{19}$ cm$^{-2}$ along its length, which is an order of magnitude below the tidal debris star-formation threshold ($4\times10^{20}$ cm$^{-2}$) found by @maybhate07. Unlike High Velocity Clouds (HVCs), whose H$\alpha$ emission can be explained by Galactic UV radiation, the MS is brighter in H$\alpha$, particularly near the MCs, suggesting another ionizing source. Global models of the MS using Smoothed Particle Hydrodynamics (SPH) and N-Body techniques have reproduced its large-scale structure with some degree of success with the then-current assumption of multiple MC orbits of the Galaxy [@moore94; @connors06; @mastro05]. The new measurement of MC proper motions by @kalliv06a and @kalliv06b led @besla07 to propose the likelihood of unbound MC orbits. Subsequently, @mastro08 presented revised results for the new orbits.
Something that the global models do not capture is the very rich, fine-scale structure revealed in recent HI MS maps such as @putman03b and @bruns05. Even finer structure was revealed by @stanim08 in a high-resolution map of the northern tip of the MS using the 305$\:$m dish at Arecibo. These observations revealed extended fine filamentary structure and clumps down to the angular resolution of the telescope ($3.5'$). They also highlighted the importance of Kelvin-Helmholtz and Thermal Instabilities in forming the clumpy structure of the MS. In addition, they revealed cooler cores in some clumps suggesting star formation potential and providing clues to halo pressure and dark matter confinement of these clumps. Processes are clearly at work on these small scales that could affect star formation, the transfer of gas to the halo, and also may provide additional drag affecting MS global dynamics.
The global simulations mentioned above can miss these processes since SPH can suppress instabilities due to smoothing [@agertz07], and N-body simulations ignore gas processes altogether. Grid-based modeling can capture these small-scale instabilities better, exemplified by simulations exploring mechanisms for excess H$\alpha$ emission in the MS [@jbh07], galaxy replenishment [@jbh09; @heitsch09], and HVCs in the halo [@quilis01]. Grid-based modeling on these scales is producing some exciting insights into the possible processes at work.
We present here a research program aimed at setting some observational constraints on these processes that act on the MS at its interface with the Galactic halo and then using them to test these models. Towards this end, we have obtained the most sensitive HI emission map of portions of the MS to date and have begun to analyze the structure and kinematics of the gas. We are in the process of obtaining high resolution, high sensitivity measurements of H$\alpha$ emission for insight into higher energy processes at the halo interface, and HI absorption to look for cold gas cores and characterize gas temperatures.
Observations: Completed, ongoing and planned
============================================
We have completed a program at the Green Bank Telescope [^1] (beamwidth $=9.1'$) where we obtained deep HI spectra across two separate $12$ deg$^2$ regions. One region (Region 1) is located in the northern tip and the other (Region 2) is $~20$ deg further up in the mid-MS region. On-the-fly (OTF) mapping and in-band frequency switching was employed. Region 1 data have been reduced with a fairly successful initial pass at baseline removal resulting in a cube with $3.53'$ square pixels, and velocity resolution of $0.161$ km$\:$sec$^{-1}$ from $-200$ to $-518$ km$\:$sec$^{-1}$. Noise was measured at $\sigma_T=4.2$ mK corresponding to column density noise $\sigma_N=1.14\times10^{17}$ cm$^{-2}$ for a $15$ km$\:$sec$^{-1}$ FWHM profile. At less than half the noise of the Galactic All-Sky Survey (GASS) [@mg09] scaled to the same profile, the map presented here is the most sensitive of the MS to date. Column density and velocity field maps are shown in Figure \[fig:region1map\].
Further plans include observations of the fine-scale signature of ionization process on selected clumps and features by obtaining deeply integrated, high-resolution H$\alpha$ observations. We will compare our results to those expected in models of ionization through Galactic UV radiation or energetic gas processes such as those proposed by @jbh07. We also plan to obtain deep absorption measurements in the direction of several background radio sources within the tip of the MS coinciding with clumps identified by @stanim08. In addition to possibly identifying cold, potentially star-forming cores, these sensitive absorption measurements through specific small-scale structures will provide useful constraints on the analysis of gas kinematics.
Results & Discussion
====================
The Region 1 HI column density image of Figure \[fig:region1map\] (left), shows the main filament extending from the lower left to the upper right, which is part of the longer filament S2 identified by @stanim08. The Region 1 velocity field (Figure \[fig:region1map\] right) shows the general velocity gradient of the main filament, but also reveals some differentiation in places, suggesting some complexity in the projected dimension. Several prominent and interesting features are labeled in Figure 1: (A) and (B) appear as head-tail clumps; (C) is a narrow filament that apparently connects to the main filament; (D) are a series of relatively dense “spokes” at the edge of the main filament; (E) is a large, apparently coherent cloud that is significantly more diffuse than other clumps of its size in the map. Most of these features (A through D) have a strong component transverse to the main filament, suggesting that this filamentary substructure may represent ram-pressure or ablation shreds coming off the main cloud. It is interesting to note that most of the transverse substructure appears in the form of short filaments, instead of diffuse cloud envelopes. In the future, these structures will be analyzed further both spatially and kinematically to see what the de-projected structures might look like and what they might say about the processes at work here. We will also statistically compare the structure of this region to simulations from @jbh07 by using Fourier Transform methods.
![From @nigra09. Left - Region 1 column density map from the GBT data. Velocity range is $-385$ to $-305$ km$\:$sec$^{-1}$. Intensity scale is $0\:$to$\:2.5\times10^{19}\:$cm$^{-2}$. Features “A” through “E” are discussed in the text. Right - Region 1 velocity field for $N_{HI}>1.5\times10^{18}$ cm$^{-2}$. See text for discussion. These images were generated with data from telescopes of the National Radio Astronomy Observatory, a National Science Foundation Facility, managed by Associated Universities. Inc.[]{data-label="fig:region1map"}](nigralf1_color){height="3.4in"}
![ A demonstration of integration techniques used for gas analysis. A portion of the main filament is modeled as a cylinder. Integrating (averaging) parallel to its axis some distance away has two clearly detectable components while a single pixel spectrum at that distance has none.[]{data-label="fig:integrate"}](nigralf2){height="2.4in"}
We are in the process of characterizing properties of the diffuse neutral gas in the MS-halo transition region. Profiles of gas properties (temperature and column density) as a function of distance from the clump centers will be compared with analytical and simulation results for various hydrodynamic processes, each of which will have a particular “signature” profile. For instance, the temperature vs. radial distance in an isolated cloud in the halo is quite different if experiencing saturated evaporation or if radiatively stabilized, as shown by @cowie77 and @mckee77.
These profiles are conventionally obtained by looking at observed spectra along a cross-section and fitting to a curve, as in @bruns01. Further from the center, brightness decreases and velocity dispersion increases, causing measurement uncertainty to increase rapidly. In order to improve upon this method, we average spatially along a symmetry dimension under the assumption that the clump symmetry extends past its apparent boundaries. For instance, an isolated circular clump in a map is assumed spherical and can be integrated along concentric rings where each column of gas is nominally the same. Similarly for a linear feature, we assume cylindrical symmetry and average along lines parallel to the filament. More advanced techniques include integrating along $N_{HI}$ contours or optimizing to fit a parametric 3D gas model.
To demonstrate the sensitivity of this approach, we show initial results from applying this method in Figure \[fig:integrate\]. Here a portion of the main filament of Region 1 is assumed to be roughly cylindrical along the axis indicated on the image. The spectra along a parallel path to the left were integrated (averaged) after removing the velocity gradient of the main filament. The spectrum sampled at a a single location, as used in a simple cross-section profile, shows no sign of a detectable gas component, while the corresponding averaged spectrum of the cylindrical model clearly shows two gaussian components, with FWHM $=10~$and$~26$ km$\:$sec$^{-1}$, respectively. Each is detected well above their $3\sigma$ detection limits (dashed line) with corresponding $N_{HI}=3.4\times10^{17}$ cm$^{-2}~$and$~1.1\times10^{18}$ cm$^{-2}$, respectively.
Although detection of these apparent gas components clearly demonstrates the sensitivity of the technique, verification as actual low column density gas is still in progress. Their shape, width and velocity are roughly consistent with MS gas, and are not likely artifacts of the baseline procedure. However, at such high sensitivity, systematic artifacts not normally seen must be considered [@nidever09]. Should either of these detections prove valid, it would suggest that a significant amount of low column density HI gas, well below thresholds of previous observations, exists in this part of the MS. If not, it will still establish unprecedented limits on a diffuse neutral gas component. In either case, this approach will provide valuable insight into the history, dynamics and hydrodynamic processes of the MS [@nigra09].
To gauge how close these detections are to the halo interface, an order of magnitude estimate of the number density is made. The filament’s scale is $\sim1^{\circ}$ on the sky. Assuming a distance of $\sim60$ kpc, the scale width is then $\sim1$ kpc$\:=3.1\times10^{21}$ cm. The number density for the $26$ km$\:$sec$^{-1}$ FWHM component is then on the order of $1.1\times10^{18}\div3.1\times10^{21}=3.5\times10^{-4}$ cm$^{-3}$. This is on the same order as the upper range of ionized halo number density estimates [@sembach03] and we can detect even weaker lines, so we may indeed be probing HI very close to the halo interface.
Conclusions
===========
The rich fine, clumpy and filamentary structure on arcminute scales revealed in modern HI maps of the MS are not reproduced in global SPH and N-body simulations. Observations and grid-based local simulations reveal that this structure is likely produced by hydrodynamical instabilities at the interface between the cool, stripped gas as it moves through the hot halo gas. These processes may be an important source of MS structure, a mechanism for ionizing the MS gas as well as a significant means of transferring gas from the MS to the halo, and eventually to the Galactic disk.
A program is in progress to place observational constraints on these processes and on the modeling of them by obtaining the most sensitive measurements to date of the MS gas in both HI and H$\alpha$ emission, as well as in HI absorption, on the periphery of the MS where it is closest to the ambient halo environment. We have already obtained maps of two regions of the MS with unprecedented sensitivity. Using these data, we have demonstrated spatial integration techniques allowing detection of HI components at densities of $\sim10^{-4}$ cm$^{-3}$, within an order of magnitude of estimated halo density upper limits. We will apply these techniques to characterize gas kinematics and then compare to models to determine what processes are operating on the MS periphery.
The authors thank Carl Heiles for his invaluable participation in the absorption measurement program. LN thanks the National Astronomy and Ionosphere Center for pre-doctoral support. JSG thanks the University of Wisconsin Graduate School for partial support of this research.
Agertz, O., et al. 2007, , 380, 963
Besla, G., Kallivayalil, N., Hernquist, L., Robertson, B., Cox, T. J., van der Marel, R. P., & Alcock, C. 2007, , 668, 949
Bland-Hawthorn, J., Sutherland, R., Agertz, O., & Moore, B. 2007, , 670, L109
Bland-Hawthorn, J. 2009, IAU Symposium, 254, 241
Br[ü]{}ns, C., et al. 2005, , 432, 45
Br[ü]{}ns, C., Kerp, J., & Pagels, A. 2001, , 370, L26
Connors, T. W., Kawata, D., & Gibson, B. K. 2006, , 371, 108
Cowie, L. L., & McKee, C. F. 1977, , 211, 135
Heitsch, F., & Putman, M. E. 2009, , 698, 1485
Hilditch, R. W., Howarth, I. D., & Harries, T. J. 2005, , 357, 304
Kallivayalil, N., et al. 2006, , 638, 772
Kallivayalil, N., van der Marel, R. P., & Alcock, C. 2006, , 652, 1213
Koerwer, J. F. 2009, , 138, 1
Mastropietro, C., Moore, B., Mayer, L., Wadsley, J., & Stadel, J. 2005, , 363, 509
Mastropietro, C. 2008, arXiv:0810.4155
Maybhate, A., et al. 2007, , 381, 59
McClure-Griffiths, N. M., et al. 2009, , 181, 398
McKee, C. F., & Cowie, L. L. 1977, , 215, 213
Moore, B., & Davis, M. 1994, , 270, 209
Nidever, D. et al. 2009, in preparation.
Nigra, L., Stanimirovi[ć]{}, S., Gallagher, J. S., III, et al. 2009, in preparation.
Putman, M. E., Staveley-Smith, L., Freeman, K. C., Gibson, B. K., & Barnes, D. G. 2003, , 586, 170
Quilis, V., & Moore, B. 2001, , 555, L95
Sembach, K. R., et al. 2003, , 146, 165
Stanimirovi[ć]{}, S., Hoffman, S., Heiles, C., Douglas, K. A., Putman, M., & Peek, J. E. G. 2008, , 680, 276
[^1]: The Robert C. Byrd Green Bank Telescope is operated by the National Radio Astronomy Observatory, which is a facility of the US National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Classical [S]{}turm-[L]{}iouville problems of $q$-difference variables are extended for symmetric discrete functions such that the corresponding solutions preserve the orthogonality property. Some illustrative examples are given in this sense.'
address:
- 'Departamento de Matemática Aplicada II, E.E. de Telecomunicación, Universidade de Vigo, Campus Lagoas-Marcosende, 36310 Vigo, Spain.'
- 'Department of Mathematics, K.N.Toosi University of Technology, P.O. Box 16315–1618, Tehran, Iran.'
author:
- 'I. Area'
- 'M. Masjed-Jamei'
title: 'A symmetric generalization of [S]{}turm-[L]{}iouville problems in $q$-difference spaces'
---
[^1]
[^2]
Introduction
============
A regular Sturm-Liouville problem of continuous type is a boundary value problem in the form $$\label{eq:1}
\frac{d}{dx} \left( k(x) \frac{dy_{n}(x)}{dx} \right) + \left(\lambda_{n} \varrho(x)-q(x) \right) y_{n}(x)=0 \qquad (k(x)>0, \varrho(x)>0),$$ which is defined on an open interval, say $(a,b)$, and has the boundary conditions $$\label{eq:2}
\alpha_{1} y(a) + \beta_{1} y'(a)=0, \quad \alpha_{2} y(b) + \beta_{2} y'(b)=0,$$ where $\alpha_{1}, \alpha_{2}$ and $\beta_{1}, \beta_{2}$, are given constants and $k(x)$, $k'(x)$, $q(x)$, and $\varrho(x)$ in (\[eq:1\]) are to be assumed continuous for $x \in [a,b]$. In this sense, if one of the boundary points $a$ and $b$ is singular (i.e. $k(a) = 0$ or $k(b) = 0$), the problem is called a singular [S]{}turm-[L]{}iouville problem of continuous type.
Let $y_{n}$ and $y_{m}$ be two eigenfunctions of equation (\[eq:1\]). According to [S]{}turm-[L]{}iouville theory [@MR922041], they are orthogonal with respect to the weight function $\varrho(x)$ under the given conditions (\[eq:2\]) so that we have $$\label{eq:3}
\int_{a}^{b} \varrho(x) y_{n}(x) y_{m}(x) dx =
\left( \int_{a}^{b} \varrho(x) y_{n}^{2}(x) dx \right) \delta_{n,m}=
\| y_{n} \|_{2}^{2} \begin{cases} 0 & n \neq m, \\ 1 & n=m. \end{cases}$$
Many important special functions in theoretical and mathematical physics are solutions of a regular or singular [S]{}turm-[L]{}iouville problem that satisfy the orthogonality condition (\[eq:3\]). For instance, the associated Legendre functions [@MR1810939], Bessel functions [@MR922041], Fourier trigonometric sequences [@MR0105586], ultraspherical functions [@MR922041] and Hermite functions [@MR922041] are some specific continuous samples. Most of these functions are symmetric (i.e. $\phi_{n}(-x)=(-1)^{n} \phi_{n}(x)$) and have found valuable applications in physics and engineering. Hence, if we can somehow extend these examples symmetrically and preserve their orthogonality property, it seems that we will be able to find new applications, which logically extend the previous established applications. Recently in [@MR2374588], this matter has been done for continuous variables and the classical equation (\[eq:1\]) has been symmetrically extended in the following form $$\label{eq:4}
A(x) \phi_{n}''(x) + B(x) \phi_{n}'(x) + \left( \lambda_{n} C(x) + D(x) + \sigma_{n} E(x)\right) \phi_{n}(x)=0,$$ where $A(x)$, $D(x)$, $E(x)$ and $(C(x)>0)$ are even functions, $B(x)$ is an odd function and $$\label{eq:sigma}
\sigma_{n}=\frac{1-(-1)^{n}}{2} = \begin{cases}
0 & n \text{ even}, \\
1 & n \text{ odd}.
\end{cases}$$ It has been proved in [@MR2374588] that under some specific conditions, the symmetric solutions of equation (\[eq:4\]) are orthogonal and preserve the orthogonality interval, in other words:
[@MR2374588]\[th:1\] Let $\phi_{n}(-x)=(-1)^{n} \phi_{n}(x)$ be a sequence of symmetric functions that satisfies the differential equation (\[eq:4\]), where $\{Ê\lambda_{n} \}_{n}$ is a sequence of constants. If $A(x)$, $(C(x) > 0)$, $D(x)$ and $E(x)$ are even real functions and $B(x)$ is odd then $$\int_{-\nu}^{\nu} P^{*}(x) \phi_{n}(x) \phi_{m}(x) dx = \left( \int_{-\nu}^{\nu}P^{*}(x) \phi_{n}^{2}(x) dx \right) \delta_{n,m},$$ where $$\label{eq:6}
P^{*}(x)= C(x) {\exp} \left( \int \frac{B(x)-A'(x)}{A(x)} dx \right) = \frac{C(x)}{A(x)} {\exp} \left( \int \frac{B(x)}{A(x)} dx \right).$$ Of course, the weight function defined in (\[eq:6\]) must be positive and even on $[-\nu,\nu]$ and the function $$A(x)K(x)=A(x) {\exp} \left( \int \frac{B(x)-A'(x)}{A(x)} dx\right) = {\exp} \left( \int \frac{B(x)}{A(x)} dx\right),$$ must vanish at $x = \nu$ , i.e. $A(\nu) K(\nu) = 0$. In this way, since $K(x) = P^{*}(x) /C(x)$ is an even function so $A(-\nu) K(-\nu) = 0$ automatically.
By using theorem \[th:1\], many symmetric special functions of continuous type have been generalized in [@MR2270049; @MR2374588; @MR2421844; @MR2601301; @MR2467682; @MR2743534; @1220.33011]. Recently in [@MASJEDAREA] we have generalized usual Sturm-Liouville problems with symmetric solutions in discrete spaces on the linear lattice $x(s)=s$, and introduced a basic class of symmetric orthogonal polynomials of a discrete variable with four free parameters [@MASJEDAREA2]. The main aim of this paper is to prove that the aforesaid extension also holds in a $q$-difference case and for a homogeneous second-order $q$-difference equation.
A symmetric generalization of [S]{}turm-[L]{}iouville problems in $q$-difference spaces
=======================================================================================
Before stating the results, we should consider some preliminaries and notations.
Let $\mu \in {\mathbf{C}}$ be fixed. A set $A \subseteq {\mathbf{C}}$ is called a $\mu$-geometric set if for $x \in A$, $\mu x \in A$. Let $f$ be a function defined on a $q$-geometric set $A \subseteq {\mathbf{C}}$. The $q$-difference operator is defined by $$D_{q}f(x)=\frac{f(qx)-f(x)}{(q-1)x}, \qquad x \in A \setminus \{0\}.$$ If $0 \in A$, we say that $f$ has the $q$-derivative at zero if the limit $$\lim_{n \to \infty} \frac{f(x q^{n})-f(0)}{xq^{n}} \qquad (x \in A),$$ exists and does not depend on $x$. We then denote this limit by $D_{q} f(0)$.
We shall also need the $q$-integral (the inverse of the $q$-derivative operator) introduced by J. Thomae [@Thomae1969] and F.H. Jackson [@JACKSON1910] —see also [@MR2128719; @MR2191786; @MR2656096]— which is defined as $$\label{eq:11}
\int_{0}^{x} f(t) \text{d}_qt = x(1-q)\sum_{n=0}^{\infty} q^n f(q^n x)\,, \qquad (x \in A),$$ provided that the series converges, and for the interval $[a,b]$ we have based on (\[eq:11\]) that $$\label{eq:10}
\int_{a}^{b} f(t) d_{q}t = \int_{0}^{b} f(t) d_{q}t - \int_{0}^{a} f(t) d_qt, \qquad (a,b \in A).$$ Relations (\[eq:11\]) and (\[eq:10\]) directly yield $$\label{eq:12}
\int_{-b}^{b} f(t) d_{q}t = b(1-q) \sum_{n=0}^{\infty} q^{n} \left( f(bq^{n})+f(-b q^{n}) \right), \qquad (b \in A).$$ This means that if $f$ is an odd function, then $\displaystyle{\int_{-b}^{b} f(t) d_{q}t=0}$. Moreover, if $b \to \infty$, (\[eq:12\]) changes to $$\int_{-\infty}^{\infty} f(t) d_{q}t =(1-q) \sum_{n=-\infty}^{\infty} q^{n} \left( f(q^{n}) + f(-q^{n})\right).$$
A function $f$ which is defined on a $q$-geometric set $A$ with $0 \in A$ is said to be $q$-regular at zero if $\displaystyle{\lim_{n \to \infty} f(xq^{n})=f(0)}$ for every $x \in A$. The rule of $q$-integration by parts is denoted by $$\label{eq:14}
\int_{0}^{a} g(x) D_{q}f(x) d_{q}x=(fg)(a)-\lim_{n \to \infty} (fg)(aq^{n}) - \int_{0}^{a} D_{q} g(x)f(qx)d_{q}x.$$ If $f,g$ are $q$-regular at zero, the $\displaystyle{\lim_{n \to \infty} (fg)(aq^{n})}$ on the right-hand side of (\[eq:14\]) can be replaced by $(fg)(0)$.
For $0 < R \leq \infty$ let $\Omega_{R}$ denote the disc $\{ z \in {\mathbf{C}} \, : \, \vert z \vert < R\}$. The $q$-analogue of the fundamental theorem says: Let $f: {\Omega_{R}} \to {\mathbf{C}}$ be $q$-regular at zero and $\theta \in \Omega_{R}$ be fixed. Define $$F(x)=\int_{\theta}^{x} f(t) d_{q}t \qquad (x \in \Omega_{R}).$$ Then, the function $F$ is $q$-regular at zero, $D_{q}F(x)$ exists for any $x \in \Omega_{R}$ and $D_{q}F(x)=f(x)$. Conversely, if $a,b \in \Omega_{R}$ then $$\int_{a}^{b} D_{q}f(t) d_{q}t=f(b)-f(a).$$
The function $f$ is $q$-integrable on $\Omega_{R}$ if $\displaystyle{ \vert f(t) \vert d_{q}t}$ exists for all $x \in \Omega_{R}$.
Let $\phi_{n}(x;q)=(-1)^{n} \phi_{n}(-x;q)$ be a sequence of symmetric functions that satisfies the $q$-difference equation $$\begin{gathered}
\label{eq:17}
A(x) D_{q} D_{q^{-1}} \phi_{n}(x;q) + B(x) D_{q} \phi_{n}(x;q) \\
+ \left( \lambda_{n,q} C(x) + D(x) + \sigma_{n} E(x) \right) \phi_{n}(x;q)=0,\end{gathered}$$ where $A(x)$, $B(x)$, $C(x)$, $D(x)$ and $E(x)$ are independent functions, $\sigma_{n}$ is defined in (\[eq:sigma\]) and $\lambda_{n,q}$ is a sequence of constants. If $A(x)$, $(C(x)>0)$, $D(x)$ and $E(x)$ are even functions and $B(x)$ is odd, then $$\int_{-\alpha}^{\alpha} W^{*}(x;q) \phi_{n}(x;q) \phi_{m}(x;q) d_{q}x =
\left( \int_{-\alpha}^{\alpha} W^{*}(x;q) \phi_{n}^{2}(x;q) d_{q}x \right) \delta_{n,m},$$ where $$\label{eq:19}
W^{*}(x;q)=C(x)W(x;q),$$ and $W(x;q)$ is solution of the Pearson $q$-difference equation $$\label{eq:20}
D_{q} \left(A(x) W(x;q) \right) = B(x) W(x;q),$$ which is equivalent to $$\frac{W(qx;q)}{W(x;q)}=\frac{(q-1)x B(x)+A(x)}{A(qx)}.$$ Of course, the weight function defined in (\[eq:19\]) must be positive and even and $A(x)W(x;q)$ must vanish at $x=\alpha$.
If the difference equation (\[eq:17\]) is written in a self-adjoint form, then $$\begin{gathered}
\label{eq:22}
D_{q} \left[A(x)W(x;q) D_{q^{-1}} \phi_{n}(x;q) \right]\\ + \left( \lambda_{n,q} C(x) + D(x) + \sigma_{n} E(x) \right) W(x;q) \phi_{n}(x;q)=0,\end{gathered}$$ and for $m$ we similarly have $$\begin{gathered}
\label{eq:23}
D_{q} \left[A(x)W(x;q) D_{q^{-1}} \phi_{m}(x;q) \right] \\ + \left( \lambda_{m,q} C(x) + D(x) + \sigma_{m} E(x) \right) W(x;q) \phi_{m}(x;q)=0.\end{gathered}$$ By multiplying (\[eq:22\]) by $\phi_{m}(x;q)$ and (\[eq:23\]) by $\phi_{n}(x;q)$ and subtracting each other we get $$\begin{gathered}
\label{eq:24}
\phi_{m}(x;q) D_{q} \left[ A(x) W(x) D_{q^{-1}} \phi_{n}(x;q) \right] - \phi_{n}(x;q) D_{q} \left[ A(x) W(x) D_{q^{-1}} \phi_{m}(x;q) \right] \\
+ \left( \lambda_{n,q}-\lambda_{m,q} \right)C(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q)\\
+\frac{(-1)^{m}-(-1)^{n}}{2}E(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q)=0.\end{gathered}$$ A simple but important idea can appear here: “The $q$-integration of any odd integrand over a symmetric interval is equal to zero". Therefore, $q$-integrating on both sides of (\[eq:24\]) over the symmetric interval $[-\alpha,\alpha]$ yields $$\begin{gathered}
\label{eq:25}
\int_{-\alpha}^{\alpha} \phi_{m}(x;q) D_{q} \left( A(x)W(x;q) D_{q^{-1}} \phi_{n}(x;q) \right) d_{q} x \\
- \int_{-\alpha}^{\alpha} \phi_{n}(x;q) D_{q} \left( A(x)W(x;q) D_{q^{-1}} \phi_{m}(x;q) \right) d_{q} x \\ + \left( \lambda_{n,q}-\lambda_{m,q} \right) \int_{-\alpha}^{\alpha} C(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x \\
+\frac{(-1)^{m}-(-1)^{n}}{2} \int_{-\alpha}^{\alpha} E(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x=0.\end{gathered}$$ Now, by using the rule of $q$-integration by parts, relation (\[eq:25\]) is transformed to $$\begin{gathered}
\label{eq:26}
\left[ A(x) W(x;q) \phi_{m}(x;q) D_{q^{-1}} \phi_{n}(x;q) \right]_{-\alpha}^{\alpha} \\
- \int_{-\alpha}^{\alpha} A(qx)W(qx;q) D_{q^{-1}} \phi_{n}(qx;q) D_{q} \phi_{m}(x;q) d_{q} x \\
-\left[ A(x) W(x;q) \phi_{n}(x;q) D_{q^{-1}} \phi_{m}(x;q) \right]_{-\alpha}^{\alpha} \\
+ \int_{-\alpha}^{\alpha} A(qx)W(qx;q) D_{q^{-1}} \phi_{m}(qx;q) D_{q} \phi_{n}(x;q) d_{q} x \\
+ \left( \lambda_{n,q}-\lambda_{m,q} \right) \int_{-\alpha}^{\alpha} C(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x \\
+\frac{(-1)^{m}-(-1)^{n}}{2} \int_{-\alpha}^{\alpha} E(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x=0.\end{gathered}$$ Since $$D_{q^{-1}} f(qx)=D_{q}f(x),$$ relation (\[eq:26\]) is simplified as $$\begin{gathered}
\label{eq:28}
\left[ A(x) W(x;q) \left( \phi_{m}(x;q) D_{q^{-1}} \phi_{n}(x;q) - \phi_{n}(x;q) D_{q^{-1}} \phi_{m}(x;q) \right) \right]_{-\alpha}^{\alpha} \\
+ \left( \lambda_{n,q}-\lambda_{m,q} \right) \int_{-\alpha}^{\alpha} C(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x \\
+\frac{(-1)^{m}-(-1)^{n}}{2} \int_{-\alpha}^{\alpha} E(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x=0.\end{gathered}$$ On the other hand, $W(x;q)$ is a symmetric solution for the Pearson $q$-difference equation (\[eq:20\]). Hence, if in (\[eq:28\]) we take $$A(-\alpha)W(-\alpha;q)=A(\alpha)W(\alpha;q)=0,$$ then to prove the orthogonality property it remains to show that $$F(m,n)=\frac{(-1)^{m}-(-1)^{n}}{2} \int_{-\alpha}^{\alpha} E(x)W(x;q) \phi_{n}(x;q)\phi_{m}(x;q) d_{q}x=0.$$ For this purpose, four cases should be considered for values $m$, $n$, which are respectively as follows:
1. If both $m$ and $n$ are even (or odd), then $F(n,m)=0$ because we have $F(2i,2j)=F(2i+1,2j+1)=0$.
2. If one of the two mentioned values is odd and the other one is even (or conversely) then $$\label{eq:31}
F(2i,2j+1)=\int_{-\alpha}^{\alpha} E(x)W(x;q) \phi_{2j+1}(x;q) \phi_{2i}(x;q) d_{q}x.$$
Since $E(x)$, $W(x;q)$ and $\phi_{2i}(x;q)$ are assumed to be even functions and $\phi_{2j+1}(x;q)$ is odd in (\[eq:31\]), its integrand would be an odd function and therefore $F(2i,2j+1)=0$. This results similarly holds for the case $m=2i+1$ and $n=2j$, i.e. $F(2i+1,2j)=0$. $\text{ }$
Some illustrative examples
==========================
In this section we consider 3 examples of the extended equation (\[eq:17\]) whose solutions are symmetric $q$-orthogonal polynomials.
For this purpose, let us define some notations related to $q$-polynomials. The $q$-shifted factorial is defined by $${\text{$(x;\,q)_{n}$}}=\prod_{j=0}^{n-1}(1-q^jx),\qquad n=0,1,\ldots,$$ the $q$-number by $${\text{$[z]_{q}$}}=\frac{q^{z}-1}{q-1}, \quad z \in {\mathbf{C}},$$ and the [basic hypergeometric series]{} is defined by $${\,{}_{r}\phi_{s}\left(\!\! \begin{array}{cc}{{\mbox{$a_1,a_2,\ldots ,a_{r}$}}}\\[-0.1ex]{{\mbox{$b_1,b_2,\ldots ,b_{s}$}}} \end{array}
\Big|\,{q};{z}\right)}=
\sum_{k=0}^{\infty}\frac{{\text{$(a_1;\,q)_{k}$}}\ldots {\text{$(a_r;\,q)_{k}$}}}
{{\text{$(q;\,q)_{k}$}}{\text{$(b_1;\,q)_{k}$}}\ldots {\text{$(b_s;\,q)_{k}$}}}
\left((-1)^kq^{\binom{k}{2}}\right)^{1+s-r}z^k.$$ Here $r,\,s\in{{\mathbf Z}_+}$ and ${\mbox{$a_1,a_2,\ldots ,a_{r}$}}$, ${\mbox{$b_1,b_2,\ldots ,b_{s}$}}$, $z$ $\in {\mathbf C}$. In order to have a well–defined series the condition $ {\mbox{$b_1,b_2,\ldots ,b_{s}$}} \neq q^{-k}$ ($k=0,1,\ldots $) is required.
Example 1
---------
Let us consider the $q$-difference equation $$\begin{gathered}
x^{2}(1-x^{2}) D_{q} D_{q^{-1}} \phi_{n}(x;q)
+qx \left(-(q^{2}+q+1) x^2+q+1 \right) D_{q} \phi_{n}(x;q)
\\+\left({\text{$[n]_{q}$}} \left( q (q^{2}+q+1)-{\text{$[1-n]_{q}$}} \right) x^{2} -q (1+q) \sigma_{n} \right) \phi_{n}(x;q) =0,\end{gathered}$$ as a special case of (\[eq:17\]). If we set $$\phi_{n}(x;q)=\sum_{j=0}^{\infty} a_{j}(n) x^{j},$$ then a solveble recurrence relation for the coefficients $\{a_{j}(n)\}_{j=0}^{\infty}$ is derived giving rise eventually to the following representation $$\label{eq:icod}
\phi_{n}(x;q)=x^{\sigma_{n}} \, {\,{}_{2}\phi_{1}\left(\!\! \begin{array}{cc}{q^{\sigma_{n}-n},q^{n+\sigma_{n}-1} \left(q^4-q+1\right)}\\[-0.1ex]{q^{2 \sigma_{n}+1} \left(q^3-q+1\right)} \end{array}
\Big|\,{q^{2}};{q^2 x^2}\right)}.$$
From Section 2, it is known that the above sequence satisfies the orthogonality relation $$\int_{-1}^{1} W_{1}^{*}(x;q) \phi_{n}(x;q) \phi_{m}(x;q) d_{q}x =
\left( \int_{-1}^{1} W_{1}^{*}(x;q) \phi_{n}^{2}(x;q) d_{q}x \right) \delta_{n,m},$$ in which $W_{1}^{*}(x;q)=C(x)W_{1}(x;q)$ is the main weight function and $W_{1}(x;q)$ satisfies the equation $$\label{eq:360}
\frac{W_{1}(qx;q)}{W_{1}(x;q)}=\frac{(q^{4}-q+1) x^2-q^3+q-1}{q^2 \left(q^2 x^2-1\right)}.$$ Up to a periodic function, a solution of the equation (\[eq:360\]) is in the form $$W_{1}(x;q)=\frac{
{\text{$(q^2 x^2;\,q^{2})_{\infty}$}}
\left(q^3-q+1\right)^{\frac{\log \left(x^2\right)}{2 \log(q)}}}
{x^2 {\text{$(\frac{q^{4}-q+1}{q^3-q+1}x^2;\,q^{2})_{\infty}$}} }=W_{1}(-x;q).$$ In this sense, note that $$\lim_{q \uparrow 1} W_{1}^{*}(x;q)=\lim_{q \uparrow 1} x^{2} W_{1}(x;q)=\frac{x^{2}}{\sqrt{1-x^{2}}},$$ which gives the weight function of the fifth kind Chebyshev polynomials [@MR2270049].
To compute the norm square value of the symmetric polynomials (\[eq:icod\]), we can use Favard’s theorem [@MR0481884], which says if $\{P_{n}(x;q)\}$ satisfies the recurrence relation $$x P_{n}(x;q) = A_{n} P_{n+1}(x;q) + B_{n} P_{n}(x;q) + C_{n} P_{n-1}(x;q), \qquad n=0,1,2,\dots,$$ where $P_{-1}(x;q)=0$, $P_{0}(x;q)=1$, $A_{n}$, $B_{n}$, $C_{n}$ real and $A_{n} C_{n+1}>0$ for $n=0,1,2,\dots$, then there exists a weight function $W^{*}(x;q)$ so that $$\int_{-\alpha}^{\alpha} W^{*}(x;q) P_{n}(x;q) P_{m}(x;q) d_{q}x=
\left( \prod_{i=0}^{n-1} \frac{C_{i+1}}{A_{i}} \int_{-\alpha}^{\alpha} W^{*}(x;q) d_{q}x \right) \delta_{n,m}.$$ It is clear that the Favard’s theorem also holds for the monic type of symmetric $q$-polynomials in which $A_{n}=1$ and $B_{n}=0$. So, if $\bar{\phi}_{n}(x;q)$ is consider as the monic form of the symmetric $q$-polynomials (\[eq:icod\]), then after some calculations they would satisfy the following three term recurrence relation $$\label{eq:45}
\bar{\phi}_{n+1}(x;q)=x \bar{\phi}_{n}(x;q) - \gamma_{n} \bar{\phi}_{n-1}(x;q), \qquad (\text{with } \bar{\phi}_{0}(x;q)=1, \quad \bar{\phi}_{1}(x;q)=x),$$ in which $$\gamma_{2m}=\frac{q^{2 m+1} \left(q^{2 m}-1\right)
\left(\left(q^4-q+1\right) q^{2 m}+\left(-q^3+q-1\right)
q^2\right)}{\left(q^4-q+1\right)^2 q^{8
m}-\left(q^2+1\right) \left(q^4-q+1\right) q^{4 m+1}+q^4},$$ and $$\gamma_{2m+1}=\frac{q^{2 m} \left(\left(q^3-q+1\right) q^{2 m+1}-1\right)
\left(\left(q^4-q+1\right) q^{2
m}-q\right)}{\left(q^4-q+1\right)^2 q^{8
m+1}-\left(q^2+1\right) \left(q^4-q+1\right) q^{4 m}+q}.$$
Note that $$\lim_{q \uparrow 1} \gamma_{n}=\frac{(4 n+2) \sigma_{n}+(n-1) n}{4 n (n+1)},$$ which coincides with [@MR2270049 Eq. (61.1)].
Therefore, the norm square value of the monic type of the $q$-polynomials (\[eq:icod\]) takes the form $$\int_{-1}^{1} \bar{\phi}_{n}^{2}(x;q) W_{1}^{*}(x;q) d_{q}x= d_{n}^{2} \int_{-1}^{1}
\frac{
{\text{$(q^2 x^2;\,q^{2})_{\infty}$}}
\left(q^3-q+1\right)^{\frac{\log \left(x^2\right)}{2 \log(q)}}}
{ {\text{$(\frac{q^{4}-q+1}{q^3-q+1}x^2;\,q^{2})_{\infty}$}} } d_{q}x,$$ where $$\begin{gathered}
d_{2m}^{2}=\frac{(q-1)^2 \left(q^2+q+1\right) \left(q^3+q^2+q-1\right) q^{m (2
m-1)-2} \left(q^3-q+1\right)^{m+1}}{ \left(q
\left(q^4-q+1\right);q^4\right){}_m} \\
\times
\frac{\left(q^2;q^2\right){}_m
\left(q^3-1+\frac{1}{q};q^2\right){}_m \left(q
\left(q^3-q+1\right);q^2\right){}_m
\left(\frac{q^4-q+1}{q^5-q^3+q^2};q^2\right){}_{m+1}}{\left(
q^4-q^2+1\right)
\left(q-\frac{1}{q^2}+\frac{1}{q^3};q^4\right){}_{m+1}
\left(q^3-1+\frac{1}{q};q^4\right){}_m
\left(q^3-1+\frac{1}{q};q^4\right){}_{m+1}} ,\end{gathered}$$ and $$\begin{gathered}
d_{2m+1}^{2}=\frac{(q-1)^2 \left(q^2+q+1\right) \left(q^3+q^2+q-1\right) q^{2
m^2+m-2} \left(q^3-q+1\right)^{m+1}}{\left(q \left(q^4-q+1\right);q^4\right){}_{m+1}} \\
\times \frac{\left(q^2;q^2\right){}_m
\left(q^3-1+\frac{1}{q};q^2\right){}_{m+1} \left(q
\left(q^3-q+1\right);q^2\right){}_{m+1}
\left(\frac{q^4-q+1}{q^5-q^3+q^2};q^2\right){}_{m+1}}{\left(
q^4-q^2+1\right)
\left(q-\frac{1}{q^2}+\frac{1}{q^3};q^4\right){}_{m+1}
\left(\left(q^3-1+\frac{1}{q};q^4\right){}_{m+1}\right){}^2
}.\end{gathered}$$
Example 2
---------
Consider the $q$-difference equation $$\begin{gathered}
\label{eq:368}
x^{2}(1-x^2) D_{q} D_{q^{-1}} \phi_{n}(x;q)
+qx\left( -{\text{$[5]_{q}$}} x^{2}+q+1 \right) D_{q} \phi_{n}(x;q)
\\+\left(
{\text{$[n]_{q}$}} (q {\text{$[5]_{q}$}}- {\text{$[1-n]_{q}$}} ) x^{2} - q (q+1) \sigma_{n}
\right) \phi_{n}(x;q) =0,\end{gathered}$$ as a special case of (\[eq:17\]).
Following the approach of example 1, we can obtain the polynomial solution of equation (\[eq:368\]) eventually $$\phi_{n}(x;q)=
x^{\sigma_{n}}
{\,{}_{2}\phi_{1}\left(\!\! \begin{array}{cc}{q^{\sigma_{n}-n},q^{n+\sigma_{n}-1} \left(q^6-q+1\right)}\\[-0.1ex]{q^{2 \sigma_{n}+1}\left(q^3-q+1\right)} \end{array}
\Big|\,{q^2};{q^2 x^2}\right)}.$$
Again, this sequence satisfies an orthogonality relation in the form $$\int_{-1}^{1} W_{2}^{*}(x;q) \phi_{n}(x;q) \phi_{m}(x;q) d_{q}x =
\left( \int_{-1}^{1} W_{2}^{*}(x;q) \phi_{n}^{2}(x;q) d_{q}x \right) \delta_{n,m},$$ where $W_{2}^{*}(x;q)=C(x)W_{2}(x;q)$ is the main weight function and $W_{2}(x;q)$ satisfies the equation $$\label{eq:3360}
\frac{W_{2}(qx;q)}{W_{2}(x;q)}=\frac{(q^6-q+1) x^2-q^3+q-1}{q^2 \left(q^2
x^2-1\right)}.$$
Up to a periodic function, a solution of the equation (\[eq:3360\]) is as $$W_{2}(x;q)=\frac{\left(q^2 x^2;q^2\right){}_{\infty }
\left(q^3-q+1\right)^{\frac{\log
\left(x^2\right)}{2 \log
(q)}}}{x^{2} \,\left(\frac{q^6-q+1
}{q^3-q+1} x^2;q^2\right){}_{\infty }}=W_{2}(-x;q).$$ Note that $$\lim_{q \uparrow 1} W_{2}^{*}(x;q)=\lim_{q \uparrow 1} x^{2} W_{2}(x;q)=x^{2}\sqrt{1-x^{2}},$$ which gives the weight function of the sixth kind Chebyshev polynomials [@MR2270049].
The monic polynomials $\bar{\phi}_{n}(x;q)$ solution of (\[eq:368\]) satisfy a three term recurrence relation of type (\[eq:45\]) with $$\gamma_{2m}=\frac{q^{2 m+1} \left(q^m-1\right) \left(q^m+1\right)
\left(\left(q^6-q+1\right) q^{2 m}+\left(-q^3+q-1\right)
q^2\right)}{\left(q^6-q+1\right)^2 q^{8
m}-\left(q^2+1\right) \left(q^6-q+1\right) q^{4 m+1}+q^4},$$ and $$\gamma_{2m+1}=\frac{q^{2 m} \left(\left(q^3-q+1\right) q^{2 m+1}-1\right)
\left(\left(q^6-q+1\right) q^{2
m}-q\right)}{\left(q^6-q+1\right)^2 q^{8
m+1}-\left(q^2+1\right) \left(q^6-q+1\right) q^{4 m}+q}.$$
In this direction we have $$\lim_{q \uparrow 1} \gamma_{n}=\frac{(4 n+6) \sigma_{n}+n (n+1)}{4 (n+1) (n+2)},$$ which coincides with [@MR2270049 Eq. (68.1)].
Thus, the norm square value of the monic sequence $\bar{\phi}_{n}(x;q)$ is derived as $$\int_{-1}^{1} \bar{\phi}_{n}^{2}(x;q) W_{2}^{*}(x;q) d_{q}x= d_{n}^{2} \int_{-1}^{1} \frac{\left(q^2 x^2;q^2\right){}_{\infty }
\left(q^3-q+1\right)^{\frac{\log
\left(x^2\right)}{2 \log
(q)}}}{\,\left(\frac{q^6-q+1
}{q^3-q+1} x^2;q^2\right){}_{\infty }} d_{q}x,$$ where $$\begin{gathered}
d_{2m}^{2}=\frac{(1-q) \left(q^5+q^4+q^3-1\right)
\left({\text{$[6]_{q}$}}-2\right) q^{m (2 m-1)-2}
\left(q^3-q+1\right)^{m+1}}{\left(q
\left(q^6-q+1\right);q^4\right){}_m} \\
\times \frac{(-1;q)_{m+1} (q;q)_m \left(q
\left(q^3-q+1\right);q^2\right){}_m
\left(q^5-1+\frac{1}{q};q^2\right){}_m
\left(\frac{q^6-q+1}{q^5-q^3+q^2};q^2\right){}_{m+1}}{2
\left(q^5+q^2-1\right)
\left(q^5-1+\frac{1}{q};q^4\right){}_m
\left(q^5-1+\frac{1}{q};q^4\right){}_{m+1}
\left(\frac{q^6-q+1}{q^3};q^4\right){}_{m+1} },\end{gathered}$$ and $$\begin{gathered}
d_{2m+1}^{2}=(1-q) \left(q^5+q^4+q^3-1\right)
\left({\text{$[6]_{q}$}}-2\right) q^{2 m^2+m-2}
\left(q^3-q+1\right)^{m+1} \\
\times \frac{(-1;q)_{m+1} (q;q)_m \left(q
\left(q^3-q+1\right);q^2\right){}_{m+1}
\left(q^5-1+\frac{1}{q};q^2\right){}_{m+1}
\left(\frac{q^6-q+1}{q^5-q^3+q^2};q^2\right){}_{m+1}}{2
\left(q^5+q^2-1\right)
\left(\left(q^5-1+\frac{1}{q};q^4\right){}_{m+1}\right){}^2
\left(\frac{q^6-q+1}{q^3};q^4\right){}_{m+1} \left(q
\left(q^6-q+1\right);q^4\right){}_{m+1}}.\end{gathered}$$
Generalized $q$-Hermite polynomials
-----------------------------------
Consider the $q$-difference equation $$\begin{gathered}
x^{2}\left( \left(1-q^2\right) x^2-1 \right)D_{q} D_{q^{-1}} \phi_{n}(x;q)
+(q+1) x \left(x^2+p\right) D_{q} \phi_{n}(x;q)
\\+\left(
q {\text{$[-n]_{q}$}} x^{2} -\sigma_{n} p
\right) \phi_{n}(x;q) =0,\end{gathered}$$ as a special case of (\[eq:17\]), with the following polynomial solution $$\label{eq:64n}
\phi_{n}(x;p;q)=x^{\sigma_{n}} {\,{}_{2}\phi_{1}\left(\!\! \begin{array}{cc}{q^{\sigma_{n}-n},0}\\[-0.1ex]{q^{2 \sigma_{n}+1} \left(-p q^2+p +1\right)} \end{array}
\Big|\,{q^2};{q^2 \left(1-q^2\right) x^2}\right)}.$$
The polynomial sequence (\[eq:64n\]) satisfies an orthogonality relation as $$\int_{-\alpha}^{\alpha} W_{3}^{*}(x;p;q) \phi_{n}(x;p;q) \phi_{m}(x;p;q) d_{q}x =
\int_{-\alpha}^{\alpha} W_{3}^{*}(x;p;q) \phi_{n}^{2}(x;p;q) d_{q}x \delta_{n,m},$$ in which $\alpha=1/\sqrt{1-q^{2}}$, $W_{3}^{*}(x;p;q)=C(x)W_{3}(x;p;q)$ is the main weight function and $W_{3}(x;p;q)$ satisfies the equation $$\label{eq:4460}
\frac{W_{3}(qx;p;q)}{W_{3}(x;p;q)}=\frac{-p q^2+p +1}{q^2+\left(q^2-1\right) q^4 x^2}.$$
Up to a periodic function, a solution of the equation (\[eq:4460\]) is in the form $$W_{3}(x;p;q)=\frac{\left(p \left(1-q^2\right)+1\right)^{\frac{\log
\left(x^2\right)}{2 \log (q)}}
{\text{$(q^2 \left(1-q^2\right)
x^2;\,q^{2})_{\infty}$}} }
{x^{2}}=W_{3}(-x;p;q),$$ where $$\lim_{q \uparrow 1} W_{3}^{*}(x;p;q)=\lim_{q \uparrow 1} x^{2} W_{3}(x;p;q)=x^{-2p}e^{-x^{2}},$$ appearing the weight function of generalized Hermite polynomials [@MR2421844 Eq. (80)].
The monic polynomials type of polynomials (\[eq:64n\]), $\bar{\phi}_{n}(x;q)$, satisfy a three term recurrence relation of type (\[eq:45\]) with $$\gamma_{2m}=-\frac{\left(p \left(q^2-1\right)-1\right) q^{2 m-1}
\left(q^{2 m}-1\right)}{q^2-1},$$ and $$\gamma_{2m+1}=\frac{\left(-p q^2+p +1\right) q^{4 m+1}-q^{2
m}}{q^2-1},$$ such that $$\lim_{q \uparrow 1} \gamma_{n}=\frac{1}{2} \left(p \left((-1)^n-1\right)+n\right),$$ exactly coincides with [@MR2270049 Eq. (79.1)].
Consequently, the norm square value corresponding to the the monic type of polynomials (\[eq:64n\]) takes the form $$\begin{gathered}
\int_{-\alpha}^{\alpha} \bar{\phi}_{n}^{2}(x;q)W_{3}^{*}(x;p;q) d_{q}x\\
= d_{n}^{2} \int_{-\alpha}^{\alpha} \left(p \left(1-q^2\right)+1\right)^{\frac{\log
\left(x^2\right)}{2 \log (q)}}
{\text{$(q^2 \left(1-q^2\right)x^2;\,q^{2})_{\infty}$}}\, d_{q}x,\end{gathered}$$ where $$d_{2m}=\frac{1}{2} q^{m (2 m-1)}
\frac{\left(-p q^2+p +1\right)^m}{\left(q^2-1\right)^{2 m}} (-1;q)_{m+1} (q;q)_m \left(-p q^3+p
q+q;q^2\right){}_m,$$ and $$d_{2m+1}=\frac{1}{2} q^{m (2 m+1)}
\frac{ \left(-p q^2+p +1\right)^m}{\left(q^2-1\right)^{2 m+1}} (-1;q)_{m+1} (q;q)_m
\left(-p q^3+p q+q;q^2\right){}_{m+1}.$$
We finally point out that if we choose $p=0$, then the weight function is reduced to the weight function of discrete $q$-Hermite I polynomials, $$W(x;0;q)=\frac{1}{\left(\left(1-q^2\right) x^2;q^2\right){}_{\infty}},$$ as well as $$\phi_{n}(x;0;q)=c_{n} h_{n}(x\sqrt{1-q^{2}};q),$$ where $h_{n}(x;q)$ denotes the discrete $q$-Hermite I polynomials [@MR2656096 Eq. (14.28.1)].
[10]{}
G. B. Arfken and H. J. Weber, *Mathematical methods for physicists*, fifth ed., Harcourt/Academic Press, Burlington, MA, 2001. William Elwood Byerly, *An elementary treatise on [F]{}ourier’s series and spherical, cylindrical, and ellipsoidal harmonics, with applications to problems in mathematical physics*, Dover Publications Inc., New York, 1959.
T. S. Chihara, *An introduction to orthogonal polynomials*, Gordon and Breach Science Publishers, New York, 1978, Mathematics and its Applications, Vol. 13.
G. Gasper and M. Rahman, *Basic hypergeometric series*, second ed., Encyclopedia of Mathematics and its Applications, vol. 96, Cambridge University Press, Cambridge, 2004.
M. E. H. Ismail, *Classical and quantum orthogonal polynomials in one variable*, Encyclopedia of Mathematics and its Applications, vol. 98, Cambridge University Press, Cambridge, 2005.
F.H. Jackson, *On $q$-definite integrals*, Q. J. Pure Appl. Math. **41** (1910), 193–203.
R. Koekoek, P. A. Lesky, and R. F. Swarttouw, *Hypergeometric orthogonal polynomials and their [$q$]{}-analogues*, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2010.
M. Masjed-Jamei, *A basic class of symmetric orthogonal polynomials using the extended [S]{}turm-[L]{}iouville theorem for symmetric functions*, J. Math. Anal. Appl. **325** (2007), no. 2, 753–775.
[to3em]{}, *A generalization of classical symmetric orthogonal functions using a symmetric generalization of [S]{}turm-[L]{}iouville problems*, Integral Transforms Spec. Funct. **18** (2007), no. 11-12, 871–883. [to3em]{}, *A basic class of symmetric orthogonal functions using the extended [S]{}turm-[L]{}iouville theorem for symmetric functions*, J. Comput. Appl. Math. **216** (2008), no. 1, 128–143.
[to3em]{}, *A basic class of symmetric orthogonal functions with six free parameters*, J. Comput. Appl. Math. **234** (2010), no. 1, 283–296.
M. Masjed-Jamei and I. Area, *A symmetric generalization of [S]{}turm-[L]{}iouville problems in discrete spaces*, Journal of Difference Equations and Applications (2013).
[to3em]{}, *A basic class of symmetric orthogonal polynomials of a discrete variable*, J. Math. Anal. Appl. **399** (2013), no. 291–305.
M. Masjed-Jamei and M. Dehghan, *A generalization of [F]{}ourier trigonometric series*, Comput. Math. Appl. **56** (2008), no. 11, 2941–2947.
M. Masjed-Jamei and W. Koepf, *On incomplete symmetric orthogonal polynomials of [J]{}acobi type*, Integral Transforms Spec. Funct. **21** (2010), no. 9-10, 655–662.
[to3em]{}, *[On incomplete symmetric orthogonal polynomials of Laguerre type.]{}*, Appl. Anal. **90** (2011), no. 3-4, 769–775.
A. F. Nikiforov and V. B. Uvarov, *Special functions of mathematical physics. a unified introduction with applications*, Birkh[ä]{}user Verlag, Basel, 1988.
J. Thomae, *Beitrage zur theorie der durch die heinesche reihe.*, J. reine angew. Math **70** (1869), 258–281.
[^1]: The work of I. Area has been partially supported by the Ministerio de Ciencia e Innovación of Spain under grants MTM2009–14668–C02–01 and MTM2012–38794–C02–01, co-financed by the European Community fund FEDER.
[^2]: The work of M. Masjed-Jamei has been supported by a grant from“Iran National Science Foundation".
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.